Refs #23. Keep Image Enable, delegate Core mode accessors, and remove widget-array mutation. Use Core Audio and Core Video with an Enable-free FPS adapter. Legacy loader IDs remain compatible. Focused CPU tests pass; Windows/browser acceptance remains pending.
391 KiB
Work Log
2026-09-07 — V3.8 Civitai feature infographic
- Used the supplied V3.7 Civitai infographic as the exact format reference and produced a V3.8 replacement at
1536x10243:2 landscape. Preserved the three-panel plus bottom-strip visual structure while replacing Tail 6 material with Review, Resume/Regenerate, simplified controls, and Spectrum/Turbo workflow messaging. Kept the conservativeIssue #13 remains openfooter. No workflow/runtime/Production edit and no publication action.
2026-09-07 — Spectrum-only optional workflow and README install/update entry
- Imported the user-supplied V3.8 Spectrum workflow into
examples/workflows/bit-exact (SHA-256E490C87E4995930F95C8FFFC4C678111D7897E74003E1962B9B4A247BA43F4AC). It parses as 32 nodes / 38 links with no missing link endpoints. The workflow body was deliberately left unchanged, including its prompt, media filenames, titles, widgets, and external-node metadata. - Kept the dependency-free V3.8 public workflow as the sole Registry workflow. Documented the Spectrum graph as the one optional GitHub workflow and the shared starting graph for a Turbo configuration after disabling Spectrum and selecting one LightX2V LoRA. Added the requested LightX2V GitHub/Hugging Face links and placed clone/pull commands at the top of README.
- Focused validation passed
11; the final Registry manifest hash freeze remains intentionally pending. SnapshotD:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-spectrum-readme-workflow-20260907_162216. No runtime, Production, commit, push, tag, or release action.
2026-09-07 — Complete V3.8 public UI reference added to English README
- Documented every visible Main, Advanced, Review/Take and Continuum Settings control, every Sampler media/model socket and output, and the complete interfaces of Load Image/Audio/Video, Reference Audios, Finalize and Second Pass. Kept third-party LoRA/Spectrum/Sage/Core nodes explicitly outside the Continuum public UI.
- Added official-source guidance for asking ChatGPT/Gemini/Grok to read the manual, a conservative Issue #13 status, built-in Size Source/Resolution behavior, 32x32 audio-listening diagnostics, native bypass semantics, duration guidance, and LightX2V FL2VA/Ref2VA Turbo file roles. No runtime or Production behavior changed.
- Copied and linked seven user-provided screenshots with exact source/destination hashes. README local-image audit:32 links/0 missing. Focused regression:120 passed; the single registry-manifest failure is expected until final package re-freeze and currently covers README plus two pre-existing dirty source hashes. Snapshot
D:\Codex\_snapshots\ComfyUI-H3-Continuum\before-v38-full-ui-readme-20260907_152253; no commit/push/tag/release.
2026-09-07 — English continuation manual 4-step live revalidation
- Revalidated the current
00058UI through Computer Use at Turbo4/Steps4/Euler/simple, Spectrum OFF, Draft 0.30 MP, fixed seed, Review Each Chunk, Progress ON, and 2x5 seconds. Fresh Chunk 1 showed Continue/Retry/Finish; selected Continue required the top-right blue Run and then sampled only Chunk 2 (43.738s), producing the complete 10-second sequence. The user-owned 8188 backend remained running. - Captured exact UI evidence and added focused public crops for review-ready actions, selected Continue, completed sequence, Back-to-Settings completion state, Render History, and explicit Regenerate From with Variation Nonce. Updated the English README to remove the superseded retry failure warning, use the new images, explain the completed-state retry meaning, and replace the old six-step timing example with the current four-step observation.
- Snapshot
D:\Codex\_snapshots\ComfyUI-H3-Continuum\before-manual-steps4-capture-20260907_142631; full evidenceD:\Codex\_test_results\ComfyUI-H3-Continuum\v38-manual-live-20260907. No commit/push/tag/release.
2026-09-07 — V3.8 Continuation Regression Final Gate / Pre-GPU freeze
- Audited V3.4 final code
d3dab7b, V3.7 release and current V3.8. Confirmed the exact transport split: current Standard+Balanced22+Audio ON uses masked Video7/Audio37 fixed prefixes; Compatibility uses one AV Reference against a fresh target. Both later-group paths remove the hard First Image keyframe and retain the original visual/Qwen tokens, so reset causation remains a GPU hypothesis rather than a code-audit conclusion. - Actual Comfy CPU
/prompt → Queue → historyshadow passed with CUDA uninitialized. It verified target/prefix/ref/mask/source SHA, seed/noise/SIGMAS, PackedLayout/MM-RoPE andposition_idsidentity for both transports; source Group1 and global RNG were unchanged. Owned11696/40416/37472 and8190/HTTP ended normally. - Prepared one fail-closed
4×8s,896×896 A/B from the corrected live00058 graph: FL2VA INT8, int8 Video VAE, Turbo4 v1.2, Euler/simple6, KJ Sage, First Image and Audio continuity ON; all optional confounders including Spectrum and R6-A OFF. Machine diff shows only Standard↔Compatibility; decoded hidden overlap22f, raw AV, boundary windows/contact sheets, PCM and resources will be captured. - Snapshot935/mismatch0, stage/source153/mismatch0, focused5/5 and syntax PASS. Frozen protocol/execution hashes are recorded in
v38-continuation-regression-final-gate-20260907/PRE_GPU_REPORT.md. Unapproved runner denial passed before run creation/listener. No GPU/backend remains, Production change or publication. The old fixed-task id was unreadable and the newly resolvedH3情報チェックdestination was rejected as a changed trust target, so no workaround was used; fresh sharing authorization and explicit GPU approval are the next boundaries.
2026-09-07 — R6-A Pre-GPU Freeze completed
- Built isolated R6 A/B graphs and runner without Production/runtime edits. A/B differ only by disabled/enabled R6 arm plus evidence/output identities. Frozen R6 formula uses immutable Group1 tail7, CPU-float64 global mean/population std, epsilon1e-12 and gain clamp0.90..1.00; only protected Video prefix can change. Sampling-input Audio is not directly edited, but generated Audio remains a primary guard.
- Reused the proven A→B fail-closed controller. Fresh A requires R3B rawAV12/12, PCM/artifact/resource/plan/normal PID-port-HTTP shutdown, and fresh evidence. Fixed-task review added decoded-quality parity: historical A full video_metrics plus PCM record must match exactly, excluding only run ID/output path; a mismatch prevents A completion and B startup.
- Final CPU run in dedicated Windows temp passed55/55, including37 prior denial cases and a 1e-9 metric-drift rejection. Upstream14/14 matched;251 execution files frozen under protocol SHA48d110db7d376622f4275125207220a997d3134c1ecdf48276cd37d65fab936a. Default temp first produced only inherited ACL setup errors. GPU_AUTHORIZATION.json remains absent; no backend/GPU, Production, commit/push/tag/release.8190 listener zero; diff check passes with line-ending warnings only.
- Report
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-pre-gpu-20260907\PRE_GPU_REPORT.md; pre-record snapshot919 files excluding inherited cache. Fixed reviews aa0cb846-a49f-44ea-a3e3-e682d6a144d5/9a9dc5e1-3363-459c-bc3e-86511dfb97e2 and91ee90f4-ebf4-4659-a329-e85747c4aac0/a5c3e54b-6ff7-4a08-84f2-a11c58a42664 accepted final R6-A Pre-GPU Gate PASS after the decoded-A correction. Explicit GPU approval is the next boundary.
2026-09-07 — R6-A actual Comfy CPU Shadow
- Ran one owned CPU Comfy Queue/history graph through public registration, Core NestedTensor and real masked AV prefix construction on R3B rawAV. OFF object identity Groups2..6; ON Group2 exact no-op and Groups3..6 changed only protected Video7 slots. Audio/masks/generated region/metadata/conditioning fixture/SIGMAS/source/reference/raw12/RNG exact.9 prototype +15 artifact checksPASS.
- CUDA remained uninitialized with tripwires; no model/VAE/noise/Sampling. OwnedPIDs9704/36796/29444,8190 and HTTP all stopped after empty Queue; user backend untouched. ACTUAL COMFY CPU WIRING PASS, but full Pre-GPU incomplete pending actual GPU A12/12 and execution invariants/freeze. Report
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-actual-comfy-shadow-20260907\REPORT.md; snapshot924 pre-r6a-shadow-record-20260907_013958. Fixed summary review accepted with no mandatory correction and retained generated-Audio indirect-effect caveat (6ae817e0-7654-4c06-a06b-75bd9eee42fc /a1bab86c-09b8-4132-aa30-91e2f85bb637).
2026-09-07 — R6-A Pin Renorm CPU design audit
- Built an isolated Production-external policy prototype using the existing diagnostic continuation hook. Frozen contract: Group1 final7 Video slots reference; Groups2+ protected Video-prefix only; current mean retained; global population std ratio clamped0.90..1.00; no amplification. Audio/masks/generated regions and protected contracts unchanged. Classified as global latent AC scaling rather than texture-only.
- Actual R3B shadow gains feeding Groups2..6 were1.0000000/0.9917270/0.9927173/0.9882362/0.9733208; Group2 unchanged, strongest measured reduction2.67%.9 focused CPU testsPASS including OFF identity/RNG, scope/immutability and abnormal inputs. CUDA uninitialized; no backend/GPU. CPU DESIGN PASS WITH CONDITIONS, efficacy unproven. Report
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-pin-renorm-cpu-20260907\REPORT.md;924-file snapshot pre-r6a-cpu-record-20260907_012541. User-authorized fixed summary review accepted classification and requested explicit degeneracy semantics; recorded CPU float64/epsilon1e-12 (05b4095a-89d0-4262-9323-d5932f3ea630 /f8e60c80-b3d1-4637-9cbb-48f490d178f6).
2026-09-07 — R5 actual GPU A/B completed, fixed5f quality FAIL
- User authorized GPU without repeated approval. Executed one frozen A/B pair only, direct GPU startup without W2 CPU bypass. A actual R3B rawAV12/12 exact and current evidence/decision/normal shutdown allowed fresh B. Both720frames/30s, PCM960000samples; B immutable anchor events2..6. Execution/integrity PASS, no OOM or nonfinite.
- Primary Laplacian drift39.86749 ->74.97015 failed; Gradient drift also worsened. Tone/color secondary metrics improved, motion/seam numeric floors and PCM numeric screen passed; no listening/full-playback PASS claimed. Fixed5f condition not promoted, no sweep or Production change, Issue13 unresolved/Release HOLD.
- All six owned PIDs absent and8190 listener absent on final recheck. Source230 py/js unchanged against924-file snapshot pre-r5-gpu-ab-20260907_002047; git diff --check passed. Final report and raw/PCM/videos retained at
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-gpu-ab-20260907. Fixed H3 information-check summary-only review accepted classification with no mandatory corrections (f62e27e4-6b3c-477b-aa0b-7a25c9d20454 /827767df-4345-48e2-b511-f36b1eeeb583).
2026-09-07 — R5-W2 CPU bootstrap avoidance and actual fixture wiring
- Implemented narrow CPU bootstrap-only kitchen availability bypass with CUDA tripwires. Fresh actual Comfy A/B ran Queue/history and PCM-observed real Finalize on synthetic inputs:720frames/960000samples, output/RNG parity, CUDA false through preexit, no new trigger. No GPU model/Sampling.
- A live denied; actual shutdown and current evidence re-read under lifecycle lock allowed CPU fixture B only. Both backends ended.15 evidence checks/37 regression testsPASS; source230 unchanged snapshot924. Scoped W2 PASS; fullR5-2 NOTYETPASS, GPU unapproved. Report
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-w2-cpu-bootstrap\REPORT.md; fixed review requested.
2026-09-07 — R5-W1 CUDA initialization first-cause isolation
- One bootstrap tripwire captured Core attention.py54 -> comfy_kitchen sage_attention.py75 capability query -> torch _lazy_init. Original initializer not called; initialized false before/after, intentional exit86. No second startup, model or Sampling. Installed files unchanged; nine offline assertions PASS and source230 unchanged versus snapshot924.
- Owned26156/33568/15832 ended;8190 absent. Report
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-w1-init-isolation\REPORT.md. Cause isolated, remediation not implemented; full R5-2 NOT YET PASS and GPU unapproved. Fixed-task review requested.
2026-09-06 — R5 actual Comfy wire HOLD
- Added isolated stage/wire Sampler and PCM Finalize adapters plus offline A/B graphs. Nine schema/difference/history-binding CPU tests passed. Actual Comfy --cpu registered nodes and accepted nonce-only prompt/history; no Sampling submission.
- Explicit CUDA guard then detected initialized CUDA despite CPU device, so stopped and did not claim CPU-only PASS. Initializer unresolved. Owned processes and8190 gone, user backend untouched. Full controller/freeze still incomplete; GPU remains unapproved.
- Report:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-comfy-wire-20260906\REPORT.md. Snapshot924 pre-r5-comfy-wire-20260906_233825; diff check0. Fixed-task review requested. No Production change/publication.
2026-09-06 — Texture-only RGB/VAE-delta real-VAE feasibility FAIL
- With explicit user approval, ran only the real H3 Video VAE preflight; no Diffusion Sampling or video generation. The fp16 VAE OOMed on the first saved22f-prefix decode on the16GB GPU, including tiled retry (peak allocated14.764GB bytes, reserved16.280GB bytes). Owned PID9440 exited and VRAM recovered.
- Fixed standalone argument parsing and retried with actual
--cpu-vae. After about5minutes owned PID4472 reached54.954GB working set,163.648GB private allocation including paging, and only6658MB available RAM. A Windows Python memory application error appeared, so only that owned process was interrupted. PID exited; available RAM recovered to56262MB, GPU free to about14.9GB, and no8190 listener/backend existed. - Static prototype remains13/13 PASS but cannot establish real-VAE feasibility. The specific decoded-RGB/VAE-delta Texture-only path is FAIL/closed because it is operationally infeasible and violates the protected no-intermediate-VAE-decode V2 contract. No
6x5efficacy pair, int8 substitution, resolution reduction, CPU forcing, Production implementation, commit or publication. - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-texture-clamp-cpu-design-20260905\REAL_VAE_PREFLIGHT_RESULT_JA.mdandREAL_VAE_PREFLIGHT_GPU_OOM.json. Added a safety guard to the private preflight script so rerun requires a new explicit approval plus--allow-heavy-preflight.
2026-09-05 — Issue13 Texture-only Feedback Clamp CPU design PASS
- Completed the read-only/source audit and an isolated CPU prototype without changing Production or installed runtime. Confirmed the existing private hook order can modify only the native hard
22f/7-slotVideo prefix copy before Sampling and restore the clean prefix beforemake_chunk_entry/ Run Storage commit; Terminal Merge remains untreated. - Froze the control metric and constants before any candidate run: decoded RGB high-pass MSE using Gaussian
5x5/sigma1.0, Group1 anchor trigger1.05x, 50% amplitude-excess correction, gain[0.90,1.00], and latent taper[1,1,1,.75,.5,.25,0]. The latent is adjusted only byencode(corrected)-encode(decoded original); no VAE round-trip replacement or Tone/Saturation correction. - CPU prototype/policy/source-order suite passed
13/13. OFF makes zero VAE calls and returns the same object; Shadow computes but does not mutate; ON is reduction-only, keeps the latest slot exact, leaves Audio/masks/generated region unchanged, and restores the clean prefix before persistence. Existing 30s baseline observation produced later-boundary gains0.9819,0.9600,0.9457,0.9557; these are observations, not efficacy evidence. - Report and prototype:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-texture-clamp-cpu-design-20260905. No backend/GPU/full pytest/Production implementation/publication. Real H3 VAE preflight and any6x5GPU pair remain separately approval-gated; fresh user confirmation is mandatory immediately before execution.
2026-09-05 — Issue13 Texture-only Feedback Clamp direction fixed
- Accepted Texture-only Feedback Clamp as the next single-intervention candidate after the Schedule-Matched 22f mitigation failed its frozen screen. Tone correction is deliberately excluded from the first experiment, and no Schedule-Matched weight sweep is planned.
- Fixed the required sequence as CPU/read-only design audit, predeclared decoded/RGB texture metric and attenuation-only limits, Default OFF parity/Shadow and clean-persistence checks, then at most one separately authorized frozen
6 x 5sA/B. - Recorded protected boundaries: only the disposable 22-frame Video prefix may be corrected; Audio and saved tails are immutable; no direct VAE round-trip replacement; latest-slot causality, Group 1, seed/noise/SIGMAS/Assembly remain unchanged; gain above
1.0is prohibited. - No audit, implementation, backend, or GPU work was started in this record-only step. The user explicitly requires a fresh confirmation before any GPU execution.
2026-08-30 - V3.7 recommended public workflow
- Created
examples/workflows/MiniMax_H3_Continuum_V37.jsonandMiniMax_H3_Continuum_V37.zipfrom the public V3.6 template. The JSON differs only at Sampler node305: its node type and S&R property useH3ContinuumSamplerV37; every graph link, widget value, Standard continuation backend, Hi-Res Fix OFF default, and V3.5 Assemble setting remains identical. - Kept Still Image Guide disconnected because it remains Experimental / Production HOLD, and did not silently enable Tail 6 or alter any Production default. Updated the English/Japanese recommended-workflow links while retaining V3.6/V3.5 compatibility templates and the untracked PDD workflow unchanged.
- Added exact-minimal-diff and ZIP-content regressions. Focused workflow/V3.7 checks are
31 passed; full regression is636 passedusing a dedicated writable pytest basetemp after the known default-temp ACL failure. The 190-entry Manifest, JSON/ZIP consistency, and public workflow links pass. No GPU rerun, runtime logic, version, release tag, commit, or push change was made. - Accepted pre-change snapshots: source
pre-v37-public-workflow-approved-source-0a7c670-20260830_125849(654 files, zero SHA-256 mismatch) and ComfyUI_W runtimepre-v37-public-workflow-approved-runtime-w-20260830_125858(362 files, zero mismatch), excluding only the known ACL-inaccessible.pytest_cache.
2026-08-30 - Restore V3.6 Masked AV infographic
- Preserved the supplied
D:\SD_PIC\Neo\Clip_95.pngbyte-for-byte asdocs/images/v36-masked-av-continuation.png(SHA-25696395c3682fb0b6375dc23eacabea6ca9ee165078ff04e94458f178eef35c4ae) and placed it directly below the English README'sV3.6 Masked AV Continuationheading. - Kept the V3.7 Tail 6 hero and V3.7 Sampler UI images unchanged. The restored image is 1536x1024 RGB and passes Pillow verification; the 188-entry Manifest has zero missing/mismatch, full regression is
633 passed, andgit diff --checkpasses. - Accepted pre-change snapshots are
pre-v36-infographic-restore-source-approved-20260830_030807(229 files) andpre-v36-infographic-restore-runtime-w-approved-20260830_030807(172 files), both at revisionc8ae126cbed2c194f369741601ef1b81b7a4f86ewith zero SHA-256 mismatch. Runtime logic, version, release tag, commit, and push are unchanged.
2026-08-30 - V3.7 Tail 6 README and release images
- Used the V3.6 release infographic
D:\SD_PIC\Neo\Clip_95.pngas the style/layout reference to createdocs/images/v37-tail6-second-pass.png, a 1536x1024 V3.7 infographic covering the exact Tail 6 suffix,100.27 s -> 64.18 s(-36.0%), Conditioning Adapter, RefineSchedule v1, Audio bit-exactness, V3.6 default compatibility, and Still Image Guide Production HOLD. - Used
docs/images/v36-sampler-node.pngas the edit reference fordocs/images/v37-sampler-node.png, updating the title to V3.7 and adding the optionalguideinput. The V3.7 infographic replaces the old remote V3.6 hero in the English README; the V3.7 node UI replaces the V3.6 node image in the later interface section. - Added the English Tail 6 performance explanation and clarified that RefineSchedule manages the exact range while the speedup comes from six evaluations instead of ten. Both PNGs pass Pillow verification; the 187-entry Manifest has zero missing/mismatch, full regression is
633 passed, andgit diff --checkpasses. Runtime logic, version, release tag, README_JA, commit, and push are unchanged. - Accepted pre-change snapshots are
pre-v370-readme-tail6-image-source-approved-20260830_025423(227 files) andpre-v370-readme-tail6-image-runtime-w-approved-20260830_025423(170 files), both at revision093475dba798f0a24a6d8291ae4b42ec90e86c79with zero SHA-256 mismatch.
2026-08-30 - V3.7.0 release publication
- Created verified pre-change snapshots of the authoritative source and ComfyUI_W runtime before release-only edits:
pre-v370-release-metadata-source-approved-20260830_022944(227 files) andpre-v370-release-metadata-runtime-w-approved-20260830_022944(170 files), both at revisionecc72dae9bad1fb909b8e2a003b8e4e5013f4a09with zero SHA-256 mismatch. - Unified
version.py,pyproject.toml, andmetadata.iniat3.7.0; updated English/Japanese release documentation, package information, validation text, and version assertions. The release message centers on the Conditioning Adapter and RefineSchedule foundations and states that Still Image Guide is Experimental / Production HOLD because of hard-anchor trajectory changes. - Runtime generation logic, public Production defaults, existing workflows, and the untracked PDD workflow were not changed.
- Final local Release Gate passed:
633 passed, 185 Manifest entries with zero missing/mismatch, Python/JavaScript syntax,git diff --check, runtime verification as3.7.0, live V3.7/V3.6 schema registration, empty queue, and exact source/runtime SHA-256 for 25 release/Production files. ComfyUI_W restarted as PID8296; evidence isD:\Codex\_test_results\ComfyUI-H3-Continuum\v370-release-runtime-restart-20260830. - Published release commit
fe4ff9c20c2cc8bb375625d1534f5673a737d1betoorigin/main, created annotated tagv3.7.0pointing to that commit, and published the non-draft, non-prerelease GitHub ReleaseComfyUI-H3-Continuum v3.7.0. The tag remains fixed at the validated release commit; this post-publication record is a separate documentation-only main commit.
2026-08-30 - V3.7 frontend parity and Still Guide input hardening
- Audited four findings from an older WIP ZIP against the current authoritative source. Confirmed and fixed the V37 compact-frontend class omission and the overly broad Still Guide IMAGE batch contract. The reported Guide/First/Last collision was already fixed by explicit Guide-role priority, and the Reference Audio
Hidden / Showbehavior was an obsolete pre-V3.5.1 expectation rather than a current regression. - Added
H3ContinuumSamplerV37to the same setup and queue-normalization paths as V3.6. Enforced exactly one Still Guide image during public payload creation and defensive source preparation. Added regressions for V37 frontend routing, batch sizes 0/2/5, defensive batch validation, and Guide collisions at physical frames 0 and 123. - Accepted snapshots are
pre-v37-frontend-batch-contract-approved-accepted-20260830_020900(source, 646 files, zero mismatch, revisionecc72dae9bad1fb909b8e2a003b8e4e5013f4a09) andpre-v37-frontend-batch-contract-runtime-w-approved-accepted-20260830_020901(ComfyUI_W runtime, 359 files, zero mismatch). Focused validation is45 passed; final full regression is633 passedwith only the knownpynvmlwarning. Runtime files were synchronized with exact SHA-256 equality. No GPU rerun, Production-default change, version change, commit, push, or release was performed.
2026-08-30 - V3.7 Conditioning Adapter and RefineSchedule pillars
- Added the internal schema-v1
RefineSchedulecontract and integrated it into context-aware Second Pass without changing public node schemas or the existing SIGMAS socket. External/Full preserve the source tensor, Tail/Partial select exact source ranges, and Second Pass records schedule/hash/noise/audio policy in its contract. First Pass Run Storage reuse remains schedule-independent; schedule differences are identified at the Second Pass contract. - Added original-source ownership for one Still Image Guide. The owner group alone retains a CPU source clone, source hash, absolute frame, and Guide schema/mode identity; target-resolution Second Pass conditioning reencodes the marked Guide from that source rather than resizing its First Pass latent fallback.
- GPU Gate PASS for correctness: 704x704 internal Tail 6 completed in six evaluations. A final same-Sage comparison against the earlier external
SplitSigmasTail 6 baseline produced Video SSIM0.982936and identical decoded PCM SHA-256 with matching 120-frame/5-second contracts. The 1152x1152 Guide + Tail 6 run reencoded the 1254x1254 Guide source to 72x72 latent geometry and completed with valid Video and 32 kHz stereo Audio. That Guide output was low-motion/freeze-detected, so the existing Still Image Guide hard-anchor quality and Production HOLD decisions are unchanged. - Final regression is
626 passed; compileall, Manifest hashes, runtime source equality, runtime verifier, andgit diff --checkpass. Evidence is underD:\Codex\_test_results\ComfyUI-H3-Continuum\v37-pillars-gpu-gate-20260830. No commit, push, release, Production-default change, or public workflow change was performed.
2026-08-28 - Sampling layout-validation measurement and no-fast-path decision
- Started from the formally accepted Phase 1 source/runtime snapshots and added Detailed-Report-only instrumentation around the existing full
validate_native_continuity_layout()call. It records calls, full validations, statuses, wall time including existing D2H/implicit synchronization, physical groups, and per-step contribution without adding explicit CUDA synchronization. Normal diagnostics modes and generation semantics are untouched. - Focused regression is
43 passed; canonical full pytest is570 passed. All 180 Manifest hashes, changed-file syntax, runtime verifier, source/runtime identity, empty queue, andgit diff --checkpass. The first full-pytest attempt is excluded because its newly named--basetempparent did not yet exist; it produced setup-only errors while 466 tests passed. Re-running after creating the parent produced the accepted all-PASS result. - GPU matrix PASS on the existing ComfyUI_W Sage backend: 384x384, 2x5-second I2VA + Continuation, Compatibility Reference Context, 4 steps, fixed seed
361040. Sage: 4 full calls / 1.889 ms total / 0.472 ms per step. Sage+Sol: 4 / 2.122 ms / 0.530 ms. Sage+Spectrum: 8 / 4.658 ms / 1.164 ms. Sage+Sol+Spectrum: 8 / 4.431 ms / 1.108 ms. All routes reported zero failures; no explicittorch.cuda.synchronize()was used. Because global--use-sage-attentionhas no observable MODEL marker, controlled prompt configuration plus the backend command line is the authoritative Sage route evidence. - Representative 576x576 / 20-step Sage PASS: 20 calls, 20 full validations, zero failures,
14.289 mstotal and0.714 ms/step. Sampling was321.632848 sand API elapsed was382.055317 s, so the validator contributes only about0.00444%of Sampling and0.00374%of API time. - Rejected the proposed validation cache/fast-path: its theoretical maximum representative gain is only about 14 ms, far below a material Production improvement. No Layout, Sampling, Terminal Merge, Run Storage, UI, metadata, workflow, schema/version, or transport behavior was optimized or changed. Evidence and the reusable live-API runner are under
D:\Codex\_test_results\ComfyUI-H3-Continuum\layout-validation-profile-phase. Commit and push were not performed.
2026-08-28 - G3 FL2VA prefix correctness repair and Phase 1 PASS
- Scoped the G3 repair to the existing post-Run-Storage FL2VA guard.
FLF_STRATEGYis now required only when Terminal Merge is active; a real-Last-Frame non-Terminal request keeps the compatible prefix already validated by Run Storage. A missing/mismatched Terminal strategy still discards the prefix, and terminal pair atomicity remains unchanged. - Added three regressions for non-Terminal prefix retention, Terminal missing-strategy rejection, and matching-Terminal retention. Focused result is
63 passed; canonical full result is568 passed. Manifest 180 entries, compileall except the known inaccessible.pytest_cachelisting notice, runtime verifier, source/runtime SHA-256, Python/JavaScript syntax, empty queue, andgit diff --checkpass. - Synchronized only
v2/sequence.pyto ComfyUI_W and restarted the same127.0.0.6:8188Sage backend with its recorded venv Python and arguments. Verified pre-change snapshots arepre-g3-fl2va-prefix-correctness-source-accepted-20260828_165400-ecc72da(632 files) andpre-g3-fl2va-prefix-correctness-runtime-w-accepted-20260828_165400-ecc72da(351 files), excluding only.gitand the known inaccessible.pytest_cache, with zero mismatch. Final accepted snapshots arephase1-correctness-pass-source-accepted-20260828_172802-ecc72daandphase1-correctness-pass-runtime-w-accepted-20260828_172802-ecc72da. - Formal GPU rerun PASS: G1 current canonical cache
2/1with Sampling 1; G2 isolated actual legacy V3.6.1"none"cache2/1with Sampling 1; G3 real-Last-Frame non-Terminal FL2VA1/2with Sampling 2 and no FLF-strategy reset; G4 Terminal Merge1/2with one physical77T/433TSampling call; G6 Run Storage Off completed with one Sampling call and no queue error. - G5 fresh 3-chunk versus matched 2→3 extension retained exact equality for all three Video/Audio latent SHA-256 values, decoded RGB24, decoded PCM, 360 frames, and 480256 PCM samples/channel. Phase 1 G1-G6 is formally PASS. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\phase1-gpu-acceptance-20260828\20260828_170034\gate_summary.json. - Public UI, metadata, workflows, schema/version, Reference/Masked transport identity, optimization paths, commit, push, and release were not changed.
2026-08-28 - Phase 1 Run Storage GPU Acceptance Gate
- Ran G1-G6 against the already-running ComfyUI_W 0.34.2 backend at
127.0.0.6:8188on RTX 5060 Ti 16 GiB. G1 current canonical I2VA 2→3 and G2 actual legacy V3.6.1"none"cache both passed at2 reused / 1 generated; G4 FL2VA Terminal Merge passed at1 / 2with one77T Video / 433T Audiophysical Sampling call; G6 Run Storage Off completed with one Sampling call and no hidden-value queue error. - G5 compared a fresh 3-chunk run with the matched 2→3 extension at 576x576, 3x5 seconds, 20 steps, and Variation Nonce 1. All three raw Video latent SHA-256 values and all three Audio latent SHA-256 values matched. Decoded RGB24 and PCM hashes also matched; both outputs are 360 frames and 480256 decoded PCM samples/channel.
- G3 failed the required negative case: real-Last-Frame non-Terminal FL2VA expected
1 reused / 2 generatedbut returned0 / 3with three Sampling calls. The old/new top-level Last Frame hash and Chunk 1 contract hash are identical, and the report first saysresuming after accepted chunk 1; the subsequent existing FLF whole-sequence guard clears that prefix because a non-Terminal request has no saved terminalflf_executionsemantic. - Overall Phase 1 status is HOLD. The G3 finding is outside the absent-hash compatibility implementation and was not repaired because this gate explicitly prohibited Sampling and Terminal Merge contract changes. No optimization, post-acceptance snapshot, commit, push, or release was performed. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\phase1-gpu-acceptance-20260828\20260828_154951\gate_summary.json.
2026-08-28 - V3.6.1 Run Storage correctness hotfix
- Canonicalized absent Last Frame hashes to
""and added a strictly limited V3.6.1 cache compatibility comparison: both top-level Last Frames must be absent, and the per-chunklast_frame_hashspelling must be the only contract difference. Real Last Frame changes and FL2VA Terminal Merge atomicity remain strict; no schema/version bump or optimization was included. - Run Storage Off now resets hidden Regenerate From/Variation Nonce widgets to
Auto/0on toggle and graph load, and defensively rewrites the queued API inputs. Backend validation remains unchanged. - Added regressions for canonical/legacy no-Last-Frame 2→3 reuse, real Last Frame invalidation, unrelated-field rejection, exact resume, Terminal Merge 1/2 atomicity, and frontend toggle/reload/queue serialization. Focused result:
64 passed; full result:565 passed; all 180 Manifest entries, Python/JS syntax, andgit diff --checkpass. - Synchronized only
run_storage.pyandweb/project_id.jsto ComfyUI_W after accepted source/runtime snapshotspre-phase1-correctness-hotfix-accepted-20260828_150500-ecc72daandpre-phase1-correctness-hotfix-comfyui-w-20260828_150659-ecc72da. Runtime verifier and source/runtime hashes pass. The first restart attempt used the resolved base CPython and stopped before listening because venv dependencies were absent; the same backend was immediately restored throughComfyUI_W\venv\Scripts\python.exe, with no double-start or queued job. - Actual ComfyUI_W 0.34.2 / RTX 5060 Ti GPU prompt
bf619577-2057-4635-8fef-2537a40295c4extended an isolated copy of legacy revision9c35c64b529a5c2cfrom 2→3 chunks as2 reused, 1 generated; first regenerated Chunk 3, Sampling169.256391s, outputD:\output\video\comfy_video\video\MiniMax_H3_00008_.mp4, and empty final queue. Exact 2-chunk prompte593b06a-9a37-4cc1-932c-0d13e55979c1then passed as2 reused, 0 generated, first regeneratednone, outputMiniMax_H3_00009_.mp4.
2026-08-28 - Current V3.6 sampler README images
- Replaced the English README's obsolete V3.4 combined workflow screenshot with separate current
H3 Continuum Sampler V3.6andH3 Continuum Assemble + Seam V3.5node images, and added the same images to the Japanese recommended-template section. - Clarified that the current Production path combines the V3.6 sampler with the V3.5 low-memory assembler. No node, workflow, schema, or runtime behavior changed.
- Snapshot:
pre-readme-v36-node-images-20260828_121500-8dec0b7(219 tracked files, zero missing or SHA-256 mismatches, revision8dec0b7ceb961102e36de4dfa735eb40c4cfa4c1). - Validation: all 180 Manifest entries PASS, all 13 unique README image links resolve, focused README/release/workflow tests
11 passed, andgit diff --checkPASS.
2026-08-28 - Run Storage partial-regeneration README guide
- Expanded the English and Japanese Run Storage documentation with a concrete normal T2VA/I2VA
2 chunks x 5 secondsprocedure: enableSave + Auto Resumebefore the first run, retain the Run Name and generation contract, selectRegenerate From = Chunk 2, and confirm1 reused, 1 generated. - Documented that Prompt Plan or other contract changes can create a new revision,
Run Storage = Offcannot be reused retroactively, regeneration operates at chunk boundaries, reassembly reevaluates Seam processing, and FL2VA Long Terminal Merge regenerates its final atomic pair together. No Production code, UI, workflow, or schema changed. - Snapshot:
pre-readme-run-storage-regenerate-guide-20260828_115840-8dec0b7(630 files, mismatch 0, revision8dec0b7ceb961102e36de4dfa735eb40c4cfa4c1). Validation: 178 Manifest entries PASS, focused README/release/workflow tests11 passed, andgit diff --checkPASS.
2026-08-28 - README ComfyUI compatibility requirement
- Added a concise English/Japanese installation note: ComfyUI
>=0.32.0is required, package validation passed on0.33.3, V3.6.1 GPU Smoke passed on0.34.0, and the latest ComfyUI release is not mandatory. - Updated the README entries in
MANIFEST.sha256. Validation passed with552 passed, the ComfyUI_W V3.6.1 runtime verifier, native PackedLayout, Fixed 3x5 Prompt Plan, andgit diff --check. The initial validation attempts were environment-only failures caused by the system Python lacking pytest and the known default pytest temp-directory ACL; the accepted run used the established ComfyUI_W Python and a writable dedicated--basetemp.
2026-08-27 - V3.6.1 Strong 39 fallback GPU Smoke
- Ran the synchronized ComfyUI_W runtime at
127.0.0.6:8188on ComfyUI 0.34.0 / RTX 5060 Ti 16 GiB with Sage Attention and--disable-pinned-memory. The public V3.6 node used Standard, Audio Continuity ON, Strong 39, T2VA 2x5 seconds, 576x576, 20 steps, fixed seed361039, Run Storage OFF, and Detailed Report. - API prompt
f5d5c36c-fa5b-4031-ba0e-349cf70ff64fcompleted in412.034s. The report explicitly resolvedReference Contextbefore execution with the expected Balanced-22 requirement reason; chunk 2 used 39 frames of fixed Reference Context, and no schema, Sampling, finite-validation, Decode, Assemble, save, OOM, or queue error occurred. - Output
D:\output\video\comfy_video\video\V361_Strong39_Fallback_Smoke_00001_.mp4is H.264576x576,240f,24fps, exactly10.000s, with AAC 32 kHz stereo audio of exactly10.000s. The queue returned to empty. V3.6.1 GPU Smoke is PASS; commit/push remain pending.
2026-08-27 - V3.6.1 mixed Prompt syntax parser hotfix (unreleased)
- Preserved all existing tracked/untracked work and created verified snapshot
pre-v361-prompt-parser-hotfix-accepted-20260827_190257at revision0d18e3abe0923502171f47f24769e957057eaf15(214 files, zero SHA-256 mismatch). The standard snapshot attempt stopped only on the known.pytest_cacheACL before any edit. - Updated only the shared Python Prompt planner and tests: Auto/Timeline now ignore standalone
---lines when Timeline syntax is active, issueH3C-P105, keep following text in the current Timeline section, and preserveH3C-P101fallback. Explicit List keeps List parsing with a mixed-syntax warning; Explicit Fixed and invalid-Timeline Fixed fallback remain literal/fail-open. - Warning reports now lead with
PROMPT PREFLIGHT WARNINGand retain resolved Timeline or List source mappings. Prompt Preview consumes the same plan/report; no Frontend Toast, JavaScript parser duplication, schema bump, or Production Sampling/Masked AV/Reference/Terminal/Assembly/Seam change was introduced. - Added mixed Auto/Timeline/List/Fixed, shared-preamble separator, five-chunk trailing separator, missing-chunk fallback, report, Preview, invalid fallback, and Run Storage revision-separation regressions. Results: focused
71 passed, related94 passed, full538 passed, compileall PASS, V3.6.0 source runtime verifier against ComfyUI_WAN PASS, andgit diff --checkPASS. GPU, runtime synchronization, version metadata, release, commit, and push were intentionally not performed.
2026-08-27 - V3.6.0 public release implementation and GPU smoke
- Added
H3 Continuum Sampler V3.6as a new public Node ID while preserving V3.5.3/V3.4 nodes and widget sequences. Public backend choices are onlyStandardandCompatibility: Standard maps to Masked AV or Masked Video according to Audio Continuity, while Compatibility maps to the original Reference Context path. Internal transport names remain private. - Updated package/version metadata, CHANGELOG, package validation/info, README English/Japanese, compact frontend handling, runtime verifier registration, and a copied V3.6 recommended workflow. The V3.5 workflow remains unchanged. Added public-node/backend/schema/workflow regressions.
- Synchronized only
nodes.py,version.py,metadata.ini,v3/driving_nodes.py, andweb/project_id.jsto ComfyUI_WAN after snapshotpre-v360-public-runtime-wan-20260827_172531; all five hashes matched. Restarted the same Sage +--disable-pinned-memorybackend and confirmed V3.6 registration withStandarddefault. - Normal API GPU Smoke using the public node completed in
65.21s: 2x5-second I2VA + Reference Image + Reference Audio, 384x384, four steps, Standard backend, Audio Seam Auto. OutputV36_Release_Smoke_Reference_AV_Standard_00001_.mp4is 240f / 24fps / 10.000s with 32 kHz stereo 10.000s audio; the report confirms Masked AV on continuation chunk 2. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v360-release-smoke\20260827_172835. - Focused release validation is
81 passed; final full source pytest is527 passed; V3.6.0 runtime verifier, JavaScript syntax, changed-filepy_compile, the regenerated 178-entry Manifest, README link validation, source/runtime SHA-256 equality for 12 Production files,git diff --check, and empty API queue all pass. Commit/push/release are the remaining publication actions; the backend remains running.
2026-08-27 - V3.6 PIG-5 subjective audio acceptance
- User completed subjective listening of
V36_PIG3_Reference_Image_Audio_Masked_AV_00001_.mp4and reported Audio PASS. No click, dropout, unnatural stereo positioning, or other audible boundary defect was reported. - PIG-0 through PIG-5 and V3.6 public-promotion eligibility are now formally PASS. This records the gate result only: published V3.5.3, public UI/defaults, version metadata, node schemas, runtime files, and workflows were not changed in this step.
- Pre-change documentation snapshot:
pre-pig5-audio-pass-20260827_171054. Commit/push were not performed.
2026-08-27 - V3.6 PIG-5 policy freeze
- User approved the final Production policy: Masked AV is the standard V3.6 continuation backend; Reference Context remains an Advanced compatibility fallback; raw transport names stay internal; Run Storage identity stays transport-scoped; Terminal Merge regeneration remains atomic per physical group; and Audio Seam Auto continues as an independent evaluation path.
- Public promotion remains HOLD only for subjective listening of the accepted PIG-3 output. A clean result for clicks, dropouts, stereo-positioning anomalies, and other audible boundary defects is sufficient to promote V3.6 without reopening PIG-0 through PIG-4. No UI, runtime, node schema, Sampling, Run Storage, or public default was changed in this policy-recording step.
- Pre-change documentation snapshot:
pre-pig5-policy-freeze-20260827_170808. Commit/push were not performed.
2026-08-27 - V3.6 Production Integration Gate PIG-0 through PIG-4
- Preserved published/default V3.5.3, every accumulated V3.6 change, and all unrelated working-tree edits. Accepted exact snapshots are
pre-v36-production-integration-pig0-accepted-20260827_161517andpre-v36-production-integration-pig2-diagnostic-accepted-20260827_162943; the standard full snapshot stopped only on the known inaccessible.pytest_cachebefore edits. - Added nondefault
continuation_transportscoping to Run Storage identity, Sampling Contract execution semantics, and generated Session settings while leaving the defaultreference_context_v1revision identity bit-for-bit compatible. Added normal/Terminal Reference Image + Reference Audio CPU regressions and transport/session/terminal identity tests. - PIG-2 real-cache testing exposed that Run Storage's temporary resume Session omitted stored
flf_execution, causing all preserved chunks to be discarded.run_storage.pynow restores execution semantics into the resume Session, andv2/sequence.pysynchronizes the reuse count after atomic Terminal-pair pruning. The accepted rerun produced0/3,3/0,1/2, and1/2reused/generated counts for Save, Auto Resume, Regenerate From Chunk 2, and Chunk 3; the two regeneration requests each made exactly one Terminal physical Sampling call and never split logical[2,3]. - PIG-3 completed 2x5-second I2VA + Reference Image + Reference Audio through Masked AV at 384x384 / 4 steps. Every physical group retained both Reference kinds, protected Video/Audio prefixes were bit-exact after finalization, and the output is finite 240f / 24fps / 10.000s with 10.000s stereo audio. Existing Audio Seam Auto produced a benign retained-boundary jump (
0.004261,0.0876xlocal p99); subjective listening remains a separate user check. - Evidence roots are
D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-production-integration\v36_pig2_masked_20260827_164303and...\pig3-reference-masked. Production integration is PASS; public/default promotion remains HOLD pending explicit UI/default-policy approval and subjective Audio Seam listening. No public UI was added, commit/push were not performed, and the backend remains running. - Final validation: PIG-focused
61 passed; canonical full pytest521 passed; changed-filepy_compile, V3.5.3 WAN runtime verifier, three source/runtime SHA-256 comparisons,git diff --check, and empty API queue PASS. Only the knownpynvmlFutureWarning remains.
2026-08-27 - V3.6-R2 shortened FL2VA Long Terminal Merge final GPU gate
- Preserved published/default V3.5.3 and every accumulated uncommitted V3.6 change. Snapshot
pre-v36-r2-short-gate-20260827_000840contains the exact runner/docs changed for this gate. Added only--phase r2-shortscheduling to the private gate runner: Seed 607 full Reference/Masked comparison with cache-only Seam OFF, followed by Seed 608 Masked-only acceptance. - Combined the existing complete Seed 606 Gate A with the shortened runs. All accepted executions preserve 3 logical chunks, 2 physical groups, Group 2 logical
[2,3],paired_timeline_v1, terminal seed/trim, and260f / 77T Video / 433T Audio. Every main MP4 is finite 640x640, 360 frames, 24 fps, and exactly 15 seconds with 15-second audio. - Seed 607 Reference/Masked Group 1 Video and Audio SHA-256 values match. Seeds 606–608 Masked finalization restores only protected 7T Video and 37T Audio prefixes to bit-exact equality; raw Core drift remains limited to one-float-scale deltas and generated regions remain sampler-owned.
- Terminal PackedLayout is
35,865 -> 32,991rows, a2,874-row /8.01%reduction. Synchronized Group 2 Sampling pairs are574.134 -> 529.933sand560.782 -> 544.413s; both favor Masked. Paired percentage median is5.31%; method medians are567.458sReference and544.413sMasked (4.06%lower). Seed 608 Masked completed in549.704sGroup 2 /801.006sAPI wall time. - Five-run monitored peak RSS median is
9.674 GiB; maximum observed device use is15.646 GiB. Terminal CUDA allocator peak is approximately4.06 GiBReference versus3.80 GiBMasked. These whole-process/device values are environment- and order-sensitive; no broad memory guarantee is inferred. - Inspected Seed 606–608 boundary sheets at nominal 5/10/15-second positions and actual 124f/243f retained boundaries. No obvious flash, collapse, direction reversal, or temporal discontinuity was found. Audio analysis found finite 15-second streams and no >=200 ms -50 dB dropout. Audio Seam Auto materially improved Seed 606 Masked, while Seed 607's 124f instantaneous jump increased despite no dropout or structural failure; the report therefore records mixed numerical seam behavior rather than claiming universal improvement.
- Final checks: canonical full pytest
516 passedwith only the knownpynvmlwarning; the first attempt's 98 setup errors were solely the inherited inaccessible Windows pytest temp root and disappeared with a new writable basetemp. WAN runtime verifier, source/runtime SHA-256 forconstants.py,v2/nodes.py,v2/prompts.py,v2/sequence.py,v3/nodes.py,v3/masked_continuation.py, and the diagnostic node,git diff --check, and empty queue pass. - Gate result: PASS. Production promotion: HOLD pending Reference variants, Run Storage/resume identity, public UI/default decisions, and broader audible Audio Seam acceptance. Evidence roots are
D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-fl2va-combined-20260826\gate-aand...\r2-short. Backend PID 15888 remains running per request. Commit/push were not performed.
2026-08-27 - V3.6 shared chunk-duration maximum 30 seconds
- Preserved all uncommitted V3.6 R1/R2/FL2VA work and changed only the shared duration range, Prompt Plan validation, documentation, and focused regressions. No per-chunk variable-duration feature, Sampling, Conditioning,
17k+5temporal math, Assembly, Seam, Run Storage, or saved-workflow socket contract changed. - Added shared constants for default
5.0, minimum4.0, maximum30.0, step0.1, and one tooltip. Applied them to all V2/V3 Production-compatible sampler schemas and Prompt Plan Preview. Values above 15 seconds are supported with a high-resolution VRAM/runtime warning;5–15sremains the recommended/validated range. - Prompt Plan creation and connected-plan validation now accept
30.0, reject30.1, and correctly map a two-chunk[0-30s]/[30-60s]Timeline. Temporal regression covers every tenth from4.0through30.0for 5/22/39-frame contexts and requires native17k+5output shapes. - The separate legacy V1
H3ContinuumJoin.extend_secondsmaximum remains 15 seconds by design; the audit found no remaining 15-second cap onchunk_seconds. - Source snapshot
pre-v36-chunk-duration-30-source-20260826_234716verifies all 209 Git-visible files. Runtime snapshotpre-v36-chunk-duration-30-wan-runtime-20260826_235423verifies the four synchronized files. Source/runtime pre-change differences inconstants.pyandv2/prompts.pywere EOL-only. - Validation: focused inherited-schema/prompt/temporal suite
89 passed; final full pytest516 passed; changed-filepy_compile,git diff --check, V3.5.3 runtime verifier, source/WAN SHA-256, live API schemas for seven V2/V3 node IDs, and empty queue PASS. The restarted WAN backend uses Sage Attention plus--disable-pinned-memoryand is intentionally left running. Commit/push were not performed. - Per the reduced GPU scope, submitted one T2VA
1x30s,384x384, one-step validation prompt. API returned prompt ID9e665f85-b32f-4b17-9734-219fb284e087withnode_errors={}; H3 model initialization/Sampling began without immediate duration/shape/Prompt Plan/OOM error. Interrupted intentionally after31.87s; no Decode/Save or 30-second output-quality claim was made. Queue returned empty. - Cancelled the FL2VA joint multi-seed continuation gate at user request. Its cache warm-up completed, but the first measured condition was interrupted before execution and is excluded from results.
2026-08-26 - V3.6-R2-2 Balanced 22f / 37T Joint AV GPU gate
- Preserved the public/default V3.5.3
reference_context_v1route and all accumulated R1/R2 work. Added only privatemasked_av_prefix_22_v1handling and diagnostic tooling; no public node/schema, V3.4 behavior,v2/sampling.py, seed, Plan, trim, Decode, Assembly, Seam, Run Storage, or saved-workflow contract changed. - Pre-change exact-file snapshots are
pre-v36-r2-2-balanced-av-source-files-20260826_214033(9 files, zero mismatch) andpre-v36-r2-2-balanced-av-wan-runtime-files-20260826_214033(3 files, zero mismatch) underD:\Codex\_snapshots\ComfyUI-H3-Continuum. The initial runtime snapshot copy omitted subdirectories and was corrected before any edit; only the verified snapshot is accepted. - CPU/static validation established the Balanced target
141f / 42T / 235T, Video prefix22f / 7T, Audio prefix37T, trim22f, net119f, and the8.333msVideo/Audio mask-boundary asymmetry. Focused result is24 passed; pre-GPU full result is510 passedwith only known warnings. - Synchronized only
v2/sequence.py,v3/masked_continuation.py, and private diagnosticnodes.pyto ComfyUI_WAN with SHA-256 equality. Restarted ComfyUI v0.34.0 with Sage Attention and--disable-pinned-memory; node registration and empty queue passed. - One-seed A/B used T2VA 2x5, 640x640, Balanced 22, Audio Continuity ON, fixed seed
238985343135605, Prompt/CLIP cache HIT, CUDA synchronization, and Exact Duration ON. Reference packed layout was20,205rows; Masked AV was17,331, eliminating2,874reference rows (14.22%). Continuation Sampling was226.373s -> 206.437s(8.81%lower) and API wall time435.907s -> 420.752s(3.48%lower). One-seed timing is not generalized. - Masked source/target Video 7T and Audio 37T prefixes match bit-exactly after scoped final restore. Core raw prefix deltas were only
2.384185791015625e-07Video and5.960464477539063e-08Audio. Both outputs are finite640x640 / 240f / 24fps / 10.000s; boundary contact sheets show no obvious visual regression. - Seam-OFF decoded audio is not intrinsically accepted: Reference jump
0.082725, Masked jump0.268163. Existing Audio Seam Auto reassembled the cached Masked latents in7.10s, did not rerun Sampling, and reduced the jump to0.029944(88.83%lower and below local p99). No new Audio Seam code was added. - R2-2 status: GPU structure/performance gate PASS with existing Audio Seam Auto, raw Seam-OFF audio documented, Production HOLD. Final full pytest is
510 passed; changed-filepy_compile,git diff --check, source/runtime SHA-256, V3.5.3 runtime verifier, and empty queue pass. Evidence is underD:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate2-20260826. Backend remains running; commit/push were not performed.
2026-08-26 - V3.6-R2-1 39f / 65T Joint AV GPU structure gate
- Synchronized only
v2/sequence.py,v3/masked_continuation.py, and the private diagnosticnodes.pyto ComfyUI_WAN after the accepted R2-0 runtime snapshot. All three source/runtime SHA-256 hashes match. Restarted the same ComfyUI v0.34.0 backend with Sage Attention and--disable-pinned-memory; the backend is intentionally left running. - Added reusable
tools/v36_r2_gpu_gate_runner.py. It removes the I2VA First Image, fixes T2VA 2x5 / 640x640 / Strong 39 / Audio Continuity ON / one seed, excludes optional conditioning and Run Storage, warms Prompt/CLIP once, and measures Reference then Masked with CUDA synchronization. A validation-only stop found Core reports the 39f AV Reference as combined kindvideo_audio; the completed Reference generation remained valid and Masked resumed without repeating it. - Reference continuation used
24,317PackedLayout rows (4,800 ref_img,130 ref_audio); Masked used19,387with no Reference rows. The exact reduction is4,930rows /20.27%. Synchronized Chunk 2 time was332.460 -> 241.888s(27.24%lower); API wall was556.058 -> 465.925s(16.21%lower). This is one-seed structural evidence, not a universal performance claim. - GPU evidence confirmed source-tail/target-prefix equality, Video mask
0:12protected and12:47generated, Audio mask0:65protected and65:263generated, no old refs/interop, and final Video/Audio Prefix SHA-256 equality. Core raw maximum deltas were2.384185791015625e-07Video and5.960464477539063e-08Audio; final restored deltas are zero. No generated region is restored. - Both primary MP4s are finite 640x640 / 240 frames / 24 fps / 10.000 seconds. Nine-frame boundary contact sheets show no gross visual snap, collapse, persistent artifact, or scene discontinuity.
- Added
tools/v36_r2_audio_boundary_analyze.py. With Audio Seam OFF, the exact 124-frame cumulative boundary is the maximum decoded single-sample jump in both files: Reference0.352289, Masked0.287521. Masked has smaller DC/RMS/Peak discontinuities but the native hard cut is not described as click-free. - Added a cache-only Audio Seam Auto probe. It reassembled the accepted Masked latents in
7.23swithout Sampling; boundary jump became0.005774(about98.0%lower), global rank fell from maximum to37.4percentile, and DC jump became0.001103. Existing Audio Seam Auto therefore handles the observed cut; no new seam algorithm was added. - R2-1 status: GPU structure/interoperability PASS, raw Seam-OFF audio cut documented, existing Audio-Seam-Auto path PASS, Production still HOLD. R2-2 22f / 37T and broader acceptance remain pending. Evidence is under
D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate1-20260826; commit/push were not performed.
2026-08-26 - V3.6-R2-0 Joint AV masked-prefix CPU/static gate
- Preserved all uncommitted V3.6-R1 work and kept V3.5.3 public/default behavior frozen. Added only the private
masked_av_prefix_39_v1research transport; no public node/schema,v2/sampling.py, V3.4, seed, Plan, trim, Decode, Assembly, Seam, or Run Storage contract changed. - Generalized
v3/masked_continuation.pyso R1 still protects only 22f / 7T video, while R2-0 copies and masks both a 39f / 12T video tail and a duration-matched 65T audio tail. Sampling-finalization restore is limited to the two protected prefixes; generated video and audio regions remain sampler-owned. - Added a Strong-39-only sequence branch using a 158-frame physical target (
47Tvideo /263Taudio), 39-frame trim, and 119-frame net output. It emits neither old Context Reference rows nor the old Spectrum interop hint and does not reapply the initial I2VA First Image after chunk 1. - Extended the private WAN diagnostic wrapper to capture audio-prefix hashes/equality, audio mask regions, source-tail matching, native audio-grid offset, and final physical-group AV prefix pairs. It remains diagnostic-only and is not a public V3.5.3 node contract.
- Added CPU regressions for exact shapes, zero/one masks, source/target immutability, invalid audio shapes/NaN, protected-prefix restoration, generated-region preservation, First Image isolation, absence of old refs/interop, and the integrated two-chunk sequence path.
- The 39f video window and 65 audio ticks are both 1.625 seconds, but native H3 source/target audio-grid phase is not silently assumed to be zero: 124f / 207T is
+1/3tick and 158f / 263T is-1/3tick. - Validation: changed-file
py_compilePASS; focused masked/temporal tests21 passed; full pytest507 passed; only the knownpynvmlFutureWarning remains. GPU R2-1 and runtime synchronization were deliberately not started in this CPU/static gate. - Accepted snapshots:
pre-v36-r2-0-joint-av-source-accepted-20260826_203022andpre-v36-r2-0-joint-av-wan-runtime-accepted-20260826_203022. Commit/push were not performed; the existing WAN backend was not stopped.
2026-08-26 - V3.6-R1 masked video-prefix continuation PoC
- Kept V3.5.3
reference_context_v1as the public/default continuation transport and added only a private V3.6-R1 switch for testing. No public widget or Production node was added; V3.4, public V3.5.3 schemas, Sampling implementation, seeds, Plan, trim, Decode, Assembly, Seam, and Run Storage remain unchanged. - Added
v3/masked_continuation.pyfor 22-frame / 7T video-prefix target construction and mask creation. The masked route copies the video tail into the next 42T target, uses video mask0for the seven protected slots and1afterward, keeps the 235T audio target empty with an all-one mask, and omits both old Context Reference conditioning and its Spectrum interop hint. - Found that Core v0.34.0 preserved the prefix inside denoising but the raw final float32 integration introduced a maximum difference of
2.38418579101562e-07. Added a narrowly scoped post-Sampling restore of the protected prefix only. Finalized prefix equality is exact bytorch.equal, max difference, and SHA-256; no generated video/audio region is restored. - Added dedicated CPU regressions, a test-runtime diagnostic node, and
tools/v36_r1_api_runner.py. Focused validation is8 passed; full pytest is500 passed;py_compileandgit diff --checkpass. Synced only the changed Production/diagnostic files to ComfyUI_WAN and verified source/runtime SHA-256 equality. - After user-authorized queue clear/interrupt, completed three normal-API A/B pairs at 640x640 / 2x5-second T2VA / 20 steps / Balanced 22. Audio Continuity, Run Storage, Spectrum, Sol, Seam, First/Last, References, and Video Guide were off; Sage Attention was enabled by the server. Every masked run passed 7T prefix exactness, mask/shape/seed/finite checks, empty second-group refs, and absent old interop. All six outputs are 240 frames at 24 fps / 10.000 seconds.
- Contact sheets at frames 112, 118, 122, 123, 124, 125, 128, 136, and 148 show no gross boundary snap, direction reversal, persistent artifact, or collapse for either route. The masked route did not show a consistent visual advantage over the accepted reference-context route across the three seeds. Decision: R1 structure PASS; visual improvement unproven; do not promote it as the Production default.
- Evidence roots:
v36-r1-gate1-postrestore-20260826,v36-r1-gate2-seed238985343135587,v36-r1-gate2-seed238985343135588, andv36-r1-boundary-sheets-20260826underD:\Codex\_test_results\ComfyUI-H3-Continuum. Output MP4s remain underD:\output\video\comfy_video\videowithV36_R1_*names. Commit/push were not performed. - Derived a simplified, annotated A/B workflow from
MiniMax_H3_Continuum_V35 (3).json:D:\Codex\_test_workflows\ComfyUI-H3-Continuum\MiniMax_H3_Continuum_V36_R1_MaskedPrefix_AB.json. Removed disconnected external AddGuide/Noise Mask/AV split-concat nodes and all non-R1 features; retained the exact Model/SIGMAS/diagnostic Sampler/Core Decode/Assembly/Save route. Static graph validation found 22 unique nodes, 24 reciprocal/type-correct links, three groups, zero errors, registered diagnostic type, and defaultmasked_video_prefix_v1. Companion explanation isV36_R1_WORKFLOW_GUIDE.md; no GPU generation was rerun for workflow serialization. - Ran the formal R1 Stress Gate as 3x5-second Strong Motion I2VA at 640x640, Balanced 22, three fixed seeds, Prompt/CLIP cache HIT, Audio Continuity/Run Storage/Spectrum/Sol/Seam OFF, and
--disable-pinned-memory. The order alternatedReference -> Masked,Masked -> Reference,Reference -> Masked. - Diagnostics caught an A/B fairness defect before acceptance: the private masked branch reapplied the original First Image on chunks 2/3 while the reference branch removed it. Interrupted/excluded those preliminary runs, restricted First Image to chunk 1 in
masked_video_prefix_v1, and added CPU coverage. Public V3.5.3 behavior was not changed. - The accepted normal API pairs were
719.800/675.284s,693.449/641.506s, and711.800/672.341sfor Reference/Masked. Masked API time was lower by6.185%,7.491%, and5.544%; continuation Sampling was lower by9.471%,12.029%, and8.309%, with paired medians6.185%and9.471%respectively. - Added real Core
PackedLayoutcapture to the diagnostic node. Reference continuation uses20,569rows including2,800 ref_imgrows; Masked uses17,769rows and no continuationcond/ref_imgrows. A CUDA-synchronized pair confirmed continuation Sampling464.97 -> 437.10s(5.995%faster), so the signal is not a host-asynchrony artifact. - All six accepted videos are 640x640, 360 frames, 24 fps, and 15.000 seconds. Every masked finalized prefix is bit-exact at both boundaries. Contact sheets show no gross snap, reversal, persistent artifact, or collapse. At the second boundary Masked had lower optical-flow jerk and lower absolute luma/saturation drift for all three seeds; flow-magnitude continuity remained mixed, so the quality result is promising but not universal proof.
- Stress evidence is under
D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r1-stress-gate-20260826\normal_fairandsync_fair. The siblingnormalruns are pre-fairness diagnostics and explicitly invalid for acceptance. Added reusabletools/v36_r1_stress_gate_runner.pyandtools/v36_r1_stress_gate_analyze.py. - Final validation: focused masked-prefix regression
5 passed; canonical full pytest500 passed; changed-filepy_compile, V3.5.3 runtime verifier,git diff --check, source/WAN hashes, and empty queue PASS. R1 Stress Gate is PASS; Production remains HOLD pending separately scoped R2/Reference/Terminal/Run Storage work. The WAN backend was intentionally left running per user instruction. Commit/push were not performed.
2026-08-26 - V3.5.3 distribution-integrity maintenance hotfix
- Kept published V3.5.2 generation and optimization behavior frozen; no Sampling, Conditioning payload, Terminal Merge, Assembly, Seam, Run Storage, Prompt/CLIP cache, Video Guide, V3.4, node, or socket contract was changed.
- Created accepted source/runtime snapshots
pre-v353-maintenance-hotfix-source-accepted-20260826_030732andpre-v353-maintenance-hotfix-comfyui-w-accepted-20260826_030731after the standard full snapshot stopped only on the known inaccessible.pytest_cache. - Repaired Conditioning Bridge workflow links
349/350from stale Sampler input slots14/15to Width/Height slots16/17; removed orphan output links171/172/173from the public V3.4, V3.4 Turbo, and V3.5 workflows. - Replaced source-only Hybrid warning renumbering with the accepted absolute public
<Picture N>behavior already present in ComfyUI_W. Added two regressions covering unavailable and missing Hybrid references. - Added an eight-workflow serialization-integrity regression covering node/link uniqueness, endpoints, slot bounds, reciprocal input/output references, and types. The focused release/hotfix result is
30 passed; canonical full pytest is495 passedwith only the knownpynvmlFutureWarning. - Regenerated the release Manifest with 173 entries, added the public-workflow integrity regression, and verified every recorded SHA-256.
- Losslessly re-encoded the unreferenced legacy
v31b-auto-resume.png; strict Pillow verification passes all 19 packaged PNGs and an independent pixel comparison found zero changed pixels. - Synchronized only
reference.py,version.py, andmetadata.inito the primary ComfyUI_W runtime and verified exact source/runtime SHA-256 equality. Runtime verification reports V3.5.3, native PackedLayout PASS, Fixed 3x5 planning PASS, and complete node registration. The incomplete ComfyUI_WAN copy was intentionally not modified. - No GPU generation was rerun because the only Python behavior change is report warning text and the remaining changes are package/workflow metadata. Commit and push were not performed.
2026-08-26 - V3.5.2 README final audit and publication authorization
- Re-audited English/Japanese README chapter order, historical/current-version wording, withdrawn controls, local links, image references, and validation claims before publication.
- Moved the Japanese product summary to the document opening and merged the duplicated English V3.5.1 Hotfix/What's New sections into one chronological
Reference Audio & Compatibility Updatesection. - Explicitly documented that the withdrawn experimental Last Queued Seed override is not included and that V3.5.2 retains standard ComfyUI Seed behavior.
- Kept the user-supplied infographic unchanged and added a precise caption distinguishing its accepted measurement labels from the final ComfyUI 0.33.3 / 172-entry packaging gate.
- Regenerated the 172-entry Manifest and verified all local README links. Post-audit full pytest is
485 passed; Manifest, runtime verifier, node registration, source/runtime hashes, queue state, andgit diff --checkremain PASS. - The user explicitly authorized committing and pushing the complete V3.5.2 release state to the main repository after this audit.
2026-08-26 - V3.5.2 final release gate
- Completed the final Stabilization & Optimization release gate without adding a generation mode or changing V3.4/V3.5 Sampling, Conditioning payloads, Terminal Merge, Assembly, Seam, Run Storage, or standard ComfyUI Seed behavior.
- Created exact pre-change snapshot
pre-v352-release-preparation-exact-20260826_022016at revisionb8ab766eb364063f8af88781e7be70ee7286fba0, containing 198 authoritative files plus the four active ComfyUI_W/ComfyUI_WAN metadata files. The standard snapshot attempt stopped only on the known repository.pytest_cacheACL. - Ran a final 576x576 / 1x5-second / 20-step T2VA API cache smoke before the release metadata change. MISS reported one encode call and
8.454306sPrompt/CLIP; HIT reported zero encode calls and0.000038s. Both used derived seed5684542245315818407and produced identical decoded video MD540e378d7c80e6bebd9d64b322c13e677and audio PCM MD54fb0a699227c4f834f7a52ba4d395ae8. - Reused the accepted Phase 4-4 Video Guide GPU A/B because no Production path changed afterward: complete source identity, 124-frame prefix, VAE/Qwen inputs, Sampling seed, decoded video, and audio PCM remain exact between legacy-v3 and optimized preprocessing.
- Bumped package, project, and runtime metadata to
3.5.2; added the Stabilization & Optimization sections to English/Japanese README and Changelog; removed the stale withdrawn Last Queued Seed claim from package information; and updated package validation to the canonical485 passedresult. - Copied the user-supplied infographic
D:\SD_PIC\Neo\Clip_91.pngtodocs/images/v352-stabilization-optimization.pngbit-exact (SHA-2560c1dfd5acff349b98ecdfb57ae50941f6da26a462dfd2aaa9b7c7b0a34abf304) and embedded it near the top of both READMEs. - Synchronized only
version.pyandmetadata.inito ComfyUI_W and ComfyUI_WAN; all four destination hashes match the source. ComfyUI_W restarted successfully and loggedH3 Continuum 3.5.2 loadedon ComfyUI0.33.3. - Final validation: compileall exit
0; full pytest485 passedwith only the knownpynvmlFutureWarning; V3.5.2 runtime verifier PASS; all 172 Manifest hashes PASS; five required V3.5 nodes registered; source/runtime hashes PASS; queue0/0; andgit diff --checkPASS. Commit, push, and release were intentionally not performed at this gate.
2026-08-26 - Stabilization Phase 4-6A Video Guide memory optimization
- Completed the read-only Phase 4 Feature Tax & Redundancy Audit and selected only Video Guide long-input memory as an implementation candidate. No Production Sampling speed change, Driving Audio lazy Decode, Audio hash special case, Session/State relaxation, Refine Context change, or V3.4 change was adopted.
- The mandatory standard snapshot stopped on the known
.pytest_cacheACL. Created verified exact snapshotpre-phase4-6a-video-guide-memory-exact-20260826_012227with six authoritative/runtime files at revisionb8ab766eb364063f8af88781e7be70ee7286fba0. - Pre-change verification found that ComfyUI_W alone contained an unrecorded Video Guide preprocess v3, while the repository and ComfyUI_WAN still had v2-equivalent content. With explicit approval, preserved and integrated the active v3
17k+5alignment and final-frame padding behavior into the repository before optimizing memory. - Replaced whole-source float32 RGB materialization, finite scan, and one-shot
byteshashing with eight-frame normalization/validation/hash updates. The source object now owns only the used prefix. The complete normalized source still contributes to the exact same SHA-256 and unused-tail NaN/Inf detection. - Deterministic 300x256x256 RGB comparison against the original ComfyUI_W v3 retained an identical full contract and hash. Peak RSS above the allocated input changed from
396.758 MiBto114.695 MiB, retained storage from225.0 MiBto93.0 MiB, and host time from0.246672sto0.257602s. - Added five CPU regressions for legacy-v3 hash byte order, full-tail identity, unused-tail finite rejection, prefix ownership, frame alignment/padding, and encoding behavior. Focused result is
17 passed; full result is485 passed. Source runtime verification and all 171 Manifest entries pass. - Completed Phase 4-4 GPU A/B through the normal API using the same 15-second / 360-frame 800x800 Video Guide, V3.5 one-chunk workflow, 576x576 output, 20 steps, seed input
46001, Sage Attention, and Hi-Res Fix disabled. Legacy and optimized paths both used one physical group, guide contract 640x640 / 124 frames, derived Sampling seed3288707263902886519, exact-duration 124->120 frames, 24 fps / 5.000 seconds, and 32 kHz stereo audio. - Reloaded the exact VHS input independently and confirmed old/new source SHA-256 and combined identity, retained 124-frame prefix, resized VAE input, and Qwen input are byte-identical. Decoded video MD5 and audio PCM MD5 are also identical (
53cd381aff184c2f8aa5910ab62be342/3653c0ec89947a928075d4ba056ff624). Representative start/middle/end frames show no gross artifact. Phase 4-6A is formally accepted; runtime was restored to the optimized source-identical file and the API queue is empty.
2026-08-26 - Stabilization Phase 3-2 synchronized GPU rebaseline
- Phase 3-2 is formally PASS. The accepted decision is to make no speculative Phase 3-3 Production optimization: measured Continuum Assembly is negligible, Core VAE/MP4 work is external, and existing Spectrum/Sol-Attn routes provide materially larger gains without changing Continuum's Sampling contract.
- Final-pass documentation snapshot:
pre-phase3-2-formal-pass-docs-20260826_004900(2 files, revisionb8ab766eb364063f8af88781e7be70ee7286fba0). - Preserved all Phase 3-1 and earlier uncommitted work. Accepted snapshot
pre-stabilization-phase3-2-profile-accepted-20260825_220920contains 479 files at revisionb8ab766eb364063f8af88781e7be70ee7286fba0; the standard snapshot attempt stopped only on the known ACL-inaccessible.pytest_cache. - Added repository-only API diagnostics that synchronize CUDA before/after measured nodes, separate true whole-workflow allocator peaks from per-node reset peaks, poll lightweight RSS/private and NVML free VRAM during each interval, and sample USS at interval boundaries. Early probes that queried CUDA from a polling thread or called
memory_full_info()at 10 Hz measurably slowed Sampling and are explicitly excluded. - Rebaselined Phase 3-1 cache-HIT Production behavior after an excluded warm-up and three measured runs. API medians are T2VA
168.069sand FL2VA379.765s; synchronized phase medians show Sampling120.516s/295.216s, Video Decode19.154s/51.634s, Audio Decode0.466s/1.040s, Assembly0.130s/0.619s, and MP4 save2.583s/7.248s. - Whole-workflow CUDA allocator peaks are
1.456 GiBT2VA and4.129 GiBFL2VA. Per-node runs independently identify Sampling as the maximum. The FL2VA contract remained 3 logical / 2 physical groups ([1],[2,3]), 360 frames, 24 fps, 15.000s; T2VA remained 120 frames / 5.000s. - Compared external attention routes on T2VA with one warm-up excluded and three-run API medians: Sage
168.069s; Sage + Sol-Attn162.090s; Sage + Spectrum124.830s; Sage + Sol-Attn + Spectrum121.147s. The initial Sol attempt failed only because the diagnostic process could not write its default Triton cache and was interrupted/excluded. The accepted rerun used a writable test cache, logged active sparse kernels, and had zero fallback. - The fastest combined route additionally passed one measured FL2VA run at
288.548sAPI with 3 logical / 2 physical groups, terminal seed/trim unchanged, 640x640, 360 frames, 24 fps, 15.000s, and audio.ffprobevalidated every selected output; contact sheets showed no gross persistent artifact. External accelerator outputs are intentionally not required to be MD5-identical. - No Continuum Production optimization was adopted. Sampling remains contract-sensitive, Core VAE Decode and MP4 save are external, and Assemble + Seam is below the 1% threshold. The diagnostic wrapper was removed from the running process and the normal ComfyUI API was restored at
127.0.0.6:8188with outputD:\output\video\comfy_video. No Production runtime synchronization was required. Evidence is underD:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase3-2-20260825; commit/push were not performed. - Focused diagnostics are
13 passed; full pytest is480 passed. Diagnosticpy_compile, all 170 Manifest hashes, source runtime verification, andgit diff --checkpass with only the knownpynvmlFutureWarning.
2026-08-25 - Stabilization Phase 3-1 Prompt/CLIP cross-run cache
- Created exact source/runtime snapshot
pre-stabilization-phase3-1-prompt-cache-20260825_211303with 19 files and preserved all earlier uncommitted stabilization changes. - Added a V3.5-only private Prompt/CLIP conditioning cache attached to the exact CLIP object. It is capped at 16 LRU entries and invalidates on prompt, exact resized First/Last pixels/shape/dtype, Last Frame presence, patch UUID, or CLIP layer. V3.4, public schemas, Sampling, Terminal Merge, Assembly, Seam, and Run Storage were not changed.
- Conservatively bypassed Reference Image/Audio/Video/Timeline and CLIP schedule/tokenizer/hook states. Cache exceptions and CUDA-resident outputs fall back to uncached encode. Tensor payloads are shared only when CPU-resident; metadata dict/list/tuple containers are copied per use.
- Added Detailed Report evidence with real
hits/misses/bypasses/encode_calls. Dedicated and integrated CPU result is25 passed; full pytest is467 passedwith only the knownpynvmland inaccessible repository pytest-cache warnings. - Synchronized only
v2/h3_builder.py,v2/sequence.py,v2/nodes.py,v3/nodes.py, andv3/driving_nodes.pyto ComfyUI_W; every SHA-256 matched. ComfyUI_W was restarted with its recorded arguments before GPU tests. - GPU/API T2VA A/B PASS: 576x576, 120 frames / 5.000s. Prompt/CLIP changed from
5.898398s(misses=1,encode_calls=1) to0.000028s(hits=1,encode_calls=0). Decoded video frame MD5 sequence and decoded audio PCM MD5 are identical. A separate base-seed-only change produced a new derived Sampling seed while retaining one HIT / zero encode calls. - GPU/API FL2VA Long Terminal Merge A/B PASS: 640x640, 360 frames / 15.000s, 3 logical / 2 physical groups (
[1],[2,3]). Prompt/CLIP changed from21.642687s(misses=4,encode_calls=4) to0.007128s(hits=4,encode_calls=0). Decoded video frame MD5 sequence and decoded audio PCM MD5 are identical. A separate base-seed-only change retained four HITs / zero encode calls, changed both normal and terminal physical seeds, and preserved the group contract. - No CUDA retention regression was observed. FL2VA warm completion RSS was 6095.8 MiB MISS versus 6117.1 MiB HIT, a bounded approximately 21.3 MiB CPU cost for four retained conditioning entries. Evidence is under
D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase3-1-20260825. Commit and push were not performed.
2026-08-25 - Stabilization Phase 2-2 Conditioning attribution
- Created exact source/runtime snapshot
pre-stabilization-phase2-2-conditioning-20260825_200800with 15 files including snapshot metadata. Existing uncommitted Phase 2-1 and all prior V3.5.1 work were preserved. - Split Detailed Report Conditioning timing into identity video VAE, Reference Image/Audio VAE, Driving Audio VAE, Video/Timeline Reference VAE, and Prompt/CLIP. The instrumentation is inactive outside the existing V3.5 Detailed Report path and never synchronizes CUDA.
- Focused validation is
29 passed; full validation is453 passedusing a writable explicit pytest--basetemp. The initial full attempt produced setup errors only because the defaultpytest-of-linksdirectory is ACL-inaccessible. Manifest, syntax, and related regressions pass. - Synchronized only
v2/sequence.py,v3/memory_attribution.py, the two related test files, andMANIFEST.sha256to ComfyUI_W; source/runtime SHA-256 matched before GPU execution. - GPU/API 1x5 T2VA accepted result: 576x576, 120 frames / 5.000s, 158.701s API, 135.970s Sampler host wall, 129.935s Sampling, 6.004392s Conditioning, and 6.004374s Prompt/CLIP. Warm-up independently measured 6.283131s Conditioning and 6.283109s Prompt/CLIP. One 3.398s execution-cache replay was excluded.
- GPU/API 3x5 FL2VA Long Terminal Merge accepted result: 640x640, 3 logical / 2 physical groups (
[1],[2,3]), 360 frames / 15.000s, 410.727s API, 346.197s Sampler host wall, 322.335s Sampling, 23.644903s Conditioning, 22.873708s Prompt/CLIP, and 0.771176s First/Last image VAE. Decode and Assemble completed with Exact Duration362 -> 360. - No optimization was implemented. Prompt/CLIP accounts for effectively all T2VA Conditioning and 96.7% of FL2VA Conditioning; the remaining measured Continuum-side work is not an optimization target. Reference-specific branches remain instrumented but require separate connected GPU evidence before any conclusion.
- Evidence is under
D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase2-2-20260825. Commit and push were not performed.
2026-08-25 - Stabilization Phase 2-1 performance attribution instrumentation
- Attempted the mandatory full snapshot first; it stopped only on the known ACL-inaccessible
.pytest_cache. Accepted exact-file snapshotpre-stabilization-phase2-exact-20260825_190303contains nine source/runtime target files at revisionb8ab766eb364063f8af88781e7be70ee7286fba0, with source/runtime Production hashes equal before modification. - Extended the existing V3.5 Detailed Report collector with non-synchronizing host timings for Preparation/Conditioning, physical-group preparation/Sampling/CPU commit plus finite validation/Refine Context capture, and Finalization. No new node, socket, widget, schema, or normal-report overhead was added.
- Terminal Merge regression confirms exactly two performance group records for logical
[1]and[2,3]. Dedicated tests also prove the collector never calls CUDA synchronization. - Focused CPU/static validation is
27 passed; full validation is451 passed; only the knownpynvmlFutureWarning remains. Runtime verification, Manifest validation,py_compile,git diff --check, and source/runtime SHA-256 equality pass. - The first API run accidentally reached the pre-existing 18:46 ComfyUI listener; the newly started process had exited on port conflict. That result is explicitly excluded. After confirming an empty queue, the exact old listener PID was stopped and the synchronized runtime was restarted with the recorded Stability Matrix arguments.
- Accepted GPU/API measurements: warm 1x5 T2VA 576x576 = 143.888s API / 134.736s Sampler host wall (Sampling 128.737s); 3x5 FL2VA Long Terminal Merge 640x640 = 396.477s API / 348.047s Sampler host wall (Sampling 323.835s, physical
[1]+[2,3]); 1x5 Hi-Res 1.2x = 330.005s and 704x704. All outputs are 24 fps with exact 5.000s or 15.000s duration and 32 kHz audio. - Warm baseline versus Hi-Res peak Working Set was 6.882 versus 7.374 GiB; peak Private Bytes was 20.655 versus 21.998 GiB. The 3x5 peaks were 9.409 GiB Working Set and 21.159 GiB Private Bytes. Global VRAM used peaked around 15.4-15.6 GiB and is recorded only as a process-external observation.
- Sampling accounts for 93.0-95.5% of measured Sampler time. CPU commit/finite validation, Refine Context, and group preparation are below the adoption threshold and will not be optimized. Conditioning is 4.4-6.9% and is eligible only for finer attribution, not an optimization change. Evidence is under
D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase2-20260825.
2026-08-25 - Stabilization Phase 0 local CI portability fix
- Froze V3.5.1 feature work and began stabilization with a test-only change. Production Sampling, Layout, Session, Run Storage, Refine Context, UI, and runtime files were not changed.
- Split the P2-5 process-memory assertion by actual platform contract: RSS is universal, USS is validated when present, and the Windows private-commit field is Windows-only.
- Verified snapshot:
pre-stabilization-phase0-ci-platform-test-accepted-20260825_124156(477 files), revisioncf92ac609341539de95675c86271fea11c85e040, excluding top-level.gitand the pre-existing inaccessible.pytest_cache. The standard snapshot attempt stopped on that known ACL before any edit. - Local focused validation is
4 passed; full validation is453 passed; syntax andgit diff --checkpass. Only the knownpynvmlFutureWarning remains. Remote Actions confirmation, commit, and push were not performed.
2026-08-24 - V3.5.1 Issue #9 Conditioning Bridge CPU implementation
- Preserved all Issue #10 changes and created accepted pre-change snapshots:
pre-issue9-conditioning-bridge-source-accepted-20260824_190823(460 files) andpre-issue9-conditioning-bridge-runtime-w-accepted-20260824_190823(337 files), excluding top-level.gitand the pre-existing inaccessible.pytest_cache. - Added
H3ContinuumConditioningBridgeV35as an Advanced MODEL + CONDITIONING interoperability node. It exposes one aligned pair per Continuum physical group, preserves complete CONDITIONING inner lists, updates target geometry, and explicitly leaves AV LATENT/noise/Audio Lock behavior to the external workflow. - Extracted shared physical-group preparation from Second Pass. The built-in path consumes each prepared group immediately, so its prior group-by-group sampling order and low-memory behavior remain intact; no V3.4 sampling, conditioning, Terminal Merge, Assembly, Seam, or Run Storage code was changed.
- Added schema, registration, 37T + 77T Terminal Merge alignment, target geometry, context identity, and non-flattening regressions. Focused result:
33 passed; full source pytest:442 passed; only the knownpynvmlFutureWarning remains. - Synchronized only
nodes.py,v3/second_pass.py,v3/conditioning_bridge_nodes.py, andtools/verify_runtime.pyto ComfyUI_W and verified per-file SHA-256 equality. Runtime verification passed package 3.5.0 registration, native PackedLayout, and Fixed 3x5 planning with the Bridge visible as a current node. The 163-entry distribution Manifest validates completely. - GPU Basic Guider / external sampler acceptance remains pending. Commit and push were not performed.
2026-08-24 - V3.5.1 Issue #10 Reference Audio CPU implementation
- Preserved the V3.4 public schema and added only internal V3.4 forwarding parameters plus two V3.5 optional sockets:
Reference AudioandReference Audio VAE. - Reused the existing Production/Core-compatible Reference Audio path without duplicate preparation or encoding. Standalone Reference Audio 1 is the only V3.5.1 scope; generated audio remains final unless the separate Driving Audio path is connected.
- Added regressions for socket append-only compatibility, published V3.5 template routes, unused-VAE behavior, required VAE validation, Run Storage revision separation/restoration, and Reference Audio retention in normal and Terminal Merge refine-context groups.
- Verified snapshots:
pre-v351-reference-audio-source-verified-ee7b993-20260824_171013(457 files) andpre-v351-reference-audio-runtime-w-verified-20260824_171013(336 files), both with zero SHA-256 mismatch. Earlier standard-script attempts are incomplete because the pre-existing.pytest_cacheACL denies traversal. - Source focused regression
62 passed; full pytest439 passed. ComfyUI_W received one preserved-tooltip Production merge plus three related tests; runtime focused result is41 passed, 1 skipped, syntax and runtime verifier PASS. The skip is only the public workflow-route test becauseexamples/workflowsis not installed in that runtime copy. - GPU tests 1-4 and required normal/Terminal long-form cases remain pending. Package version and public README remain V3.5.0.
- User authorized publication. To avoid pushing the canonical branch's four private/local commits, copied only Production, three regression files, and Manifest into the clean isolated clone at public
origin/main. Regenerated the 161-entry Manifest from public Git index blobs, correcting 36 pre-existing public-tree hash mismatches. Full isolated-clone pytest remained439 passed; commita704aa961fc2ba0c19aef6f2ecfdcb46fa9384b2was pushed and remote-read back exactly.
2026-08-24 - V3.5 recommended template workflow
- Added
examples/workflows/MiniMax_H3_Continuum_V35.jsonas the recommended V3.5 template while retaining every V3.4 template. - Verified the V3.5 Sampler -> optional Hi-Res Fix -> Core Video/Audio Decode -> Assemble + Seam V3.5 route, including
refine_context, Video VAE, Reference Images 1-3, and Video Reference socket placement. - The template defaults Hi-Res Fix to OFF and Assemble to Auto. First/Last/Reference still images share one 0.3 MP control; VHS Video Reference sizing remains independent.
- Updated English/Japanese README, Changelog, project state, and distribution Manifest. No Production Python or runtime file was changed.
- Validation: JSON parse and semantic route audit PASS; all Manifest hashes PASS; full pytest
430 passed; ComfyUI_W runtime registration, native PackedLayout, and Fixed 3x5 prompt-plan verification PASS;git diff --checkPASS.
2026-08-22 - E2-8A / E2-8B1 implementation
- Added additive physical-group Second Pass metadata to V3.4 Assembly Plan output.
- Added static validation, target-geometry update, audio passthrough, and namespaced refine-seed helpers.
- Added CPU regression coverage for 24ch, B/T/group preservation, terminal 260f preservation, H/W upscale, prompt policy, and audio passthrough.
- Production Second Pass sampling, LBH integration, GPU validation, UI, and node registration remain unimplemented.
- Validation: Second Pass contract
8 passed; related V3.4 regressions94 passed; full suite257 passed. - Updated two legacy mocked Assembly Plans with required geometry fields; Production generation code was not changed by this test-only correction.
Append concise completed-work records here. Do not paste full console logs or duplicate PROJECT_STATE.md.
2026-08-20 - Project guidance organization
- Established the three-file structure:
AGENTS.md,PROJECT_STATE.md, andWORKLOG.md. - Added reusable snapshot and validation scripts under
tools. - Kept the existing uncommitted Hybrid reference work untouched.
- No tests were run for this documentation and tooling-only change.
2026-08-21 - Hybrid FLF Chunk Contract v1
- Replaced the single global FLF sample/split path with a 0.4 MP guide and separately sampled FL2VA chunks.
- Kept Reference Images persistent in the guide and every output chunk.
- Preserved the existing V3.4 UI and non-FLF execution paths.
- GPU behavior remains unverified.
2026-08-21 - Restore successful FLF terminal baseline
- Recorded GPU visual baselines: PASS 00029/00031/00032; FAIL 00034/00035.
- Reverted only the FLF terminal strategy from
context_bound_terminal_bridge_v2tocontext_bound_terminal_v1. - Removed the synthetic interpolated terminal keyframe; retained the continuation-context boundary anchor and supplied Last Frame anchor.
- No UI, Hybrid presentation, Reference, Run Storage, Driving Audio, Video Reference, or prompt-policy changes.
- Synced the modified
v2/sequence.pyto the activeComfyUI_Wruntime. - Runtime GPU retest intentionally pending user execution.
2026-08-21 - GPU result 00036 classified FAIL
- User confirmed that 00036 is a full failure, not a partial pass.
- The 5 s boundary was stable, but the sequence failed near 8 s with a bright ghost/flash and discontinuous loss of the reference fox.
- Recorded that whole-sequence continuity is the acceptance criterion.
- Recorded that the
context_bound_terminal_v1restoration remains unresolved for terminal Hybrid FLF+Reference behavior. - No code or runtime file was changed for this classification update.
2026-08-21 - GPU result 00037 classified FAIL
- Confirmed a second full failure after 00036.
- Embedded workflow: official FL2VA pruned INT8 ConvRot model, FL2V Turbo 8-step LoRA, 8 steps, Euler/simple, Spectrum Off, 2 x 5 s, 640x640, one reference image, Video Reference Efficient 0.4 MP.
- The fox remained stable across the 5 s seam, then became translucent and disappeared with a bright ghost/flash near 7.5-8 s.
- Rollback decision: do not use the immediately previous bridge-v2 implementation because its 00034/00035 results failed; identify and restore the exact accepted 00029/00031 baseline.
- No code or runtime file was changed in this classification step.
2026-08-21 - 00038 FAIL; withdrew terminal-v1 restoration
- User classified 00038 as FAIL and requested rollback.
- Confirmed embedded workflow used Ref2VA pruned INT8 ConvRot, FL2V Turbo 8-step LoRA, Euler/simple 8 steps, Spectrum Off, 2 x 5 s, randomized seed, and Video Reference Efficient 0.4 MP.
- Verified repository and ComfyUI_W
v2/sequence.pymatched before modification. - Snapshotted repository/runtime sequence files plus project records to: D:\Codex_snapshots\ComfyUI-H3-Continuum\pre-rollback-after-00038-fail-20260821_032550
- Restored only
v2/sequence.pyfrom the exact pre-restoration snapshotpre-restore-flf-success-baseline-repo-20260821_030316. - Synced the restored file to ComfyUI_W. No UI or unrelated feature files were changed.
- The restored bridge-v2 implementation remains unapproved; this action only reverses the failed terminal-v1 restoration.
2026-08-21 - Corrected rollback target to exact 00031 state
- User clarified that the rollback target is the code state that generated PASS
MiniMax_H3_Continuum_V31_00031_.mp4. - Correlated filesystem times: 00031 completed at 02:32:59; the pre-bridge snapshot was captured at 02:35:53.
- Restored only
v2/sequence.pyfrompre-flf-terminal-bridge-v2-repo-20260821_023553to the repository and active ComfyUI_W runtime. - Active FLF strategy is
context_bound_terminal_v1; bridge v2 is no longer active. - No UI or unrelated feature files were changed. GPU retest was not run.
2026-08-21 - M1 Hybrid diagnostics
- Used ComfyUI_W as the authoritative source after confirming repository/runtime differences.
- Created and verified a runtime snapshot before edits.
- Implemented display-only Hybrid labels and global Picture-number warnings.
- Documented Hybrid Picture numbering and added focused regression tests.
- Runtime generation behavior and public UI were intentionally left unchanged.
- Tests were not run in this implementation step.
2026-08-21 - FLF duplicate boundary anchor removal
-
Preserved the active ComfyUI_W M3 display-label and M4 Last Frame session checks while synchronizing
v2/sequence.pyto the repository. -
Changed the FLF contract to
context_bound_terminal_last_only_v2so Run Storage cannot reuse chunks produced by the previous strategy. -
Removed only the final-chunk call that duplicated the continuation-context tail as a boundary keyframe.
-
Kept the 22-frame continuation context and the single supplied Last Frame anchor retained by
prepare_conditioning(). -
No UI, public sockets, Assembly, audio, Spectrum, or general continuation behavior changed.
-
Automated and GPU validation were not run in this implementation step.
-
2026-08-22 00:49:57 Terminal Merge Physical Seed v2 implemented locally after snapshot pre-change-20260822_004423. Changed v2/sequence.py and run_storage.py; added seed and revision regressions. No commit or push.
-
2026-08-22 00:51:45 Validation: 220 passed, 1 pre-existing failure in test_v2_sequence_orchestration because conditioning_display_label is absent from conditioning.py. All Physical Seed v2 and Run Storage revision tests passed. GPU generation not run.
2026-08-22 - Terminal Merge staged diagnostics
- Snapshot: D:\Codex_snapshots\ComfyUI-H3-Continuum\pre-terminal-merge-diagnostics-20260822_020527
- Added tests-only deterministic capture/comparison helpers for stages A-E.
- Added unit coverage for noise, PackedLayout/position_ids, sampled-stage ordering, numerical tolerance, and bit-exact 243f Terminal split/recombine.
- No Production code, UI, Conditioning, Seed v2, Terminal Merge, Assembly, or Seam behavior changed.
- Commit/push not performed.
2026-08-22 - Standalone Core vs Continuum GPU diagnostics
- Added a standalone diagnostic node without changing Production runtime semantics.
- Added capture helpers and regression tests.
- Installed to ComfyUI_W as a separate custom node and verified source/runtime SHA-256 equality.
- Registration import: PASS (
H3CoreContinuumGpuDiagnostic). - Targeted tests: 6 passed.
- Full suite: 232 passed, 1 pre-existing unrelated failure (
conditioning_display_labelmissing for unfinished M3 wiring). - GPU execution: not run.
- Commit/push: not performed.
2026-08-22 - Terminal Merge physical decode groups
- Preserved logical chunks for Session and Run Storage.
- Recombined terminal logical AV latent slices bit-exactly before external Core VAE Decode.
- Added assembly_plan.decode_groups so one terminal physical sample is decoded once.
- Assembly hardening and Video Seam now operate at physical decode-group boundaries.
- Sampling, Physical Seed v2, Conditioning, UI, sockets, and Run Storage schema were unchanged.
- Synced implementation to ComfyUI_W.
- Pending validation: targeted pytest, full pytest, and GPU Core-vs-Continuum decode comparison.
2026-08-22 - Long Terminal Merge prompt pairing
- Snapshots:
pre-long-terminal-merge-repo,pre-long-terminal-merge-comfyui-w, andpre-long-terminal-merge-comfyui-wanunderD:\Codex\_snapshots\ComfyUI-H3-Continuum. - Removed only the final-prompt-hash equality gate from Terminal Merge eligibility.
- Added a scoped physical prompt policy: identical prompts remain byte-identical; distinct final logical prompts are rebased to
[0-5s]/[5-10s]for one terminal physical sample. - Added scoped Run Storage execution semantics and 3-logical-to-2-physical decode-group regression coverage.
- No UI, public sockets, model selection, Spectrum, Assembly, Seam, or global Run Storage schema changes.
- Focused regression result: 22 passed; GPU generation not yet run for the new 3x5 path.
- Runtime correction:
run_storage.pywas also synchronized to ComfyUI_W and ComfyUI_WAN after the first GPU attempt exposed an oldRunStorageController.prepare()signature. Run Storage plus Terminal Merge regressions: 47 passed. - GPU PASS:
MiniMax_H3_Continuum_V31_00008_.mp4exercised 3x5s FL2VA withterminal_pair=2-3,physical_frames=260,sampling_passes=1for the pair,paired_timeline_v1, two physical decode groups, and final 360 frames / 15.000 seconds. - Core comparison
MiniMax_H3_00005_.mp4reproduced the same short terminal still period. Long Terminal Merge was formally accepted; terminal stillness is Core-equivalent FL2VA convergence and is not a Continuum repair target. - Final release validation: full pytest
249 passed;tools/verify_runtime.py --comfy-root D:\StabilityMatrix\Data\Packages\ComfyUI_Wpassed V3.4 registration, native PackedLayout, and Fixed 3x5s prompt-plan checks. GPU acceptance is recorded above.
2026-08-23 - E2-8C experimental T2VA Second Pass
- Snapshots:
pre-change-20260823_004307for the repository andcomfyui-w-before-e2-8c-20260823_004341for the active runtime files. - Added a generic physical-group Second Pass node; LBH remains an external upstream latent upscaler.
- Reused the existing prompt encoder and ComfyUI-native
sample_chunk()path with workflow-supplied SIGMAS. - One namespaced refine seed is used per physical group; temporary refined audio is discarded and first-pass audio objects are returned unchanged.
- Existing V3.4 Sampling, Terminal Merge, Assembly, Seam, Run Storage, and public Sampler sockets were not changed.
- CPU regression and GPU 1x5 T2VA acceptance remain pending after implementation.
2026-08-23 - E2-8C Second Pass GPU acceptance
- E2-8C is formally PASS after 1x5 T2VA, 3x5 T2VA, and 3x5 FL2VA Long Terminal Merge GPU runs.
- Final acceptance artifact:
E2-8B2_01_1x5_T2VA_LBH3D_00011_.mp4; embedded workflow used three 5-second timeline prompts, supplied First/Last images, the FL2VA INT8 ConvRot model, no enabled LoRA,res_multistep, First Passsimple20 steps / denoise 1.0, external LBH FP16 CUDA upscale from 640x640 to 768x768, and Second Passsimple4 steps / denoise 0.3. Execution time was 17 minutes 44 seconds. - Long Terminal Merge preserved
37T + 77Tand two physical groups: group 1 logical[1]; group 2 logical[2,3]withpaired_timeline_v1. Each physical group ran exactly one Second Pass sampling call with its own reported refine seed. - Final media was 768x768, 360 frames, 24 fps, and 15.000 seconds. First-pass audio LATENT objects were returned bit-exact; Core Video/Audio Decode and Assemble completed successfully.
- The 5-second and 10-second boundaries had no major flash, ghosting, jump, or temporal collapse. The final approximately 0.5-second low-motion interval was accepted as natural Last Frame convergence.
- V3.4 Sampling, Terminal Merge, Assembly, Seam, and Run Storage were unchanged. No code or runtime file was modified during GPU acceptance classification.
- Final validation after documentation update: full pytest
260 passed in 4.41s.
2026-08-23 - V3.5 Continuum Latent Resize
- Snapshots:
pre-v35-latent-resize-repo-20260823_153334(345 files) andcomfyui-w-before-v35-latent-resize-20260823_153336(308 files). - Added
H3ContinuumLatentResizeV35, a list-aware thin wrapper over ComfyUI Corecommon_upscalefor the standard Continuum Hi-res fix path. - The node exposes Core nearest-exact, bilinear, area, bicubic, and bislerp interpolation plus
scale_by; it preserves LATENT metadata, physical-group order, and B/C/T while changing only spatial H/W. - LBH and other learned latent upscalers remain optional external alternatives. No upscaler was embedded into Continuum.
- Updated runtime registration verification to include the existing V3.5 Second Pass and the new Latent Resize node.
- Synced only
nodes.pyandv3/latent_resize_nodes.pyto ComfyUI_W and verified SHA-256 equality. - Validation: Latent Resize tests
5 passed; full pytest265 passed in 5.09s; ComfyUI_W runtime registration, native PackedLayout, and Fixed 3x5s prompt-plan verification passed. GPU generation with the new resize node remains pending. - V3.4 Sampling, Terminal Merge, Conditioning, Assembly, Seam, Run Storage, and public Sampler sockets were unchanged. Commit/push not performed.
2026-08-23 - V3.5 one-node Hi-Res Fix
- Snapshots:
pre-v35-hires-fix-repo-20260823_174946(349 files) andcomfyui-w-before-v35-hires-fix-20260823_174947(310 files). - Added
H3ContinuumHiResFixV35as the normal Main workflow: ON performs Core-compatible spatial resize followed by the existing physical-group Second Pass; OFF returns the exact first-pass video/audio LATENT lists and original Assembly Plan object. - MODEL/CLIP/SAMPLER/SIGMAS are lazy inputs, so the node does not request Second Pass dependencies while disabled.
- Preserved
H3ContinuumSecondPassV35as the Advanced external-latent bridge required by Issue #8 andH3ContinuumLatentResizeV35as a standalone Utility. Removed only the(Experimental)suffixes from their display names. - No V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, Run Storage, or Sampler socket behavior changed.
- Validation: V3.5 focused tests
13 passed; full pytest270 passed in 3.82s; ComfyUI_W runtime registration, native PackedLayout, and Fixed 3x5 prompt-plan verification passed. GPU execution through the new wrapper remains pending. Commit/push not performed.
2026-08-23 - V3.5 Second Pass audio-lock candidate
- Snapshots:
pre-v35-second-pass-audio-lock-repo-20260823_203840(357 files) andcomfyui-w-before-v35-second-pass-audio-lock-20260823_203843(312 files). The earlier 45-file repository snapshot was incomplete after a.pytest_cachepermission error and is not the accepted snapshot. - Added
v3/refine_sampling.pyand switched only V3.5 Second Pass to direct Core sampling with seed-derived random video noise/mask 1 and zero audio noise/mask 0. No inverse noise scaling or manual sigma compensation is applied. - Kept V3.4
sample_chunk(), UI sockets, Conditioning, Terminal Merge, Assembly, Seam, and Run Storage unchanged. First/Last/Reference conditioning remains out of scope. - Validation: V3.5 focused tests
19 passed; Core minimal-mask expansion PASS for representative 37T video and H3 audio shapes; ComfyUI_W venv full pytest273 passed in 3.98susing a writable dedicated basetemp (pynvmlFutureWarning only). - GPU A/B against the fixed 00015 artifact condition remains pending. This CPU result does not establish that audio re-noise was the artifact's primary cause. Commit/push not performed.
2026-08-23 - V3.5 no-resize Second Pass GPU isolation
- Snapshot:
pre-v35-resize-bypass-diagnostic-workflow-verified-20260823_211203(356 repository files excluding the inaccessible.pytest_cache, plus the SHA-256-verified 00017 source MP4). - Created
MiniMax_H3_Continuum_V35_NoResize_SecondPass_Diagnostic.jsonby replacing the Hi-Res Fix wrapper with a directH3ContinuumSecondPassV35connection and adding a status preview. No Production code or runtime file was changed. - Reused the exact cached 576x576 first-pass latent from the 00017 condition. Second Pass remained
res_multistep,simple10 steps / denoise 0.35, refine seed 42; only spatial resize and 1152x1152 re-entry were removed. - GPU execution succeeded as prompt
aa1faf7f-7353-49bd-8efb-23b1cf9c2462. The report confirmed video random noise/mask 1, audio zero noise/mask 0, shape(1, 24, 37, 36, 36), one physical group, one sampling pass, and first-pass audio LATENT passthrough. - Output
V35_NoResize_SecondPass_Diagnostic_00001_.mp4is 576x576, 120 frames, 24 fps, and 5.000 seconds. Twenty representative frames were visually clean with none of 00017's persistent cyan/orange/white line or plate artifacts. - Result: audio-lock does not repair the 2x artifact and random audio re-noise is not its primary cause. The direct Second Pass path passes; the remaining failure is scoped to the 2x spatial-resize/high-resolution re-entry path. Interpolation method versus scale/resolution remains to be isolated.
- Commit/push not performed.
2026-08-23 - V3.5 spatial method/resolution GPU isolation
- Held the cached 576x576 first-pass latent,
res_multistep,simple10 steps / denoise 0.35, refine seed 42, prompt, audio, and Assembly Plan fixed. Only resize method or target H/W changed. - FAIL
V35_NearestExact2x_Diagnostic_00001_.mp4(c35f5c9d-e1b7-43aa-8b09-7f4994aa6225): nearest-exact 2x, 1152x1152, latent(1, 24, 37, 72, 72). The same persistent colored line/plate artifact family remained, so bilinear-specific channel mixing is not the primary cause. - PASS
V35_NearestExact_576to768_Diagnostic_00001_.mp4(676b23cb-79c8-4d03-9e24-0ecfa2288128): nearest-exact 4/3x, 768x768, latent(1, 24, 37, 48, 48). Twenty representative frames were clean. - PASS
V35_Bilinear_576to768_Diagnostic_00001_.mp4(9a7b7383-06dd-4972-bb6c-82c010e64b1e): bilinear 4/3x, 768x768, latent(1, 24, 37, 48, 48). The target artifact did not recur. - FAIL
V35_Bilinear_576to1024_Diagnostic_00001_.mp4(b0530d14-fb2c-42d4-b62c-08d73ec13c8d): bilinear 16/9x, 1024x1024, latent(1, 24, 37, 64, 64). Persistent line/plate artifacts returned across the sequence. - Every run completed Core Decode and Assemble as 120 frames / 24 fps / 5.000 seconds with the audio-locked report and first-pass audio LATENT passthrough. Visual acceptance, not execution validity, separates PASS from FAIL.
- ComfyUI Core source defines H3
BASE_SHORT_EDGE = 768andMAX_PIXELS = 768 * 1344. The observed square boundary is consistent with the native canvas: 768 passed; 1024 and 1152 failed. The exact transition between 768 and 1024 remains untested. - Added four importable diagnostic workflow JSON files under
D:\Codex\_test_workflows\ComfyUI-H3-Continuum. No Production code, runtime file, V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, or Run Storage behavior changed. - Commit/push not performed.
2026-08-23 - V3.5 H3 native-canvas Auto default
- Snapshots:
pre-v35-native-canvas-auto-repo-20260823_215429(356 files, excluding the inaccessible.pytest_cache) andpre-v35-native-canvas-auto-comfyui-w-20260823_215429(309 files). Relevant source/runtime files matched by SHA-256 before the change. - Preserved the existing
H3ContinuumHiResFixV35node ID, sockets, widget order, and saved-workflow behavior. The new defaults arebilinearandscale_by=0.0; zero means H3 native-canvas Auto, while every positive value retains the previous manual scale multiplier. - Auto calls ComfyUI Core
adapt_canvas()and uses an exact-target resize helper so non-square Core targets do not depend on a rounded uniform multiplier. It never shrinks an already-larger latent. Manual targets beyond Core native canvas emit a status warning only; no clamp or hard stop was introduced. - Added exact-target, multi-group metadata, native target, no-shrink, landscape 1024x576 -> 1344x768, manual-compatibility, and warning coverage. V3.5 focused tests:
28 passed; full pytest:277 passedwith only the knownpynvmlFutureWarning. - Synchronized only
v3/hires_fix_nodes.pyandv3/latent_resize_nodes.pyto ComfyUI_W and verified SHA-256 equality. Runtime registration reportedupscale_method=bilinear,scale_by=0.0, and minimum 0.0 after restart. - GPU acceptance prompt
26a7709d-0e4a-4c3f-ad34-3424e3dc7c69producedV35_NativeAuto_576to768_Acceptance_00001_.mp4. The report selected 576x576 -> 768x768, sampled one(1, 24, 37, 48, 48)physical group once, retained first-pass audio bit-exact, and completed Core Decode/Assemble as 120 frames / 24 fps / 5.000 seconds. - Ten-frame visual review found none of the 1024/1152 colored line/plate artifacts. Decoded video matched the accepted manual bilinear 576 -> 768 run exactly (
PSNR = inf). The importable acceptance workflow isD:\Codex\_test_workflows\ComfyUI-H3-Continuum\MiniMax_H3_Continuum_V35_NativeAuto_576to768_Acceptance.json. - V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, Run Storage, and public Sampler sockets were unchanged. Commit/push not performed.
2026-08-24 - V3.5 context-aware conditioning implementation and GPU isolation
- Verified snapshots before modification:
pre-v35-refine-context-repo-verified-20260823_233600(356 repository files excluding.gitand the inaccessible.pytest_cache) andpre-v35-refine-context-comfyui-w-verified-20260823_233632(309 runtime files with the same exclusions). - Added runtime-only
v3/refine_context.py, a new publicH3ContinuumSamplerV35, physical-group capture immediately before First Pass sampling, and optionalrefine_context/video_vaeinputs on Second Pass and Hi-Res Fix. V3.4 retains its original five outputs; V3.5 appends context as output six after Driving Audio. - Context-aware Second Pass validates the capture against the Assembly Plan, re-encodes target First/Last keyframes through the Video VAE, resizes only marked Continuum video context, bypasses prompt re-encoding, and uses one call-local Continuum MODEL clone per physical group. Missing, incomplete, or target-adaptation-unavailable context warns and uses the legacy prompt-only path; corrupt typed context remains a contract error.
- Kept the existing direct refine sampling contract unchanged: video random noise/mask 1, audio zero noise/mask 0, workflow SIGMAS once, one derived seed and sampling call per physical group, and the original first-pass audio LATENT objects returned unchanged.
- Review found no P0/P1 integration defect. Known non-blocking risks are CPU RAM growth from immutable conditioning copies for long/reference-heavy runs, bilinear context resize when an Advanced external learned upscaler is used, and a still-required CLIP socket even when complete context bypasses prompt encode.
- Validation: targeted integration
40 passed; full pytest293 passed;git diff --checkPASS; source runtime verifier and ComfyUI_W runtime registration/native PackedLayout/Fixed prompt-plan checks PASS. Only the knownpynvmlwarning remained. - Synced the changed production files to ComfyUI_W after the verified snapshot. Eight files match source by SHA-256.
v3/driving_nodes.pydiffers only by the pre-existing runtime Video Reference tooltip addition, which was deliberately preserved; runtimepy_compileand registration passed. - Context-aware GPU A/B prompt
0adff3b7-1552-47e9-a167-7b6e4010b0a5used the same FL2VA 1x5 inputs and seeds as the prior artifact run, 576x576 -> 864x864 bilinear 1.5x,res_multistep,simple10 steps / denoise 0.35, refine seed 42. Report: one physical group,conditioning_source=refine_context, two target keyframes VAE re-encoded,(1, 24, 37, 54, 54), audio passthrough.V35_ContextAware_FL2VA_576to864_d035_00001_.mp4still shows the colored line/plate artifact family, so conditioning mismatch is not the primary cause. - Resize-only prompt
75249bbe-306a-45f9-a139-a79c0b17e536reused the cached First Pass and producedV35_ResizeOnly_FL2VA_576to864_bilinear_00001_.mp4. It is 864x864 / 120 frames / 24 fps / 5.000 seconds; colored lines are absent, but severe temporal/person/background ghosting proves that direct spatial interpolation is not a valid general H3 latent-upscale backend. - Pixel/VAE roundtrip prompt
2a03a1d4-2355-4d5a-a237-0ec3e0c09761used First Pass Core Decode -> pixel Lanczos 864x864 -> H3 Video VAE Encode -> the same context-aware denoise 0.35 Second Pass. Sampling, Core Decode/Audio Decode, and Assemble succeeded.V35_PixelRoundtrip_FL2VA_576to864_d035_00001_.mp4is 864x864 / 120 frames / 24 fps / 5.000 seconds and is visually clean: no colored lines, plates, or resize ghosting, with the orange jacket, subject identity, motion, and terminal composition retained. - Pixel/VAE roundtrip 2x prompt
de11e383-8b3d-4d8a-806f-1ac3550091f1used the same cached First Pass, pixel Lanczos 1152x1152, H3 Video VAE Encode, context-awareres_multistep/simple10-step denoise 0.35 Second Pass, and refine seed 42. It completed without OOM on the 16 GiB GPU. Report: one physical group,(1, 24, 37, 72, 72), one sampling pass, two target keyframes VAE re-encoded, first-pass audio passthrough, then successful Core Decode/Audio Decode and Assemble.V35_PixelRoundtrip_FL2VA_576to1152_d035_00001_.mp4is 1152x1152 / 120 frames / 24 fps / 5.000 seconds and is visually clean across twelve representative time points: no colored lines/plates, resize ghosting, or temporal collapse, with subject identity, orange clothing, motion, and composition retained. - Result: retain the Advanced external-latent Second Pass bridge for Issue #8; do not promote plain Latent Resize as the Main Hi-res backend. The evidence-backed next code phase is a low-memory sequential pixel/VAE roundtrip inside Hi-Res Fix, one physical group at a time, followed by the existing context-aware Second Pass. That backend change is pending explicit approval. Commit/push not performed.
2026-08-24 - V3.5 Main safe Pixel/VAE Hi-res Fix implementation and GPU acceptance
- Pre-change snapshots:
pre-v35-pixel-vae-main-repo-20260824_004310(159 files / 3,817,668 bytes) andpre-v35-pixel-vae-main-comfyui-w-20260824_004310(132 files / 1,660,144 bytes). Source/runtime hashes matched for the relevant pre-existing production files before modification. - Added
v3/pixel_vae_resize.pyand changed only the V3.5 MainH3ContinuumHiResFixV35path to perform one physical group at a time as H3 Video VAE Decode -> bounded 8-frame CPU pixel resize -> same H3 Video VAE Encode -> existing context-aware Second Pass. Default Main controls are Lanczos andscale_by=2.0; zero remains native-canvas Auto. - Enabled Main execution requires the lazy Video VAE and has no direct-latent interpolation fallback. OFF remains exact video/audio/Assembly Plan identity and does not evaluate VAE or sampling dependencies. Advanced Second Pass and the standalone Latent Resize Utility were not changed.
- The first integrated GPU attempt (
46f775ca-093c-4c1f-b5ce-1aa9a8f8616f) stopped at the explicit output-shape contract before Second Pass. Root cause: passing[B,F,H,W,C]directly to CoreVAE.encode()makes Core's 4D IMAGE crop path treat F as a spatial axis. The helper now calls Encode per B with[F,H,W,C], validates each[1,C,T,H,W]result before concatenation, and restores B explicitly. No latent fallback ran during the failed attempt. - Validation after the Core-boundary correction: focused pixel/VAE and Hi-Res tests are covered by the full
308 passed;py_compile,git diff --check, and the ComfyUI_W runtime verifier all pass. Independent review found no remaining P0/P1/P2 defect. Runtimev3/pixel_vae_resize.pymatches source SHA-256B7C4A35A5E73F528492BC6946B97B687C412F1D3F5F976C19D45C6DC20288C9Aafter restart. - Integrated Main GPU acceptance prompt
c5dbe55c-c846-47e8-9f49-8bb167887236used FL2VA 1x5, 576x576 -> 1152x1152 pixel Lanczos, context-awareres_multistep,simple10 steps / denoise 0.35, and refine seed input 42. Report: one physical group, one sampling pass,(1, 24, 37, 72, 72), two target keyframes re-encoded, and first-pass audio LATENT passthrough. Core Video/Audio Decode and V3.4 Assemble+Seam completed in 14:18. V35_MainSafePixelVAE_Core4D_FL2VA_576to1152_d035_00001_.mp4is 1152x1152 / 120 frames / 24 fps / 5.000 seconds with 32 kHz stereo audio. All decoded video and audio frame hashes exactly match the accepted external Decode -> Lanczos -> Encode -> Second Pass baseline. Visual contact-sheet review found none of the prior colored line/plate or resize-ghosting artifacts. The shared still interval from 3.666667 seconds is the same FL2VA Last Frame convergence in both outputs.- Host-memory use remains material: observed peak was about 47.7 GiB working set / 61.3 GiB private bytes, with about 7.3 GiB system RAM available during the run. The Main backend is accepted for this 1x5 FL2VA 2x condition only; 3x5 Terminal Merge 2x remains pending separate GPU/resource acceptance.
- V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, Run Storage, public V3.4 sockets, Advanced Second Pass, and Latent Resize Utility were unchanged. Commit/push not performed.
2026-08-24 - Phase 2A P2-0 Windows file-backed Tensor PoC
- Verified pre-change snapshot:
D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-phase2a-p2-0-backend-poc-verified-20260824_022242(374/374 source files, SHA-256 mismatch 0, revisiond77a467796c048908f8b7b83736e226e736e0f98). The earlierpre-phase2a-p2-0-backend-poc-20260824_022217attempt is incomplete because the repository.pytest_cacheACL denied traversal and is not the accepted snapshot. - Added only the unregistered standalone PoC
tools/p20_file_backed_tensor_poc.pyand its tests. Production nodes and ComfyUI_W/ComfyUI_WAN runtime packages were not modified. - Selected
torch.from_file(shared=True)and implemented strict backing-path/sidecar validation, alias-aware Storage lifetime, a process-global manager, retryable deletion, dead-process stale cleanup, live/young retention, root-scoped thread/process locking, and fail-closed invalid metadata handling. - Independent review reproduced and drove fixes for Windows sharing-violation retry, dropped unpublished allocations, allocation/cleanup races, GC/root-lock re-entry, cross-thread ABBA lock inversion, and close-after-cancel fd reuse. Finalizers now enqueue only; actual cleanup uses root-lock -> manager-lock ordering.
- Validation: ComfyUI_W focused
26 passed; full repository334 passed; standalone probe PASS (bit_exact=true, 480/480 bytes, alias-live deferred=1, final-alias reclaimed=1, managed files remaining=0). ComfyUI_WAN independently passed the same 26 tests and probe under Python 3.12.12 / PyTorch 2.13.0+cu130. Only the knownpynvmlFutureWarning was emitted. - Deferred by design: Production Assemble integration, real ComfyUI cache/requeue and VHS/Save compatibility, RAM/Disk full-assembler equivalence, cross-process application integration, and representative 9-15 GiB sequential-write/RSS acceptance. These remain P2-2 through P2-5 gates; P2-1 Memory Attribution is next.
- Commit/push not performed.
2026-08-24 - Phase 2A P2-1 through P2-3 CPU/static acceptance
- Verified the authoritative source and both runtime copies, preserved all prior uncommitted work, and created
pre-phase2a-p2-1-memory-attribution-20260824_030858followed by the accepted implementation snapshotD:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-phase2a-p2-2-disk-backed-20260824_032923(166/166 source files, zero SHA-256 mismatch, revisiond77a467796c048908f8b7b83736e226e736e0f98). - Completed P2-1 with fail-soft, read-only V3.5 Memory Attribution at sequence start/before CPU commit/after CPU commit/completion, including deduplicated retained AV/refine-context storage, projected final IMAGE sizes, RSS/system RAM, and CUDA allocated/reserved values. No synchronization, peak reset, cache clear, forced GC, model unload, or V3.4 default change was added.
- Completed P2-2 with the Production adapter
v3/file_backed_buffer.py: manual RAM/Disk-backed selection only, no Auto and no fallback, process-global alias-aware lifetime management, lazy root-first stale cleanup, and Video-only mapping while Audio stays in RAM. - Completed P2-3 with independent V3.5 assembly/seam modules. The final target IMAGE is allocated once and receives physical decode groups in order through direct trim/pad/final-anchor spans. Exact Duration and Terminal Merge contracts are retained without a final full IMAGE
cat,stack,clone, orcontiguousmaterialization. - Reused the existing V3.4 Video Seam analysis and math while limiting temporary copies to the affected 1-4 boundary frames and anchors. Preserved Audio Seam and cumulative sample alignment in RAM. Disk-backed publication is delayed until late Driving Audio/report validation completes; regressions cover early/late abort, alias lifetime, and dead-process stale recovery.
- CPU regressions cover RAM/Disk bit-exact output; Seam Off/Analyze/Auto/Auto 2; Exact ON/OFF; trim, pad, and preserved final frame; 1x5, 2x5 Terminal Merge, and 3x5 Long Terminal Merge; 32 kHz rounding; Audio Seam; physical ordering; Auto rejection; and error/lifetime paths.
- Final canonical validation: full pytest
412 passedwith only the knownpynvmlFutureWarning; changed Productionpy_compilePASS;git diff --checkPASS. ComfyUI_W runtime verification passed node registration, native PackedLayout, and Fixed 3x5 prompt plan; focused runtime regression141 passed. - Synchronized only 19 P2-related source-or-preserved-merge files to primary ComfyUI_W and verified every SHA-256 comparison. Preserved its pre-existing Video Reference tooltip delta. ComfyUI_WAN was left unchanged rather than receiving a broken partial V3.5 dependency set.
- Stopped at the requested P2-3 boundary. Still pending for P2-4/P2-5: real cache/requeue lifetime, Preview/VHS/Save, ESC cancel, 9-15 GiB stress/RSS, downstream copy behavior, GPU generation, long-duration GPU acceptance, and Auto backend. Commit/push not performed.
2026-08-24 - Phase 2A P2-4 ComfyUI cache/downstream integration
- Accepted verified snapshots: source
pre-phase2a-p2-4-comfy-integration-source-verified-20260824_101909(415 files, SHA-256 mismatch 0, revisiond77a467796c048908f8b7b83736e226e736e0f98) and ComfyUI_Wpre-phase2a-p2-4-comfy-integration-comfyui-w-verified-20260824_101938(324 files, mismatch 0). The two standard-script attempts and one intermediate verification attempt are incomplete because the existing.pytest_cacheACL denies traversal and are not accepted snapshots. - Added bounded 8-frame interrupt polling to only the V3.5 direct IMAGE writer. ComfyUI's normal interrupt is checked after Disk-backed allocation and between copy batches, so immediate or mid-copy ESC uses the unpublished-allocation cleanup path. RAM/Disk output equality is unchanged; focused interrupt/P2-3 regression was
47 passed. - Added the reusable headless integration probe
tools/p24_comfy_integration_probe.py. It registers deterministic decoded CPU fixtures only, then executes the real installed ComfyUI_WPromptExecutorCLASSIC cache, Core Save/Preview nodes, VHS h264-mp4, Core interrupt handling, and separate-process stale cleanup. No model/VAE/sampler/CUDA/GPU path is used. - Actual cache evidence: an intentional downstream failure retains the completed mapped IMAGE; unchanged requeue reports assembler node 2 cached and leaves one backing pair; Save/Preview/VHS requeue reports nodes 1/2/4/5/6 cached without rewriting outputs; cache eviction followed by replacement allocation keeps pair count bounded at one; final pair count is zero.
- Downstream evidence: Core Save 120 PNGs, Core Preview 120 PNGs, and VHS H.264/AAC succeeded.
ffprobeverified 16x16, 120 frames, 24/1 fps, 5.000-second video/container duration, and 32 kHz audio. Core interrupt producedexecution_interruptedwith zero unpublished files. Crash child left one pair; clean-start child removed it. - Final runtime artifacts:
D:\Codex\_test_results\ComfyUI-H3-Continuum\P24-runtime-W-final-20260824_1105\p24_result.json,p24_prompt_graphs.json, andoutput\P24\VHS_00001-audio.mp4(39,006 bytes). The expected intentional downstream-error traceback is part of the probe and not a failure. - Final validation: full pytest
415 passed; ComfyUI_W P2-focused89 passed; runtime verifier,py_compile,git diff --check, and source/runtime SHA-256 checks pass. Synced onlyv3/assembly_v35.py, the P2-4 probe, and its test. The knownpynvmlFutureWarning is unchanged. - P2-4 is formally PASS and work stops here. P2-5 still covers representative 9-15 GiB/RSS stress, full-resolution and long-duration GPU execution, manual UI cancellation under large I/O, downstream memory peaks, and any future Auto backend policy. Commit/push not performed.
2026-08-24 - V3.5 compact UI parity
- Verified source/runtime
web/project_id.jsequality, preserved all existing uncommitted work, and accepted snapshotspre-v35-compact-ui-source-verified-20260824_104010(420 files) andpre-v35-compact-ui-runtime-w-verified-20260824_104048(330 files), both with zero SHA-256 mismatch. Standard snapshot attempts failed only on the pre-existing inaccessible.pytest_cache. - Added the V3.5 Sampler and Assemble+Seam class IDs to the existing frontend compact/settings integration. This hides Report Detail/debug/preview/strict compatibility and the internal Auto Resume ID, while retaining the existing conditional Run Name/Regenerate/Variation Nonce behavior.
strict_compatibilitywas not removed from the backend schema; legacy loading remains possible, while the frontend continues to force the ignored value tofalse. No backend or generation behavior changed.- Added
tests/test_v35_compact_ui_frontend.py. Dedicated regression2 passed, JavaScript module syntax PASS, and canonical full pytest417 passedusing a fresh writable basetemp. Only the knownpynvmland repository pytest-cache warnings remain. - Synchronized only the frontend file and dedicated regression test to ComfyUI_W; both source/runtime SHA-256 comparisons match, and the runtime-side dedicated regression is
2 passed. - P2-5 was not started. Commit/push not performed.
2026-08-24 - Phase 2A P2-5 long-output acceptance
- Preserved every existing uncommitted change and accepted verified snapshots
pre-phase2a-p2-5-gpu-acceptance-source-verified-20260824_105605(422 files) andpre-phase2a-p2-5-gpu-acceptance-runtime-w-verified-20260824_105626(331 files), both with zero SHA-256 mismatch after excluding the pre-existing inaccessible.pytest_cache. - Added the reusable source-side
tools/p25_gpu_acceptance_probe.pyand focused tests. The probe runs ComfyUI_W headlessly through its installedPromptExecutor, reconstructs the accepted 3x5 FL2VA Long Terminal Merge physical decode groups from Run Storage revision1e30d8bf2154bbe7, and uses the installed Core Video/Audio VAE Decode and Production V3.5 assembler. No UI automation or new model/sampler generation was used. - Fresh-process RAM/Disk 640x640 A/B passed: 3 logical chunks, 2 physical groups (
37T + 77T), 360 frames at 24 fps / 15.000 seconds, 32 kHz stereo audio, identical Exact Duration362 -> 360and audio correction-2667, identical Seam Auto report, finite output, identical video SHA-25695d5479bc5fee9c235ece7996c7c88dcc60981625dbedc330344403016ca9b9d, and identical audio SHA-256eec0c6ff378c5bfcc3e71d02bab14230dea02a57e18ce1212f1dc1be4c88a3f9. - Core Save/Preview each consumed 360 mapped frames. VHS H.264/AAC was verified as 640x640, 360 frames, 24 fps, 15.000 seconds, and 32 kHz. Same-input requeue cached every node through assembler/digest without growing beyond one backing pair; cache release reclaimed it.
- A post-allocation Core interrupt produced
execution_interruptedand left zero managed files. The P2-4 crash/restart probe was rerun in the P2-5 environment: one crash-stale pair was created and the next process removed it on first manager access. Downstream-error retention/recovery and final cache pair count also remained correct. - Ran a 1536x1536x360 float32 stress producing a 10,192,158,720-byte / 9.49 GiB final IMAGE. RAM and Disk-backed outputs were hash-identical and finite. Disk-backed retained one file only while cached and zero after release.
- Corrected the measurement contract: Windows RSS includes resident file-mapped pages, so stress RSS was about 10.44 GiB for both backends and is not a valid standalone proof of commitment savings. RAM private/USS measured 13,322,362,880 / 11,045,199,872 bytes; Disk-backed measured 3,122,954,240 / 855,834,624 bytes. The mapped backend therefore reduced anonymous private commitment by about 9.50 GiB and USS by about 9.49 GiB, matching the output size.
- Validation: focused acceptance regression
58 passed; full pytest420 passed; probepy_compileandgit diff --checkPASS. The first pytest attempt failed only because the inherited default temp/cache roots denied access; reruns used freshD:\Codex\_pytest_tempbases with cacheprovider disabled. Only the knownpynvmlFutureWarning remains. - P2-5 is formally PASS for manual RAM/Disk-backed V3.5 Assembly. Auto backend, arbitrary downstream full-copy behavior, new 3x5 sampler/model generation, and 3x5 Main Hi-res Fix 2x GPU/resource acceptance remain separate. No Production runtime file required synchronization. Commit/push not performed.
2026-08-24 - Phase 2A P2-6 Auto Buffer Backend
- Preserved all existing uncommitted work and accepted snapshots
pre-phase2a-p2-6-auto-source-verified-20260824_112744(426 files) andpre-phase2a-p2-6-auto-runtime-w-verified-20260824_112744(331 files), both with zero SHA-256 mismatch after excluding the pre-existing inaccessible.pytest_cache. - Added
Autoto only the independent V3.5 Assemble backend. New V3.5 assembler nodes default to Auto; saved manualRAMandDisk-backedvalues remain compatible. V3.4 paths and sockets were not changed. - Fixed policy: RAM requires final IMAGE
<= 4 GiBand available physical memory>= final IMAGE + max(4 GiB, 10% total RAM). Otherwise Auto selects Disk-backed. Missing memory counters select Disk-backed. Disk requires final IMAGE plus a 2 GiB reserve; measurement/capacity/allocation failure is explicit with no RAM fallback. - Added detailed decision reporting and CPU regression for safe RAM choice, size-limit Disk choice, memory-pressure Disk choice, missing counters, disk shortage, manual compatibility, actual selected writer, bit-exact output, and backing lifetime.
- Synchronized only
v3/file_backed_buffer.py,v3/assembly_v35.py, and a preserved-runtime merge ofv3/driving_nodes.pyto ComfyUI_W. The first two match source SHA-256; the third matches the prepared merge exactly and retains the pre-existing runtime Video Reference tooltip addition. - Headless real acceptance
P26-auto-ram-long-terminal-20260824_113400used accepted 3x5 FL2VA Long Terminal Merge physical latents plus Core GPU Video/Audio Decode. Auto selected RAM for the 1,769,472,000-byte final IMAGE and reproduced the accepted P2-5 640x640 / 360f / 15.000s video/audio hashes exactly. - Headless stress acceptance
P26-auto-disk-stress-20260824_113610selected Disk-backed for the 10,192,158,720-byte / 9.49 GiB final IMAGE, reproduced the accepted stress hashes, retained one pair only while cached, and reclaimed all managed files after release. - Validation: focused regression
66 passed; full pytest428 passed; source/runtimepy_compile, runtime verifier, andgit diff --checkPASS. Only the knownpynvmlFutureWarning remains. Commit/push not performed.
2026-08-24 - V3.5 release-candidate preparation
- Preserved all existing uncommitted changes and accepted verified snapshots
pre-v35-release-preparation-source-verified-20260824_115858(426 files) andpre-v35-release-preparation-runtime-w-verified-20260824_115858(331 files), both with zero SHA-256 mismatch. - Updated
version.py,pyproject.toml,metadata.ini, Changelog, package information/validation, and English/Japanese README to3.5.0. Added an automated metadata-consistency regression. - Promoted the prior local Hi-Res connection memo to
docs/V35_HIRES_FIX.md. The public guide distinguishes standard integrated Pixel/VAE Hi-Res Fix from the external-latent Advanced Second Pass Bridge, documents workflow-side SIGMAS, and warns that direct high-ratio latent interpolation is not the accepted Main path. - Classified public artifacts. Production modules, applicable CPU tests, the public guide, and Production-required
tools/p20_file_backed_tensor_poc.pyare included inMANIFEST.sha256. Generated media/results and diagnostic workflows are excluded; P2-4/P2-5 probes remain repository-only validation tools. - Targeted release/V3.5 regression is
75 passed; canonical full pytest is430 passed. Source and synced ComfyUI_W runtime verifier both report V3.5.0, native PackedLayout PASS, Fixed 3x5 prompt-plan PASS, and expected node registration. Only the knownpynvmlFutureWarning remains. - Synchronized only release metadata, READMEs, Manifest, the public V3.5 guide, and its metadata regression to the already snapshotted ComfyUI_W runtime, with per-file SHA-256 equality. Runtime-specific Production differences were not overwritten.
- Representative release GPU revalidation used the accepted Long Terminal Merge revision through Core GPU Decode and Assemble V3.5 Auto. Result: PASS, 3 logical / 2 physical groups (
37T + 77T), 640x640, 360f, 24 fps, 15.000s, exact accepted video/audio hashes, finite outputs, and zero managed backing files after release. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\V35-release-candidate-auto-long-terminal-20260824\p25_backend_result.json. - Main Hi-Res Fix remains Experimental because 3x5 2x and Reference/Hybrid-specific GPU acceptance are pending. Commit/push not performed.
2026-08-24 - V3.5 pending GPU acceptance
- Preserved all uncommitted work. Verified pre-documentation snapshots are
pre-v35-pending-gpu-results-source-verified-20260824_130824(431 files) andpre-v35-pending-gpu-results-runtime-w-verified-20260824_130824(334 files), excluding.gitand the pre-existing inaccessible.pytest_cache. - Started ComfyUI_W headlessly on
127.0.0.1:8188and submitted workflows through its HTTP API. Test GPU was an NVIDIA GeForce RTX 5060 Ti with 16,311 MiB VRAM. No Production code was changed. - Main Hi-Res Fix 3x5 FL2VA 2x used 576x576 First Pass, Lanczos 1152x1152 target,
res_multistep, and workflowsimple / 10 steps / denoise 0.35. First Pass retained the Long Terminal Merge contract: 3 logical chunks, 2 physical groups (37T + 77T), logical[2,3]shared terminal sample, and 362 pre-assembly frames. The 37T Second Pass group completed 10/10; the terminal 77T group failed at its first Second Pass inference with CUDA OOM (15.93 GiB limit, 10.04 GiB allocated, 1.49 GiB requested). Result: FAIL on the tested 16 GiB GPU; prompt ide6042730-1618-4967-9e3b-06125ff8ee35. - Hybrid FL2VA + Reference 1x5 integrated Main Hi-Res Fix 2x passed. It used one captured context-aware physical group, one Second Pass sampling call, Lanczos 576-to-1152, actual refine seed
2946484852083492747, and bit-exact first-pass audio LATENT passthrough. Core Decode / Assemble produced 1152x1152, 120 frames, 24 fps, 5.000 seconds, 32 kHz stereo. Visual contact-sheet inspection found no persistent colored line/plate artifact, flash, or temporal collapse. OutputV35_RC_HybridReference_MainHiRes_1x5_576to1152_d035_00001_.mp4, SHA-256E5B27CABCD5FBC1422F169AC6E75570EF4D54010CD04A3439852D8DED15C1384; prompt idc48dd1de-dd45-4896-b39e-45478d971194. - Direct
H3ContinuumSecondPassV35Hybrid/Reference 1x5 also passed independently at 576x576. It reportedconditioning_source=refine_context, actual refine seed5765580629374593532,(1, 24, 37, 36, 36), one sampling pass, temporary refined audio discard, and bit-exact first-pass audio passthrough. Core Decode / Assemble produced 120 frames / 5.000 seconds; visual inspection was clean. OutputV35_RC_HybridReference_MainHiRes_1x5_576to1152_d035_00002_.mp4(filename retained from the reused workflow prefix), SHA-25667311A797CD812584B4E2CCC21D7DF83AE8C1F1D72AB8BF5FA1C142D00894D2D; prompt id01b8a2a6-245d-4d29-bc4d-0e92426308d2. - The Hybrid prompt omitted angle brackets around
Picture 3, generating a non-blocking H3C-P103 warning. Connected Reference conditioning remained active; this is accepted for execution/transport but not a warning-free public prompt example. - Final validation after documentation synchronization: full canonical pytest
430 passed; ComfyUI_W runtime verification passed package 3.5.0 registration, native PackedLayout, and Fixed 3x5 planning;git diff --checkpassed. README English/Japanese, the V3.5 guide, package validation, and Manifest match ComfyUI_W by SHA-256. - Remaining acceptance: Main 3x5 2x on more than 16 GiB VRAM or a lower-memory terminal 77T implementation, and Reference/Hybrid durations longer than 1x5. Hi-Res Fix remains Experimental. Commit/push not performed.
2026-08-24 - V3.5.0 formal README and release publication
- Preserved all accumulated uncommitted V3.4 Terminal Merge, E2-8, Phase 2A, V3.5, test, and release-preparation work. Verified snapshots are
pre-v35-formal-readme-release-source-verified-20260824_141358(431 files) andpre-v35-formal-readme-release-runtime-w-verified-20260824_141359(334 files), excluding.gitand the pre-existing inaccessible.pytest_cache. - Reorganized English/Japanese README around the two V3.5 additions: Continuum-aware Second Pass / experimental integrated Hi-Res Fix, and low-memory Assemble + Seam. Added concise wiring, the workflow-side SIGMAS baseline, OFF/no-Hi-Res performance behavior, 16 GiB GPU acceptance limits, and an explicit V3.4 saved-workflow compatibility statement.
- Added separate Hi-Res Fix and Second Pass node images without BasicScheduler, wires, or group labels. The Hi-Res image uses the accepted
lanczos / 2.00setting. Image generation/editing was limited to isolating the supplied UI and aligning that displayed option; README text remains authoritative for connections and SIGMAS. - Added the validated 9.49 GiB stress numbers to both READMEs: private memory 12.41 -> 2.91 GiB and USS 10.29 -> 0.80 GiB, with hash-identical RAM/Disk output. Explicitly documented that this is host-RAM/private-commitment reduction, not sampler VRAM reduction, and that Disk-backed can affect final assembly I/O rather than model sampling speed.
- Updated Changelog, package information, package validation, V3.5 guide wording, and the 159-entry Manifest. Full canonical pytest is
430 passed; every Manifest hash,git diff --check, and ComfyUI_W runtime verification pass. Nine changed public files/assets were synchronized to ComfyUI_W with per-file SHA-256 equality. - User explicitly authorized committing and pushing the complete V3.5.0 release state to
origin/main.
2026-08-24 - V3.5.1 Last Queued Seed reuse frontend fix
- Preserved all existing Issue #9/#10 and earlier uncommitted work. The standard snapshot script was attempted first and failed only on the known ACL-inaccessible top-level
.pytest_cache; accepted SHA-256-verified snapshots arepre-v351-last-queued-seed-fix-source-accepted-20260824_210526(467 files) andpre-v351-last-queued-seed-fix-runtime-w-accepted-20260824_210526(339 files), excluding.gitand.pytest_cache, both with zero mismatch at source revisionee7b9937dc9bda5b4e55b012e8e725b7afc5f3aa. - Added session-only V3.5 sampler Seed state in
web/last_queued_seed.js:last_queued_seed,auto_updated_seed,previous_control_mode, andauto_update_pending.beforeQueuePromptcaptures the submitted Seed;Randomize -> Fixedrestores it only when the displayed value still equals the exact ComfyUI automatic After Generate update. A manual Seed edit clears pending restoration. Linked/missing Seed inputs clear the capture. - No Workflow JSON state, new widget, backend change, cache manipulation, Run Storage relaxation, Sampling/Second Pass/Terminal Merge/Assembly/Seam change, V3.4 behavior change, Node ID change, or display-name change was introduced. Seed restoration only restores normal cache eligibility; an absent cache regenerates normally with the same Seed.
- Added
tests/test_v351_last_queued_seed_frontend.pywith an actual Node.js module harness covering automatic restore, manual override, consecutive Randomize queues, Before Generate/no-update behavior, Fixed baseline, and linked Seed cancellation. Focused frontend regression is5 passed; JavaScript syntax passes; canonical full pytest is445 passedwith only the knownpynvmlFutureWarning; source Manifest has 165 entries with zero missing or mismatched hashes;git diff --checkpasses. - Synchronized only
web/project_id.jsand newweb/last_queued_seed.jsto ComfyUI_W; both match source SHA-256. The runtime-specific Manifest received only the matching frontend entry updates; its pre-existing unrelated divergence from the canonical source Manifest was not normalized or overwritten. Runtime JavaScript syntax, compact-UI regression (2 passed), package registration, native PackedLayout, and Fixed 3x5 planning pass. RealRandomize -> Fixed -> Second Pass ONbrowser/cache acceptance remains pending; ComfyUI_W was not restarted or reloaded. Commit/push not performed.
2026-08-24 - V3.5.1 Hi-Res Fix Core canvas alignment
- Preserved all existing Issue #9/#10/Seed Fix working-tree changes. The standard snapshot script again encountered only the known top-level
.pytest_cacheACL denial; accepted SHA-256-verified snapshots arepre-v351-hires-core-alignment-source-accepted-20260824_213303(470 files) andpre-v351-hires-core-alignment-runtime-w-accepted-20260824_213303(340 files), excluding.gitand.pytest_cache, with zero mismatch at source revisionee7b9937dc9bda5b4e55b012e8e725b7afc5f3aa. - Diagnosed the GPU failure
shape '[1, 24, 1, 1, 21, 2, 21, 2]' is invalid for input of size 44376: manual576x576 * 1.2usedround(36 * 1.2) = 43, while Core H3 requires a 32-pixel canvas grid / even latent dimensions for its2x2spatial patchifier. - Added Core-derived manual target alignment in
v3/hires_fix_nodes.py. The failing case now produces704x704/44x44; rectangular1024x576 * 1.2produces1216x704/76x44. No Sampling, Conditioning adaptation, Terminal Merge, Assembly, Seam, Run Storage, node schema, ID, or display name changed. - Added direct and integrated square/rectangular regressions. Focused Hi-Res/Second Pass result is
48 passed; full source pytest is448 passedwith only the knownpynvmlFutureWarning. An initial full run's 97 setup errors were solely the existing inaccessible global pytest temp root and disappeared with an explicit writable--basetemp. - Synchronized only
v3/hires_fix_nodes.pyto ComfyUI_W and verified exact SHA-256 equality. Runtime verifier passed package registration, native PackedLayout, and Fixed 3x5 planning. The running ComfyUI process was not restarted, so GPU re-acceptance remains pending after a deliberate restart. The attempted Seed cache run was not valid because its second prompt was queued before the first completed; logs show First Pass20/20ran again. Commit/push not performed.
2026-08-25 - V3.5.1 Reference Audio conditional UI and naming
- Preserved all existing Issue #9/#10/Seed/Hi-Res working-tree changes. The standard snapshot attempt failed only on the known ACL-inaccessible top-level
.pytest_cache; accepted SHA-256-verified snapshots arepre-v351-reference-audio-conditional-ui-source-accepted-20260824_230702-ee7b993(470 files) andpre-v351-reference-audio-conditional-ui-runtime-w-accepted-20260824_230702(340 files), excluding.gitand.pytest_cache, with zero mismatch. - Added
web/reference_audio_ui.js. V3.5 Reference Audio sockets default toHiddenwhen unlinked, can be restored withShow, retain their original input positions, auto-open for connected saved workflows, and cannot be hidden while linked. No graph link is removed. The UI widget is attached only after Core graph configuration and removed from serialized Workflow widget values to avoid reload shifts. - Clarified only human-facing labels/tooltips:
Video Guide Frames,Driving Audio,Driving Audio VAE,Reference Audio (Optional), andReference Audio VAE (Optional). Backend input keys, ordering, types, Node IDs, conditioning paths, Sampling, Terminal Merge, Assembly, Seam, and Run Storage contracts are unchanged. - Live Chrome/ComfyUI testing on the already-running
127.0.0.6:8188instance found and corrected two integration-specific issues that the initial mock did not expose: current Core keeps widget descriptors after the Reference Audio sockets innode.inputs, and UI injection before graph configuration can shift saved widget values. Final live acceptance passedHidden,Hidden -> Show -> Hidden, original-position restoration, connected auto-Show, and linked hide protection. Temporary workflows were closed without saving; the template workflows were not modified. - Added/updated real JavaScript harness coverage for Core-style non-terminal inputs, disconnected
-1, workflow serialization, labels, show/hide, link protection, callback chaining, and absent-input recovery. Focused validation is20 passed; canonical full pytest is450 passed; JavaScript syntax,git diff --check, and ComfyUI_W runtime verification pass with only the knownpynvmlFutureWarning.
2026-08-25 - V3.5.1 release documentation and naming
- Renamed the displayed video-guide inputs to
Video Guide FramesandVideo Guide Size; backend keys, input order, Node IDs, graph links, and saved-workflow compatibility remain unchanged. - Expanded English/Japanese README guidance for V3.5.1 Reference Audio, Conditioning Bridge, Last Queued Seed Reuse, and the distinct roles of still Reference Images, Video Guide Frames, Driving Audio, and Reference Audio.
- Updated package version metadata, Changelog, package information, package validation, release tests, project state, and distribution Manifest for V3.5.1.
- User explicitly authorized committing and pushing the complete V3.5.1 release state after validation. The supplied screenshots still showed the old display label, so no stale screenshot was added to the release documentation.
- Replaced the initially supplied stale/shifted screenshots with three user-approved normal-state images showing
Video Guide Frames,Reference Audio Inputs: Hidden, andShow; both README editions reference the same assets. - Final release checks: canonical ComfyUI_W-root pytest
450 passedwith only the knownpynvmlwarning; JavaScript syntax andgit diff --checkpass; all 172 Manifest entries and README image/workflow references verify. - Added the user-supplied LBH + Conditioning Bridge workflow as a separate public connection example and a lightweight SVG flow chart to both README editions. The recommended template remains unchanged, and Core nodes retain their standard display titles.
- Synchronized only the relevant implementation/tests to ComfyUI_W and verified per-file SHA-256 equality. No ComfyUI process was started or stopped. Commit/push not performed.
2026-08-25 - Issue #11 conditional size-control documentation
- Confirmed that Issue #11 reports a discoverability/documentation gap rather than missing V3.5.1 inputs: Compact UI hides
Reference SizeandVideo Guide Sizeuntil their matching image inputs are connected. - Added matching English/Japanese README guidance covering the conditional display, the
Video Reference SizetoVideo Guide Sizedisplay-name clarification, internal aspect-preserving resize before VAE encoding, the no-automatic-upscale behavior, and the limited cases where external resize is useful. - No node schema, backend key, workflow, runtime implementation, Sampling, Conditioning, Terminal Merge, Assembly, Seam, or Run Storage behavior changed.
- Pre-change snapshot:
pre-issue11-readme-clarification-source-accepted-20260825_001938(478 files), excluding.gitand the known ACL-inaccessible.pytest_cacheafter the standard snapshot script encountered that existing ACL condition.
2026-08-25 - V3.5.1 public-state cleanup
- Updated the project scope rule from the obsolete V3.4 public baseline to V3.5.1, preserving V3.4 node IDs, contracts, and saved-workflow compatibility.
- Aligned
PACKAGE_VALIDATION.txtwithPROJECT_STATE.md: Last Queued Seed Reuse has automated frontend validation, while real UI/cache acceptance remains pending. - Posted implementation answers to GitHub Issues #9, #10, and #11. Closed #9 Conditioning Bridge and #10 Reference Audio as completed; retained #11 as open until its local README clarification is published.
- No implementation code, node schema, workflow, runtime, Sampling, Conditioning, Terminal Merge, Assembly, Seam, or Run Storage behavior changed.
- Pre-change snapshot:
pre-v351-public-state-cleanup-source-accepted-20260825_004006(478 files), excluding.gitand the known ACL-inaccessible.pytest_cacheafter the standard snapshot script encountered that existing ACL condition. - Validation: release metadata
2 passedwith the ComfyUI_W venv, all 172 Manifest entries match SHA-256, andgit diff --checkpasses. The system Python lacks pytest; no product-test failure occurred.
2026-08-25 - GitHub Actions dependency and Core stub repair
- Kept Production runtime code and
requirements.txtunchanged. Addedpsutilonly to the GitHub Actions test-install command for the P25 probe tests. - Added a test-file-local
comfy_extras.nodes_minimax_h3stub with onlyCANVAS_MULTIPLE = 32, allowing manual Hi-Res canvas tests to exercise the real Production import path without installing ComfyUI Core in CI. - Focused Hi-Res Fix/P25 result is
16 passed; full result is450 passedwith only the knownpynvmlFutureWarning. The accepted full run used a dedicated writable--basetempafter the known global Windows pytest temp ACL caused setup-only errors. - No Production runtime contract, Hi-Res Fix algorithm, Node ID, V3.4/V3.5 compatibility path, or runtime dependency changed.
- Pre-change snapshot:
pre-ci-dependency-stub-fix-source-accepted-20260825_005009(478 files), excluding.gitand the known ACL-inaccessible.pytest_cacheafter the standard snapshot script encountered that existing ACL condition. - Confirmed the prior README clarification was already on public
mainand closed GitHub Issue #11 as completed. Issues #9, #10, and #11 are now all closed with implementation answers.
2026-08-25 - V3.5.1 Reference Audio permanent-socket compatibility hotfix
- The standard snapshot script stopped only on the known ACL-inaccessible
.pytest_cache. Accepted snapshots arepre-v351-reference-audio-permanent-sockets-source-accepted-20260825_024425_613(478 files) andpre-v351-reference-audio-permanent-sockets-runtime-w-accepted-20260825_024426_476(350 files), excluding.gitand.pytest_cachewith matching source/destination file counts. - Removed all frontend
addInput/removeInputbehavior forreference_audio_1andreference_audio_vae, the UI-onlyReference Audio InputsHidden/Show widget, its connection callbacks, and its serialization guard. Both sockets now remain permanently present exactly as defined byH3ContinuumSamplerV35.INPUT_TYPES. - Retained only human-facing label normalization for Video Guide Frames, Driving Audio, and Reference Audio. Node IDs, backend socket keys/order/types, Sampling, Conditioning, Last Queued Seed Reuse, V3.4 behavior, and all other conditional widgets are unchanged.
- Added a real JavaScript save/reload regression that serializes the complete 20-value V3.5 widget sequence, reloads it positionally, and verifies every value remains aligned while both Reference Audio sockets retain their names and order. The bundled LBH/Conditioning Bridge example now has the same aligned 20-value sequence and both permanent sockets. Focused validation is
13 passed; full validation is452 passed; JavaScript syntax andgit diff --checkpass. The only warning is the knownpynvmlFutureWarning. - Synchronized only
web/project_id.jsandweb/reference_audio_ui.jsto ComfyUI_W and verified exact source/runtime SHA-256 equality. The two matching runtime Manifest entries were updated without normalizing its pre-existing unrelated 30 missing/divergent entries. No ComfyUI process was started, stopped, or restarted; live workflow save/reload acceptance is left to the already-running user environment. - Removed the obsolete Hidden/Show comparison article and its two screenshots. Replaced the input overview with the user-supplied permanent-socket view and added the user-supplied full sampler view showing the corrected widget alignment.
2026-08-25 - Optional Windows pinned-memory README tip
- Added a short English-only optional tip for
--disable-pinned-memorynear the measured memory section. It records the local RTX 5060 Ti 16 GB / RAM 64 GB0.4 MP, 2 x 5sresult (5m 23s, approximately 6.1 GB sampler RSS and 8.6 GB after assembly) without presenting the flag as a default, universal crash fix, or guarantee against Shared GPU Memory use. - Pre-change snapshot:
pre-readme-pinned-memory-tip-source-accepted-20260825_031002_500(477 files), excluding.gitand the known ACL-inaccessible.pytest_cache.
2026-08-25 - Last Queued Seed promptQueued observer candidate
- Replaced the unsuccessful widget callback,
afterQueued, andonWidgetChangedwrapping approach instead of stacking another mechanism on top.beforeQueuePromptnow stages the V3.5 Randomize seed, the officialpromptQueuedevent accepts it after a successful queue, and a pending-only 50 ms observer reads currentnode.widgetsuntil it restores the accepted seed onFixedor cancels on a manual seed change. - Queue validation failures cannot update
last_queued_seed; a later single-queue attempt replaces any unaccepted candidate. Restoration changes the visible widget value directly and never mutates the API prompt behind the UI. - Accepted exact-file snapshots are
pre-last-seed-observer-source-files-20260825_175117(4 files),pre-last-seed-observer-runtime-files-20260825_175117(2 files),pre-last-seed-observer-docs-20260825_175652(2 files), andpre-last-seed-observer-runtime-manifest-20260825_175724(1 file). The standard full snapshot attempt stopped only on the known ACL-inaccessible.pytest_cachebefore any edit. - Focused frontend/compact-UI validation is
9 passed; full source pytest is453 passedwith only the knownpynvmlFutureWarning. JavaScript syntax passes, andweb/last_queued_seed.jsplusweb/project_id.jsmatch ComfyUI_W by SHA-256. Live UI transition and First Pass cache reuse remain pending; commit/push were not performed.
2026-08-25 - Last Queued Seed experiment withdrawn
- Real-browser Randomize-to-Fixed pairs repeatedly submitted different base seeds, including
00043(432308459616145, Randomize) and00044(238985343135586, Fixed). Each file's Workflow/UI seed matched its submitted Prompt seed, proving that the restoration itself failed rather than metadata being stale. - Removed only the Last Queued Seed frontend integration, its dedicated module, and its simulated regression.
H3 Continuum Sampler V3.5again delegates Seed andcontrol after generateentirely to ComfyUI. Reference Audio UI, compact UI settings, and all unrelated Production paths are unchanged. - GitHub review found a valid alternative pattern in rgthree's dedicated Seed node: it removes the native control, finalizes the Prompt/Workflow seed immediately before queueing, and exposes an explicit
Use Last Queued Seedbutton. That design is materially different from transparently intercepting an existing Sampler's Randomize-to-Fixed transition and was not added to Continuum. - Accepted exact-file snapshots are
pre-remove-last-queued-seed-source-files-20260825_184039(10 files) andpre-remove-last-queued-seed-runtime-files-20260825_184039(4 files). The standard snapshot attempt stopped only on the known ACL-inaccessible.pytest_cachebefore any edit. - Focused compact/reference-audio/release validation is
8 passed; full source pytest is450 passedwith only the knownpynvmlFutureWarning. JavaScript syntax and runtime verification pass; the seed-only removal is synchronized to ComfyUI_W. Commit/push were still pending at this checkpoint.
2026-08-27 - V3.6.1 non-Balanced Audio Continuity fallback
- Confirmed the public V3.6 contract mismatch: Standard selected the 22-frame-only Masked AV transport whenever Audio Continuity was ON, while Fast, Strong, and Auto could resolve to 5 or 39 frames and reach an internal contract stop; Strong 39 also had an explicit Long Terminal Merge rejection.
- Added pre-execution routing in
v3/driving_nodes.py: Balanced 22 stays Masked AV, Audio Continuity OFF stays Masked Video, and Fast 5 / Strong 39 / Auto with Audio Continuity ON use accepted Reference Context from the start. Compatibility is unchanged. Status and tooltip diagnostics describe the resolved transport without rewriting UI values. - Added routing, report, unchanged-path, and legacy Reference Run Storage identity regressions. Focused validation is
108 passed; final full source pytest is543 passed; all 178 Manifest entries, compileall, andgit diff --checkpass. V3.6.1 version/README/CHANGELOG/package metadata now include this fix together with the already-pushed Prompt Parser hotfix. - Pre-change snapshots are
pre-v361-continuity-fallback-source-accepted-20260827_194154-e626053(598 source files) andpre-v361-runtime-w-sync-accepted-20260827_195452(350 runtime files). Synchronizedv2/prompts.py,v3/driving_nodes.py, andversion.pymatch ComfyUI_W by SHA-256; the V3.6.1 runtime verifier passes. Representative GPU Smoke remains pending; commit/push were not performed.
2026-08-27 - Issue #13 contrast / oily drift investigation
- Preserved all existing V3.6.1 release/fallback changes. The standard snapshot attempt encountered only the known ACL-inaccessible
.pytest_cache; accepted source snapshot ispre-issue13-drift-investigation-accepted-20260827_204413-e626053with 598 files at revisione62605389302192394446645cf3483fff08dfbf2. - Added two non-Production diagnostic tools:
tools/issue13_drift_gate_runner.pyfor matched five-case ComfyUI API execution andtools/issue13_drift_analyze.pyfor 25-frame-per-chunk color metrics and a contact sheet. They are not imported or registered by the package. - GPU matrix completed on ComfyUI_W 0.34.0 / RTX 5060 Ti 16 GB: Fixed Standard, same-text List Standard, same-text Timeline Standard, varied List Standard, and Fixed Compatibility. Every case produced 600 frames / 25 seconds at 384x384 without execution error, NaN/Inf, or OOM.
- Fixed, same-text List, and same-text Timeline are decoded-stream bit-identical and use identical resolved prompt hashes. Prompt Format itself and repeating the same text through a different parser mode are rejected as the cause.
- Compatibility / Reference Context reproduced cumulative contrast and high-saturation-tail growth; Standard Masked AV did not. This supports legacy Reference Context visual self-conditioning as the most likely cause under the controlled condition. The exact reporter workflow is still unavailable, and model/sampler/seed generality was not claimed.
- Evidence is under
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-drift-20260827_204827, includingISSUE_13_INVESTIGATION_REPORT.md, five-case prompts/histories,analysis.json,chunk_metrics.csv, andcontact_sheet.png. - Diagnostic
py_compilepasses; full pytest is543 passedwith only the knownpynvmlFutureWarning; the ComfyUI_W V3.6.1 runtime verifier andgit diff --checkpass. Production, frontend, schema, version, commit, push, and release were unchanged.
2026-08-27/28 - Audio Continuity Research Gate
- Preserved all existing V3.6.1 working-tree changes and created accepted snapshot
pre-audio-continuity-research-gate-accepted-20260827_231806-e626053(602 files, excluding.gitand the inherited ACL-inaccessible.pytest_cache). - Added tests for 24 fps / 40 Hz grid offsets, cumulative PCM boundaries, 39f/65T State retention, Audio40 end alignment, and non-mutating Seam diagnostics. No Production implementation was changed.
- Added private research-only Audio40 routing to the diagnostic node plus API/output analysis tools. Ran six 3x5-second GPU outputs: music, dialogue, and ambient, each at 37T and 40T with fixed seed, 576x576, 20 steps, Balanced 22, Audio Continuity ON, and Seam OFF.
- All outputs completed and retained 360f / 15.000s / 32 kHz stereo contracts with finite latents. 40T added only six packed rows but did not improve audio boundaries consistently: Ambient improved, Dialogue was mixed, and Music immediate jumps worsened. Production remains at Audio37T.
- Generated machine-readable summaries, finalized-output metrics, boundary contact sheets, and 12 listening clips under
D:\Codex\_test_results\ComfyUI-H3-Continuum\audio-continuity-research-20260827_2335. The detailed report isAUDIO_CONTINUITY_RESEARCH_REPORT.md. - Final validation: focused
33 passed, related115 passed, full552 passed; compileall, V3.6.1 ComfyUI_W runtime verifier, diagnostic source/runtime SHA-256, empty queue, andgit diff --checkPASS. No commit/push/version/release was performed; ComfyUI_W remains running.
2026-08-30 - V3.7 Turbo 0.3 MP motion-graphics preview benchmark
- Accepted a user-reviewed V3.7.0 I2VA + Continuation preview at 480x640, 24 fps, 2x10 seconds, 480 final frames, with Standard Balanced 22, Audio Continuity On, Sage On, Spectrum Off, and Second Pass Off. Video/Audio generation, decode, exact-duration assembly, and output completed normally; the user observed clear 0.3 MP output with no obvious artifacts.
- The exact MP4-embedded graph used
minimax_h3_fl2va_pruned_int8_convrot.safetensorsplus the non-LightXminimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensorsat strength 1.0, Euler/simple, and 8 evaluations per physical group. It must not be labeled as a LightX result. - Sampling was
270.800908 stotal (125.619138 s+145.181843 s). Continuum reported359.54 sinternal total; the user-observed total was378.87 s. Typography was too condensed and the character too static/phone-centric, both classified as prompt-design findings rather than runtime correctness failures. - Evidence is fixed under
D:\Codex\_test_results\ComfyUI-H3-Continuum\v37-turbo-03mp-motion-graphics-20260830, including the output, First Frame, Detailed Report, contact sheet, MP4-embedded execution workflow/API prompt, later edited workflow, machine-readable summary, and SHA-256 list. The MP4-embedded metadata is the execution source of truth because the Downloads workflow contains a later prompt. - Pre-change snapshot:
pre-v37-turbo-03mp-benchmark-record-approved-excluding-pytest-cache-20260830_151823(659 files, excluding.gitand the inherited ACL-inaccessible.pytest_cacheafter the standard snapshot script stopped on that directory). No Production/runtime/workflow/version behavior changed; no commit or push was performed.
2026-08-30 - V3.8 Easy Mode v1.1 Phase 2
- After explicit approval, fixed the v1.1 decisions: 10-second-centered Duration resolution, Run Storage Off by default, no Scheduler socket, decimal MP, Native long-edge cap priority, and Audio VAE excluded from the required Easy contract.
- Added
H3ContinuumEasyV38as a five-output facade overH3ContinuumSamplerV37. It resolves Prompt/Auto format, 4-480-second total Duration, First Image aspect or 16:9 fallback, Draft/Balanced/Native/Custom resolution, Balanced 22, Standard continuation, Audio Continuity On, and existing Run Storage arguments without copying Sampling, Scheduler, VAE decode, continuation, or seed derivation logic. - Added the Easy frontend for Randomize/Fixed mapping to native
control_after_generate, conditional Custom MP and Run Name visibility, and hidden stable project identity. Added focused Resolver/schema/registration/V3.7-mapping/frontend tests. The approved 10/15/20/30-second FL2VA mappings are explicitly covered as non-merged because the Production Terminal Merge contract remains limited to multiple 5-second chunks. - Snapshot
pre-v38-easy-phase2-approved-excluding-pytest-cache-20260830_170711-0a7c670contains 659 source files with a matching relative path set;.gitand the inherited ACL-inaccessible.pytest_cacheare excluded. The standard snapshot attempt stopped on that cache before any edit. - Validation: focused
29 passed; related V3.7/V3.6/frontend63 passed; full source665 passed; all 193 Manifest entries, changed Python and JavaScript syntax, andgit diff --checkPASS. Only the knownpynvmlwarning remains. No runtime sync/restart, GPU execution, Minimal Workflow, Easy References helper, package-version change, commit, push, or release was performed. - After separate approval, snapshotted the ComfyUI_W runtime as
pre-v38-easy-runtime-w-sync-approved-20260830_181803-runtime-370(365 files, path-set match), synchronized the Easy Node registration/backend/frontend and focused test with exact SHA-256 equality, and restarted the same venv/Sage backend on127.0.0.6:8188. Live API registration returns display nameH3 Continuum Easy V3.8, the approved sockets, First/Last optional inputs, five outputs, and an empty queue. - The runtime was already a deliberately partial package: its pre-change 190-entry Manifest had 56 existing missing/hash differences. Replacing it with the source 193-entry Manifest did not make that partial runtime verifiable, so the runtime-local Manifest was restored from snapshot instead of expanding the sync to unrelated docs/tests/assets. No GPU generation, package-version change, commit, push, or release was performed.
2026-08-30 - V3.8 Easy References still-image scope
- Added public helper node
H3ContinuumEasyReferences/H3 Continuum Easy Referenceswith three permanent optional IMAGE sockets and oneH3_CONTINUUM_EASY_REFERENCESoutput. No dynamic sockets or Audio/Video inputs were added. - Added one optional
referencessocket toH3ContinuumEasyV38. The bundle expands directly to V3.7reference_image_1,reference_image_2, andreference_image_3; the helper does no resize or conditioning. Existing V3.7 preprocessing remains the single owner of gap compaction, Match Output geometry, conditioning, and<Picture N>numbering. - Focused coverage includes no Reference, 1, 1+2, 1+2+3, Reference 2 only, First+Reference, First+Last+Reference, output geometry, Picture numbering, and direct V3.7 mapping parity. Results: focused
41 passed, related Reference/Audio/Timeline100 passed, full source677 passed; 193-entry Manifest, Python syntax, andgit diff --checkPASS. Only the knownpynvmlwarning appeared. - Accepted source snapshot:
pre-v38-easy-reference-images-phase2-approved-20260830_184252-0a7c670(666 files, exact relative-path match, excluding.gitand.pytest_cache, revision0a7c6703b7715689c7aab4177e44e5fcd318fe5b). Accepted exact runtime-file snapshot:pre-v38-easy-reference-images-runtime-sync-20260830_185057-runtime(3 files). - Synchronized only
nodes.py,v3/easy_nodes.py, andtests/test_v38_easy_mode.pyto ComfyUI_W with exact SHA-256 equality. After the active generation completed and ComfyUI_W restarted, live API registration confirmed the helper display name, all three fixed IMAGE inputs, the custom output type, the Easyreferencesinput, and an empty queue. - Production Sampling, V3.7 direct inputs, Reference preprocessing, frontend JavaScript, package version, and Minimal Workflow are unchanged. No GPU gate, commit, push, or release was performed.
2026-08-30 - V3.8 Easy Mode Phase 3 GPU acceptance and Minimal Workflow
- Added reusable diagnostic tool
tools/v38_easy_gpu_gate_runner.py, starting from successful Easy API historyb89c22a0-7bb9-432b-9132-0f906652e4f6. Its seven-case dry-run fixed one seed and generated explicit Sage, Sage+Sol, Sage+Spectrum, and Sage+Sol+Spectrum routes plus 15/20/30-second functional cases before any GPU submission. - GPU Gate passed 7/7 on ComfyUI_W / RTX 5060 Ti 16 GiB. The four identical I2VA+Reference 10-second profile runs completed in 302.979s / 211.546s / 146.037s / 142.920s. First+Last+Reference 15s completed in 272.906s, Reference Image 2 only 20s in 371.483s, and no-Reference T2VA 30s in 514.388s.
- Duration mapping, reference-gap compaction,
<Picture 2>/<Picture 3>hybrid numbering, conditioning mode, exact output geometry, 24 fps frame count, duration, and 32 kHz stereo contracts passed. No Picture warning, execution error, interruption, NaN/Inf report, Decode failure, Assembly failure, or OOM occurred. - Highest device-wide VRAM use was 15.73 GiB and minimum available system RAM was 1.63 GiB. The result is accepted for correctness but records limited 16 GiB headroom. Full evidence and the report are under
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-easy-phase3-gpu-gate-20260830_202000. - Created official
examples/workflows/MiniMax_H3_Continuum_Easy_V38.jsonfrom the user-provided Reference workflow. It has 26 nodes after removing execution-inactive Group Bypassers and Notes, approved Easy defaults, one active Reference, optional Reference 2/3 and Last inputs, explicit Picture prompt text, and standard Core display titles. Added four static workflow regressions. - Final validation: Easy-focused
45 passed, full source682 passed, 195-entry Manifest, runner syntax, andgit diff --checkPASS. Only the knownpynvmlwarning remains. Pre-change snapshot:pre-v38-easy-phase3-gpu-gate-approved-excluding-pytest-cache-20260830_201345-0a7c670(666 files, zero path/hash mismatch, excluding.gitand.pytest_cache, revision0a7c6703b7715689c7aab4177e44e5fcd318fe5b). Production/runtime code, package version, commit, push, and release were unchanged.
2026-08-30 - V3.8 Phase 3 status correction: not final or frozen
- The user explicitly rejected treating the Phase 3 result as a feature freeze, Finalization Phase, final V3.8 workflow, or final 3.8.0 specification. The 7/7 GPU PASS and 26-node workflow remain candidate evidence only.
- Feature scope, further phases, UI/workflow changes, Documentation, version 3.8.0, commit, push, and release remain undecided until separately approved. No Audio/Video feature deferral or V3.8.0 cutoff is fixed by Phase 3.
- Updated only current-state wording and appended this correction. No implementation, Workflow JSON, GPU evidence, test result, Manifest, runtime, version, or publication state changed. Pre-correction snapshot:
pre-v38-phase3-status-not-final-correction-excluding-pytest-cache-20260830_211557-0a7c670(670 files, zero path/hash mismatch, excluding.gitand.pytest_cache).
2026-08-30 - Easy Load Audio and Driving Audio prototype
- After explicit implementation approval, added generic
H3EasyLoadAudio/H3 Easy Load Audioby subclassing current CoreLoadAudio; no Core load, validation, fingerprint, upload/player, or file handling code was copied. Added one shared native-Bypass frontend helper and refactoredH3 Easy Load Imageonto it without changing its IMAGE/MASK contract. - Added optional
Driving AudioandDriving Audio VAEsockets to the Easy facade and mapped them directly to the unchanged V3.7driving_audio/audio_vaearguments. Reference Audio, Video Guide, Production Sampling, V3.7 direct schemas, package version, and release state were not changed. - Focused Backend/Frontend coverage passed
64tests before GPU. It covers Core delegation, native mode synchronization, ON/OFF/reload filename/player preservation, Image regression, no Audio, Audio+VAE mapping, Audio-only existing validation, and VAE-only normal generation. Full post-GPU regression passed694tests; 203/203 Manifest entries, syntax,git diff --check, runtime verifier, live registration, and empty Queue pass with only the knownpynvmlwarning. - Runtime source sync covered exactly six implementation files with SHA-256 equality. After ComfyUI_W restart, live API reported
H3 Easy Load Audio, requiredaudio, outputAUDIO, and Easy optional inputsfirst_frame,last_frame,references,driving_audio,audio_vaewith an empty queue. - GPU evidence passed 6/6: 10s OFF, 10s ON, 20s/2-chunk ON, and 10s ON -> OFF -> ON. All cases produced exact 24 fps / 32 kHz stereo media; OFF returned to generated audio with Audio VAE still connected, ON preserved deterministic saved PCM across repeated 10-second runs, 20-second audio had no boundary-delta spike, and no NaN/Inf/OOM occurred. Peak device VRAM was 15.05 GiB. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-easy-driving-audio-gpu-gate-20260830_223000\gpu-2. - Updated the candidate Easy workflow to 27 nodes with one default-bypassed H3 Audio loader and shared Core Audio VAE. Removed role-specific custom titles from the H3 Easy image loaders; group names carry roles, while loader node names remain recognizable. No Core node title was changed.
- Accepted snapshots: source
pre-easy-load-audio-driving-audio-approved-excluding-pytest-cache-20260830_220541-0a7c670(676 files) and runtimepre-easy-load-audio-driving-audio-runtime-w-approved-excluding-pytest-cache-20260830_221217-runtime(372 files), both verified with zero mismatch. No commit, push, tag, or release was performed.
2026-08-31 - V3.8 Main Sampler Resolution Preset prototype
- After explicit approval, extracted the accepted Easy Aspect/Preset/Custom-MP/Native-768 calculation into shared
v3/resolution.pyand retained Easy compatibility aliases with exact output parity. - Added public facade
H3ContinuumSamplerV38without changing V3.7. V3.8 removes onlywidth/heightfrom its schema, resolves geometry fromAspect,Resolution Preset,Custom MP, and optional First Image, then calls unchangedH3ContinuumSamplerV37.run()with resolved dimensions. Run Storage and refine/Second Pass continue to identify source geometry by final width/height. - Extended existing
project_id.jsProduction handling for V3.8 and added only Custom MP visibility. Added focused parity/schema/identity/frontend tests and a reusable three-case GPU mini-Gate runner; no official workflow or documentation surface was changed. - Validation passed: focused
97, frontend retry40, full source729; all 206 Manifest entries, Python/JavaScript syntax, the runtime verifier, andgit diff --checkpassed with only the knownpynvmlwarning. GPU passed 3/3 at 736x416 Draft Landscape, 576x1024 Balanced Portrait, and 768x1024 Native Auto + First Image; all outputs were 120 frames / 5.000s / 24 fps / 32 kHz stereo with finite audio and non-black/non-collapsed video samples. - Synchronized five implementation files to ComfyUI_W with exact SHA-256 equality. Used only a dedicated port-8190 backend for live registration/GPU execution, then stopped its PID tree and verified the Stability Matrix port-8188 instance remained alive. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-main-resolution-gpu-gate-20260830_234000. No version bump, commit, push, tag, or release was performed.
2026-08-31 - H3 Continuum Load Video Gate
- After explicit approval, added
H3ContinuumLoadVideo/H3 Continuum Load Videounder the existing Easy category. The node exposes only Enable Video, Core video upload/select, and Force Rate, with standard IMAGE/AUDIO outputs; no Utility category or Easy-facade embedding was added. - The implementation calls current Core
InputImpl.VideoFromFile(...).get_components()directly. Fraction-based frame selection preserves nominal duration for 30->24, 24->30, and 29.97->24 without changing audio. OFF resolves neither the file nor Core decode and blocks both outputs silently; missing audio blocks only AUDIO. - Validation passed with
18focused Easy-loader tests,10actual-Core focused tests, and739full tests. Manifest208/208, Python syntax, runtime verifier, source/runtime hashes, real Core decode, live object-info placement schema, andgit diff --checkpass with only the knownpynvmlwarning. - Synchronized only
nodes.pyand newv3/easy_video_nodes.pyto ComfyUI_W after runtime snapshot. A dedicated 8190 launcher/backend verified exact input/output registration and was then stopped; user-owned 8188 remained alive. No V3.7, workflow, README, CHANGELOG, version, GPU Sampling, commit, push, tag, or release change was made.
2026-08-31 - Load Video native Bypass appearance
- After explicit approval, connected the existing serialized Enable Video widget to ComfyUI native modes through the shared Easy Bypass helper. OFF selects mode 4 and therefore the standard theme Bypass purple; ON selects mode 0. External Bypass and workflow reload synchronize back without hardcoded colors or widget-array changes.
- Added
web/easy_load_video.jsplus regressions for OFF/ON, external mode changes, widget preservation, serialization, reload, and the absence of polling/hardcoded node colors. Focused Easy loaders passed20; full source passed741; Manifest209/209, Python/JavaScript syntax, runtime hashes, andgit diff --checkpass with only the knownpynvmlwarning. - Synchronized only the changed shared helper and new video frontend extension to ComfyUI_W after exact runtime snapshot. Backend ExecutionBlocker, Core decode, Force Rate, V3.7, workflows, version, commit, and push remain unchanged.
- The active user-owned 8188 backend was not restarted or stopped and already exposes the new extension through
/extensions; a ComfyUI page reload is sufficient to load the behavior.
2026-08-31 - V3.8 Review / Smart Regenerate Phase B pure resolver
- After Phase A approval, added only the pure
v3/review_control.pypolicy module andtests/test_v38_review_control.py; Production execution, Run Storage I/O/lifecycle, frontend, GPU workflows, schema/contract versions, and Manifest were not changed. - Fixed Full/Review/Continue/Regenerate/Finish decisions, one-physical-group review units, atomic final Terminal Merge pairs, review-head validation, one-shot intent, non-identity pause metadata, projected partial-prefix state, and explicit validation for missing heads, Smart/Manual conflicts, Run Storage Off, impossible terminal prefixes, and nonce overflow.
- Added the necessary
latest_branch_regenerate_frominput alongside latest effective nonce. Pure tests prove interrupted/in-progress resume and review-ready Continue retain the existing branch, repeated Smart Regenerate advances1 -> 2 -> 3, and Continue after nonce 2 remains on its regenerate boundary instead of returning to nonce 0. - Snapshot
pre-v38-review-phase-b-pure-resolver-approved-excluding-pytest-cache-20260831_012741-0a7c670verified 701 files with zero path/hash mismatch. Validation passed: focused36, related152, full777, and changed-file Python syntax; only the knownpynvmlwarning remains. No commit, push, release, runtime sync, backend restart, or GPU execution was performed.
2026-08-31 - V3.8 Review / Smart Regenerate Phase C physical execution cap
- After explicit approval, added only the Production-internal
max_new_physical_groupscap throughv2/sequence.py,v2/nodes.py, andv3/nodes.py.Noneretains full execution; a positive limit counts only newly sampled physical groups after reuse and is not exposed by any public node schema. - The stop boundary is after a complete normal chunk or complete Terminal Merge pair, including validation, entry/state update, and any Run Storage commit. Limited execution returns a valid non-empty contiguous prefix with matching session, latents, state, report, and assembly plan; no future dummy chunks are produced.
- Corrected V3 result handling to distinguish configured chunks from completed entries. Partial prefixes do not masquerade as one-chunk complete runs, and Last Frame preservation remains disabled until the configured sequence is complete. Partial refine capture stays explicitly unsupported/incomplete.
- Added
tests/test_v38_review_execution_cap.pycovering C0-C3, T0-T2 atomic behavior, normal and FL2VA full/incremental parity, Run Storage commit ordering, invalid caps, schema invisibility, and the Second Pass limitation. - Validation passed: focused
17, related216, full794, Python syntax, andgit diff --check; only the knownpynvmlwarning remains. Six protected lifecycle/frontend/Manifest/plan files match the accepted Phase C snapshot by SHA-256. - Snapshot
pre-v38-review-phase-c-physical-cap-approved-excluding-pytest-cache-20260831_015840-0a7c670contains 707 files with zero path/hash mismatch, excluding.gitand the inherited ACL-inaccessible.pytest_cache. Run Storage lifecycle, Phase B resolver, frontend/UI, GPU, workflows, versions, commit, push, and release remain unchanged; Phase D was not started.
2026-08-31 - V3.8 Review / Smart Regenerate Phase D Run Storage lifecycle
- After explicit approval, connected Phase B and Phase C only through Production-internal arguments. Run Storage now persists intentional
review_readyseparately frominterrupted, whilein_progressremains the commit-time status and all configured chunks producecomplete. - Added non-identity review metadata for the atomic review unit, pause reason, branch boundary, effective nonce, and review-control version without changing Run Storage schema 2, Sampling Contract 5, sampling identity, or existing seed derivation.
- Added deterministic validated-head selection within the existing nonce lineage. Ordering is effective nonce, persisted update timestamp, and revision ID; validation covers contract identity, Prompt/Reference/Driving Audio lineage, contiguous chunks, stored-file SHA, review geometry, Terminal Merge atomicity, branch boundary, and nonce. Corrupt newest candidates safely fall back to the latest valid head.
- Consecutive Smart Regenerate persists
1 -> 2 -> 3; Continue, Finish Remaining, abnormal restart, and Review-to-Full retain the resolved branch as specified. Review Finish stores the actual final generated group; Full Run complete does not gain review metadata. Legacy revisions still resume normally but cannot become Smart Regenerate heads. - Added
tests/test_v38_review_run_storage.pyand one Phase C integration case. Validation passed: focused25, Phase C combined43, related269, full820, Python syntax,git diff --check, and Manifest213/213; only the knownpynvmlwarning remains. - Snapshot
pre-v38-review-phase-d-lifecycle-approved-excluding-pytest-cache-20260831_021636-0a7c670contains 711 files with zero path/hash mismatch. Phase E frontend/UI, workflows, GPU, runtime sync, versioning, release documentation, commit, push, tag, and release were not started.
2026-08-31 - V3.8 Review / Smart Regenerate Phase E UI and Queue intent
- Exposed Generation Mode and Review Action only on the normal V3.8 sampler, appended after all existing V3.8 widgets to preserve prior positional values. V3.7 and Easy remain schema-invariant; frontend controls do not derive storage heads, chunk targets, nonce, branch boundaries, Terminal Merge, or physical limits.
- Audited ComfyUI frontend 1.49.6 and used the formal widget
afterQueuedcallback that runs only after successful prompt submission. The existingbeforeQueuePrompthook preserves Regenerate/Finish in the submitted prompt; only the widget resets to Continue afterward. Failed submission preserves the one-shot value, saved one-shot values normalize on workflow load, and no queue API monkey patch was added. - Full -> Review enables Save + Auto Resume when Off; Review -> Full retains it. Full Run submits Continue regardless of a stale action, while the backend resolver independently ignores stale actions. Smart plus Manual Regenerate remains a backend validation error without frontend target rewriting.
- Added backend Review Status formatting for ready units, Terminal Merge, Smart Regenerate, Finish Remaining, and completion, plus the explicit unsupported partial-review Second Pass warning. Sampling, Run Storage lifecycle/identity, seed derivation, and Terminal Merge execution were not changed.
- Validation passed: focused
86, related362, full827, Python/JavaScript syntax, andgit diff --check. Snapshotpre-v38-review-phase-e-ui-approved-excluding-pytest-cache-20260831_025634-0a7c670verified 715 files with zero missing/extra/hash mismatch. Runtime sync, live UI, GPU, Phase F, workflow/version/release text, commit, push, tag, and release were not performed.
2026-08-31 - V3.8 Review / Smart Regenerate Phase F GPU Gate STOP
- After explicit execution approval, verified the Phase E source/runtime state, created exact source/runtime snapshots, synchronized the seven changed runtime implementation files plus the current runtime verifier, and started a dedicated Sage backend on
127.0.0.6:8190. Runtime verifier and Review UI harness 7/7 passed; the user-owned 8188 backend was not stopped or restarted. - Added
tools/v38_review_phase_f_gpu_gate_runner.pyas Gate-only tooling. It intentionally sends identical fixed-seed V3.8 sampler inputs on consecutive Review Continue queues and does not add a test-only cache nonce. It captures Run Storage manifests, per-chunk video/audio tensor SHA-256, decoded RGB/PCM SHA-256, media contracts, and GPU/RAM metrics. - G0 Full T2VA 3x10 seconds at Draft 736x416 and 20 steps completed three chunks in 1053.884s. Output was 720 frames, 30.000s, 24 fps, and 32 kHz stereo. Peak device VRAM was 15.472 GiB, peak process RSS 46.555 GiB, and minimum system available RAM 0.575 GiB; no OOM or NaN/Inf occurred.
- G1 Queue 1 generated exactly Chunk 1, saved
review_ready, and exactly matched G0 Chunk 1 for seed, prompt hash, video/audio latent SHA-256, and plan. Queue 2 completed in 6.075s withH3ContinuumSamplerV38node 305 explicitly listed underexecution_cached; it performed no Sampling, left the manifest mtime unchanged, and remained at one chunk /review_ready. - Phase F was stopped at the first Production failure as required. G2-G5, full post-Gate regression, and release promotion were not run. Review Each Chunk and Smart Regenerate are HOLD pending a Review-specific ComfyUI cache invalidation / Queue intent contract and fixed-seed Continue / repeated-Regenerate integration coverage. Production code was not changed.
- Evidence is
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-gpu-gate-20260831_101000\PHASE_F_STOP_REPORT.md. The dedicated PID 28872 was stopped only after its Queue was empty; port 8190 became unavailable and the user-owned port 8188 remained HTTP 200. No workflow/version/release text, commit, push, tag, or release change was made.
2026-08-31 - V3.8 Review Phase F-R1 cache recovery
- Added
IS_CHANGEDonly toH3ContinuumSamplerV38: Full Run returns stableFalse; Review Each Chunk returnsfloat("NaN"). The method accepts the normal node inputs without changing their values. V3.7, Easy, Sampling, Run Storage, seed/prompt identity, frontend queue behavior, and manifests were not changed. - Added
tests/test_v38_review_cache.pywith direct policy coverage and an actual ComfyUI CoreCacheKeySetInputSignature/IsChangedCachesubprocess harness. Review Continue, Regenerate Current, and Finish Remaining invalidate on every Queue; Full Run remains cache-stable for all actions. Focused passed8, related passed259, and full source passed835; the runtime verifier, Python/JavaScript syntax, all217Manifest entries, andgit diff --checkpassed. - Added
tools/v38_review_phase_f_recovery_runner.py. It resumes the stopped G1 run, reuses the exact failed Queue 2 prompt, rejects sampler node 305 inexecution_cached, verifies one-new-physical-group progression and exact stored-prefix preservation, then compares final G1 to the saved G0 baseline down to decoded RGB/PCM hashes. - Recovery Queue 2 passed at two chunks /
review_readyin 430.143s with Chunk 1 preserved and only Chunk 2 generated. The identical Queue 3 prompt on the same backend passed at three chunks /completein 450.204s with Chunks 1-2 preserved and only Chunk 3 generated. G0/G1 seed, prompt hash, plan, video/audio latent, decoded video/audio, and media-contract parity are exact. Peak device VRAM was 15.43 / 15.49 GiB. - Evidence is
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-r1-20260831_111111\gpu-recovery\recovery_summary.json. Source/runtime snapshots arepre-v38-review-phase-f-r1-cache-invalidation-approved-excluding-pytest-cache-20260831_110319-0a7c670andpre-v38-review-phase-f-r1-runtime-sync-excluding-pytest-cache-20260831_110749-runtime, both exact excluding inherited.pytest_cache. - Synchronized only
v3/driving_nodes.pyto ComfyUI_W. The dedicated PID 13952 on 8190 was stopped after Queue empty. Port 8188 was unreachable before the run and was not touched. Smart Regenerate G2 was not run; no workflow/version/release text, commit, push, tag, or release was performed.
2026-08-31 - V3.8 Smart Regenerate Phase F-G2 GPU Gate
- Ran only G2 on a fresh fixed-seed T2VA 3x10-second Review Run Storage; G0/G1 were not repeated. Initial Q1/Q2 established a two-chunk nonce 0 baseline, then R1/R2/R3 regenerated only Chunk 2 and advanced nonce
0 -> 1 -> 2 -> 3. - All four Chunk 2 seeds and video latent hashes were unique. Chunk 1 seed, prompt hash, video latent, and audio latent remained exact across every regeneration. Each Queue had one sampler pass, sampler node 305 was never cached, all stored latents were finite, and no OOM/NaN/Inf/Decode/Assembly error occurred.
- Canonical revisions advanced
4762ea4bd196c3f3 -> a2b5a5c96b64bf3f -> 61c28a5b4322b6eb -> bd82c8661d8863ab. Stopped/restarted the dedicated 8190 backend after nonce 3; the Production disk probe restored revisionbd82c8661d8863ab, prefix 2, review unit 2, branch 2, and nonce 3. Continue preserved Chunks 1-2 exactly, generated only Chunk 3, retained the nonce 3 branch, and completed the sequence. - Sampling/API seconds: Q1 254/311.379, Q2 323/404.905, R1 323/404.718, R2 324/406.180, R3 324/404.887, Continue 327/457.637. Peak VRAM was 15.59 GiB, maximum process RSS 48.24 GiB, and minimum available RAM 3.16 GiB.
- The first diagnostic attempt stopped after Production R1 because the inherited evidence helper resolved reused prefix files under the current revision instead of their
storage_revision_id. Stopped the dedicated backend, corrected only the external evidence runner, validated it against the nonce 1 branch, and reran the complete Gate from a fresh Run Storage name. Production code remained unchanged. - Post-Gate regression passed:
835tests, Manifest217/217, runtime verifier, 57 Python files, 7 JavaScript files, eight source/runtime hash pairs, andgit diff --check. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g2-20260831_113611\PHASE_F_G2_PASS_REPORT.mdandgpu_retry1/g2_summary.json. Snapshotpre-v38-review-phase-f-g2-gpu-gate-20260831_113611-0a7c670contains 724 exact source files excluding.git/.pytest_cache. - Stopped only the dedicated 8190 backend after Queue empty. Port 8188 was already unreachable and was not touched. G3, workflow/version/release text, commit, push, tag, or release were not performed.
2026-08-31 - V3.8 Finish Remaining Phase F-G3 GPU Gate
- Ran G3 with separate fresh fixed-seed T2VA 3x10-second Full and Review Run Storage names. Full completed three Sampling groups; Review Queue 1 generated only Chunk 1 and saved
review_ready; Finish Remaining exactly reused Chunk 1, sampled only Chunks 2 and 3, and finalized all three chunks ascomplete. - Full/Finish parity passed for every chunk's seed, Prompt hash, video/audio latent tensor SHA, and plan, plus decoded RGB24 video, float32 32 kHz stereo PCM, 720 frames, 736x416 geometry, 24 fps, and 30.000-second duration. The normal Review branch remained boundary 0 / nonce 0 with final review unit 3.
- Sampling/API seconds were Full 252+324+325 / 1035.089, Review Queue 1 247 / 293.268, and Finish Remaining 321+319 / 759.298. Peak VRAM was 15.57 GiB, maximum process RSS 48.53 GiB, and minimum available RAM 4.18 GiB.
- The first diagnostic attempt stopped because the external runner incorrectly compared whole safetensors container SHA across separate revisions. Required tensor/identity/plan fields were already exact. Corrected only the evidence runner and reran the complete Gate under fresh
retry1names; Production code was unchanged. - Post-Gate regression passed:
835tests, Review one-shot UI7/7, Manifest217/217, runtime verifier, Python/JavaScript syntax, eight source/runtime hash pairs, andgit diff --check. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g3-20260831_130650\PHASE_F_G3_PASS_REPORT.mdandgpu_retry1/g3_summary.json. - Snapshot
pre-v38-review-phase-f-g3-finish-gpu-gate-20260831_130526-0a7c670verified 724 exact files excluding.git/.pytest_cache. Stopped only the dedicated 8190 backend after Queue empty; 8188 remained HTTP 200 and untouched. Optional Smart-branch Finish Remaining, G4, G5, commit, push, tag, and release were not performed.
2026-08-31 - V3.8 FL2VA Terminal Smart Regenerate Phase F-G4 STOP
- Created exact pre-G4 snapshot
pre-v38-review-phase-f-g4-terminal-smart-xhigh-gpu-gate-20260831_142609-0a7c670with 724 files and zero missing/extra/hash mismatch, then passed 82 focused Review/Run Storage/physical-group preflight tests. - G4-A Full passed its baseline checks: FL2VA First+Last, 3x5 seconds, 3 logical / 2 physical groups, Terminal pair Chunks 2–3 sampled once with physical seed
8293933729523758687, and final 544x544 / 360f / 15.000s / 24fps / 32kHz stereo output. - G4-B Queue 1 generated only Chunk 1, saved
review_readyunit1-1, did not apply Last Frame, ran exactly one uncached physical sample, and produced a valid 120f / 5.000s prefix. Seed, prompt hash, plan, revision, and sampling contract exactly matched Full. - Stopped on the required exact-latent comparison: Full/Review Chunk 1 video SHA
d6da9472.../140064e6...; audio SHAdab3e1ee.../4742f4d9.... Tensor deltas are nonzero, so this is not a container serialization artifact. - Source analysis narrowed the remaining boundary to separately recomputed FL2VA image inputs:
video_vae.encode(first_image)and Qwen3-VL image conditioning. No intermediate hashes exist in this evidence, so no Production fix was guessed or applied. Proposed next step is a focused cold/warm/restart encoder reproducibility diagnostic and then a content-addressed, contract-keyed fix only for the proven unstable subpath. - G4-C/D, nonce progression, restart persistence, G5, and post-PASS regression were not run. Dedicated 8190 was stopped after Queue empty; user-owned 8188 was not operated. No workflow/version/release text, commit, push, tag, or release change was made.
- Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g4-20260831_143138\PHASE_F_G4_STOP_REPORT.mdandgpu\g4_summary.json.
2026-08-31 - G4 FL2VA Conditioning Determinism Extra High diagnostic
- After the G4 parity STOP, created an exact 724-file pre-diagnostic snapshot. The official snapshot script first encountered the known ACL-denied
.pytest_cache; the accepted snapshot excluded only.git/.pytest_cacheand verified zero missing, extra, or hash-mismatched files. - Added an evidence-only API runner and loaded an external diagnostic node through
--extra-model-paths-config; no Production package or runtime file was instrumented. Tensor fingerprints include original dtype, shape, and exact contiguous CPU raw bytes. P4-P7 canonical payloads contained no unsupported types. - Without Sampling, collected P0-P7 for Cold A, Warm B, Full-preparation, Review-preparation, and Restart C. Every checkpoint was bit-exact in all comparisons. P2 Video VAE and P5 Qwen/CLIP therefore did not meet their focused-repeat trigger conditions; their normal cold/warm/context/restart executions were already exact.
- Classified CASE E. This rejects the earlier VAE/CLIP cache hypothesis. The next root-cause boundary is target/noise/mask/sigma assembly, first H3 model-call input/output, or later sampler-loop behavior under the actual Full Terminal then Review ordering. No fix was implemented.
- Dedicated 8190 was restarted once for D2 and both instances were stopped after Queue empty. 8188 remained externally unreachable and was not operated. Production Critical files, workflow, version, commit, push, tag, and release were unchanged.
- Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-g4-conditioning-determinism-20260831_150837\G4_CONDITIONING_DETERMINISM_STOP_REPORT.mdandg4_conditioning_determinism_summary.json.
2026-08-31 - G4 Sampling Determinism Extra High diagnostic
- Created and SHA-verified the 725-file pre-change snapshot after the official snapshot helper encountered the known
.pytest_cacheACL denial. Added only an evidence runner and external diagnostic collector; no Production or runtime file was instrumented. - Completed Fresh Full/Fresh Review, Full -> Review, and Review -> Review. S0-S7, M0-M3, P7/S7 parity, and final sampled AV were exact. C/D showed only S8 model residency changing from CPU/no current weight-patch UUID to CUDA/current UUID; this did not change the first model output or final latent.
- Rejected the first C collector result because immediate CUDA-to-CPU M hashing introduced host synchronization. The accepted collector defers M hashing until Sampling completes and matches a current public-node no-hook control exactly.
- The public-node no-hook control reproduced the historical Full hashes but not the historical divergent Review hashes; current Review matched Full. The historical G4 mismatch is non-reproducible in this Gate, so no step trace, accelerator A/B, or Production fix was run. The old G4 runner's automatically started Queue 2 was interrupted on dedicated 8190 after Queue 1 passed.
- Validation:
835 passed, Manifest217/217, runtime verifier, Python/JavaScript syntax, Critical source/runtime hashes, andgit diff --check. Dedicated 8190 was stopped after Queue empty. Port 8188 remained HTTP 200 during the diagnostic, but a final read-only post-report check found no listener; this task never stopped, restarted, or otherwise operated 8188. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-g4-sampling-determinism-20260831_154505.
2026-08-31 - Phase F-G4R FL2VA Terminal Repeatability Extra High Gate
- Created and verified the 731-file pre-G4R source snapshot after the official helper met the known
.pytest_cacheACL denial. Production Critical source/runtime hashes were exact before execution. An initial sandbox-launched 8190 attempt was rejected and not counted because Run Storage lock creation was denied before Sampling; Production was not changed. - Ran three accepted independent Fresh-8190 public/no-hook cycles. Each Full -> new Review Queue 1 pair matched seed, Prompt hash, Video latent SHA, and Audio latent SHA exactly; R1/R2/R3 all reproduced
d6da9472.../dab3e1ee.... Every Review generated one uncached physical group, retained unit1-1, and left Last Frame unapplied. - A separate final cycle passed Full, Review Queue 1, and Terminal Queue 2. Queue 2 reused Chunk 1, sampled the Chunks 2-3 Terminal pair once, atomically committed both entries, and matched Full for all logical latent/plan fields, assembly/physical grouping, decoded RGB, decoded PCM, duration, fps, and audio format.
- One Terminal Smart Regenerate passed target Chunk 2, Chunk 1 exact reuse, nonce
0 -> 1, one Terminal physical sample, both logical variations changed, both entries in new revisionbaf61e416796683f, and no partial pair. The evidence runner's only initial false check required a report sentence that applies to partial-pair reset, not direct boundary-2 Smart Regenerate; only the evidence assertion was corrected and the completed GPU evidence was reassessed. - Final classification: Phase F-G4 = PASS with prior transient non-reproducible parity event. The historical mismatch remains recorded. Peak VRAM
15.54 GiB, peak RSS47.50 GiB, minimum available RAM4.38 GiB. Post-Gate validation passed835tests, Manifest217/217, runtime verifier, Python173/173, JavaScript7/7, source/runtime SHA8/8, andgit diff --check. - Dedicated 8190 and launcher were stopped after Queue empty; 8188 remained listening and untouched. G5, Production changes, workflow/version/release text, commit, push, tag, and release were not performed. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g4r-20260831_175742.
2026-08-31 - Phase F-G5 Review / Smart Regenerate × Driving Audio GPU Gate
- Created and hash-verified a 738-file pre-G5 snapshot after the standard helper encountered the known
.pytest_cacheACL denial. Production Critical source/runtime SHA pairs were exact, and no Production or runtime file was changed. - Ran the fixed 20-second T2VA Driving Audio sequence as Full, Review Queue 1/2, Smart Regenerate, restart persistence, and a separate Queue 1 -> Finish Remaining flow. Full/Review/Finish matched exactly for both chunks' seed/Prompt/Video latent/Audio latent/plan, assembly, decoded RGB, and final PCM. Output was 480 frames, 24 fps, 20.000 seconds, 32 kHz stereo, and 640,000 samples/channel.
- Smart Regenerate targeted Chunk 2, retained Chunk 1 exactly, advanced nonce
0 -> 1, changed Chunk 2 Video/Audio latent, and preserved final Driving Audio PCM SHAea81cc52.... Restart restored revisiondfdadc00aa25c4a7, boundary 2, nonce 1, unit2-2, and the complete canonical head. Finish Remaining reused Chunk 1 and sampled only Chunk 2. Every Review action executed the uncached Sampler. - The fixed source is FLAC 32 kHz stereo, 20.974938 seconds, SHA
ad9d942c...; therefore the optional 30-second auxiliary branch flow was not run and no substitute audio was introduced. Peak VRAM15.56 GiB, peak RSS45.85 GiB, minimum available RAM5.01 GiB; 16 GiB passes with small headroom. - Regression:
835 passed, Manifest217/217, runtime verifier, Python173/173, JavaScript7/7, Critical SHA8/8, andgit diff --checkPASS. Phase F-G5 Review / Smart Regenerate × Driving Audio GPU Gate = PASS. Final Integration, commit, push, version, tag, and release were not performed. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g5-20260831_203817.
2026-08-31 - V3.8 Workflow Cleanup Gate
- Reworked only public workflows and workflow tests. Easy V38 is a 20-node Core+Continuum Strict Minimal graph with rgthree Turbo LoRA, both Easy-Use cleanup nodes, KJ Sage/ChunkFeedForward, PreviewAny, and their links removed. All optional media loaders remain connected and native-bypassed for a runnable T2VA default; Core BasicScheduler is 20 steps.
- Added the 19-node normal
MiniMax_H3_Continuum_V38.jsonwith the V3.8 sampler, 24 positionally aligned widgets including appended Full Run/Continue controls, and connected optional First/Last/Reference 1-3/Driving Audio paths. Both public V3.8 workflows reject non-Core third-party dependencies and advanced/acceleration nodes through focused lint and JSON/link/widget save-reload tests. - The dedicated 8190 KJ-workflow-node-free 10-second T2VA smoke passed at 736x416, 240 frames, 24 fps, exact 10.000 seconds, and 32 kHz stereo. Sampling/Decode/Assembly/Save completed without OOM or NaN/Inf; peak VRAM was 14.96 GiB, peak RSS 44.83 GiB, minimum available RAM 6.30 GiB, and API time 312.03 seconds. The existing backend-global
--use-sage-attentionoption remained enabled; no KJ/Sage patch node was present in the active 14-node API graph. - Source Gate passed
32; full source passed841. The first external runner call failed before Queue submission only because itsclient_idargument was missing; the external evidence runner alone was fixed and the accepted smoke rerun passed. Production files are exact against the 738-file pre-Cleanup snapshot. Performance Optional was not created. Dedicated 8190 was stopped; user 8188 PID 35120 was untouched. Final Integration rerun, README/CHANGELOG/version, commit, push, tag, and release were not performed. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-workflow-cleanup-gate-20260831_223800.
2026-08-31 - V3.8.0 Release Preparation Gate
- Created and hash-verified
pre-v380-release-preparation-excluding-pytest-cache-20260831_233524-0a7c670: 742 exact files excluding only.gitand the inherited ACL-inaccessible.pytest_cache. The standard snapshot helper stopped at that known ACL boundary before any project edit; the accepted fallback verified every included SHA-256. - Unified the active package-version sources at
3.8.0inversion.py,pyproject.toml, andmetadata.ini, and updated only the corresponding release assertions. V3.7 node names, old CHANGELOG history, schemas, compatibility documentation, Production implementation, and both accepted V3.8 workflows were not changed. - Updated README.md, README_JA.md, CHANGELOG.md, PACKAGE_INFO_JA.txt, and PACKAGE_VALIDATION.txt for Easy/normal V3.8 workflows, aspect-aware Resolution presets, Review/Smart/Finish controls, Driving Audio, loader bypass UX, the partial-Review Second Pass limitation, and conservative 16GB wording. README links to both public V3.8 workflows resolve. Natural-Japanese quick/tech lint completed; all 23 info findings were reviewed and consciously retained where they reflect technical names, historical version sections, or existing document style.
- Validation passed:
841tests, Manifest219/219, runtime verifier as3.8.0, Python174/174, JavaScript7/7, README relative links41/41, version consistency9/9, final protected/non-authorized SHA254/254, andgit diff --check. The first pytest invocation hit only the inherited user-Temp ACL failure; rerunning withD:\Codex\_tmpproduced the accepted841 passedresult. - Final Integration remains PASS; G4 remains
PASS with prior transient non-reproducible parity event, G5 remains PASS, and all internal evidence is retained. Release Preparation did not operate 8188/8190 and did not commit, push, tag, publish, or create a GitHub Release. Candidate tag:v3.8.0.
2026-09-01 - Easy Loader Bypass Enable foreground
- Added one shared
drawWidgetswrapper inweb/easy_bypass_toggle.js. ComfyUI performs its normal bypassed draw first, then the wrapper redraws only the configured Enable widget at full opacity. The native purple/transparent node appearance, normal widget coordinates, hit-test, callback, andnode.modesource-of-truth contract are preserved for Image, Audio, and Video. - Added a three-loader Node.js drawing regression that verifies the bypass pass at alpha 0.2, the Enable-only foreground pass at alpha 1, clickability, immediate return to mode 0, and no duplicate foreground pass while enabled. Focused tests passed
21; full source passed842. - Created exact source/runtime snapshots and synchronized only the common helper to ComfyUI_W with SHA
c2883272b05a9a9146b5c6cf614ddae8a3e6893cd910902023aae61360195289. No process was stopped or restarted. Package docs and validation text now describe the visible/clickable Enable row; commit, push, tag, and release remain unperformed.
2026-09-01 - V3.8 Reliability Diagnostics
- Added read-only post-generation Frame Alignment, Audio Latent, Core compatibility, and Visual conditioning diagnostics only to the existing normal
H3ContinuumSamplerV38. The current Resolution Preset, Review controls, public schema/widget positions, Frontend, V3.7 Production path, Sampling, Conditioning, Run Storage, Assembly Plan, State/Session, RefineSchedule, and Second Pass remain unchanged. - Extracted the existing Reference Image size decision and exposed the existing Video Guide size/frame decisions as shared pure resolvers used by both Production preprocessing and diagnostics. Visual load is peak per physical group: Reference areas plus Video Guide spatial/temporal ratio, no chunk/reuse multiplier, and no Still Guide numeric weight. Still Guide is reported separately. Diagnostics do not resize, encode, clone, hash, read pixels, or transfer devices.
- Added
31reliability tests covering current V3.8 schema preservation, 120->124 pad, cap/trim, normal/Terminal physical groups, Video/Audio metadata and phase advisories, Audio Tensor identity, Core exception preservation, Low/Medium/High boundaries, Still exclusion, Basic/Detailed/Off formatting, and preservation of every non-status output. Focused passed167; formal full validation passed873; Manifest221/221, runtime verifier, Python syntax, and four source/runtime SHA pairs pass with only the knownpynvmlwarning. - Accepted snapshots are source
pre-v38-reliability-diagnostics-approved-excluding-pytest-cache-20260901_004946-0a7c670(742 exact files) and runtimepre-v38-reliability-diagnostics-runtime-20260901_010833(three exact pre-existing files; new module absent). Synchronized only the four changed Production files to ComfyUI_W without stopping or restarting any process. GPU generation, commit, push, tag, and release were not performed.
2026-09-01 - Reliability GPU baseline and Phase 5 Ref Encode Cache
- Restarted only ComfyUI_W backend
127.0.0.6:8188and passed Phase 1-4 no-guide plus Video Guide GPU smoke at 2x5 seconds / 576x576 / fixed seed / Audio Continuity ON. Both produced 240 frames, 10.000 seconds, 24 fps, 32 kHz stereo, two physical groups, and clean V3.8 reliability reports. Saved the exact 747-file accepted pre-Phase5 snapshot. - Implemented the approved V3.8-only VAE-scoped CPU LRU in
v3/ref_encode_cache.py, integrated Reference Image / Video Guide / safe Reference Audio wrappers, and kept public schema, Run Storage identity, V3.7 paths, Sampling, decode, and non-approved conditioning types unchanged. - Fixed one GPU-preflight-only internal signature omission after the first submission stopped before Sampling. Added a direct Production-facade regression. Focused validation passed 56; full validation passed 898; Manifest passed 223/223.
- Fresh-process Video Guide A/B passed: Cold
miss=1/store=1, Warmhit=1; encode timing 26.656295 -> 0.006424 seconds; API 472.253 -> 437.818 seconds. Both Video latent tensors, both Audio latent tensors, seeds, normalized Assembly Plan, physical group count, and output media contract matched exactly. Peak observed device use was 15.11 GiB; no OOM, NaN/Inf, advisory, or Prompt warning occurred. - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-reliability-phase1-4-gpu-baseline-20260901andD:\Codex\_test_results\ComfyUI-H3-Continuum\v38-ref-encode-cache-gpu-gate-20260901. Accepted Phase5 snapshot:v38-phase5-ref-encode-cache-pass-excluding-pytest-cache-20260901_024700-0a7c670. No commit, push, tag, or release was performed.
2026-09-01 - Phase 6 Chunk Boundary Diagnostic
- Added a V3.8 post-generation, metadata/shape-only structural diagnostic in
v3/reliability_v38.pyplus 22 dedicated tests. Basic is silent on PASS; Detailed reports the structural-only contract and one line per logical boundary. - Focused 53 tests passed for Fast/Balanced/Strong native context grids, visible geometry, gap/overlap, mapping coverage/order/duplication, physical output counts, Terminal shared-group classification, Second Pass contract parity/mismatch, and strict read-only behavior. Full validation passed 920 tests.
- The required GPU Gate passed at 3x5 seconds / 576x576 / Balanced 22 / fixed seed / 4 steps. Pre/post API time was 171.193 / 169.188 seconds; both boundaries were gap 0, context 22f / 7T / 37T, PASS. Three Video and three Audio raw latent hashes, seeds, and Assembly Plan were exact; only Status gained boundary diagnostics.
- Runtime pre-change snapshot
pre-v38-phase6-boundary-runtime-20260901_1240preserved the exact prior reliability module. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-chunk-boundary-gpu-gate-20260901. The optional extra Terminal GPU rerun was not repeated; Terminal structure is unit-covered and prior G4R Production GPU evidence remains valid. No commit, push, tag, or release was performed. - Final combined validation passed
tools/validate.ps1, 920 tests, compileall, runtime verifier, 54 public workflow/UI tests, JavaScript 7/7, Manifest 224/224, nine source/runtime hashes, protected Phase5 parity 267/267, andgit diff --check. Remaining messages are only the inherited.pytest_cacheACL notice and knownpynvmlFutureWarning.
2026-09-01 - Phase 5/6 continuation closure
- Clarified the existing Phase 5 source-contract test name so it explicitly covers both full-reuse cache avoidance and partial-regenerate entry into the cached Reference encode stage; Production and runtime files were unchanged. The cache-focused suite passed 25 tests, full pytest remained 920 passed, Manifest remained 224/224, and
git diff --checkpassed. Port 8188 remained stopped and no frontend/backend process was operated.
2026-09-01 - Advanced Phase A0 Baseline Freeze / Capability Gate
- Formally accepted V3.8 Phase 1-6 and created snapshot
v38-phase1-6-accepted-baseline-pre-a0-20260901_125719-0a7c670: 755 files with zero missing, extra, or SHA-256 mismatch, excluding only.gitand the inherited ACL-inaccessible.pytest_cache. - Recorded three non-blocking, unverified GPU checks: an additional Phase 5/6 Terminal Merge run, Reference Image Cache Cold/Warm A/B, and Reference Audio Cache Cold/Warm A/B. Reference Audio remains a future canary if its VAE/device path changes.
- Re-ran 74 Second Pass, RefineSchedule, Refine Context, Hi-Res, and masked-transport tests. Six source/runtime pairs matched. On ComfyUI Core
0.34.2, a direct CPU probe confirmed NestedTensor Video random-noise/mask-1 and Audio zero-noise/mask-0 packing, expansion, and shape-preserving unpacking. - Phase A0 = PASS. The current Video-refine / bit-exact first-pass Audio passthrough and External/Full/Tail/Partial RefineSchedule contracts are frozen. Production, Frontend, public schema, workflow, and runtime were unchanged. Phase A1 remains a no-implementation specification audit of Video Only, Audio Only, and Video + Audio, with current Video Only bit-exact compatibility as its first Gate.
2026-09-01 - Advanced Phase A1 Refine Target / Video Only parity
- Created snapshot
pre-v38-a1-refine-target-video-only-20260901_132227-0a7c670before implementation: 755 files, zero missing/extra/SHA-256 mismatch, revision0a7c6703b7715689c7aab4177e44e5fcd318fe5b. - Added four source/test files only:
v3/refine_target.py,v3/targeted_refine_sampling.py,v3/targeted_second_pass.py, andtests/test_v38_refine_target.py. Existing V3.5 sampling/orchestration, RefineSchedule v1, Production facade, Frontend, public schema, workflows, and runtime were not edited. refine_target_contract v1records Video Only, Audio Only, and Video + Audio stream policies plus independent seed namespaces. Only Video Only is executable in A1; future modes stop before Sampling. The target-aware sampler and group runner delegate Video Only to the unchanged accepted V3.5 functions and add target/execution identity without changingsecond_pass_contract.version=1or existing Assembly Plan fields.- Dedicated parity passed 9 tests; the related Second Pass/RefineSchedule suite passed 44; full regression passed 929. Exact checks cover Video latent SHA-256, Audio object identity, legacy refine seed, SIGMAS identity, group order, Sampling count, status, Assembly Plan existing fields, and unchanged V3.5 public schema. Manifest is 228 entries.
- Phase A1 Video Only parity = PASS. No GPU run was needed because the new path is internal and delegates unchanged Production sampling; it is not registered or publicly reachable. A2 Audio Only / Video + Audio execution was not implemented. No commit, push, tag, or release was performed.
2026-09-01 - Advanced Phase A2 Audio Only CPU/static Gate
- After the standard snapshot stopped only on the known
.pytest_cacheACL, accepted exact snapshotpre-v38-a2-audio-only-static-gate-excluding-pytest-cache-20260901_134820-0a7c670was created at revision0a7c6703b7715689c7aab4177e44e5fcd318fe5b: 768 files and zero missing, extra, or SHA-256 mismatch. - Extended only the internal target-aware modules and tests. Audio Only now sends Video zero noise/mask 0 and Audio seeded-random noise/mask 1, uses the Audio target seed namespace per physical group, adopts sampled Audio, and returns the exact original Video LATENT objects after discarding Core's temporary sampled Video.
- Mocked Core deliberately changed Video, but the final Video object identities and SHA-256 values remained exact. Audio shape/finite checks, RefineSchedule SIGMAS identity, target seeds, physical-group order, one atomic Terminal pair sample,
second_pass_contract.version=1, and normalized Assembly Plan parity all passed. Video Only parity and the unchanged V3.5 public schema remain covered. - Focused A1+A2 result is 13 passed; related Second Pass/RefineSchedule is 48 passed; full pytest is 933 passed. Manifest is 229/229; compileall/runtime verifier and
git diff --checkpass with only the known cache ACL notice andpynvmlwarning. - A2 CPU/static = Experimental PASS / Production HOLD. Upstream ComfyUI H3 denoise-mask PR #15988 remains unresolved for Production promotion. No GPU canary, runtime sync, public node, Hi-Res Fix, Frontend, Run Storage, workflow, version, commit, push, tag, or release action was performed. Video + Audio remains contract-only.
2026-09-01 - Advanced Phase A2 Audio Only GPU Canary
- Preserved exact pre-GPU snapshots
pre-v38-a2-audio-only-gpu-canary-source-excluding-pytest-cache-20260901_141200-0a7c670(771 source files) andpre-v38-a2-audio-only-gpu-canary-runtime-w-excluding-pytest-cache-20260901_141201(387 ComfyUI_W runtime files). Synchronized only the three internal A2 modules with exact source/runtime hashes; public node registration and schema remained unchanged. - Added isolated evidence tooling under
tools/and loaded its temporary node through an external custom-node path on dedicated 8190. The first launch stopped before backend startup because this Core does not support--extra-custom-node-paths; it performed no Sampling. The accepted launch used--extra-model-paths-configwith acustom_nodesentry. - G1 1x5-second Tail 6, G2 3x5-second Balanced 22, and G3 FL2VA Terminal Merge all passed. Every first/refined Video latent SHA matched exactly and every Video LATENT object/tensor identity remained exact; every Audio SHA changed with shape preserved. Temporary sampled Video was reported discarded for every physical group.
- G2 retained
[1] + [2] + [3]; G3 retained[1] + [2,3], refining the terminal pair as one physical Audio group without a partial pair. Outputs were 120/360/360 frames, exact 5/15/15-second containers, 24 fps, and 32 kHz stereo. PCM was finite, non-silent, non-collapsed, within duration tolerance, and clean at both 5/10-second structural boundaries. - API time was 132.653 / 359.819 / 444.114 seconds; peak VRAM was 15.448 / 15.537 / 15.676 GiB and maximum status-reported backend RSS was 41,465.5 MiB. No OOM, NaN/Inf, decode, assembly, sampling, or duration error occurred.
- Post-Gate validation passed 933 tests, Manifest 229/229, runtime verifier, source/runtime SHA 3/3, syntax, and
git diff --check. Dedicated PID 19084 was stopped after Queue empty; 8188 and the frontend were untouched. - Final classification: A2 GPU Experimental PASS / Production HOLD because upstream PR #15988 remains unresolved. A3, public nodes, Hi-Res Fix, Frontend, Run Storage, workflows, version, commit, push, tag, and release were not changed. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a2-audio-only-gpu-canary-20260901_141853\A2_GPU_CANARY_PASS_REPORT.md.
2026-09-01 - Advanced Phase A3a Video + Audio internal Gate
- Created and hash-verified
pre-v38-a3a-video-audio-internal-excluding-pytest-cache-20260901_144711-0a7c670: 801 exact files, excluding only.gitand the inherited inaccessible.pytest_cache. The standard snapshot helper stopped at that known ACL before edits. - Implemented internal Video + Audio execution under the existing
refine_target_contract v1: both streams receive seeded-random noise and minimal mask 1 from oneh3-continuum-refine-av-v1physical-group seed, and both sampled outputs are adopted. The shared Core handoff preserves exact SIGMAS and one Sampling call per physical group. - The 3-logical / 2-physical mocked Gate retained
[1] + [2,3], sampled the terminal pair once, preserved shapes, finite values, group order, target seeds, RefineSchedule identity, all prior Assembly Plan fields, andsecond_pass_contract.version=1. Invalid shape and NaN/Inf cases for either adopted stream stop before commit. - Existing behavior remained separated: Video Only still delegates bit-exact to V3.5; Audio Only still discards temporary sampled Video and returns original Video objects. Existing V3.5 public node/schema and registered-node set are unchanged.
- Focused A1/A2/A3a passed 19; related Second Pass/RefineSchedule passed 62; full pytest passed 939. Manifest is 230/230; compileall/runtime verifier pass with only the known
.pytest_cacheandpynvmlwarnings. - Final classification: A3a Video + Audio internal = CPU/static Experimental PASS. No active runtime sync, GPU run, A3b public node, Frontend, workflow, Hi-Res Fix, Run Storage, version, commit, push, tag, or release action was performed.
2026-09-01 - Advanced Phase A3b Selective Refine public node
- After the standard snapshot helper stopped only on the inherited
.pytest_cacheACL, created and SHA-verifiedpre-v38-a3b-selective-refine-excluding-pytest-cache-20260901_151920-0a7c670: 804 source files, mismatch zero, excluding only.gitand.pytest_cache. - Added only the public
H3ContinuumSelectiveSecondPassExperimentaladapter, its mapping/display entry, public-node allowlist/runtime-verifier registration, and A3b tests. The fixed UI labels resolve explicitly to the existingvideo_only,audio_only, andvideo_audiocontracts; unknown labels fail before the targeted runner. - Video Only continues through the accepted V3.5-compatible internal route, Audio Only retains Video object passthrough and Production HOLD, and Video + Audio retains A3a sampled-output adoption with GPU unverified. The adapter preserves physical-group lists, including one atomic terminal
[2,3]group, and only prefixes target/experimental status text. - Focused A3b passed 10 tests; combined Second Pass/RefineSchedule passed 72; full pytest passed 949. Compileall, JavaScript syntax 7/7, source runtime verifier,
git diff --check, and the 231-entry Manifest pass. - Final classification: A3b Selective Refine public node = CPU/static Experimental PASS. Existing V3.5/V3.8 public schemas, Sampling, Frontend, workflows, active runtime, backend/frontend processes, GPU execution, A3c, version, commit, push, tag, and release were not changed.
- Submitted the completed Gate to the pinned
H3情報チェックtask and incorporated its reply. The review formally accepted A3b with no mandatory fix, retained the four target/node classifications already recorded, and required A3c to begin with a minimal source/runtime sync followed by a stop-on-mismatch Video Only versus V3.5 GPU parity Gate. A3c was not started.
2026-09-01 - Advanced Phase A3c Selective Refine GPU Matrix
- Created and hash-verified exact pre-Gate snapshots after the standard helper stopped only at the inherited
.pytest_cacheACL: sourcev38-a3b-accepted-pre-a3c-excluding-pytest-cache-0a7c670-20260901_153752(807files) and ComfyUI_W runtimepre-v38-a3c-runtime-w-excluding-pytest-cache-20260901_153758(393files), both mismatch zero. Synchronized onlynodes.py,v3/refine_target.py,v3/second_pass_nodes.py,v3/targeted_refine_sampling.py,v3/targeted_second_pass.py, andtools/verify_runtime.py; all six source/runtime SHA-256 pairs matched. - Loaded an external evidence-only probe on dedicated
127.0.0.6:8190; the user-owned8188backend and frontend were not stopped, restarted, or changed. G1 first compared the public Selective Video Only path against unchanged legacy V3.5 in one prompt from the same First Pass and passed exact Video/Audio latent SHA, Audio object/tensor passthrough, refine seed, effective SIGMAS, Sampling count, shape/finite, and normalized Assembly Plan parity. - G2 Audio Only 1x5 retained the exact Video SHA/object and changed Audio. G3 Video + Audio 1x5 changed both streams. G4 Video + Audio 3x5 retained three physical groups and passed boundary/PCM/Assembly checks. G5 FL2VA retained
[1]+[2,3]; the terminal pair stayed one atomic physical group with one targeted Sampling call and no partial pair. All media passed exact frame/duration, 24 fps, 32 kHz stereo, finite/non-silent/non-collapsed PCM, and no OOM/NaN/Inf/decode/assembly error. - API seconds were
221.155 / 90.713 / 92.557 / 448.460 / 550.807; maximum device use was15.613 GiB, actual backend peak Working Set44.08 GiB, and minimum system available RAM2.594 GiB. 16GB GPU passed with low VRAM/RAM headroom under the 3-chunk/Terminal workload. The evidence runner's launcher-PID RSS field is not used as backend RSS; the backend lifetime peak was read from the actual port-8190 Python PID. - Post-Gate validation passed
949tests, Manifest231/231, compileall, runtime verifier, JavaScript7/7, critical source/runtime SHA6/6, evidence-tool syntax, andgit diff --check. Evidence isD:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a3c-selective-refine-gpu-matrix-20260901_154935\summary.json. - Final classification: A3c Selective Refine = GPU Experimental PASS. Video Only remains a Production-compatible candidate; Audio Only and Video + Audio remain Production HOLD; the public node remains Experimental. ComfyUI_W's pre-existing
3.7.0package-label files were outside the approved minimal six-file GPU sync and are known runtime packaging debt, not an A3c functional failure; resolve them before public release or formal runtime-package finalization. Production, Frontend, workflows, package source version, commit, push, tag, and release were unchanged. - After confirming an empty Queue, stopped only the dedicated 8190 backend PID
15948and its venv launcher PID36556; 8190 has no listener. User-owned 8188 remains listening as PID12636and was not operated. - The pinned
H3情報チェックreview formally accepted A3c with no mandatory code fix or repeat GPU run. Created and SHA-verified accepted snapshotv38-a3c-gpu-experimental-pass-accepted-excluding-pytest-cache-20260901_162409-0a7c670:811files, mismatch zero. - The review fixed the A4 stop contract: non-selected physical groups must perform no MODEL clone, conditioning, Sampling, or seed consumption and must preserve Video/Audio object identity; Terminal
[2,3]cannot be split; selected-group seeds must match All mode; Assembly Plan and Run Storage contracts cannot change. A4 was not started.
2026-09-01 - Advanced Phase A4 Chunk-scoped Refine GPU Matrix
- Created and verified source snapshot
pre-v38-a4-chunk-scoped-refine-excluding-pytest-cache-20260901_174621-0a7c670(811files) and runtime snapshotpre-v38-a4-gpu-runtime-w-excluding-pytest-cache-20260901_1845(393files), both mismatch zero. Implemented a pure scope resolver and appended All/Physical Group/Logical Chunk plus the 1-based index to the Experimental Selective node while keeping All as the A3c-compatible default. - The filter skips non-selected groups before conditioning adaptation/encode, MODEL clone, seed derivation, and Sampling and preserves their exact Video/Audio LATENT objects. Logical 2/3 in a Terminal pair selects the complete pair. Scope stays out of Assembly Plan and Run Storage; V3.5, Production Sampling, V3.8 Sampler, RefineSchedule, Frontend, workflows, Hi-Res Fix, and package version remain unchanged. Added the requested legacy A3c saved-workflow fixture proving missing tail widgets restore All/1.
H3情報チェックaccepted the CPU/static Gate without a Sampling fix and prescribed the five-case GPU Matrix. Dedicated 8190 then passed All A3c parity, Physical Group 2, Logical Chunk 2, Terminal Logical 2, and Terminal Logical 3. G2/G3 and G4/G5 respective outputs matched exactly; Terminal 2/3 used seed12172123750489447764, one atomic Sampling, and identical Video/Audio latent plus decoded media hashes.- Two runner-only false STOPs were resolved without Production edits:
_runtime_started_atwas excluded only from cross-run Plan value comparison while preserving exact field-set checks, and Terminal Plan fields were compared to the A3c Terminal rather than non-Terminal baseline. Evidence sampling was reused rather than repeated. - Peak VRAM was
15.566 GiB, peak backend Working Set43.693 GiB, and minimum available RAM2.157 GiB; all outputs were 360 frames / 15 seconds / 24 fps / 32 kHz stereo with healthy PCM. Post-GPU full pytest passed968; focused passed61; Manifest234/234, runtime verifier, Python/JavaScript syntax, source/runtime SHA4/4, andgit diff --checkpass. - Queue was empty before stopping only dedicated backend PID
35164and launcher PID15964. Port 8190 has no listener; user 8188 PID12636and the frontend were untouched. Final state: A4 GPU Experimental PASS, upstream Audio-target Production HOLD retained, A5 not started, and no commit/push/tag/release. - Submitted the final GPU report to pinned
H3情報チェックand incorporated its approval: A4 is formally GPU Experimental PASS, both runner-only false STOPs are non-product failures, no mandatory correction/repeat/extra Gate remains, and work stops before A5. - Created and hash-verified accepted snapshot
v38-a4-gpu-experimental-pass-accepted-excluding-pytest-cache-20260901_194500-0a7c670:819files, mismatch zero, excluding.gitand.pytest_cache.
2026-09-01 - Issue #13 R2 long-form continuation diagnostic
- Reviewed the latest public 60-second report and kept A5 unstarted. After the standard snapshot helper stopped on the inherited inaccessible
.pytest_cache, created verified fallback source snapshotpre-issue13-r2-reinvestigation-excluding-pytest-cache-20260901_191610-0a7c670(819files, mismatch zero). Also preserved the exact external runtime diagnostic node aspre-issue13-r2-runtime-diagnostic-20260901_192148(7files, mismatch zero). - Extended only the external/evidence diagnostic node and added
tools/issue13_r2_gate_runner.py,tools/issue13_r2_analyze.py, andtests/test_issue13_r2_tools.py. Production H3 imports/nodes/runtime files were not changed. Focused diagnostic validation passed11tests before GPU execution. - On dedicated ComfyUI_W
127.0.0.6:8190, G1 Standard / Masked AV6 x 10 seccompleted in448.687 secand produced 1440 frames / exact 60 seconds / 24 fps / 32 kHz stereo. It showed a related Chunk 1-to-6 rise in luma std+0.028048, saturation p95+0.090196, contrast+0.072273, and sharpness+74.9052. - G2 matched the old R1
5 x 5 secStandard condition, completed in198.755 sec, and exactly reproduced the prior decoded RGB24 stream (1e655f35cfb7b93082831d275e3ee6b5) and recomputed float32 PCM stream (5da19da4744431f49701bb1ba818f157). It did not reproduce the same cumulative contrast/saturation/sharpness rise. - Every final Video/Audio protected prefix was bit-exact: G1
5/5, G24/4, maximum difference zero. Core sampling can introduce only floating epsilon before the existing scoped restore; final stored prefixes are exact. No OOM, NaN/Inf, decode, assembly, or duration error occurred. - Formal classification after
H3情報チェックreview: Issue #13 R2 = controlled related reproduction under long-form Standard; cause unresolved; R2 complete at G1/G2; Production unchanged; A5 HOLD. Duration/trajectory-dependent continuation feedback is supported but not proven. No broad extra GPU matrix is required; reporter exact workflow is the next R3 input. - Final validation passed focused
5, full pytest973, Manifest234/234, runtime verifier, Python syntax, JavaScript7/7, A4 Production source/runtime SHA4/4, diagnostic source/runtime SHA1/1, andgit diff --check. Accepted snapshot:issue13-r2-diagnostic-complete-excluding-pytest-cache-20260901_194903-0a7c670. - After Queue
0/0, stopped only dedicated backend PID29868and launcher PID6772; 8190 no longer listens. 8188 and the frontend were not operated. Removed only the temporary runtime input after path/SHA verification; the identical source remains under the evidence tree with SHA-25688a3861f.... - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r2-20260901_192500\ISSUE_13_R2_REINVESTIGATION_REPORT.md. Commit, push, tag, release, and GitHub Issue mutation were not performed.
2026-09-01 - Issue #13 R2.5 staged cause isolation
- Preserved the authoritative source as
pre-issue13-r25-cause-isolation-excluding-pytest-cache-20260901_200300-0a7c670(826files, mismatch zero), the external runtime diagnostic aspre-issue13-r25-runtime-diagnostic-20260901_202340(7files, mismatch zero), and the pre-download model inventory aspre-r25-turbo-model-add-20260901_201506. Installed the requested official MiniMax H3 FL2V Turbo 4-step v1.0 LoRA after verifying the Comfy-Org file SHA-256c396a9a06f58399e9df9754b18299818d84a2ddd371724ba48fe4a41221437dc. - Added evidence-only per-group Seed/SIGMAS/Noise/Prompt-embedding capture, continuation-input RGB proxy analysis, the staged R2.5 runner, and nine focused regressions. Production Sampling, Conditioning, Assembly, Run Storage, Frontend, workflows, and version files were not changed.
- Dedicated 8190 GPU screening used the supplied close-up image (
9cbed938...), 576x864, fixed seed, Euler/simple, Balanced 22 masked AV continuation, Audio Continuity ON, and slightly evolving per-chunk prompts. Turbo cases completed as3x5,3x10, and6x5; the first two did not show the full composite degradation, while later6x5groups combined increased contrast/highlight clipping/edge hardness with visible facial gloss. - Leading classification is C: continuation-count / cumulative-feedback dependent, with the explicit limitation that the approved six-chunk prompt schedule is also longer. The
6x5continuation-input proxy contrast rose0.8940 -> 0.8959 -> 0.9029 -> 0.9131 -> 0.9219, while every accepted Video/Audio prefix stayed bit-exact (5/5); prefix/Assembly corruption was not observed. The staged stop rule omitted4x10/5x10/6x10. - Motion LoRA isolation compared 3x5 normal 20-step Euler/simple OFF/OFF versus H3 Motion Booster V0.2 + BUNNY Motion Continuity Repair v1.0 ON/ON with identical seeds, SIGMAS, prompt embeddings, and trigger text. Both LoRAs were actually attached (
208 patches) and increased motion, but did not reproduce the target composite. Because6x5reproduced without either Motion LoRA, the pair is not a necessary cause; single-LoRA runs were omitted. - All five GPU runs completed without OOM, NaN/Inf, decode, or assembly errors. API seconds were
482.119 / 1038.856 / 939.591 / 903.117 / 975.867; full source regression passed977, focused passed9, and the source-package3.8.0runtime verifier passed. Accepted completed source snapshotissue13-r25-cause-isolation-complete-excluding-pytest-cache-20260901_215430-0a7c670contains833files with zero hash mismatch. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r25-20260901_202457\ISSUE13_R25_CAUSE_ISOLATION_REPORT.md. A5, commit, push, tag, release, and GitHub Issue mutation remain untouched. - After Queue
0/0, stopped only dedicated 8190 backend PID23804and launcher PID17664; no 8190 listener remains. No 8188 process or frontend was operated. Deleted only the SHA-verified temporary ComfyUI_W input copy and retained the original source image. - Submitted the self-contained completion report to pinned
H3情報チェックand incorporated its approval. R2.5 is formally cause-isolation PASS with leading continuation-count/cumulative-feedback hypothesis, root cause unresolved, Motion LoRAs not necessary but long-run amplification unresolved, no mandatory correction/repeat Gate, Production unchanged, and A5 HOLD. Reporter exact-workflow R3 is the primary resume condition; a single all-chunks-identical-prompt6x5confirmation is the fallback only if that workflow does not arrive.
2026-09-02 - Issue #13 R2.6 Fixed Prompt / Fixed Prefix isolation
- Preserved the dirty authoritative source before R2.6 as
pre-issue13-r26-fixed-prompt-excluding-pytest-cache-20260901_233817-0a7c670(833files, mismatch zero) after the standard helper encountered only the inherited inaccessible.pytest_cache. Preserved the four-file external diagnostic runtime before teacher forcing aspre-issue13-r26-fixed-prefix-runtime-diagnostic-20260902_000105(4files, mismatch zero). - Added only evidence tooling: an all-six-chunks-identical-prompt
screen_6x5_turbo_fixed_promptcase and aFixed Group 1 Taildiagnostic continuation-source mode. The latter captures Group 1 Video/Audio continuation context and injects it into Groups 2–6; no Production package path imports or exposes this mode. - Dedicated 8190 completed normal recursive A in
915.064 secand Fixed Prefix B in881.397 sec. Both outputs are 720 frames / exact 30 seconds / 24 fps and completed without OOM, NaN/Inf, decode, or assembly error. 8188 and the Frontend were untouched; dedicated 8190 is stopped. - With the prompt fixed, recursive A retained saturation
+0.020101, sharpness+171.377, gradient+0.041037, and highlight+0.004487. Fixed Prefix B reduced them to+0.004639 / +56.616 / +0.011058 / +0.001967, or76.9% / 67.0% / 73.1% / 56.2%suppression. Prompt progression is not necessary; recursive continuation feedback is the dominant amplification pathway under this controlled condition, while the smaller initiating bias remains unresolved. - A/B contract comparison matched all six seeds, SIGMAS, Video/Audio noise, prompt embeddings, sampler, and model options. Groups 1–2 inputs/outputs matched exactly; Groups 3–6 diverged only after continuation-source divergence. All five teacher-forced boundaries selected the fixed source correctly. Its immediate-previous-tail prefix result is intentionally false from Group 3 and is not a Production prefix/Assembly failure.
- Final regression passed focused
6, full pytest979, Manifest234/234, runtime verifier, Python4/4, JavaScript7/7, A4 Production source/runtime SHA4/4, diagnostic source/runtime SHA1/1, andgit diff --check. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r26-20260901_234141\ISSUE13_R26_FIXED_PROMPT_TEACHER_FORCING_REPORT.md. - Created and hash-verified accepted snapshot
issue13-r26-dominant-feedback-complete-excluding-pytest-cache-20260902_002924-0a7c670:833files, mismatch zero, excluding only.gitand the inherited inaccessible.pytest_cache. - Submitted the self-contained R2.6 report to pinned
H3情報チェックand incorporated its approval. R2.6 is formally PASS with no mandatory correction or additional GPU Gate; A5 remains HOLD. The accepted wording isrecursive continuation feedback strongly supported as the dominant amplification pathway, not a unique-root-cause claim. Fixed Prefix remains diagnostic-only, contrast did not rise in this same-prompt pair, other model/Core/conditioning/Motion-LoRA contributions are not fully excluded, and the single fixed 576x864 condition is not generalized to every setting. - Formal accepted state: Issue #13 R2.6 complete / same-prompt degradation persists / fixed-prefix substantially suppresses degradation / recursive continuation feedback strongly supported as the dominant amplification pathway / initiating root cause unresolved / Production unchanged / A5 HOLD. No additional GPU run, Production change, commit, push, tag, release, or GitHub Issue mutation was performed.
2026-09-02 - Issue #13 R2.7 one-shot Soft Re-anchor A/B
- Preserved the dirty authoritative source as
pre-issue13-r27-soft-reanchor-excluding-pytest-cache-20260902_004140-0a7c670(292non-cache source files in the fallback snapshot scope, mismatch zero) and the four-file external diagnostic runtime aspre-issue13-r27-soft-reanchor-runtime-diagnostic-20260902_004140(4files, mismatch zero). Production was not changed. - Added only evidence tooling: diagnostic
soft_reanchor.py, oneSoft Re-anchor Group 4mode, one runner case, expanded decoded boundary/skin-highlight analysis, and focused tests. The helper decodes the current Group 3 and Group 1 reference tails, matches only global Y/U/V mean/std with a 50% blend, re-encodes an exact-shape temporary Group 4 conditioning copy, and never changes accepted latent, State, Audio, seed, prompt, SIGMAS, crop, resize, geometry, or sharpening. - Before GPU Queue, runtime object-info revealed a stale Python bytecode schema with only the prior two modes. The empty dedicated 8190 backend was stopped, only the two diagnostic file timestamps were refreshed, and restart exposed the required third mode. No Production or user process was involved.
- Dedicated 8190 completed A recursive in
939.830 sand B one-shot Soft Re-anchor in884.677 s. Both produced exact 720-frame / 30-second / 24-fps / 32-kHz-stereo media with six physical groups and six Sampling calls. All A/B groups matched seed/SIGMAS/noise/prompt; Groups 1–3 Video/Audio outputs matched exact SHA. B recorded exactly one Group 4 event, preserved both input contexts and shape, passed Audio through, and caused no OOM/NaN/Inf/decode/assembly error. - B reduced Saturation buildup
32.1%and reversed ordinary highlight clipping, but worsened luma std252.7%, Sharpness11.4%, Gradient1.1%, and the skin-highlight proxy5.7%. The apparent uncapped/capped aggregates are diagnostic appendix only because highlight overcorrection dominates them; no valid composite mitigation score was established. Group 3→4 flow discontinuity rose16.5%, although no major visible seam/identity/trajectory collapse appeared. - Initial classification was R2.7 one-shot Soft Re-anchor = FAIL as a mitigation candidate. It provides a partial color/highlight signal only and does not justify Production, Adaptive, repeated, alternate-strength, or UI work. Production remains unchanged and A5 remains HOLD.
- Regression passed focused
10, full pytest983, Manifest234/234, runtime verifier, Python5/5, JavaScript7/7, A4 Production source/runtime SHA4/4, diagnostic SHA2/2, andgit diff --check. Peak VRAM was14.934 GiB, peak backend Working Set42.166 GiB, minimum available RAM2.867 GiB. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r27-20260902_005002\ISSUE13_R27_SOFT_REANCHOR_AB_REPORT.md. - Queue was empty before stopping only dedicated backend PID
31592and launcher PID11600; 8190 is closed. Port 8188 and the Frontend were not operated. Commit, push, tag, release, and GitHub Issue mutation were not performed. - Submitted the complete Gate to pinned
H3情報チェックand incorporated its approval. R2.7 is formally FAIL with no mandatory fix or extra GPU Gate.recursive-prefix parityis intentionally not applicable after the one-shot event and is not corruption. R2.6's dominant-amplification-pathway conclusion remains unchanged; R2.7 demonstrates that global decoded appearance statistics alone are insufficient to suppress it. - Final accepted snapshot:
issue13-r27-soft-reanchor-fail-accepted-excluding-cache-20260902_013547-0a7c670(293files, mismatch zero; excludes.git,.pytest_cache,__pycache__, and.pyc).
2026-09-02 - Issue #13 R2.8 Continuation Conditioning Depth / Strength Gate
- Preserved the dirty authoritative source as
pre-issue13-r28-continuation-depth-strength-excluding-cache-20260902_020343-0a7c670(293files, mismatch zero) after the standard snapshot helper stopped only on the inherited.pytest_cacheACL. Preserved the active runtime before sync aspre-issue13-r28-runtime-20260902_021351(4files, mismatch zero);continuation_policy.pydid not exist there before this Gate. - Audited Balanced 22 as an H3-grid 22-pixel-frame tail represented by 7 Video latent slots plus 37 Audio ticks under nested denoise masks. Added a private default-
Nonediagnostic callback path, a dedicated external R2.8 node/policy, runner, analyzer, and focused tests. Public Production schemas/workflows cannot activate the hook, and default behavior is unchanged. - Reused accepted R2.6-A as Baseline. B1 effective 13f kept the last 4 Video slots active and fully denoised the older 3 while leaving Audio 37T unchanged. B1 worsened Contrast
331.02%, Sharpness56.98%, Gradient7.52%, and highlight clipping132.45%; it failed by individual metrics and visual review. Per the approved stop rule, effective 9f was not run. - C1 retained the 22f transport and applied only Core's fractional Video protected-prefix denoise mask
0.25= 25% denoise / 75% preserved; Audio remained fully protected. Sharpness and skin-highlight improved, but Saturation, Contrast, Gradient, ordinary highlights, and face/pose/composition preservation worsened. C1 failed, so the B+C combination was not run. - Both dedicated 8190 candidates completed six physical groups/six Sampling calls with Video/Audio prefix integrity
5/5, exact 720-frame/30-second/24-fps/576x864/32-kHz-stereo outputs, and no OOM/NaN/Inf/decode/assembly/duration error. API totals were897.447 / 864.321 s; peak device memory was14.854 GiB, minimum available RAM0.639 GiB. Queue was empty before stopping only the dedicated 8190 listener; 8188 and Frontend were not operated. - Post-Gate validation passed focused
21, full pytest989, Manifest234/234, runtime verifier, Python syntax, JavaScript7/7, source/runtime SHA5/5, andgit diff --check. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r28-20260902_022000\ISSUE13_R28_CONTINUATION_CONDITIONING_REPORT.md. - Submitted the complete report to
H3情報チェックand incorporated its approval. It confirmed R2.8 FAIL, no mandatory correction/additional GPU Gate, correct omission of 9f and Combination, and the precise25% denoise mask / 75% preserved / Audio fully protectedwording. Aggregate screening means remain diagnostic only; the verdict rests on individual metrics plus visual review. - Formal state: R2.8 complete / effective 13f FAIL / 9f not run / 22f Video protected-prefix 75% preservation FAIL / combination not run / recursive-feedback hypothesis remains supported / Production behavior unchanged / A5 HOLD. Commit, push, tag, release, and GitHub Issue mutation were not performed.
- Accepted completion snapshot:
issue13-r28-depth-strength-fail-accepted-excluding-cache-20260902_025646-0a7c670(297files, mismatch zero; excludes.git,.pytest_cache,__pycache__, and.pyc).
2026-09-02 - Advanced Phase A6 Temporal Window Refine CPU/static Gate
- Confirmed by code audit and
H3情報チェックreadiness review that A6 does not depend on the held A5 Reference Video Influence Schedule. Preserved the dirty authoritative source aspre-v38-a6-temporal-window-contract-excluding-cache-20260902_115158-0a7c670(297files, mismatch zero) after the standard snapshot helper encountered only the inherited.pytest_cacheACL. - Added pure versioned
v3/refine_window.pyand appendedTime Window,Window Start (sec), andWindow End (sec)to the existing Experimental Selective Second Pass node. The resolver uses the final visible output timeline and maps exact-duration output frames back through natural logical ranges into physical-local frames, H3 Video slots, and Audio ticks. Continuation trim prefixes are never selected; Terminal Merge remains one atomic physical group. - Wired all three Experimental targets to temporal denoise masks plus explicit bit-exact restoration outside the selected window. Groups with no intersection are skipped before conditioning work, MODEL clone, seed derivation, and Sampling, while prior no-window/All behavior remains unchanged. No intra-group compute-pruning claim is made.
- Added 15 focused tests covering clamping, final-frame mapping, prefix exclusion, terminal atomicity, AV masks, bit-exact restore, skipped groups, Assembly Plan non-mutation, and saved-workflow defaults. Related A3/A4/V3.5 regression passed
82; the complete suite passed1004with the ComfyUI_W Python and a dedicated writable--basetemp. Manifest236/236, runtime verifier, Python syntax, JavaScript7/7, andgit diff --checkpassed. H3情報チェックformally accepted the CPU/static Gate with no mandatory code correction and prescribed five GPU cases. Snapshotted the A4-matching runtime aspre-v38-a6-gpu-runtime-w-20260902_122034(4existing files, mismatch zero;refine_window.pyabsent), synchronized only the five A6 implementation files with SHA5/5, and loaded the existing external A3c probe on dedicated 8190. 8188 and Frontend were not operated.- The GPU Matrix passed all five cases: legacy/no-window Video Only parity; Video Only one-group partial; Video+Audio one-group partial;
3x5cross-group window with Group 3 exact reuse; and FL2VA Terminal[2,3]with one-sided visible window, Group 1 exact reuse, and one atomic terminal Sampling. Every selected range changed as expected, every out-of-window Video slot/Audio tick and continuation prefix was bit-exact, skipped groups retained object/tensor identity, and absolute seeds/SIGMAS matched target-specific All baselines. - The initial G2 result stopped only because the evidence runner compared Video Only's namespace-dependent seed and field set to Video+Audio. All G2 product checks were already true. Corrected only the runner to target-specific baselines, revalidated G1/G2 without resampling, and continued G3-G5. No Production fix was made.
- API seconds were
210.845 / 90.768 / 92.013 / 378.361 / 478.395; peak VRAM15.514 GiB, peak backend Working Set42.004 GiB, minimum available RAM3.403 GiB. Media and PCM health passed without OOM/NaN/Inf/decode/assembly/duration errors. Evidence isD:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a6-temporal-window-refine-gpu-matrix-20260902_122356\summary.json. - Post-GPU validation passed full pytest
1004, Manifest236/236, runtime verifier, Python/evidence-runner syntax, JavaScript7/7, source/runtime SHA5/5, andgit diff --check. Queue was empty before stopping only dedicated backend PID12808and launcher PID13648; 8190 is closed. Commit/push/tag/release were not performed. A6 is GPU Experimental PASS; A5 and Audio-target Production HOLD remain; A7 was not started. - Submitted the complete GPU report to
H3情報チェックand incorporated its formal approval: no mandatory correction, additional GPU Gate, or resampling is required. It accepted the runner-only initial G2 STOP diagnosis and clarified that G4[4,6)intersected Groups 1/2 only while Group 3 was exactly reused. Formal close wording is A6 Temporal Window Refine = GPU Experimental PASS / final-visible-timeline scoped / group-level skip supported / intra-group compute pruning not claimed / outside-window output explicitly restored bit-exact / Terminal Merge atomic / Production unchanged / A5 HOLD. - Final accepted source snapshot:
v38-a6-gpu-experimental-pass-accepted-excluding-cache-20260902_125452-0a7c670(300files, mismatch zero; excludes.git,.pytest_cache,__pycache__, and.pyc).
2026-09-02 - Advanced Phase A7a Packed-row Observability CPU/static Gate
- Audited the current serial physical-group Sampling path and ComfyUI_W Core 0.34.2 H3
PackedLayout. Confirmed that actual packed rows combine text, keyframe/reference conditioning rows, target Audio rows, and target Video rows, while Continuum already commits groups serially and preserves Terminal Merge as one physical group. - Submitted the readiness audit to pinned
H3情報チェック. It approved a read-only A7a with no public widget or execution policy and required actual Core rows, fail-soft diagnostics, fixed order/Terminal atomicity, no allocator mutation, no VRAM/time prediction, and later GPU SHA parity. - The standard snapshot helper stopped only on the inherited
.pytest_cacheACL. Created and verified fallback snapshotpre-v38-a7a-packed-row-observability-excluding-cache-20260902_131424-0a7c670:300source files,300snapshot files, zero SHA mismatch. - Added
v3/packed_row_planner.pyand Detailed-only lifecycle hooks inmodel_patch.pyandv2/sequence.py. The wrapper observes the Core-produced layout before normal passthrough; the planner deduplicates repeated per-step observations, reports topology changes as unavailable/advisory, and cannot replace a Sampling result. - Added 13 focused tests for exact row accounting, malformed/unknown layouts, deterministic group order and peak ties, Terminal metadata, fail-soft behavior, wrapper passthrough, forbidden allocator calls, Detailed-only exposure, and unchanged public schema. Corrected one initial import cycle by moving the optional planner import into the wrapper call path; no Sampling behavior changed.
- Validation passed focused
13, related72, full pytest1017, Manifest238/238, runtime verifier, Python compile, JavaScript7/7, andgit diff --checkwith only inherited line-ending warnings. Runtime sync and GPU Matrix remain pending H3 information-check review; 8188 and Frontend were not operated. No commit/push/tag/release was performed, A7b action policy remains HOLD, and A5 remains HOLD. - Submitted the completed CPU/static Gate to pinned
H3情報チェックand incorporated its formal approval. There is no mandatory correction or extra GPU case. The required term ispeak-row candidate, notpeak group; G4/G5 do not fail merely because adding conditioning rows leaves the same candidate. The approved next Gate is G1 1x5 T2VA, G2 3x5 continuation, G3 FL2VA Terminal[1]+[2,3], G4 three Reference Images, and G5 high-row Video Guide/Reference Video on dedicated 8190.
2026-09-02 - Advanced Phase A7a Packed-row Observability GPU Matrix
- Preserved the active ComfyUI_W runtime as
pre-v38-a7a-gpu-runtime-w-20260902_132746: both existing files matched the source andv3/packed_row_planner.pywas absent before synchronization. Synchronized onlymodel_patch.py,v2/sequence.py, andv3/packed_row_planner.py; source/runtime SHA later passed3/3. - Started only a dedicated Core 0.34.2 backend on
127.0.0.6:8190with Sage. The first launcher exited before listening or generation because the isolated user directory did not exist; after creating that directory, the dedicated backend started normally. No Production/runtime defect was involved, and 8188 plus Frontend were not operated. - G1
1x5T2VA passed same-runtime Basic/Detailed parity with candidate 1 /12,436actual rows. G23x5passed fixed order[1,2,3], max concurrent 1, candidate 2 /14,112rows, and exact latent/Assembly parity against the accepted A6 pre-A7 First Pass. G3 FL2VA passed groups[1]+[2,3], candidate 2 /26,836rows, exact parity, and atomic Terminal[2,3]. - G4 three Reference Images passed with candidate 1 /
14,499rows and1,008condition-video rows. G5 124-frame Efficient 0.4 MP Video Guide passed with candidate 1 /29,706rows and14,800condition-video rows. Both same-runtime Basic/Detailed runs retained exact Video/Audio latent SHA, decoded RGB24/PCM, and Assembly semantic identity. - All Detailed observations exactly partitioned Core
seq_len, contained one target Video and Audio segment, preserved execution order and Terminal atomicity, and emitted observation-only memory plus the no-prediction/no-policy disclaimer. No OOM, NaN/Inf, Decode, Assembly, duration, or PCM failure occurred. Detailed API seconds were75.025 / 237.375 / 294.769 / 101.829 / 218.484; peak VRAM was15.531 GiB, peak backend RSS47.881 GiB, and minimum available RAM4.301 GiB. - The first G1 evidence runner STOP came only from hashing volatile
_runtime_started_atinside the complete Assembly JSON. Product latent/decode parity was already exact. Snapshotted the runner before both corrections, excluded only that established volatile field for cross-run semantic identity, added resume support, and reassessed G1 without re-Sampling. Production code was unchanged. - Post-GPU validation passed full pytest
1017, Manifest238/238, runtime verifier, Python/evidence-runner syntax, JavaScript7/7, source/runtime SHA3/3, andgit diff --checkwith only inherited line-ending warnings. Queue was0/0before stopping only the command-line/parent/listener-verified dedicated backend and launcher PIDs; no 8190 listener remains. - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a7a-packed-row-gpu-matrix-20260902_133600\summary.json. Submitted the complete result to pinnedH3情報チェックand incorporated its approval: no mandatory correction, additional GPU Gate, or re-Sampling is needed; reuse of the accepted A6 pre-A7 baseline for G2/G3 is sufficient. - Formal close: A7a Packed-row Observability = GPU Experimental PASS / actual Core PackedLayout rows observed read-only / peak-row candidate derived deterministically from largest observed physical-group seq_len / max concurrent physical groups = 1 / Terminal Merge atomicity preserved / Sampling and outputs bit-exact / memory values observation-only, not VRAM-time prediction / no execution policy / no public widget / Production unchanged / A7b HOLD / A5 HOLD. No commit, push, tag, or release was performed.
- Final accepted source snapshot:
v38-a7a-gpu-experimental-pass-accepted-final-excluding-cache-20260902_141417-0a7c670(303files, mismatch zero; excludes.git,.pytest_cache,__pycache__, and.pyc).
2026-09-02 - Advanced Phase A7b Memory Action Policy v1 GPU Matrix
- Preserved the dirty authoritative source before edits as
pre-v38-a7b-action-policy-excluding-cache-20260902_143407-0a7c670(303files, mismatch zero; excludes.git,.pytest_cache,__pycache__, and.pyc). Preserved the exact active ComfyUI_W runtime before synchronization aspre-v38-a7b-gpu-runtime-w-20260902_154625(2/2existing files, mismatch zero;v3/memory_action_policy.pywas absent). - Added a pure versioned action-policy contract, an Experimental policy node, and an Experimental V3.8 sampler variant. The only executable action is explicit opt-in one-step Reference reduction. Default
Disableddelegates to the existing sampler exactly. MODEL unload and backend/attention switching are status advisories only; external MODEL wrappers, Production Sampling, Core PackedLayout/extra conditions, public V3.8 schema, and package version remain unchanged. - CPU/static Gate passed focused
49, related120, full pytest1027, Manifest240/240, runtime verifier, Python syntax, JavaScript7/7, andgit diff --check. The pinnedH3情報チェックaccepted this Gate and approved G1-G5 without a mandatory fix. - Synchronized only
nodes.py,v3/driving_nodes.py, andv3/memory_action_policy.pyto the dedicated ComfyUI_W runtime; source/runtime SHA passed3/3. Dedicated Core0.34.2backend on127.0.0.6:8190loaded both Experimental nodes while the existing sampler remained unchanged. - G1 Disabled matched the accepted A7a baseline bit-exact for Video/Audio latent, decoded RGB/PCM, normalized Assembly, sampler/SIGMAS links, and seed inputs. G2 Reference Image 1
Max Identity -> Match Outputreduced rows15,144 -> 13,128; G3 short Video GuideBalanced 0.6 -> Efficient 0.4reduced14,230 -> 13,639; G4 three Reference Images reduced20,546 -> 14,498; G5 124-frame Video Guide reduced37,790 -> 29,319. - Active cases retained Reference count/frame count, fixed seed, SIGMAS, physical/Assembly semantics, and the same external MODEL wrapper. Effective geometry, conditioning hash, output variation, and Run Storage global identity changed as intended. No auto-unload or backend switch executed; no OOM, NaN/Inf, Decode, Assembly, duration, or PCM failure occurred.
- API seconds were
102.649,117.286 / 100.437,112.951 / 108.319,160.027 / 99.069, and359.271 / 257.536(1,417.546 stotal). Peak VRAM was15.516 GiB, peak backend Working Set45.724 GiB, and minimum available RAM5.479 GiB. - Observed packed-row reduction and the shorter warm second-run API times are fixed-case evidence only, not a guaranteed VRAM or speed reduction contract; cache warmth and backend state remain uncontrolled performance factors.
- Post-GPU validation passed full pytest
1027, Manifest240/240, runtime verifier, Python/evidence-runner syntax, JavaScript7/7, source/runtime SHA3/3, andgit diff --check. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a7b-memory-action-gpu-matrix-20260902_154825\summary.json. - Queue was empty before stopping only the verified dedicated backend PID
36400and launcher PID38804; 8190 HTTP/listener are closed. Removed only the two reproducible A7b Reference test videos from ComfyUI_W input after shutdown. Port 8188 and Frontend were not operated. - Submitted the complete GPU report to pinned
H3情報チェックand incorporated its approval. A7b v1 is formally GPU Experimental PASS with no mandatory correction, additional GPU Gate, or re-Sampling. The untested standaloneMatch Output -> Balanced 0.6 MPcase is optional because the resolver chain and CPU tests cover it. - Created and hash-verified accepted source snapshot
v38-a7b-gpu-experimental-pass-accepted-excluding-cache-20260902_162542-0a7c670:306source files,306snapshot files, mismatch zero..git,.pytest_cache,__pycache__, and.pycwere excluded; the inherited.pytest_cacheACL produced only the known enumeration warning. - Formal state: A7b v1 Memory Action Policy = GPU Experimental PASS / explicit opt-in Reference downgrade only / default Disabled bit-exact / unload advisory-only / backend switch advisory-only / external wrappers and Production unchanged / A5 HOLD. No commit, push, tag, or release was performed.
2026-09-02 - Advanced Phase A8a Adaptive Continuity Observer GPU Matrix
- Preserved the dirty authoritative source before edits as
pre-v38-a8a-adaptive-continuity-observer-excluding-cache-20260902_173233-0a7c670(306files, mismatch zero) and the active ComfyUI_W runtime before synchronization aspre-v38-a8a-gpu-runtime-w-20260902_174958. - Added pure
v3/adaptive_continuity.py, Detailed-only observation inv2/sequence.py, focused tests, and an evidence-only GPU Matrix runner. The deterministic trace records resolved group/context/transport/Terminal/boundary facts plus recommendation, confidence, reason, fallback, and reuse state. Every record isexecution_applied=false; it cannot change any Production input, tensor, identity, or execution path. - CPU/static validation before GPU passed focused
107, full pytest1041, Manifest242/242, runtime verifier, Python/JavaScript syntax, andgit diff --check. Synchronized onlyv2/sequence.pyandv3/adaptive_continuity.pyto ComfyUI_W, with source/runtime SHA exact2/2. - G1
1x5T2VA passed with one initial/no-context record and exact latent/Assembly/decoded parity. G23x5passed with groups[1,2,3], context0/22/22f, Production transportinitial/masked_av_prefix_22_v1, and exact parity. G3 reused the accepted Issue #136x5 same-promptTurbo/Euler/8-step configuration and matched all six Video/Audio physical latent SHAs, Seed/SIGMAS/prompt signatures, and five protected prefixes while emitting six non-applied records. - G4 FL2VA passed with observer groups
[1]and atomic[2,3], no split, and exact latent/Assembly/decoded parity. G5 passed Review Queue 1, an exact dedicated-backend restart, Queue 2 prefix replay plus Chunk 2 generation, and Queue 3 prefix replay plus Chunk 3 completion. Saved trace facts matched entry metadata, no new Run Storage schema/identity keys were added, and only existing status reporting contains observer text. - The G3 runner's initial HTTP 400 exposed only a missing required diagnostic-wrapper argument; the evidence runner supplied the already accepted
Recursivevalue. Two later G3 stops were evidence-parser assumptions after Sampling had completed; saved history/media were reused for comparison without re-Sampling. No Production code was changed for any runner correction. - API seconds were G1
104.448, G2257.070, G4313.145, and G5 Q1/Q2/Q382.325 / 130.589 / 134.518; G3 took approximately945seconds. Overall peak VRAM was15.588 GiB, maximum status-reported RSS46.478 GiB, and minimum available RAM2.979 GiB. No OOM, NaN/Inf, decode, assembly, duration, PCM, cache-skip, or Terminal/replay error occurred. - Post-GPU validation passed full pytest
1041, Manifest242/242, runtime verifier, Python syntax/compileall, JavaScript7/7, source/runtime SHA2/2, andgit diff --check. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a8a-adaptive-continuity-gpu-matrix-20260902_180928\A8A_ADAPTIVE_CONTINUITY_GPU_PASS_REPORT.md. - Submitted the complete Gate to pinned
H3情報チェックand incorporated its formal approval. It requires no mandatory fix, additional GPU Gate, or re-Sampling and accepts: A8a Adaptive Continuity Observer = GPU Experimental PASS / read-only deterministic decision trace / execution_applied=false / restart and Run Storage replay verified / Terminal atomicity preserved / no Issue #13 mitigation claim / Production, Run Storage, State, Session and Assembly unchanged / A8b execution policy HOLD / A5 HOLD. - Queue was empty before stopping only the verified dedicated 8190 backend and launcher. 8188 and Frontend were not operated. No commit, push, tag, or release was performed.
- Final accepted source snapshot:
v38-a8a-gpu-experimental-pass-accepted-final-excluding-cache-20260902_190700-0a7c670(309files, mismatch zero; excludes.gitand ignored cache/bytecode files).
2026-09-02 - A8b/A5 HOLD release contracts fixed
- Recorded the user's formal HOLD-release criteria without changing code, runtime, schema, workflows, or package behavior. Pre-change snapshot:
pre-v38-a8b-a5-hold-release-contract-20260902_191322-0a7c670(309files, mismatch zero). - A8b requires the accepted A8a
execution_applied=falsebaseline, one intervention type only, Seed/SIGMAS/physical grouping preservation, exclusion of the failed Issue #13 13f/75%-preserved/Soft-Re-anchor/Production-Fixed-Prefix approaches, Issue #13 R3 completion or proven path independence, and a successful Shadow/Replay Gate before GPU Sampling. - Any initial A8b implementation remains Experimental and Default OFF; existing Production must stay bit-exact, public schema change needs separate approval, and Sage/Sol/Spectrum wrappers remain external and unchanged.
- A5 requires reporter-exact Issue #13 R3 and a Reference Video causal classification. Unrelated permits GO; amplification-only requires a bounded safe Experimental range; necessary/strong cause retains HOLD. OFF remains Production-bit-exact, initial execution may alter Reference Video influence only, and
3x5plus long-form drift metrics are mandatory. - A5 and A8b must not be implemented concurrently. Formal current state remains A8a GPU Experimental PASS / A8b HOLD / A5 HOLD. No test rerun, runtime/backend/frontend operation, commit, push, tag, or release was performed for this documentation-only update.
2026-09-02 - V3.8 Product Direction Freeze and public-surface read-only audit
- Formally adopted two product layers (
Main / Production,Advanced) and three pillars (Continuation Quality,Non-destructive Production,Open Integration). V3.8 is the current supported product; V3.7 and earlier are historical Release/tag downloads rather than public Node-ID compatibility requirements in V3.8. - Audited the current Python mappings, V3.8/Easy/Second Pass/assembler schemas, frontend widget handling, Review/Run Storage contracts, public workflows, and packaging boundary without changing Production code. The source exposes 40 registered IDs; the proposed V3.8 surface is six searchable nodes, with all other IDs assigned to Labs, registration removal, or historical releases.
- Main UI/Schema is UI-only GO if the existing Python input keys and positional widget order remain unchanged. A real one-field Duration contract would require migration; the first implementation should retain
chunksandchunk_secondsand may display their computed total. Existing hidden-widget handling retains values rather than removing widgets. - Confirmed the Non-destructive backend already supports Continue/implicit accept, Regenerate Current, Smart Regenerate, Regenerate From, Finish Remaining, automatic resume, prefix reuse, branch boundary, and nonce. Render History/Take selection and structured accepted/reused/pending visualization remain product gaps.
- Kept Hi-Res Fix outside the planned standard workflow, retained Second Pass as the external latent-processor bridge, and classified Assembly as mandatory with optional Seam behavior. The first Finalize surface should reuse the current assembler ID rather than add another node.
- Created pre-change snapshot
pre-v38-product-direction-public-surface-audit-excluding-cache-20260902_202606-0a7c670(309/309, mismatch zero) after the standard helper met the known.pytest_cacheACL denial. Audit reports are underD:\Codex\_test_results\ComfyUI-H3-Continuum\v38-product-direction-freeze-20260902_202941. - This phase changed only direction documentation and its Manifest entry. Runtime sync, Backend/Frontend operation, GPU Sampling, public registration/display names, workflow/schema implementation, commit, push, tag, and release were not performed. P1a/P2 remain Deferred; A5/A8b remain HOLD.
- Submitted the completed audit to the pinned
H3情報チェックand incorporated its approval. Formal result: Product Direction Freeze = PASS / Main UI boundary = UI-only GO / Public Surface Cleanup = GO / export allowlist preferred over physical deletion / packaging cleanup allowed / backend schema and Production behavior unchanged. The review additionally requires an explicit unknown-node migration notice for non-exported IDs, presentation-only Basic/Production state, and retention of every runtime-imported legacy module and production helper in the release archive.
2026-09-02 - V3.8 Public Surface Cleanup
- Created and verified source snapshot
pre-v38-public-surface-cleanup-excluding-cache-20260902_210308-0a7c670(309/309, SHA mismatch zero) before implementation. The standard snapshot helper encountered only the established.pytest_cacheACL denial, so the verified fallback excluded.git,.pytest_cache,__pycache__, and bytecode. - Replaced the package-root registration merge with an explicit six-ID allowlist while retaining every imported implementation module. Renamed the supported display surface to Sampler V3.8, Finalize, Load Image, Load Audio, Load Video, and Second Pass; the three loader categories now sit directly under
MiniMax H3/Continuum. - Added frontend-only
H3 Continuum View = Basic | Production. View changes hide or reveal existing widgets without deleting widgets/sockets, changing widget values, normalizing Run Storage state, or removing prompt inputs. The V3.8 workflow savesBasicand keeps its originalwidgets_valuescontract. - Added
.comfyignore, leaving one Registry workflow and excluding historical workflows plus tests/tools/diagnostics/cache artifacts. Added release/migration documentation covering the unknown-node boundary and historical tags. Updated README, Japanese package guidance, changelog, runtime verifier, and registration tests. - Focused validation passed
18; full validation passed1046tests with cache disabled and dedicatedD:\Codex\_pytest_tmp\v38_public_surface_full_20260902_2206. Source runtime verification, JavaScript syntax,git diff --check, and Manifest245/245also passed.tools/validate.ps1itself was not accepted as a green wrapper because its fixed default pytest cache/global temp locations hit the known Windows ACL denial (911 passed,135setup errors); there was no product-test failure. - Before runtime synchronization, all seven existing target files matched the pre-change source; the V3.8 workflow was absent. Preserved them as
pre-v38-public-surface-runtime-w-20260902_222000(7/7, SHA mismatch zero), then synchronized only eight public-surface files to ComfyUI_W. Post-sync source/runtime SHA passed8/8, and a direct installed-runtime import confirmed exactly six class mappings and six exact display names. - No backend, port 8190, port 8188, Frontend, GPU Sampling, commit, push, tag, or release operation was performed. Production Sampling, backend schemas, widget order/values, sockets, Run Storage, State/Session, and Assembly contracts remain unchanged; A5 and A8b remain HOLD.
- Submitted the completion evidence to pinned
H3情報チェックand incorporated its formal approval: Public Surface Cleanup = PASS, mandatory code corrections none, additional GPU Gate none. The cleanup may close. Release/Registry readiness remains HOLD pending one later non-GPU package-staging dry-run that proves the real archive excludes historical workflows/tests/tools/diagnostics, retains the current workflow plus required metadata/docs/runtime-imported legacy modules, and imports with exactly six public nodes from an isolated staging directory.
2026-09-02 - V3.8 Release / Registry Package Staging Re-Gate
- Preserved the authoritative dirty source before packaging edits as
pre-v38-release-package-fix-filtered-20260902_224003-0a7c670(312/312, SHA mismatch 0). The standard snapshot helper stopped before edits only on the inherited.pytest_cacheACL denial. - Narrowed the Registry package to the two runtime-required tools modules and the single V3.8 workflow; excluded historical examples, tests, internal development documents, and the source-tree Manifest. Added a separate staged-package
REGISTRY_MANIFEST.sha256while retaining a valid247/247development Manifest. - Focused tests passed 6; full pytest passed 1047 with the known
pynvmlFutureWarning. Official Comfy CLI 1.18.0 validate and pack exited 0. Existing E702 future-compatibility warnings are non-blocking and remain outside this packaging-only fix. - Final ZIP contains 139 files, SHA-256
52115DA30B55F91699D01352580CE3A76F462EF27701C31183201597568A50EB. Source/stage hashes pass139/139; Registry Manifest passes138/138with missing/extra/mismatch 0; isolated import and public export/display pass6/6. Historical examples, internal docs, tests, and source Manifest are absent. - The designated
H3情報チェックtask found no mandatory blocker and formally accepted Package Staging Re-Gate = PASS / Release & Registry readiness = GO / Public Surface Cleanup remains PASS. Production/runtime/backend/frontend/GPU Sampling were unchanged. No commit, push, tag, publish, or release was performed.
2026-09-03 - V3.8 Emergency Dialogue / Audio Release Blocker Gate
- Preserved the dirty authoritative source as
pre-v38-emergency-dialogue-audio-release-blocker-fix-filtered-20260902_233838-0a7c670(314/314, SHA mismatch zero) after the standard helper encountered only the known.pytest_cacheACL denial. Ten relevant source/runtime files matched before the Gate. - Ran a dedicated wrapper-free V3.8
2x5, 24 fps, 384x384, Balanced 22, Audio Continuity ON, Standard, List-prompt case on127.0.0.6:8190. Raw Group 2's protected 22f prefix carried no ASR dialogue; Video and Audio latent prefixes were bit-exact. The raw retained suffix and Finalize Chunk 2 both transcribed all eight requested words, and Finalize matched raw Group 2 after the 29,333-sample trim bit-exact. The report was not reproduced; neither CASE A nor CASE B applied. - Repeated the same graph with Core Reference Audio connected and no Spectrum/Sol/FBC/Turbo/model-wrapper node. Both physical groups completed without matrix mismatch or packed-row error. Classified the Continuum/Core basic path PASS and retained the external/obsolete wrapper hypothesis for the reporter-specific crash.
- No Production fix or candidate resampling was performed. The same subject/studio remained continuous at the boundary with no major seam; music was not part of the minimum case and was not evaluated. GPU API times were
113.725 / 77.690 s; peak VRAM15.807 GiB, peak backend RSS44.156 GiB, minimum system available3.940 GiB. - Post-Gate validation passed focused
147, full pytest1047, Manifest247/247, runtime verifier, JavaScript7/7, source/runtime SHA10/10, andgit diff --check. The current filtered source matched the pre-fix snapshot314/314with changed/missing zero before this state-record entry. - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-emergency-dialogue-audio-release-blocker-20260902_234859. Queue was empty before stopping only dedicated 8190. The temporary runtime probe was hash-verified, preserved under evidence, and removed; 8188 and Frontend were untouched. - Formal result: Emergency Dialogue/Audio Release Blocker Gate = PASS / Production fix not required / Package Staging Re-Gate remains PASS / Release & Registry readiness = GO on current evidence. No commit, push, tag, release, or external post was performed.
2026-09-03 - V3.7.0 Civitai Reproduction / V3.7.1 Hotfix Gate
- Extracted annotated tag
v3.7.0(fe4ff9c20c2cc8bb375625d1534f5673a737d1be) into an isolated ComfyUI 0.34.2 base with226/226Git blobs exact. Created the required dirty-source snapshotpre-v370-civitai-reproduction-hotfix-gate-filtered-20260903_003041-0a7c670(314/314, mismatch zero) after the standard helper encountered only the known.pytest_cacheACL denial. - Ran wrapper-free dedicated-8190 GPU checks for Dialogue, Reference Audio, and Fixed 3x5 character/music continuity. The exact V3.7.0 versus V3.8 Dialogue condition was bit-exact for protected Video/Audio prefixes, raw Group 1/2 Audio, and Finalize Audio; both retained all eight requested words. Reference Audio completed without matrix mismatch. Character identity reset was not observed.
- Seam-OFF semantic output exposed decoded sample jumps
0.094740 / 0.415353; Assembly-only reuse with the public default Audio Seam Auto reduced them to0.042853 / 0.034786. This does not prove a semantic music restart, which remains inconclusive. No reported hard blocker reproduced, so no V3.7.1 candidate or Production change was made. - Historical full pytest passed
632; runtime verification and native PackedLayout passed. Peak VRAM was15.519 GiB. Evidence isD:\Codex\_test_results\ComfyUI-H3-Continuum\v370-civitai-reproduction-hotfix-20260903_003300. Queue was0/0before stopping exact dedicated 8190; 8188 and Frontend were untouched. No commit, push, tag, release, or external post was performed.
2026-09-03 - V3.8 Non-destructive Production Phase N1
- Audited the existing Review resolver, Run Storage lifecycle, Sampler Status, Main frontend, and saved V3.8 workflow. Confirmed exact existing mappings for Continue, Smart Regenerate Current, Finish Remaining, manual Regenerate From, persisted nonce, prefix reuse, and atomic Terminal Merge; no new backend behavior was needed.
- Added four non-serialized Production-only frontend helpers to
web/project_id.js: a truthful pre-Queue Run Plan and three action shortcuts. The shortcuts set only existing widgets, require the normal Queue control, and leave post-Queue current/accepted/reused/regenerated/pending truth to the existing Sampler Status. - Regression added saved-workflow serialize/configure exclusion, Basic/Production visibility, action mapping, Run Storage enablement, Auto Regenerate From safety, queue-input parity, nonce preservation, and reload coverage. One-shot reset still uses the existing accepted-Queue callback.
- Focused
105, full1048, JavaScript7/7, Manifest247/247, Registry Manifest138/138, runtime verifier/public nodes6/6, source/runtime JS SHA1/1, protected backend/schema contracts10/10unchanged, andgit diff --checkpassed. GPU was unnecessary because Sampling was untouched and prior accepted Review GPU Gates already cover Continue/Regenerate/Finish/Terminal/restart. - Source snapshot
pre-v38-n1-main-production-ux-filtered-20260903_012203-0a7c670is314/314exact; runtime snapshotpre-v38-n1-runtime-w-20260903_013138is1/1exact. Synchronized only the frontend JS without operating 8188, 8190, or the live Frontend. N1 PASS; N2 Branch/Take remains separate. No commit, push, tag, release, GitHub, or Civitai action was performed.
2026-09-03 - V3.8 N2 Branch / Take read-only audit
- Confirmed old Smart Regenerate Takes remain in distinct Run Storage revisions and can be enumerated from existing project/manifest data. Current canonical selection nevertheless always chooses the highest valid nonce; restart restores disk history but no selected old revision.
- Proved the blocking identity boundary: one existing reroll boundary/nonce cannot encode a selected earlier Take at one group and a fresh nonce beginning at a later group. Exact revision selection alone would also need a non-ambiguous persisted selector and a way to invalidate only downstream records without corrupting chunk-contract validation.
- Stopped under the explicit rules because N2 PASS requires a Run Storage format/provenance or Sampling Contract extension. Did not add a cosmetic history UI, mutate old revisions, or implement backend selection. Focused tests passed
109; Manifest247/247, Registry Manifest138/138, pre-audit source314/314unchanged, andgit diff --checkpassed. - Snapshot:
pre-v38-n2-branch-take-render-history-filtered-20260903_014541-0a7c670. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-n2-branch-take-audit-20260903_015018. N2 HOLD / no implementation. No runtime/backend/frontend/GPU/package/GitHub/Civitai/commit/push/tag/release action was performed.
2026-09-03 - V3.8 N2b Branch Provenance Contract v1
- Preserved the authoritative dirty source as
pre-v38-n2b-branch-provenance-contract-v1-filtered-20260903_020411-0a7c670with314/314files and zero SHA mismatch after the standard snapshot helper encountered only the established.pytest_cacheACL denial. - Added immutable physical-group revision provenance and Run Storage Schema v3. Each revision records its parent, exact physical group, nonce/seed identity, payload lineage, source storage revision, and branch cut.
project.jsonnow stores the canonical head/chain and group-level active revision map. - Kept Run Storage v2 non-destructively readable through synthesized legacy ancestry while separating new v3 storage identity from the unchanged Sampling Contract v5 identity. Finalized manifests are written before the canonical project head; incomplete/interrupted revisions are ignored and corrupt project heads recover from validated manifests.
- Added Production-only transient Render History / Takes controls. Previous/Next select a stored Take, Use This Take changes canonical selection without Sampling, and Continue From Here preserves the selected prefix then regenerates only later physical groups through the existing Queue and Sampling path. Terminal Merge remains atomic.
- Added the formal contract document and regression coverage for migration, ancestry, A/B preservation, exact selected-prefix reuse, restart, active selection, wrong selector/parent rejection, incomplete revision exclusion, corrupted-head fallback, Run Storage OFF, saved-workflow reload, and N1 behavior.
- Focused N2b validation passed
101; full pytest passed1062; public-surface follow-up passed7; JavaScript7/7, Python syntax, development Manifest250/250, Registry Manifest140/140, source runtime verifier/public export6/6, protected contract parity9/9, andgit diff --checkpassed. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-n2b-branch-provenance-20260903_024344. No runtime synchronization, live backend/frontend operation, GPU Sampling, commit, push, tag, release, GitHub, or Civitai action was performed. - Formal result: N2b PASS / N2 implementation boundary closed. State/Session, Assembly Plan, Sampling Contract, Seed/SIGMAS, public node count, and unrelated development tracks remain unchanged. Release / Registry readiness returns to HOLD until the separate package staging dry-run is repeated against the N2b-inclusive source.
2026-09-03 - V3.8 N2c Render History / Take UX Completion
- Created the required pre-change snapshot
pre-v38-n2c-render-history-take-ux-filtered-20260903_025506-0a7c670(317/317, SHA mismatch zero) after the standard helper encountered only the established.pytest_cacheACL denial. - Expanded the existing transient Production
Render History / Takesfield into a multiline catalog derived only from validatedproject.jsonBranch Provenance. It displays every physical-group Take with number/count, short revision, nonce, parent, Selected/Canonical markers, and canonical Head; Terminal Merge is labeled as an atomic group range. - Previous/Next now walks the full Take catalog without changing canonical state. Use This Take and Continue From Here still set the existing hidden Run Storage v3 action and require normal Queue. The UI shows the selected ancestry's exact reusable prefix, later stored groups to regenerate, and the pending Queue action.
- Added clean no-history/404 handling, stale-selector clearing, safe Schema v2 legacy/read-only display, canonical-head default selection, and refreshes for Production view load, Run Storage identity change, and execution completion. Transient serialization guards and Basic hiding remain intact.
- Focused N1/N2b/N2c/public validation passed
110; full pytest passed1064. Final Manifest, runtime verifier, public surface, protected-contract, JavaScript, and diff evidence is stored underD:\Codex\_test_results\ComfyUI-H3-Continuum\v38-n2c-render-history-take-ux-20260903_030627. - N2c PASS. Only frontend UX, its tests, Manifests, and internal state/worklog records changed. Run Storage Schema v3, Branch Provenance v1, public Schema/widget order/sockets, Sampling v5, State/Session, Assembly, backend/runtime, and GPU generation remain unchanged. Release Track remains HOLD pending an N2c-inclusive staging Gate; no commit, push, tag, release, GitHub, Civitai, Open Integration, or unrelated-track work was performed.
2026-09-03 - V3.8 O1 Open Integration Contract
- Preserved the authoritative dirty source in
pre-v38-o1-open-integration-contract-filtered-20260903_032012-0a7c670(317/317, SHA mismatch zero) after the standard snapshot helper encountered only the established.pytest_cacheACL denial. - Audited Second Pass, physical-group/Terminal Merge ownership, decoder-to-Finalize types, ModelPatcher integration, accelerator compatibility checks, Run Storage separation, existing tests, and release documentation. No public Schema, Sampling, State/Session, Run Storage, Assembly Plan, or Core patch requirement was found.
- Formalized one open boundary in
docs/V38_OPEN_INTEGRATION_CONTRACT.md: external processors may alter only finite Video LATENT spatial geometry while retaining B/C/T, physical-group order/count, atomic Terminal Merge, matching first-pass Audio, and the original plan. Finalize remains decoder-class neutral; Sage/Sol/Spectrum/TensorRT lifecycles remain external. - Added the only missing structural guard: incoming Second Pass Video and Audio LATENT tensors now reject NaN/Inf before Sampling. Valid payload construction, SIGMAS, RefineSchedule, conditioning, seeds, temporary-Audio discard, and bit-exact first-pass Audio passthrough are unchanged.
- Added dedicated O1 regression for external processor pass/fail cases, plan immutability, Audio object identity, public schemas, decoder-neutral Finalize, and forbidden optional-accelerator/Run Storage imports. Focused tests passed
80; full pytest passed1075; public/release tests passed20; JavaScript7/7; runtime/public export6/6; protected contracts16/16; development Manifest252/252; Registry Manifest141/141; Python compile andgit diff --checkpassed. - GPU was not run because no valid-input processor or Second Pass Sampling behavior changed. Source runtime, 8188/8190, live Frontend, package staging, commit, push, tag, release, GitHub, Civitai, and unrelated tracks were untouched. O1 PASS; Release / Registry readiness stays HOLD pending a separate O1-inclusive staging Gate.
- Accepted snapshot:
D:\Codex\_snapshots\ComfyUI-H3-Continuum\accepted-v38-o1-open-integration-filtered-20260903_033231-0a7c670. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-o1-open-integration-20260903_033231.
2026-09-03 - V3.8 AUDIO-R1 Multi Reference Audio 1/2/3
- Created the required filtered pre-change snapshot
pre-v38-audio-r1-multi-reference-audio-filtered-20260903_113210-0a7c670(319/319, SHA mismatch zero) after the standard helper encountered only the established.pytest_cacheACL denial. - Audited active ComfyUI 0.34.2 Core.
MiniMaxH3ReferenceToVideoaccepts zero to threeref_audios; tokenizer presentation assigns<Audio N>in order;PackedLayoutand_cond_audio_rowspack all audio blocks in the same order. No Core incompatibility or Sampling-contract stop condition was found. - Added the public
H3 Continuum Reference Audioshelper and appended one V3.8-onlyAudio References (Optional)bundle socket. Three sources share one Core Audio VAE but are validated, hashed, cached, and encoded independently. Gap-free connection order keeps helper labels equal to Core prompt ordinals. - Generalized only the Reference Audio adapter and conditioning builder to carry one legacy asset or an ordered tuple of up to three assets. Helper/legacy simultaneous use fails clearly. The legacy one-audio source contract, identity SHA, conditioning block, V3.7 schema, and upstream graph SHA remain bit-exact.
- Added ordered bundle identity/Run Storage provenance, helper-route graph fingerprinting, prompt warnings for all active ordinals, and a native three-audio PackedLayout check that validates
4+6+8=18packed reference rows. Run Storage Schema v3 and Sampling Contract v5 remain unchanged. - Added focused regressions for public 7/7 registration, V3.8-only socket append, helper save/schema shape, 1/2/3 ordering, gap rejection, legacy conflict, individual VAE encoding, one-audio parity, tokenizer/ref block order, prompt warnings, ordered identity, helper graph routing, and V3.7 isolation.
- Validation passed focused
70, full pytest1086, development Manifest253/253, Registry Manifest141/141, runtime/public export7/7, native PackedLayout multi-audio self-test, protected contract parity16/16, Python compile, andgit diff --check. Accepted snapshot:accepted-v38-audio-r1-multi-reference-audio-filtered-20260903_115310-0a7c670(320/320, mismatch zero). Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-audio-r1-multi-reference-audio-20260903_114811. - AUDIO-R1 PASS. GPU was not run because Core-native multi-audio presentation and packed rows are directly verified while Sampling, Seed/SIGMAS, physical groups, Terminal Merge, Driving Audio, State/Session, Assembly, and Second Pass remain unchanged. No runtime sync, 8188/8190 operation, V3.7 backport, follow-on compatibility/continuity work, commit, push, tag, or release was performed. Release Track stays HOLD pending an AUDIO-R1-inclusive staging Gate.
2026-09-03 - Issue #19 Chunk Boundary Continuity Reproduction / Cause Isolation
- Created the required filtered pre-Gate snapshot
pre-issue19-chunk-boundary-reproduction-filtered-20260903_130348-0a7c670(320/320, mismatch zero) after the standard helper encountered only the established.pytest_cacheACL denial. - Ran dedicated wrapper-free
8190GPU cases at 576x576, 4 steps, seed190019, Standard/Balanced 22, Audio Continuity ON, and Run Storage/Reference/Driving OFF:1x5,2x5,3x5,2x10, plus matched-seed explicit-prompt2x5. Optional3x10was omitted after no new anomaly through2x10and minimum system available RAM of about0.004 GiB. - Confirmed Video/Audio latent prefix bit-exactness at
5/5boundaries. The severe reported A/V cut did not reproduce, but separate physical-group Video VAE decode made the decoded RGB overlap non-bit-exact at5/5and raised visible-join MAE by10.657–21.899%; Finalize had multimetric P99 outliers0/5and introduced no new jump. - Audio showed no dropout or abnormal cut. Existing Audio Seam Auto reduced the mild five-second OFF-only splice signal. Reanalysis of accepted FL2VA
[1]+[2,3]evidence found no physical or Terminal-internal cut and preserved atomicity. P1/P2 gave the same classification. - Compared Issue #19 with Issue #13 and kept them separate: #13 is cumulative recursive-feedback amplification at later groups, while this weak decode-context effect appears immediately and did not accumulate through three groups. Fixed Prefix was not promoted or rerun.
- Focused tests passed
72; full pytest passed1086; development Manifest253/253, Registry Manifest141/141, runtime/public7/7, source snapshot320/320, andgit diff --checkpassed. Queue was0/0before stopping exact8190; the temporary runtime probe was removed;8188and Frontend were untouched. - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue19-chunk-boundary-continuity-20260903_130500. The designatedH3情報チェックtask approved ISSUE19 PARTIAL REPRODUCTION, no mandatory correction, no extra GPU for Gate closure, and no Production change. Further work requires reporter-exact reproduction or saved-latent separate-vs-consistent-context decode A/B. Release readiness remains HOLD; no commit, push, tag, or release was performed.
2026-09-03 - V3.8 Final Package Staging Re-Gate
- Preserved the authoritative 320-file dirty working tree as
pre-v38-final-package-staging-regate-filtered-20260903_141442-0a7c670with320/320files and zero SHA mismatch after the standard snapshot helper encountered only the established.pytest_cacheACL denial. - The first latest-tree pack contained 142 files and revealed a packaging-only exclusion gap for historical docs/images and repository-only config. Extended
.comfyignoreto retain only the current N2b/O1/V3.8 release contracts; regenerated the Registry Manifest and updated the Development Manifest hashes. No Production implementation changed. - Official
comfy-cli 1.18.0validate/pack exited zero. The final 103-file ZIP is 349704 bytes with SHA-2569675373D8BD37A345D4467BF3BDB6E869EA3BFE051B730E316AE338A60B0D97F. Source/stage/archive hashes are exact; Registry102/102, Development253/253, banned/internal/duplicate/unsafe/zero-byte/secret/user-path findings zero. - Stage-only import passed with no repository fallback and public surface
7/7. AUDIO-R1, O1, N2b/N2c assets, exactly two required runtime tools, and only the current V3.8 workflow are present. Focused tests passed126, full pytest1086, JavaScript7/7, runtime verifier, and protected contracts16/16. - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-final-package-staging-regate-20260903_141519. The designatedH3情報チェックtask approved RELEASE PACKAGE GATE PASS / Release & Registry readiness GO, with no mandatory blocker or additional Gate. GPU, runtime/backend/frontend operation, commit, push, tag, release, and publish were not performed.
2026-09-03 - V3.8 Resolution / Size Source UX Refactor
- Ran the required snapshot helper; it stopped only on the established
.pytest_cacheACL denial. Created and hash-verified filtered snapshotpre-v38-resolution-size-source-ux-refactor-filtered-20260903_152559-0a7c670(320/320, missing/extra/mismatch zero). - Confirmed that V3.8 Aspect is a facade-only resolver input and that Production receives only resolved Width/Height. Replaced the visible Aspect concept with
Size Source = First Image | Manual, appended Width/Height without reordering existing widgets, and retained the hidden legacy Aspect path for saved-workflow and API compatibility. - First Image reuses the existing MP/Native resolver and updates read-only dimensions when exact frontend image geometry is available. Manual accepts exact 32-aligned dimensions and ignores Preset. Missing First Image fails open to the displayed Manual canvas with an explicit backend Status line; Reference Image and Video Guide are never inferred as size sources.
- Old Auto migrates to First Image. Old Landscape/Portrait/Square migrate to Manual with the former exact Draft/Balanced/Native/Custom resolved canvas. The current public workflow now serializes the complete Take plus Size Source fields and retains the same normal generation graph.
- Added CPU/static and JavaScript regression for landscape, portrait, square, 3:2, 2:3, image replacement, missing image, Manual standard/non-standard aspects, three presets, alignment/range validation, legacy API parity, frontend migration, disabled/editable widgets, save values, V3.7 isolation, FL2VA-safe fallback, and public
7/7. - Final validation passed focused
107, full pytest1117, JavaScript7/7, Development Manifest253/253, Registry Manifest102/102, public7/7, workflow19nodes /25links / orphan0, protected contracts16/16, Python compile, andgit diff --check. Sampling Contract v5, Seed/SIGMAS, physical groups, Run Storage v3, State/Session, Assembly, Driving Audio, AUDIO-R1, O1/Second Pass, and Terminal Merge remain unchanged. GPU was not required because only resolved Width/Height differ by explicit user selection. - Accepted snapshot:
D:\Codex\_snapshots\ComfyUI-H3-Continuum\accepted-v38-resolution-size-source-ux-final-20260903_160000-0a7c670(320/320, mismatch zero). Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-resolution-size-source-ux-20260903_153907. Resolution / Size Source UX Refactor = PASS; Release / Registry readiness = HOLD pending a new staging Gate. No runtime synchronization, 8188/8190 operation, commit, push, tag, release, or publish was performed.
2026-09-03 - V3.8 Resolution UX ComfyUI_W Runtime Sync
- After the user observed the legacy Aspect menu, confirmed that ComfyUI_W still carried pre-refactor runtime files. Created exact filtered runtime snapshot
pre-v38-resolution-size-source-runtime-w-sync-20260903_163000(199/199, mismatch zero). A first narrow four-file sync demonstrated that the installed runtime predated N2b and failed server import specifically on missingbranch_provenance.py; the standalone verifier had resolved the source checkout and was therefore insufficient evidence for this mixed runtime. - Corrected the mixed runtime by synchronizing all
102/102files from the current Registry Manifest, with missing and hash mismatch both zero. Restarted only the listener and parent whose command lines exactly matched ComfyUI_Wmain.py --listen 127.0.0.6; the server then reported all seven public IDs through liveobject_info, served all seven H3 frontend extensions, exposedsize_source/width/height, and servedproject_id.jswith the new Size Source and legacy migration logic. Commit, push, tag, release, and publish were not performed.
2026-09-03 - V3.8 Test Output Archive Location
- Recorded the user-moved V3.8 test video and artifact location as
D:\output\video\comfy_video\v38(868 files at verification). The alternative literal spellingD:\output\video\comfy\_video\v38does not exist. Snapshot before this documentation-only update:pre-record-v38-test-output-path-20260903_175500-0a7c670(320/320, mismatch zero). No output file was moved, renamed, deleted, or modified.
2026-09-03 - V3.8 Release Hygiene
- Created the required authoritative-source snapshot
pre-v38-release-hygiene-20260903_181000-0a7c670atD:\Codex\_snapshots\ComfyUI-H3-Continuumwith source/snapshot320/320and zero SHA mismatch before documentation or integrity-record changes. - Synchronized
README.md,README_JA.md,INSTALL_JA.txt,PACKAGE_VALIDATION.txt, andAGENTS.md. The normal path remains Sampler V3.8 → Core Decode → Finalize → Core Save; Hi-Res/Fix and older V3.7 material is explicitly historical. Render History / Takes is documented as a non-auto-queue Production action. - Regenerated hashes from the established lists without changing package membership: Development Manifest
253/253and Registry Manifest102/102, both missing/mismatch/duplicate0. Registry policy checks found tests0, tools2, and workflows1. - Full source pytest baseline remains
1117 passed, 0 failed; JavaScript syntax and Python compileall passed. Runtime verifier passed with all seven display names. Tracked Markdown relative links0broken; tracked PNG signatures24/24valid. - Built a temporary manifest-isolated candidate (
102/102, missing/extra0) and imported it without repository fallback. It loaded from the staged path and confirmed public export/display7/7. - Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-release-hygiene-20260903_201500\PACKAGE_RELEASE_HYGIENE_REPORT.md. No runtime/backend/frontend/GPU operation, commit, push, tag, release, or publish was performed. Release / Registry readiness remains HOLD until the separate official package-staging validate/pack/archive gate.
2026-09-03 - Run Storage Windows Atomic Replace Reliability Fix
- Snapshot:
D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-run-storage-win-atomic-fix-filtered-20260903_211500-0a7c670(321/321, SHA mismatch0); the standard helper stopped only on the established.pytest_cacheACL denial. - Reproduced the prior B1 Run Storage failure in both
D:\Codex\_tmpand a fresh test root: completed temporary JSON replacement failed atos.replace()with Windows access denied. The failure was not a stale assertion. - Added Windows-only bounded retry for WinError/errno 5 and 32 after the temp write handle is closed and fsynced. Retry delays are 20/50/100/200/400 ms (770 ms total); replacement remains atomic. On exhaustion, a diagnostic Run Storage error records source/destination existence, readonly state, parent, and native error values. Existing canonical manifests are never deleted or overwritten non-atomically.
- Added regression for transient retry and permanent failure preservation. Focused atomic/B1 passed
3; B1 was run through 20 fresh native filesystem roots without a stopped iteration; Run Storage/Branch Provenance tests passed78; full pytest passed1120with0failures in 49.00 seconds. - Post-fix verification: Development Manifest
253/253, Registry Manifest104/104, runtime verifier/public7/7, JavaScript syntax, andgit diff --checkpass. Sampling v5, Schema v3, State/Session/Assembly, Branch/Take semantics, Seed/SIGMAS, public schema, GPU/runtime/backend/frontend, commit, push, tag, release, and publish were not changed. Release / Registry readiness remains HOLD pending the official package staging gate.
2026-09-03 - ComfyUI_W Reference Audio and Installer Sync
- Created verified filtered pre-sync snapshot
pre-sync-comfyw-release-hygiene-filtered-20260903_203500-0a7c670(320/320, SHA mismatch0) after the standard snapshot helper stopped only on the established.pytest_cacheACL denial. - Synchronized the proven ComfyUI_W fixes into source: Reference Image plus bundled Audio 1/2/3 now creates three ordered Core presentation items and three ordered
minimax_refsaudio blocks; legacy single Reference Audio remains supported. Added a direct regression for the mixed image/audio path. - Moved the exact seven-node verification map into reusable runtime-verifier validation; package staging now includes
tools/verify_runtime.py, andinstall_windows.ps1explicitly stops if the installed verifier is missing. Included existing runtime dependencyv3/reference_video.pyin the Registry set. - Focused regression passed
18; source runtime verifier passed Public7/7; 104-file stage-only import passed Public7/7; Development Manifest253/253and Registry Manifest104/104both have zero mismatch;git diff --checkpassed with inherited line-ending warnings only. No runtime/backend/frontend/GPU operation, commit, push, tag, release, or publication occurred. Release / Registry readiness remains HOLD pending official package staging.
2026-09-03 - Official Package Staging Re-Gate
- Snapshot before recording the result:
pre-official-package-staging-gate-record-filtered-20260903_211009-0a7c670, source/snapshot321/321, SHA mismatch0; the normal helper encountered only the established.pytest_cacheACL denial. - Official
comfy-cli==1.18.0was installed and configured only in isolated temporary directories.comfy node validateandcomfy node packexited0against a Git-index-isolated clone of the authoritative dirty working tree; no source file or source Git index changed. - The archive SHA-256 is
B9278516D0A95745ED6EFFF7ABC693574F52E2DA9CCC2F667E3A599725E9A673. It expands to Registry payload104/104plus the Manifest itself; source/stage/archive missing, extra, and mismatches are all0. Stage-only runtime verifier passed public7/7and native PackedLayout reference-audio rows. - Packaged runtime tools are exactly
tools/__init__.py,tools/p20_file_backed_tensor_poc.py, andtools/verify_runtime.py; installer validation requires the staged verifier. Only the V3.8 public workflow is included (19 nodes, 25 links, no dangling/duplicate or legacy experimental nodes). Tests, caches, internal docs/state, source Manifest, local paths, and recognizable credential patterns are absent. - RELEASE PACKAGE GATE PASS / Release & Registry readiness = GO candidate. No commit, push, tag, release, Registry publication, runtime/backend/frontend/GPU operation, or Production change occurred. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-official-package-staging-regate-20260903_210529\PACKAGE_STAGING_REGATE_REPORT.md.
2026-09-04 - V3.8 Intuitive Main UI Prototype
- Preserved exact source/runtime pre-change snapshots and left the user-provided
ComfyUI-H3-Continuum38_924.zipunchanged as an additional rollback artifact. - Replaced the hidden Basic/Production discovery path with a transient, non-serialized Main facade and visible Advanced disclosure. Added context-sensitive Video Guide and Reference Image size controls, Review-mode action buttons, and collapsed Render History while preserving every backend widget and queue intent.
- Focused UI regression passed
87; full pytest passed1120; JavaScript syntax,git diff --check, runtime source/runtime SHA sync, live frontend route, runtime verifier, and public7/7passed. Synced onlyweb/project_id.jsandREGISTRY_MANIFEST.sha256to ComfyUI_W. No backend restart, GPU, commit, push, tag, release, or publish was performed. Release readiness is HOLD pending user visual acceptance and a later package staging Gate.
2026-09-04 - V3.8 Intuitive Main UI Refinement
- Preserved nine exact source/runtime files in
pre-v38-intuitive-ui-refinement-20260904_153802; source and ComfyUI_Wproject_id.jsmatched before the refinement. - Replaced ambiguous facade wording with
Full Video,Review by Chunk, andResume & History; replaced raw booleans with explicit labels; kept Seed behind one visible Advanced Settings disclosure. Added compact integer/decimal rendering without changing stored numeric values. - Review actions now use canonical
project.jsonrevision status. The three plain-language actions appear only for backendreview_ready, selection remains visible until normal Queue, and empty Render History stays hidden. Backend intent values, serialization, auto-Queue policy, public schema, Sampling, and storage contracts are unchanged. - Focused UI/release checks passed
18; full pytest passed1120/1120. Development Manifest remains253/253, Registry Manifest remains104/104, JavaScript syntax andgit diff --checkpass. Synced onlyweb/project_id.jsandREGISTRY_MANIFEST.sha256to ComfyUI_W; source/runtime hashes match, the live route returned HTTP 200 with all new markers, and runtime verifier passed public7/7. - The required
H3情報チェックhandoff was attempted once and rejected by the task-safety boundary because this exact internal report payload was not considered explicitly authorized for that destination. It was not retried or routed indirectly. No Backend restart, GPU, commit, push, tag, release, or publish action was performed.
2026-09-04 - V3.8 Complete Mouse-over Help
- Preserved the dirty authoritative source and active ComfyUI_W runtime in SHA-verified filtered snapshots before editing; the standard snapshot helper again stopped only on the established
.pytest_cacheACL denial. - Added context-sensitive Main help for Generation, Chunks, Seconds, Output Size, Resolution, Custom MP, Width/Height, Resume & History, Reference Image Size, Video Guide Size, and Advanced Settings. First Image now explicitly identifies I2VA/FL2VA behavior; Manual explicitly identifies T2VA and no-First-Image use.
- Added help for all visible Advanced, Review, and Render History controls and non-empty metadata tooltips for every required/optional V3.8 Sampler input. The implementation writes both
widget.tooltipandwidget.options.tooltip, and refreshes selection-dependent text without changing serialized values. - Focused UI/schema tests passed
85; full pytest passed1124; JavaScript syntax passed7/7; Development Manifest254/254and Registry Manifest104/104had zero missing/mismatch/duplicate entries. No Sampling, Run Storage, State/Session, Assembly, Branch/Take, Seed/SIGMAS, GPU, commit, push, tag, release, or publish change was made. Runtime sync and live visual confirmation remained next; release readiness stayed HOLD. - Synchronized only
web/project_id.js,v3/nodes.py,v3/driving_nodes.py, andREGISTRY_MANIFEST.sha256to the snapshotted ComfyUI_W runtime; source/runtime SHA matched4/4, the live extension route returned HTTP 200 with all new help markers, and the standalone runtime verifier passed Public7/7. The existing user backend was not restarted or stopped. Its in-memoryobject_infotherefore remains the pre-sync schema until the next user restart. Eight unrelated Registry mismatches were already present in the pre-change runtime snapshot and were not overwritten in this scoped UI task.
2026-09-04 — V3.8 mock-faithful Main UI
- Snapshotted the exact source (
327files) and active ComfyUI_W runtime (105files), each with zero snapshot hash mismatches, before changing the presentation layer. Preserved the existing dirty working tree and the user-provided rollback ZIP. - Rebuilt the V3.8 Main facade around the approved mock: effective-input badges, automatic T2VA/I2VA/FL2VA mode, combined Video Length and chunk detail, conditional output sizing, friendly Quality/Run/Progress labels, a Queue-ready card, and one Advanced disclosure. Added the amber review-ready state with three explicit actions and no automatic Queue. Public schema, sockets, stored values/order, and protected backend contracts were not changed.
- Added effective-input regression coverage so bypassed/disabled loaders are not presented as active. Focused UI/schema passed
86/86; full pytest passed1124/1124. The temporary live-UI Edge attempt was stopped after Windows reused an existing Edge process group; only processes carrying the exact temporary profile were terminated and no user tab was operated. - The dedicated ComfyUI validation backend on port
8190was stopped after inspection. Cleanup confirmed no listeners on8190,8188, or the temporary debugging port9223. No GPU run, commit, push, tag, release, or publish was performed. Release readiness remains HOLD pending a new package staging Gate. - The required
H3情報チェックhandoff was attempted once with a concise completion report and was rejected by the task-safety boundary because this implementation/runtime information did not have destination-specific sharing approval. It was not retried or routed indirectly.
2026-09-04 — V3.8 explicit Main controls
- Replaced the input badges, inferred Input Mode, total-duration selector, and automatic output presentation with direct controls: Chunks, Seconds per Chunk, read-only Total Length, Size Source, Resolution or Width/Height, Run, and Progress. Total Length never writes settings; Size Source never changes because an optional input was disconnected or bypassed.
- Added a non-blocking
Check First Imagesummary when First Image sizing has no active First Image, including the exact deterministic fallback canvas. Preserved the canonical backend-driven review-ready state, three Queue actions, Advanced disclosure, serialization, and every protected backend contract. - The standard snapshot helper stopped only at the known
.pytest_cacheACL. SHA-verified filtered pre-change snapshots preserved ten exact source files and two exact ComfyUI_W runtime files. Focused UI/schema passed86/86; full pytest passed1124/1124with only the existingpynvmldeprecation warning. - No backend was started, no GPU test was run, and no commit, push, tag, release, or publish was performed. Release readiness remains HOLD pending a new package staging Gate.
- The required
H3情報チェックcompletion handoff was attempted once without local paths and was rejected by the task-safety boundary because this UI/contract report lacked destination-specific sharing approval. It was not retried or routed indirectly.
2026-09-04 — V3.8 Main core controls and card alignment
- Restored Prompt Format, Continuity, Base Seed, Control After Generate, and Audio Continuity as the first five readable Main facade controls while keeping their original serialized widgets hidden in place. Advanced now contains only the remaining technical controls.
- Corrected Ready, missing-input, and Review custom-card drawing to use the declared widget height rather than a smaller canvas callback value. Added deterministic draw regression proving both headline and detail remain inside the rounded border.
- The standard snapshot helper stopped only at the known
.pytest_cacheACL; ten source files and two ComfyUI_W runtime files were preserved in SHA-verified filtered snapshots. Focused UI/schema passed86/86; full pytest passed1124/1124with only the existingpynvmlwarning. - No backend was started and no GPU, commit, push, tag, release, or publish action was performed. The previous destination-specific handoff rejection remains unchanged, so the same report was not retransmitted without new authorization. Release readiness remains HOLD pending package staging.
2026-09-05 — V3.8 review navigation correction
- Added transient
Back to SettingsandReturn to Reviewcontrols to the canonical chunk Review screen. Settings can now be inspected or changed without losing the pending Review, selected action, Take selection, or canonical branch; normal Queue remains the only execution trigger. - A new canonical
review_readyrevision still opens Review automatically. While Settings are open, the Main status card remains amber and states which chunk is awaiting Review. No public/backend/storage/sampling contract changed. - Preserved eight exact source/runtime files in
pre-v38-review-back-to-settings-20260905_002018-0a7c670before editing. No backend was started; no GPU, commit, push, tag, release, or publish action was performed.
2026-09-05 — V3.8 single-chunk Review guidance
- Corrected the misleading
Ready to Queuepresentation forRun = Review Each ChunkwithChunks = 1. The Main card and relevant mouse-over help now explain that no next chunk exists and instruct the user to setChunksto2or more before Queue. - Kept the accepted explicit-control policy: no automatic Chunks change and no new frontend execution block. Backend Review/Run Storage/Sampling contracts are unchanged.
- The mandatory full snapshot helper stopped only on the known repository
.pytest_cacheACL; the accepted filtered snapshot preserves ten exact source/runtime files with10/10hash matches. No backend was started and no GPU, commit, push, tag, release, or publish action was performed. - Focused UI/help/Manifest regression passed
20/20; full pytest passed1124/1124; JavaScript syntax, Development Manifest254/254, Registry Manifest104/104, andgit diff --checkpassed. - Focused UI passed
16/16; full pytest passed1124/1124; JavaScript syntax, Development Manifest254/254, Registry Manifest104/104, runtime verifier/public surface7/7, source/runtime SHA2/2, andgit diff --checkpassed. Synchronized onlyweb/project_id.jsandREGISTRY_MANIFEST.sha256to ComfyUI_W; no backend restart is required for this JavaScript-only change, but the browser must reload the extension.
2026-09-05 — V3.8 Review completion refresh and exact user guide
- Verified the user's 5-second output metadata and embedded prompt:
Chunks = 2,Seconds per Chunk = 5,Run = Review Each Chunk, and Run Storage enabled. The matching Run Storageproject.jsonwas valid andreview_readyfor Chunk 1, proving that the first 5 seconds were intentional and the missing action panel was a frontend refresh failure. - Added a global
execution_successrefresh for all loaded V3.8 Sampler nodes, while retaining the node-local execution callback as a fallback. It performs an immediate Run Storage read plus one 250 ms follow-up and never changes Review intent or starts Queue. - Replaced the terse Review explanation in both READMEs with exact, one-step-at-a-time
2 × 5s = 10 secondsinstructions using the current visible UI labels. The guide covers the first 5-second Queue, video review, the three action buttons, the required second Queue, final 10-second output, navigation, and troubleshooting. - The standard snapshot helper stopped only on the known
.pytest_cacheACL. The accepted filtered snapshot preserved ten exact source/runtime files with zero hash mismatches. No backend was started, stopped, or restarted; the existing user-controlled ComfyUI_W backend remained untouched. No GPU, commit, push, tag, release, or publish action was performed. - Focused Review/UI tests passed
86/86; full pytest passed1126/1126. JavaScript syntax, Development Manifest254/254, Registry Manifest104/104, runtime verifier/public surface7/7, source/runtime SHA2/2, live extension delivery, empty Queue, andgit diff --checkpassed. Synchronized onlyweb/project_id.jsandREGISTRY_MANIFEST.sha256to ComfyUI_W; the browser must reload once to activate the new listener.
2026-09-05 — V3.8 Review fixed-seed safety gate
- Audited the actual stored Review project and found 36 manifests with 36 distinct Base Seeds; Base Seed is part of the Run Storage global contract. This proved that
Control After Generate = randomizecan create another Chunk 1 revision instead of continuing the accepted prefix. - Added an explicit Main-card warning and pre-Queue guard requiring
Control After Generate = fixedonly for V3.8Review Each Chunk. Full Run and historical nodes are unaffected;Try this chunk againstill creates its variation through the existing Review nonce contract. - Extended Run Storage UI refresh from successful execution to successful, failed, and interrupted terminal events. Updated both immediate README walkthroughs with the fixed-seed prerequisite and recorded the required later deliverable: an image-led beginner guide using only exact visible UI labels and one action per step.
- The standard snapshot helper stopped only on the known
.pytest_cacheACL. The accepted filtered snapshot preserves eleven exact source/runtime files with zero mismatches. No backend was started or stopped, and no GPU, commit, push, tag, release, or publish action was performed. - Focused Review/UI regression passed
116/116; full pytest passed1127/1127; JavaScript syntax, Development Manifest254/254, Registry Manifest104/104, runtime verifier/public surface7/7, source/runtime SHA2/2, live served-extension delivery for the guard and all three terminal events, andgit diff --checkpassed. Synchronized onlyweb/project_id.jsandREGISTRY_MANIFEST.sha256to ComfyUI_W.
2026-09-05 — V3.8 Review seed-lineage UI fix
- Reproduced the failed continuation from persisted data: the two latest Review attempts used different Base Seeds (
588314452213700and922887055962422) and different lineage hashes, so both safely became Group 1 rather than incorrectly reusing an incompatible prefix. - Changed only the V3.8 frontend presentation/control layer. The exact ComfyUI
control_after_generatelinked tobase_seedis now authoritative; selecting or restoringReview Each Chunk, pressing a Review action, or choosing a Take continuation fixes that control before Queue. The existing pre-Queue rejection remains as defense in depth. - Added linked-widget precedence and Review-mode auto-fix regressions, and synchronized the affected public-surface expectation. Related tests passed
166/166; full pytest passed1128/1128with the ComfyUI_W venv. The system-Python full run reached1127 passed / 1 environment failurebecause that Python has no CUDA; the same real-Comfy cache test passed under the correct runtime venv. - Updated Development and Registry Manifest hashes, synchronized only
web/project_id.jsandREGISTRY_MANIFEST.sha256to ComfyUI_W, verified source/runtime SHA equality, and passed the standalone runtime verifier with Public7/7. The existing user backend was offline and was not started or stopped. Browser reload is required after its next start. No GPU generation, commit, push, tag, release, or publish was performed.
2026-09-05 — Issue13 R3B initial GPU mitigation screen
- Completed four isolated current-V3.8 cases atseed131325:6x5 baseline,3x10,6x5 plus native clean Reference,6x5 Audio Continuity OFF. Each generated720frames/30s AV with exit0;21 physical groups total. B/C failed the frozen mitigation screen. D reduced Laplacian drift41.83% but worsened saturation drift51.25%, so remains a mixed diagnostic rather than a Production fix. No solution or long-duration quality guarantee was declared.
- Video prefix parity17/17; Audio-ON parity12/12. A/D first raw AV outputs are exact; reconstructed Core CPU noise hashes also match6/6, explicitly not captured Sampling telemetry. Corrected only the external analyzer's estimated boundary times to actual Assembly frame positions. Preserved failed/mixed outcomes, workflows, raw latents, AV files, metrics and visual comparisons.
- The historical4step LoRA was unavailable, so all cases use the same current8step LoRA and current Core. Reporter-exact R3, A5/A8b, multiseed/long-form, semantic AV and wrapper promotion gates remain unresolved. GPU-audit-only sharing to the fixed H3情報チェック task was explicitly authorized; completion review is next.
- Pre-GPU source115/115 unchanged; only this state/history record was added after the screen. The standard snapshot helper encountered the known .pytest_cache ACL; verified filtered snapshots preserved115 initial files and2 immediate pre-record files. Observer CPU tests and git diff --check passed. Registry104/104 passed; Development Manifest has a pre-existing stale REGISTRY_MANIFEST.sha256 hash (1/254 mismatch), not repaired in this GPU track. Full pytest was not rerun.
- Dedicated8190 was stopped only after empty Queue/owner checks; exact PIDs36076/28936, listener/HTTP and monitor exit verified. User8188 remained reachable and untouched. No runtime synchronization, Production change, commit, push, tag, release or publish. Report:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-20260905\ISSUE13_R3B_INITIAL_GPU_SCREEN_REPORT_JA.md.
2026-09-05 — Issue13 R3B matched Ref2VA Reference OFF/ON screen
- Read the fixed H3情報チェック initial-screen review before starting this pair. It supported matched Ref2VA testing and did not promote the mixed AudioOFF result. User GPU-audit-only sharing authorization remains limited to that destination.
- Executed two isolated Ref2VA+Ref2V8step,seed131325,6x5s cases with identical First Image,prompt,AudioON,Standard/Balanced22 and576x864. Only Reference Image1 connection differed (apart from evidence/output labels). Both720f/30s AV cases completed:888.343s OFF,918.375s ON. Primary Laplacian drift13.1606 versus27.7282 failed the frozen mitigation criterion; secondary improvements did not override it. No mitigation accepted or long-form claim made.
- Recorded actual public-input SIGMAS equality, twelve existing pre-sampling conditioning snapshots, first-keyframe parity, and identical Reference payload across all6 ON groups. Video/Audio prefix parity10/10 each and CPU reconstructed noise6/6 passed; reconstruction is not actual-noise telemetry. Inspected contact sheets and all10 sampled boundaries; full motion naturalness, semantic listening and lip-sync remain unqualified.
- Stage/source Registry104/104 and live Public7/7 passed. CPU observer tests passed; the first latent-analysis command used a Python without safetensors, then passed using existing ComfyUI_W venv with CUDA disabled for CPU analysis. No package install or Production/runtime edit. Standard state snapshot hit the known .pytest_cache ACL; verified2-file filtered state/history backup was used after the successful116-file experiment snapshot.
- Dedicated8190 launcher5792/child3228 were stopped at15:13:36JST after empty-queue/ownership checks; both PIDs, listener, HTTP and monitor exited. User8188 untouched. Monitor191 readings,min availableRAM2.576GiB,no interrupt. No commit/push/tag/release/publish. Report:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-ref2va-20260905\ISSUE13_R3B_REF2VA_GPU_REPORT_JA.md. Pair completion review is next; Issue13 unresolved,A5/A8b and Release HOLD retained.
Ref2VA screen completion review received
- Received/read the fixed H3情報チェック reply in turn
b916ed28-cb5f-4587-8456-4a7bff915049. It endorsed experiment PASS / mitigation screen FAIL, found no required conclusion correction, and supported not promoting unpromising candidates to multi-seed/long-form. It did not independently view local media or authorize a HOLD override. - Final local checks: experiment AST15files, observer CPUexit0, git diff --check exit0, Registry104/104, live Public7/7. Evidence SHA inventory was98files before adding this review receipt. Development's pre-existing stale Registry-manifest hash remains separate; full pytest not rerun. Overall Issue13 goal remains incomplete. Await reporter-exact conditions or explicit approval for a single experimental intervention; no new GPU phase started after review.
2026-09-05 — Isolated Video context-noise exception approved
- After the specific A8b exception question, user replied "許可、許可。" This authorizes a single isolated weak tapered Video-prefix-noise A/B, not a Production change or general HOLD release. Audio, masks, depth, Seed/SIGMAS/grouping remain unchanged. CPU/Shadow/Replay and fixed-task review precede GPU.
- Prepared Registry104/104 isolated source and reused prior runners/observers. Source documentation backup3/3 verified at
D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-issue13-r3b-noise-filtered-20260905_153601; clean experiment snapshot completed. No runtime/model/frontend edits, GPU launch, commit/push/tag/release/publish at this point.
2026-09-05 — R3B Video context-noise isolated GPU pair completed
- CPU policy10/10, existing A8a/Review/Branch69, observer, raw Shadow5/5 passed. Received fixed H3情報チェック review8bab4c9c-6cf0-4c27-87a7-13d7866b6046 and corrected Gaussian0.05 to a std multiplier, not maximum amplitude; recorded no clipping, per-slot RMS/max-abs and prefix/Audio hashes. Reviewer acknowledged correction in d39c12ec-87cb-4a4c-bcf3-ec8f2112de13. Underlying formula unchanged before GPU.
- One FL2VA8step seed1313256x5s OFF/ON pair, only weak tapered Video context noise enabled for ON. Both720f/30s completed876.398s/830.900s. Valid experiment, but mitigation screen FAIL: Laplacian increase39.8675→43.9942 and saturation drift+47.25%; contrast/highlight/shadow partly improved. Motion proxy fell34.08%, with repetitive poses in sampled images. No complete-freeze claim or general noise-family rejection; no accepted candidate.
- Fresh OFF matches initial raw baseline Video/Audio12/12. First AV group OFF/ON exact; actual packed Core noise/SIGMAS/derived seeds6/6, Assembly and conditioning parity; clean prefix Video/Audio10/10 each; intended intervention5/5, newest-slot delta0. Source/stageRegistry104/104, livePublic7/7,10assets rehashed unchanged, AST16, git diff --check and observer CPU PASS. Full pytest not rerun.
- Standard full-source snapshot hit known cache ACL; verified filtered docs3/3 and clean experiment snapshot used. Later exact-file snapshot updated only AGENTS entry in Development Manifest; its pre-existing REGISTRY_MANIFEST mismatch remains1/254. No Production, user runtime, model or frontend edits. Final state/history exact-file snapshots at16:29JST before recording.
- Initial backendPID18204 exited before server/GPU because isolated user directory was absent. Preserved startup logs; removed only an empty unsubmitted case directory after checking0entries; created isolated dirs and relaunched. Successful8190 launcher1972/child4128 stopped16:26:44JST after ownership/empty-queue checks; PID/listener/HTTP and monitor exit verified. Monitor195readings,min availableRAM3.064GiB,no interruption. User8188 untouched. No commit/push/tag/release/publish.
- Report:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-noise-20260905\ISSUE13_R3B_NOISE_GPU_REPORT_JA.md. Eight current-environment runs/45physical groups now, no solution. Approved single pair complete; completion review next, no further GPU or Production work authorized by this narrow exception.
Context-noise completion review received
- Fixed H3情報チェック reply4832f123-abeb-47de-9d97-56e25840a79c received/read: experiment PASS /0.05 condition screen FAIL, no mandatory correction or additional GPU. The reviewer checked the report, not local media independently. It supported retaining motion suppression as a concern rather than asserting freeze, and no general rejection of noise approaches.
- No automatic new strength/taper, other intervention, multi-seed/long-form or Production promotion. Further experiments require their appropriate explicit authorization; reporter-exact R3 still needs reporter inputs. Issue13 goal remains incomplete.
- Final evidence sealing checks115source files (only4 documentation/manifest paths changed), Registry104/104, AST17including sealing helper and diff --check PASS. All Codex backend/monitor/runner sessions have exited; user8188 untouched. GPU review receipt and final SHA inventory saved locally. No publication.
2026-09-05 — R3B non-Turbo20 Reference OFF/ON GPU pair completed
- Read-only fixed-task review0570a6de-3286-4cda-9f48-2ee4ee841340 supported the design; then user explicitly authorized exactly two non-Turbo20step30s Reference OFF/ON runs. Reused current Registry104-file source stage and fixed scene/seed1313256x5AudioON. Verified pre-change snapshots115files plus exact state/history files; no Production or installed runtime edits.
- Removed the LoRA loader only from the private executable test graph, set20steps, and retained native Reference Image conditioning only in ON. CPU verified passive cloned-MODEL telemetry does not mutate arguments, returned objects or RNG. Frozen graph differences are Reference socket, evidence case and output filename only. No context-noise policy copied into this phase.
- OFF1771.584s and ON1854.748s completed720f/30s AV each. Experiment PASS / mitigation screen FAIL: Laplacian delta25.1840→39.2554 (+55.87%); saturation and highlight guards also failed, although contrast drift improved58.05% and gradient/shadow were nearly unchanged. Do not interpret as a55.87% global quality loss, a general Reference failure, or Turbo-only causality.
- Actual Core noise/seeds/SIGMAS/masks/targetshapes6/6 match;21SIGMAS unchanged,0modelpatches in both. Video/Audio clean prefixes10/10 each,FirstImage keyframe exact,nativeReference present identically in6ON groups,Assembly exact except starttime. ON motion proxy0.4987 vsOFF0.5330; sampled repetitive poses do not prove freeze. Semantic AV/natural motion remain unqualified.
- Registry104/104, livePublic7/7 before/after,10asset hashes unchanged,AST18,observerCPU,diff-check PASS. Fullpytest not rerun; known Development Registry-hash mismatch remains1/254. Source changes onlyPROJECT_STATE/WORKLOG after fresh exact-file snapshots. Results at
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-ref2va-standard20-20260905. - Own8190 launcher29056/backend29368 stopped19:14:23JST after exactownership/emptyqueue checks. PID/listener/HTTP,conhost32972 andmonitor all exited; user8188 unchanged.370monitor samples,minRAM0.995GiB,nointerrupt. OuterPowerShell LASTEXITCODE misuse emitted a false shutdown failure after success; separate recorded process checks confirmed all exited. No extra GPU/restart.
- R3B totals10runs/57groups without an accepted mitigation. Pair complete; fixed H3情報チェック completion review next. No automatic new tests, Production changes, commit/push/tag/release/publish.
Standard20 completion review received
- Fixed H3情報チェック turn a9db6c24-27cc-4b15-b0c6-da908eb1702f (reply eba69b5f-4d9b-4de5-bb28-fa73b63e8659) received and fully read. It supported experiment PASS / mitigation screen FAIL / Issue13 unresolved / no promotion or additional GPU; no required rerun. Reviewer read the report only, not local media independently.
- After exact-file snapshots, clarified that Reference affects generation from Group1: this screen compares within-trajectory start-to-end drift, not the same initial generated video with a late Reference toggle. No numeric or verdict change. Full receipt saved in COMPLETION_REVIEW.md; remaining work requires reporter-exact inputs or a new separately approved single hypothesis.
2026-09-05 — R3B whole-investigation report and ChatGPT direction review
- Consolidated four completed R3B phases:10successful30s runs/57physical groups, no accepted mitigation. Re-read four reports and historicalR2.6, rechecked all ten MP4 SHA, and aggregated equal_5s_windows (especially3x10). Existing evidence/report files retained unchanged. No new GPU, backend, production code or runtime operation.
- Exact-file pre-change snapshots for PROJECT_STATE/WORKLOG each1file, new report target0files, then pre-review report snapshot3files were verified. Source Registry104 files remained unchanged. Report distinguishes experimental graph/noise changes from zero Production mitigation, strong Fixed Prefix diagnostic evidence from production unsuitability, and measured drift from total perceptual quality.
- Sent self-contained report only to fixed H3情報チェック under GPU-audit sharing permission. Review turn4ee9d332-985c-49b5-8f3c-1e3c5186d3d2/reply7326b44b-f2b4-4828-a1f6-d8011f5113c3 received/read; supports substantive investigation progress, no fix, initiating bias and motion-preserving mitigation unresolved. No mandatory rerun. Review is document-only, not independent local-media inspection.
- Direction proposals: reporter-exact condition differential, saved-data temporal/region audit, then one same-prefix controlled intervention design if warranted. No candidate method or next GPU is authorized by review. Restrict exclusions to tested conditions; time ordering alone is not causal proof. Report/metric matrix/full review receipt saved under
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-consolidated-20260905. - Japanese report lint and manual evidence/structure review completed; one rhythm warning retained to preserve precise technical scope. Full pytest was not rerun for this documentation-only work. Existing Development Manifest REGISTRY hash mismatch remains separate; A5/general A8b/Release HOLD and no commit/push/tag/release/publish continue. Stopped after report/review, no new investigation phase.
2026-09-05 — Schedule-Matched Disposable Prefix integrated A/B
-
Explicit proposal execution and user three-step instruction; no algorithm change/sweep. Isolated104-file source copy, only ordinary sequence sample dispatch altered; pergroup defaultOFF controller and same savedGroup1. CPU14+protected85 PASS; real graph replay preflight verifies wrapper registration/clean persistence. Fixed-task preGPU review26665e59 received/read; A rawAV12/12 reproduction required before B.
-
A801.79s/B779.90s each720f/30s,10newgroups total. A rawAV12/12 exact vs prior. Actual noise/seed/SIGMAS/base mask/shape5/5 match; conditioning/Assembly match; B40forward evaluations maskvalid; cleanAV exports10boundaries exact. Experiment PASS / mitigation FAIL: edge delta39.8675→56.3675 (+41.39%), contrast/highlight failed; saturation/gradient/shadow passed. Flow mean0.510734→0.534240 (+4.60%), not full-motion/semanticAV certification. No60s/multiseed/promotion.
-
Assets10/sourceRegistry104/isolatedPython83 unchanged,Public7. Production/runtime untouched; source state/history only after exact-file snapshots. Backend13280/26320,listener,monitors stopped; user8188 idle/reachable/unmodified. Preflight audit lookup, port-query ACL/failed duplicate launcher and audit path separator issues corrected/documented without extraSampling. Exact-file snapshot helper avoids nestedpytest.git count mismatch.
-
Report
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-schedule-prefix-20260905\RESULT_JA.md; machine evidence/videos retained. Completion review next. GeneralHOLD/no commit/push continue. -
Fixed-task completion review faf8917a-b718-48e6-b40c-b380cc68c17a / reply26bc256c-bf10-49e8-b131-0d01e2740844 received/read: experimentPASS/this22f screenFAIL, no mandatory correction, close pair. Report-only review. Japanese lint failed on existing uv cache; skill manual checklist and final readthrough used. git diff --check passed with existing CRLF warnings. No additional work phase or GPU.
2026-09-06 — V3.8 beginner manual design and capture-script draft
- Added
docs/V38_BEGINNER_MANUAL_PLAN_AND_CAPTURE_SCRIPT_JA.mdas the canonical restart document for a later image-led V3.8 beginner manual. It defines the reader, chapter architecture, one-action-per-step writing rule, capture prerequisites, fixed environment record, stable Shot IDs for Quick Start/Review/Regenerate/Take/inputs/Advanced/troubleshooting, image naming, Computer Use procedure, bug-report template, and completion criteria. - Reused the current exact
README_JA.md2 × 5sReview flow rather than creating a competing procedure. Final capture is explicitly deferred until the public UI, standard workflow, and launch treatment of the active drift investigation are stable; the image directory is not created until capture begins. - The standard snapshot helper stopped only on the known
.pytest_cacheACL. The verified filtered snapshotD:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-manual-plan-capture-script-filtered-20260906_022733preserved918/918files excluding.gitand.pytest_cache. - Documentation planning only: no UI capture, runtime synchronization, backend start/stop, GPU generation, package staging, commit, push, tag, release, or publication.
2026-09-06 — Issue13 R4-M1 CPU/read-only Memory Ownership Audit
- Reclassified the prior result as
R4-3 INCOMPLETE — SOFT RESOURCE THRESHOLD EXCEEDED; R4 remains a continuation candidate. Historical Groups 2-5 Sampling completed without OOM, allocation error, NaN/Inf, packed-layout/AV failure, or backend crash. - Audited the R4 transport branch, guide module, sequence ownership, session/state, observer and saved evidence. No R4-specific major leak, unbounded history, tensor cache, per-conditioning tensor multiplication, Refine Context capture, or observer tensor accumulation was found. Intentional full physical CPU entries and latest state are common to A/B and unchanged.
R4_M1_CPU_OWNERSHIP_CHECK.pyPASS verified guide/fresh-target weak-reference release and module-global ownership. Existing isolated R4 CPU tests passed4/4; Python compile passed. No GPU was invoked.- Identified that the former RSS exception was raised inside the Sampling wrapper before the sampled Group 5 result returned to sequence, so normal Group cleanup never ran. The 46.80 GiB value is Sampling-immediate, not post-cleanup evidence.
- Prepared only the isolated R4 observer/test copy for an approved rerun: RSS 46 GiB and low GPU free are Soft Warnings; exact Sampling-before/after and Group-cleanup snapshots, available RAM, pagefile, Windows commit, GPU use/free and elapsed time are recorded; hard resource stops are deferred until cleanup where possible.
- Report:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-native-guided-overlap-20260906\R4_M1_MEMORY_OWNERSHIP_AUDIT_JA.md. Production, installed runtime, algorithm/guide/overlap/SIGMAS/noise/seed/sampler/Assembly Plan remain unchanged. No commit, push or release. R4-3 GPU rerun requires fresh explicit approval.
2026-09-06 — Issue13 R4-M1 review receipt and R4-M2 Pre-GPU freeze
- Received/read the fixed-task review supplied in the active task: R4-M1 PASS, no mandatory correction, ready to present Pre-GPU conditions. Formal sequence is R4-0 PASS WITH CONDITIONS -> R4-1 PASS -> R4-2 PASS -> R4-3 INCOMPLETE / SOFT RESOURCE THRESHOLD -> R4-M1 PASS.
- Added the mandatory rerun conclusion for Group 5: compare B-A RSS both Sampling-immediate and post-cleanup, then state whether the previous +1.63 GiB disappears, shrinks, or remains.
- Froze the unchanged one-pair A/B, Dynamic RAM parity, revised Hard Stop/Soft Warning policy, exact new artifact destinations, resource schema, execution isolation, and code/protocol/shared-input hashes in
R4_M2_PRE_GPU_FINAL_CHECK_JA.mdandR4_M2_PRE_GPU_FINAL_CHECK.json. - All new case/raw/video targets were absent. Current small code/protocol hashes match the freeze; large model/CLIP/VAE/LoRA assets remain the R4-2 hashes and must be recomputed after execution approval but before prompt submission. No backend or GPU was started. Production/installed runtime, commit, push and release remain untouched. Pre-GPU fixed-task review was then requested; its acceptance is not GPU approval.
- Fixed-task review turn
72fc3888-2e94-4911-a531-39598e855576, reply13c5d2dc-0013-436c-8d23-3c5bafd48ab0returned PASS WITH CONDITIONS. Added the sole mandatory correction: backend health failure is mechanically defined by process exit, three consecutive 30-second API failures, or a 10-minute multi-signal no-progress window. Also made fresh-process run-internal RSS deltas primary. Corrected freeze SHA-256 isb31daaf81c00534df8811755d47b956ee27c16b34e00772722427fdaaef3c959. No GPU authorization is inferred.
2026-09-06 — Issue13 R4-M3 GPU rerun stopped at A parity Gate
- User gave explicit GPU approval. The first A
rerun1stopped before model load/Sampling because the isolated observer targeted a facade withoutrun_sequence. Preserved that evidence, snapshotted the observer, corrected only its hook/restore target tov2.nodes.run_sequence, and passed compile/static checks plus R4 focused CPU tests4/4. Production/runtime and R4 transport were untouched. - Ran fresh-backend Standard A as
A_standard_6x5_s131325_rerun2. It completed five new physical groups in780.011 sand produced exact 720-frame / 30-second / 24-fps / 576x864 / 32-kHz-stereo media. No OOM, allocation, NaN/Inf, crash, layout, AV, or duration fault occurred. - The frozen pre-B parity requirement failed: shared Group 1 Video/Audio were exact, but Group 2-6 Video/Audio tensors differed from the R3B baseline (
2/12exact). Seeds, prompt and geometry remained fixed. B, retry, sweep and quality verdict were therefore prohibited and not run. Group 5 B-A memory comparison remains unavailable. - A Group 5 RSS before/immediate/cleanup was
41.715 / 41.854 / 41.846 GiB; run minimum available RAM2.989 GiB, peak commit72.219%. Queue0/0was confirmed before stopping only dedicated listener PID34336and launcher PID34940; port/HTTP 8190 are offline and 8188 was not operated. - Report:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-native-guided-overlap-20260906\R4_M3_GPU_RERUN_RESULT_JA.md. No commit/push/tag/release. - Fixed-task review turn
c476af69-2f07-4dfb-a495-fea7f881199e/ reply7259d478-7029-4077-9e8c-94d94f83ce78acceptedINCOMPLETE / A BASELINE PARITY STOPwith no mandatory correction. It directs a bit-exact reproducibility cause audit before any more GPU A/B; no additional GPU authorization is inferred.
2026-09-06 — Issue13 R4-M3A Group 2 first-divergence audit
- Used only saved R3B/R4 evidence and CPU tensor reads. No backend or GPU was started; Production, R4 parameters, A/B execution and public contracts were unchanged.
- Normalizing the four diagnostic-only prompt differences produced identical prompt graph SHA
3fe71b354e8f160808930105de25966d8fb4858ae7cde3a4d82169f55955b15d. Group 1 raw AV and Group 2 protected prefix also match bit-exact. - Located the first contract divergence at Assembly Plan
chunks[0].context_frames: Production R3B initial plan5, R4 shared-Group-1 observer plan0. This is a harness metadata mismatch but is excluded as the tensor-output cause because earlier R4 Aretry2used the same0and matched all R3B Group 1-6 Video/Audio tensors. - Located the first saved tensor divergence at Group 2's generated region: Video latent
T=7onward and Audio latentT=37onward. Video generated max absolute difference was3.096475124359131; Audio was0.06531018018722534. - Derived seed, shapes, scheduler config and deterministic CPU reconstruction of noise/target/masks agree. Actual conditioning/SIGMAS/pre-sampler/first-forward/per-step SHA were never captured by the deliberately post-only R3B observer, so internal A/B/C classification and GPU runtime/kernel causality remain unresolved.
- Report and machine evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-m3a-parity-audit-20260906.uvlint could not use the user uv cache/managed-Python directories; the natural-japanese manual checklist and final read-through were used instead. No commit/push/tag/release. - The first fixed-task dispatch was rejected before transmission because the earlier authorization covered GPU-audit results only. After the user separately authorized this CPU-audit sharing, review turn
b9306a6b-b2da-486e-80d8-0b054eb80f76/ reply7e67d0bf-783b-4a52-a2d7-d8d9cb2f6c16accepted the verdict and B STOP. It clarified thatcontext_frames=0is not a sufficient cause by itself and required the next Group-2-only instrumentation to include model/wrapper runtime state in addition to tensor boundaries.
2026-09-06 — Issue13 R4-M3B Group 2 diagnostic instrumentation frozen
- Created an isolated observer and two isolated stages for the historical R3B control and R4-M3A Standard harness. The observer records hashes and compact metadata only, restores every runtime hook in
finally, deliberately terminates after a durable Group 2 trace, and never retains raw tensors. - Froze the comparison boundary order from actual conditioning and noise through first diffusion forward, every sampling step and Group 2 generated regions. Added runtime model/wrapper/attention/Dynamic RAM/PyTorch-CUDA state required by the fixed-task M3A review.
- Prepared normalized 21-node prompts using the same saved Group 1 AV. R3B uses entry
0/ Assembly Plan5; R4 uses entry0/ plan0. The runner requires tokenR4-M3B-GPU-APPROVED, refuses overwrite or non-empty queues, and forbids the R4 case unless the R3B control matches both historical generated-region SHA values. - CPU checks passed after the review corrections:
12 passed, Python compile PASS, isolated imports expose Public 7 plus trace node, R3B source-stage104/104, R4330/330, all missing/extra/mismatch0/0/0, normalized prompts exact, raw outputs absent and cache files zero. The only test warning is the environment's existingpynvmldeprecation notice. - The automated Japanese linter could not initialize the existing broken
uvcache. Applied the skill's manual checklist and final read-through instead; the report retains precise technical terms and avoids unsupported conclusions. - Formal verdict:
R4-M3B CPU INSTRUMENTATION PASS / GPU APPROVAL PENDING. No backend/GPU, Production/runtime edit, R4 B, retry, sweep, commit, push, tag or release. Result root:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-m3b-first-divergence-20260906. - After separate user authorization, sent the M3B result to
H3情報チェック. Review turn49e76964-038a-4894-a9c5-0e4a9685d85e/ reply8e48c551-6772-4c8c-ab2b-1dfe0fcf03b0accepted the verdict and required all-forward tracing plus an explicit hook-transparency Gate. Added call/step indices for every diffusion forward, dynamic full-forward comparison, protected-input/wrapper/model-options/callable/hook-registry restoration checks, return-identity and no-RNG CPU guards. Both isolated stages were resynchronized and revalidated. Fresh user GPU approval remains mandatory.
2026-09-06 — Issue13 R4-M3B Group 2 GPU trace found no current divergence
- User explicitly approved the frozen GPU diagnostic. Preflight found user 8188 absent, dedicated 8190 absent, GPU free
14,879 MiB, system available RAM43.951 GiB, and pagefile use1.53%; all frozen asset/runtime hashes were reverified before submission. - Ran R3B control Group 2 first. It reproduced historical generated Video SHA
38e8eb8f...a188and Audio SHA2b482cfc...12d8exactly in122.256 s, with eight diffusion forwards/steps and hook transparency PASS. - Stopped the R3B backend after Queue
0/0, then started a fresh R4-M3A Standard harness backend. Its Group 2 generated Video/Audio hashes matched the R3B control exactly in119.468 s; all eight forward records and hook-transparency checks completed. - Ordered CPU comparison checked 150 boundaries from conditioning/latent/noise/SIGMAS through every forward/step and generated output: different
0, missing0, first divergencenull; runtime model options, wrappers, memory and PyTorch/CUDA records match. - Corrected comparator-only capture-location noise after snapshot: the raw
model_optionsrecord containedstage_r3bversusstage_r4absolute module names, but source SHA/name/qualname were identical. Normalization ignores only that module path under the identity guard. Capture contract reverted to and retains frozen SHA39affe50...19c2; tests now pass13. - Formal verdict:
R4-M3B GPU TRACE PASS / NO DIVERGENCE REPRODUCED / PRIOR M3A CAUSE UNRESOLVED. The prior R4-M3A mismatch remains unexplained and was not reproduced. This protocol did not run R4 Native Guided Overlap B, quality A/B, retry, sweep or final video. - Both dedicated backends were stopped: R3B
7608/35932, R435984/39160; Queue0/0, 8190 LISTENING0, HTTP offline. User 8188 was not operated. No Production/runtime/contracts edit, commit, push, tag or release. - Report/machine evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-m3b-first-divergence-20260906\R4_M3B_GPU_RESULT_JA.md,R4_M3B_GPU_RESULT.json,R4_M3B_FIRST_DIVERGENCE.json, both raw traces and backend logs. A new direction decision is required before any further GPU execution. - Fixed-task completion review turn
d9289f52-2ed6-49f5-937a-73ca63e6efe3/ replyebb530f8-372d-4707-8636-84d3dc306980accepted the verdict without mandatory correction. It requires conservative wording: this one-pair exact result is notbaseline recovery confirmed. Next priority is a separately designed fresh-backend Standard Group-2 repeatability Gate before any R4 Native Guided Overlap quality A/B.
2026-09-06 — Issue13 Standard Group 2 repeatability Gate PASS
- With fresh user GPU approval, executed the frozen Standard Group 2 trace on three independent fresh 8190 backends. Every run reproduced the R3B Video/Audio generated-region hashes, all
150/150ordered boundaries,8/8forward evaluations and runtime/wrapper records exactly; hook transparency passed3/3and first divergence remainednull. - Reverified all R4-M2 model/Core/KJ Sage/Group 1 asset hashes. No OOM, NaN/Inf, unexpected backend crash, Production/runtime/contract edit, R4 B, full video, commit, push, tag, or release occurred.
- Stopped each owned backend after Queue
0/0and verified its PIDs, 8190 listener, and HTTP were gone. 8188 was absent and untouched. - Formal result:
STANDARD GROUP 2 REPEATABILITY PASS / PRIOR M3A DIVERGENCE NOT REPRODUCED. Evidence is underD:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-standard-group2-repeatability-20260906. - Fixed task review turn
0da950e6-2fc3-4c75-8a69-9daac82c9374, reply1d3c4727-9259-4832-8219-ae227a9477b1accepted the result. Next order is frozen full Standard A6x5parity first; only an exact R3B raw AV12/12result may unlock one fresh-backend R4 B run. - 2026-09-06: Completed the frozen Issue #13 R4 final GPU A/B after explicit approval. Fresh Standard A reproduced R3B raw AV
12/12exactly, was stopped at Queue0/0, and Fresh R4 B then completed6x5with exact30.000sVideo/Audio and no hard resource or integrity failure. R4 was formally rejected: Laplacian drift+39.8675 -> +74.1887, Video join2.1225 > 1.6598, Audio join3.3896 > 1.6069, highlight+282.9%, and gradient+74.2%; saturation improved63.0%and motion floor passed. Fixed task review acceptedGPU A/B VALID / FAIL / DO NOT PROMOTEwith no mandatory correction and limited the conclusion to this frozen one-scene/one-seed condition. Both owned backends were stopped; 8190 is offline, 8188 was absent/untouched, Production/runtime/contracts and commit/push/tag/release are unchanged. Per user direction, free-VRAM/RSS numbers are advisory for future approved GPU runs; actual OOM/allocation/backend/integrity failure is the stop boundary. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-final-ab-20260906. - 2026-09-06: Completed Issue #13 R5-0 Dual-Role Anchor / Tail CPU architecture audit without backend/GPU. Rejected a Group 1 frame-zero keyframe because its Core position collides with the masked previous-tail target head. Selected the narrow R5-1 prototype route: immutable Group 1 head
5f / 2 video slotsas one Video-onlyminimax_refsblock plus unchanged previous22f Video / 37T Audiomasked AV tail. Frozen Group 1 source SHA was unchanged; Core rows were21,501 -> 22,473(+4.52%), anchor Audio rows were zero, and coexistence with existing refs/final keyframe was structurally packable. VerdictPASS WITH CONDITIONS; quality, motion, wrapper behavior, and user-reference semantics remain unproven. Production/contracts/Release/A5/general A8b unchanged; snapshotpre-r5-cpu-audit-record-20260906_194942(924files). Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-dual-role-anchor-audit-20260906. - 2026-09-06: Fixed task
H3情報チェックreview turn3f6fc599-b9b0-4ca9-8e5e-8910cce205d6, reply3179d562-3414-41bd-8e38-d91ec4bf03a5accepted the R5-0PASS WITH CONDITIONSverdict. Before R5-1, hidden-reference ordering/numbering/presentation semantics, separate PackedLayout row/position evidence, and immutable run-lineage Group 1 ownership across replay/branch/Review are mandatory. No GPU/backend or Production change followed from this review. - 2026-09-06: Completed R5-1 isolated Default-OFF prototype and 88 CPU tests plus 10 actual-Core-layout cases. Source 230 code files remain identical to pre-change snapshot (924 files). Review accepted PASS WITH CONDITIONS: Core shifts target Video/Audio absolute positions by +8.333333, so R5-2 first needs position-matched control feasibility or explicit compound-intervention classification. ON disk-restart lineage recovery remains outside this prototype. Two Core imports initialized CUDA and exited without model/Sampling/backend; final source-definition runner verified CUDA uninitialized. Fixed review turn
87d13aed-685c-4b5c-b13b-22a484da8766, replyf64a95b6-219d-4f8c-9073-006290220483. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-1-prototype-20260906. No commit/push/release. - 2026-09-06: Completed R5 position-control CPU audit using unmodified Core definitions without module imports. Real/zero Reference layouts match, but zero refs are active noise/projection/attention conditioning; missing-latent refs leave 972 rows unsatisfied. Reclassified R5 under the user's second option as reference plus native AV positional shift; pure appearance causality is not claimed. Source immutable, CUDA uninitialized, no backend/GPU Sampling or Production edits. Snapshot
pre-r5-position-audit-20260906_203627924 files; result directoryissue13-r5-position-control-20260906. - Completion review
57811644-8575-4e2d-87cd-930a7065ad45/5ac513f9-ba7e-4814-87c4-d42bff3f2518accepted Position-Matched Control Audit PASS. Limit the absence of inert controls to inspected native candidates, freeze Standard A/R5 B only, retain Video/Motion/Identity/Video join/Audio join/Audio continuity guards. R5-1 conditional PASS remains; GPU unapproved. - 2026-09-06: R5-2 Audio CPU preparation completed: 25 NumPy-only tests passed, PCM numeric thresholds and actual sample cuts frozen, capture identity/nonoverwrite tested, assets8/8 SHA verified, CPU code/protocol244 hashes saved. Production230 code files unchanged versus pre-r5-2-audio-gate-20260906_223418 snapshot924; diff check0. Full R5-2 remains incomplete pending isolated runner integration and pre/post duration evidence. No GPU/backend/commit/push. Fixed H3情報チェック handoff was rejected by auto-review as outside GPU-only sharing scope; no report was delivered. Explicit CPU report-sharing permission is required before next phase. Japanese report lint could not initialize existing uv cache (os error183); manual checklist/read-through used without adding dependencies. Evidence root:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-2-pre-gpu-20260906. - 2026-09-06: After explicit CPU-sharing approval, prior report was delivered/reviewed (b738494e-a202-4736-a9c8-a2c6380d110c / bf27b4da-9d2c-4313-95dc-4d43b1f04867). Then executed authorized CPU real Assembly/Finalize integration: 85 tests PASS, natural958667 +1333 exact tail repeat ->960000, no-observer/observer PCM SHA identical, CUDA uninitialized, hidden padding/copy corruption detected. Source230 unchanged, snapshot924 pre-r5-pcm-integration-20260906_224938, artifacts23 frozen with synthetic PCM fixtures. Review 26c56b98-8172-4f3c-bd58-ff3d391d851e / 771b11bf-9d7f-49f1-88c9-bdcd8d16335a accepts CPU integration sub-Gate PASS only; R5-2 remains INCOMPLETE for runner integration, actual SHA/shutdown E2E, final execution freeze. GPU/backend/Production/commit/push unchanged. Evidence:
D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-pcm-integration-20260906. - 2026-09-06: CPU runner evidence evaluator/Finalize adapter implemented in isolated issue13-r5-runner-control-20260906; 37 tests PASS, raw12 real historical baseline hashed, PCM artifacts directly recomputed, stale/mismatch/missing/OS-error dispatch denied. Source230 unchanged, snapshot924 pre-r5-runner-control-20260906_230811, freeze14 mismatch0, diff check0,8190 absent. Review fa12ae03-03e4-430e-a2eb-26408a82ab1b / cde10501-5542-4b83-b075-864125170228 accepted sub-Gate only; actual launcher lifecycle, atomic completion publication, exclusive ownership and real shutdown E2E remain before full PreGPU PASS. No GPU/backend/model launch/commit/push. Non-sampling service E2E proposed by reviewer but not executed under current no-backend scope.
- 2026-09-06: On continuation instruction, tested minimal CPU HTTP service lifecycle (not ComfyUI/model/GPU):8 PASS. Exclusive Windows lock/controller fresh identity/atomic completion/decisionSHA/real PID-port-HTTP shutdown verified. Initial fsync read-only Errno9 fixed only in isolated diagnostic, failure cleaned A and did not launch B. Final A14284/18348/38780 and B25704/40888/39560 exited;8190=0. Source230 unchanged, snapshot924 pre-r5-launch-lifecycle-20260906_231926, diff0. Review fd337f43-4b58-494a-bbf6-6339635fb0d7 /913264ed-3b1b-42f2-8aaa-d8acd33930dc accepts CPU Lifecycle PASS only; actual Comfy prompt/history/resource/shutdown and R5 graph adapters remain. Evidence root issue13-r5-launch-lifecycle-20260906. No GPU/Production/commit/push.
- 2026-09-07: R6-A independent pre-GPU correction closed signed-Laplacian/blur acceptance, seam-threshold inconsistency and actual Sampling-evidence gaps. Isolated correction46 and reused controller/lifecycle45 tests PASS; staged policy's original9 included. Source230 code unchanged, execution259/upstream14 hash freeze exact; protocol79bbe9a5b671ed2d6d2c66d0d549e759ab28bef1b4a67938cd0e80a3049a442d. Fixed-task summary review requested before authorized single A/B GPU launch. Research snapshot353 and source-record snapshot919 verified. Earlier55-test freeze is superseded, not evidence for the new observer. Report issue13-r6a-pre-gpu-20260907/CORRECTION_GATE_REPORT.md. No Production/GPU/commit/push at this recording point.
- 2026-09-07: After fixed correction review PASS9b3c7cf2/70782869 and explicit user GPU authorization, Fresh A completed13m04s; rawAV12/12 and historical decoded/PCM exact, normal PID/port/HTTP shutdown PASS. Conditional B started, then user requested stop and A-only analysis. B was interrupted at Group2, last log2/8; launcherexit1, all six owned PIDs gone,8190/HTTPoff, GPU_AUTHORIZATION revoked. B quality untested, not FAIL. A-only CPU frame/seam/PCM analysis confirms residual late texture strengthening (Lap+39.8675 units=+16.04%), strongest seam15.0833s, quiet RMS~-54.9dBFS intact PCM; no full playback/listening claim. Evidence USER_PAUSE.json and A_ANALYSIS/REPORT.md; source snapshot919 pre-r6-A-analysis-record-excluding-cache-20260907_034446. Review requested; no B resume, Production change, commit or push.
- 2026-09-07: Fixed A-analysis summary review c4a59d67-786c-40ab-9567-5e33f662d9bf / f1b2f17a-a988-43c1-91b7-be16f79cbe39 accepted baseline-reproducibility PASS and R6-efficacy-unjudged classification. Preserved limits: late edge strengthening is observed, not causal proof; static joins and PCM validity do not establish perceptual playback/listening PASS. B remains user-paused; all owned backends ended.
- 2026-09-07: User authorized B continuation ahead of newly supplied Turbo/Spectrum workflows. NewB0b0b607953934a6bb2e7ed7142784ae6 completed13m08s after fail-closed reuse ofA93d9852bd2fb40c9bce257a4cf054ba3 and resume CPU53/53/freeze289/fixed review542d0853/3bfd0ff3. InterruptedB preserved. Frozen numeric screens allPASS: Lap drift39.86749->13.36942,Gradient72.91% reduction,Audio reasons[]; scope/noise invariants exact. Strong endpoint benefit concentratesG6; peak excursion improvement15.15%, saturation+4.10% and seam+9.39% worse within guards. Formal numeric candidate only, perceptual review pending, one seed, no Production/Issue13-resolution claim. Backend normalshutdown/PID/port/HTTP complete, new workflows copied only. Report issue13-r6a-pre-gpu-20260907/FINAL_REPORT.md; completion review requested. Source snapshot919 pre-r6-B-completion-record-excluding-cache-20260907_063034; inherited cache ACL caused ordinary helper failure before filtered snapshot. No commit/push/release.
- Completion summary review25ee8534/9747f470 accepted NUMERIC candidate only; full30s playback/listening must precede next GPU phase. Source230 hashes unchanged, diff check0, ownedB41648/40316/36912 absent and8190 listener0. No independent local-media review by reviewer, no new GPU or Production promotion.
- 2026-09-07: Completed isolated R6-A Turbo4v1.2 / actual6-step Euler/simple / Draft0.30MP adapted
6x5sA/B. Fresh sharedG1; A OFF/B ON; effective544x544, empty prompt, AudioON/Balanced22/KJSageON. Execution contract PASS:720f/30s and PCM960000 each, exact Assembly, complete5-record traces, resource hard/contract0, normal A/B shutdown and8190 offline. R6 evaluations G2=1 required no-op,G3=1 no-op,G4=.9,G5=.95440975,G6=.9; rawAV exact G1-3 and differs G4 onward. Numeric promotion evidence NEGATIVE/HOLD: endpoint Lap+8.28%,Gradient+11.85%,Contrast+113.44%,Shadow+71.07% worse; highlight99.14% and mean-saturation20.20% better; motion~1.6% lower and seam metrics warn worse. Empty-prompt semantic excursions limit Issue13 interpretation; no perceptual A/B pass, no invalidation of separate Turbo8 result, no Production/promotion. Fixed review bd47c1b1/3cfbb3e0 accepted this narrow verdict without correction. Reportissue13-r6a-turbo4-6step-0p3mp-20260907\TURBO4_6STEP_0P3MP_REPORT.md; snapshot919pre-r6-turbo4-report-record-filtered-20260907_081550; no commit/push/tag/release.
2026-09-07 — Review continuation GPU reuse and timing check
- Reused the user-supplied
MiniMax_H3_00058_.jsonworkflow exactly by SHA, added a short I2VA motion prompt, and changed the Continuum execution controls to2 x 5s,Review Each Chunk, andSave + Auto Resumefor the live check. Turbo4 v1.2, Euler/simple 6 steps, Draft 0.30 MP, First Image, Balanced 22, and Audio Continuity ON remained fixed. - Queue 1 reported
0 reused / 1 generated, one sampled physical group and 6 total sampling steps, producing 120 frames / 5 seconds in 87.11 seconds. Queue 2 reported1 reused / 1 generated, again one sampled physical group and 6 steps, producing 240 frames / 10 seconds in 96.80 seconds. - Stored Chunk 1 remained exact across both Queues by file SHA, manifest SHA, Video latent SHA, Audio latent SHA, seed, prompt hash, and plan. GPU Sampling time was 47.38 seconds for Queue 1 and 48.38 seconds for Queue 2. The second Queue therefore sampled only Chunk 2; its similar total time came from decoding, finalizing, and saving the complete 10-second sequence again.
- Added this explanation, the measured example, and the Detailed Report counters to the English
README.mdbeside the beginner Review procedure.README_JA.mdis intentionally deferred until the final Japanese translation pass. - Related Review tests passed
71; Markdown diff check passed. The first test invocation reached 66 passes but five fixtures hit the inherited Windows pytest Temp ACL, then all 71 passed with a fresh workspace--basetemp. The dedicated backend stopped at Queue empty; owned PIDs exited, port 8190 closed, and HTTP went offline. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-resume-gpu-check-20260907.
2026-09-07 — Illustrated English manual draft; live capture pending
- Replaced the two-chunk-only README walkthrough with a general 11-section step-by-step continuation manual. Included review/retry/finish differences, fixed seed, Back to Settings/Return to Review, Take branches, explicit regeneration, extension compatibility, and Terminal Merge limits. Corrected the prior timing explanation: the residual wall time includes other workflow work and was not individually attributed to decode/finalize/save. Detailed Report now points to the actual ComfyUI setting label.
- Copied six user-provided UI crops unchanged into
docs/images/v38-manual/; source/destination SHA checks PASS. Six PNG links and all new local section anchors PASS; README diff whitespace check PASS. README_JA SHA remainsF8104F76277C3A937DF153181EA13C555D7186FA88EE8AF39868FCF922EAB4AB. Four scoped pre-edit documents were snapshotted and verified after the standard helper hit the existing cache ACL. - User offered the running
127.0.0.6:8188ComfyUI /00058/ Turbo four-step workflow for live screenshots and timing. Both Computer Use kernels failed withfailed to write kernel assets/ Windows error 3, including after resets, before UI access. Workspace and TEMP paths exist; underlying internal path is unidentified. No new generation, screenshot capture, user-backend stop, runtime/code edit, package build, commit, or push. Live validation and final review remain pending; current scope and recovery point are recorded in PROJECT_STATE and the capture plan.
2026-09-07 — Computer Use recovered; live retry failure discovered
- Used existing Chrome00058/user backend8188 after UI workflow export and five-file scoped snapshot. Kept Turbo4v1.2/4steps/Euler/simple/Spectrum/KJ/Draft0.30MP/empty prompt; changed seed control to fixed and Run to Review Each Chunk (Progress auto-enabled). Initial UI Run produced120frames/5s in75.232s, actual review panel confirmed; sampler time30.447705s. Saved video00064.
- Try this chunk again failed before sampling with
Regenerate Current requires an existing reviewed unit. Compared full prompts: only review_action changed. Stored manifest still review_ready/unit1 and chunk file SHA matches. Compatible-head failure is observed; underlying cause unresolved. No more GPU sampling and no code/runtime modification. User backend remains running with queue0/0. - Added four real screenshot crops, corrected execution instructions to top-right blue Run (distinct from in-node Run), and added a draft-validation warning instead of claiming retry PASS. Source/runtime run_storage difference is confined to existing Windows atomic-replace retry and is not established as causal. Full evidence/report:
D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-manual-live-20260907\REPORT.md. README_JA/Production/package/commit/push unchanged. Remaining live captures and final handoff remain pending.
2026-09-07 — Retry diagnosis instrumented; waiting for user-owned backend restart
- Approved repair investigation independently read the saved chunk on CPU: saved-contract self-validation and compatible-head lookup both PASS, prefix1/review_ready, CUDA false before/after. Actual retry-contract divergence remains unknown; runtime weight/wrapper changes are a hypothesis only.
- Added bounded, hash-only contract-section diagnostics on missing-head errors without relaxing identity checks or altering Sampling. Added five CPU regression tests covering private-value omission, CPU RNG/input preservation, no persisted writes or unsafe reuse, no saved run, and successful lookup bypass. Relevant suites162 PASS; diff check PASS.
- Source four-file snapshot/SHA4 verified; full installed runtime181-file snapshot succeeded. AST-verified narrow one-file diagnostic deployment deliberately excludes the pre-existing source-only Windows atomic retry change; runtime SHA
16f0bb17b863dc0b1aef95656deae9a4783338b35145f010bec95a661feae6df. New imports zero. User backend8188 remains running, queue0/0; diagnostic requires its restart. No new GPU execution, no underlying-fix PASS, no commit/push. See evidenceRETRY_DIAGNOSTIC_STATUS.md; manual remains a validation draft.
2026-09-07 — Identity repair and completed-review UI follow-up
- Confirmed LoRA set-order and cold/warm VAE dtype false identity changes, repaired scoped observation without altering model weights or Sampling;169 focused tests passed. Runtime run_storage SHA59593daf4e2021bad2e6849427aae95a439220745f582b406ea1d240b19ebe45. User restart then initial/continue succeeded; revision99684d4a78b5855b kept Chunk1 exact and generated only Chunk2. Retry GPU not yet proven.
- User reported no screen transition. Actual endpoint canonical revision was complete/unit2 while JS allowed only review_ready. Fixed UI to retain completed retry (only with valid unit), history, settings and return controls; hide no-op Continue/Finish. Existing backend already handles completed retry; no schema/transport/status change. Deployed matching JS SHAB571FD4E061B113931D5470DD9844CAD77390B5227CC0EF05576ECCDA9662812 after7+7web and3record snapshots. Added real full-JS lifecycle test covering1..6chunks and surrounding states.
- Focused128 passed; additional UI15 passed/manifest1failed; full1139 passed/2failed (manifest hashes and process-global CUDA initialization assertion in isolated-passing identity test). Full-suite PASS not claimed. Browser reload blocked by auto-review due to unsaved tab; did not bypass or submit another Queue. Await user save/reload or explicit reload approval, then actual Try this chunk again GPU gate and completion handoff. Evidence COMPLETED_REVIEW_UI_REPAIR.md in v38-manual-live-20260907. No commit/push/release.
2026-09-07 — Provenance reuse bypass isolated and repaired
- User requested the actual Chunk1 review screen. Manual Regenerate From Chunk1 through currentUI queued29e95cfc; incorrectly reused2/generated0 under a new nonce1 revisioncb575f0c6aa00705 (video00070,52.35s). Stopped further Runs. Confirmed provenance reuse omitted regenerated chunk-contract checks, unlike ordinary prefix reuse.
- Added opt-in sampling-contract enforcement for prepare's provenance reuse from the branch boundary; compare original producing revision, preserve inherited prefix and same-nonce interrupted resume, leave explicit selected-Take semantics and stored-self-validation intact. Added9 on-disk tests; review/backend107 PASS. Fixed blank canvas history text using custom draw, frontend19 PASS plus actual draw assertion PASS. No masks/seeds/noise/Audio/Assembly/schema changes.
- Snapshot runtime181/source scoped4, verified; narrow2-method AST stage PASS, no import changes. Deployed run_storage0E402AD9 and JS7B318BE7 with queueempty. User-owned backend NOT restarted; restart required before testing. Target remains Chunk1 review screen, no automatic Chunk2. Current original revision99684d4a78b5855b Chunk1 SHA E96ED58CB17829F5E218B2F54E67F226C0B2912040681B3077B35DA3F81E11B4,Chunk2 B8E74670B5BFE388D511F46F621C2D235335F17FF4E59BA80CB181DF7290BBE1 recorded before the attempt. Saved videos not deleted. No publication or next research phase.
2026-09-07 — V3.8 Continuation Regression Final Gate A/B/C GPU complete
- Completed Fresh A Standard/Masked AV, Fresh B Compatibility/Reference Context and Fresh C Standard+R6-A at 896×896, 4×8 seconds, seed595713724167856, Turbo4v1.2, Euler/simple/6, Sage, Balanced22, Audio ON. Each produced768 frames, PCM1024000, raw Video4/Audio4 and clean8190 shutdown; A/B/C Group1 is bit-exact and hard-stop count is zero.
- C applied gain1.0 at Groups2/3/4 and is a valid no-op. A/C trace boundaries, raw AV8/8, PCM and decoded AV are bit-exact. This establishes GPU no-op parity but does not test R6-A efficacy. A/B both softened rather than reproducing Issue13 sharpness ratchet; B increased saturation/skin-highlight and triggered PCM waveform-edge review at8s. Keep Standard default, Compatibility unpromoted, R6-A Experimental/Default OFF, Issue13 open.
- First A attempt remains Harness FAIL only: observer misread the valid1ch mask as24ch latent; Production unaffected. Observer-only correction passed CPU8/8 and syntax before fresh execution. Evidence:
v38-continuation-regression-final-gate-20260907\FINAL_ABC_REPORT_JA.md. User backend8188 untouched/queueempty; no Production/public schema/commit/push/tag/release change.
2026-09-07 — V3.8 prelaunch CPU total audit; Release HOLD
- Audited the dirty authoritative source without altering Production/workflows. Guarded full pytest1,149 PASS/2 FAIL of1,151,99 modules; syntax Python226/JS7 PASS. Actual Core0.34.5 CPU registration7/7, public callable schema, native PackedLayout, prompt plan and both V3.8 workflow graphs PASS. The cache-test failure is its unqualified Core startup; identical harness passes four cache assertions with explicit CPU/W2 bootstrap. Kept the original red full-suite evidence.
- Found stale root node.zip (75 files,29 current package files absent, isolated import ModuleNotFoundError branch_provenance), Registry hash4/source Manifest7 mismatches, and unsynchronized source-only Windows atomic replace retry in installed run_storage. Other two runtime hash differences are whitespace-only/AST-equal. Source documentation links resolve;44 doc references are outside Registry payload.
- Independently reproduced CPU-mode Detailed Report requesting CUDA lazy init via hardening.py mem_get_info. First suite blocked14 memory-probe calls plus1 unqualified Core startup; standalone strict tripwire confirmed CPU-mode trigger without actual initialization. Audit processes ended CUDA-uninitialized. Existing user backend received only GET queue/schema reads; no new backend, model load, GPU Sampling, Queue, cancel or restart.
- Spectrum32-node/38-link workflow remains exactly user-supplied. Core19-node/25-link graph and Spectrum socket types match; frontend-only note/bypasser nodes are not misclassified as missing backend nodes. Corrected only the external diagnostic checker to honor Core MatchType rather than treating it as a fixed AUDIO mismatch.
- Evidence/report: D:\Codex_test_results\ComfyUI-H3-Continuum\v38-prelaunch-cpu-audit-20260907\CPU_PRELAUNCH_REPORT_JA.md. Record-only snapshots1+1 verified under before-cpu-prelaunch-audit-state/worklog-20260907_182042. No implementation/package rebuild/runtime deployment/commit/push/release. Next scope is narrow diagnostic fix plus final package/runtime parity freeze, not algorithm changes.
2026-09-07 — Correct ZIP scope; fix CPU diagnostics and release hashes
- User clarified backup ZIPs are unrelated to release. Identified the actual prior test target as the source checkout's node.zip, not Continuum.zip/Continuum3-8.zip; withdrew the inference that its failure blocks launch. Preserved all ZIPs and workflows. Added an explicit correction to the historical audit report.
- Changed hardening.py's diagnostic-only guard from CUDA available to CUDA already initialized. Added3 CPU regression tests (counter/RNG/tensor safety, mock initialized counters, failure reporting) and made the real-Core cache subprocess explicitly CPU-only with initialization tripwires and a process-local Kitchen INT8 capability bypass. Source/runtime snapshots verified before changes.
- Focused60 PASS; full1,154 PASS in55.60s with no deselection and no recorded CUDA initialization requests. Updated Registry104 declared hashes and source Manifest254 hashes. Constructed105-file isolated Registry payload without ZIPs; actual CPU import/public7 nodes/PackedLayout/memory report/atomic JSON roundtrip PASS.
- Narrow runtime sync proof: only hardening.py and run_storage.py are semantically different, new dependency only stdlib time; two other byte differences are AST-identical whitespace. Deployment guard found user queue1/0 and aborted before copying. No restart, cancellation, GPU Sampling, workflow mutation, ZIP change, commit, push, or external publication. Runtime disk sync remains pending idle queue.
2026-09-07 — Correct README accuracy without changing workflows
- Implemented the six reviewed documentation corrections in README.md and README_JA.md: Create Video in the save path; nonce0 auto/positive fixed/interrupted resume versus review Retry; seven-node export boundary with retained old IDs; Core Audio Switch versus rgthree; current Advanced and Manual sizing UI; latest CPU1,154 PASS separated from historical527/GPU evidence. Scoped duplicate historical compatibility passages to their original releases instead of removing history.
- Six one-file pre-change snapshots verified (before-readme-correction-*-20260907_195135); four further one-file snapshots before final consistency edits. Refreshed only affected README/Registry-manifest hashes, preserving declared payload membership. Registry104/104 and source Manifest254/254 match.
- Final focused CPU tests90/90 PASS in the guarded environment; all recorded exits remain CUDA-uninitialized with zero blocked initialization requests. Local README links53/53, balanced Markdown fences and whitespace checks PASS. Evidence: D:\Codex_test_results\ComfyUI-H3-Continuum\v38-readme-correction-20260907\final.xml and CUDA guard logs.
- No workflow, image, ZIP, generation/runtime code, Sampling, backend lifecycle or publication change. Runtime deployment remains a separate pending action; no GPU-quality or release promotion.
2026-09-07 — Prepare approved V3.8 main publication
- User approved the supplied Spectrum graph as generic V38 JSON plus matching ZIP; updated README links, package declarations, migration and regression expectations without touching either asset. Both README openings retain git clone/pull and point to LightX2V and the workflow downloads. Existing V3.8 source and manual images are included in this publication; unrelated node.zip is excluded, while the exact historical V3.7 ZIP is required by source regression tests.
- Verified per-file source snapshots. Added exact-byte Git attributes for manifest/ZIP reproducibility, not generation changes. Initial full guarded CPU run: 1,154 passed/1 failed; failure is optional SageAttention import probing CUDA in the real-Core CPU cache harness, stopped before initialization. Isolated that optional import in the test child only and reran focused checks; evidence focused.xml alongside the preserved failing full.xml. Production, installed runtime, user backend, GPU and supplied workflow bytes unchanged.
- User requested immediate upload and no further workflow tests. Complete only publication integrity checks and the explicitly authorized main commit/push; do not claim another full-suite PASS or Registry/GPU acceptance. Remote/main was equal to local HEAD before staging. Publication hash/result are returned to the user after push.
2026-09-08 — Review restart UI hotfix, user-authorized main publication
- Confirmed the UI keeps an old saved review after seed/prompt/reference edits, allowing an incompatible Retry. Dedicated CPU integration had already shown ordinary Queue correctly starts from chunk1 for changed inputs; 4x15s synthetic Full/Review sampling inputs matched. The rendering-static report is not solved by this change.
- Added browser-only generation-input snapshots (including connected inputs) to withdraw stale actions without granting backend reuse. Stale queued Review intent returns to normal Continue/Next compatibility checks. Added a transient, non-destructive Start again from Chunk 1 button using existing manual boundary/automatic nonce controls, consumed once after Queue. Preserved manual boundaries, saved takes, serialized widget order and all Python runtime contracts.
- Added real-JS lifecycle assertions for edits/undo, upstream prompt/media/bypass/reconnection, edits during generation, stale queue intent, direct manual regeneration, and one-shot restart. Focused frontend suite15 PASS after adapting extraction fixtures and simulating each completion's real queued settings. No GPU/model load/backend restart/workflow mutation/runtime deployment. Snapshots9 files verified. Final CPU checks, reviewer handoff and main push recorded in the task evidence folder and response.
- Final lifecycle hardening also preserves edits during a delayed initial history fetch and prevents a queued baseline from another Run Name being used. Focused61 PASS; full1155 PASS in46.62s; Registry105/source256 declared hashes PASS; all recorded CUDA exits false/blocked0. H3情報チェック returned PASS / PUSH TO MAIN OK after delayed delivery (user also explicitly permitted review waiver for this hotfix only). Reviewer noted updated_utc must remain a write-time revision signal; current history reads do not write it. Publication scope remains main source only, no new tag/Release/Registry assets; 15-second static/noise unresolved.
2026-09-08 — Local mode-selection setup repair
- Reproduced the user's Full -> Review pre-generation screen takeover in a real-JS test before editing runtime logic. Added a transient setup-selection token; old saved completion no longer hides the settings on this user action. Explicit Return to Review and newly queued review results can reopen review. Mode selection preserves saved progress and does not force a new run or erase takes. Queue from setup cannot inherit a hidden old Retry/Finish; manual regeneration remains supported.
- New cases cover complete/partial saved state, Storage Off/On, normal ready card, serialized values, history refresh before/after Queue, delayed first history, prior-flight completion, different Run Name, explicit return and subsequent review completion. Existing facade fixture now supplies Queue and completion rather than equating injected history with execution. Focused90 PASS; final full1155 PASS/48.44s; first red reproduction and first full1154/1 fixture failure preserved. Syntax and whitespace checks pass; package/source hash declarations refreshed without changing membership.
- Seven one-file source snapshots verified. Evidence: D:/Codex/_test_results/ComfyUI-H3-Continuum/v38-mode-settings-20260908. Fixed H3情報チェック received self-contained report; deployment pending review. User backend GET only at127.0.0.6:8188, queue0/0; no GPU, Queue, restart, workflow change or publication.
- Deployment completed after fixed reviewer returned PASS / local deployment OK (self-contained report review, not independent file inspection). Backed up installed JS1 file, deployed only web/project_id.js, verified source=runtime=live HTTP200 SHA d84877206cc1142dbfe7fce3f7e9c1b6edb8a22cae5ab117a1407f9435f6dea7. Live object_info lists H3ContinuumSamplerV38 from custom_nodes.ComfyUI-H3-Continuum. Browser visual confirmation remains for user hard refresh. No commit/push, browser reload, backend lifecycle or GPU execution.
2026-09-08 — Review boundary lifecycle correction, CPU Gate
- Reclassified the user's latest run from “Chunk 1 completed” to saved-history reuse: prompt/history showed zero new physical groups, all three logical chunks reused, and the previous canonical
completerevision retained. This disproved the prior presentation-only completion claim. - Corrected the first Queue after an explicit Full Video -> Review Each Chunk selection. If canonical saved review history exists, the API payload alone starts a non-destructive automatic-nonce branch at Chunk 1; serialized/visible settings and saved Takes are preserved. Sampler execution clears the transient setup state, reloads backend truth, and subsequent Continue actions inherit the branch for Chunk 2 and Chunk 3.
- Added exact backend and real-frontend
3 x 5slifecycle coverage. Focused Review/UI regressions passed123; full CPU regression passed1156/1156in50.69s. Registry/source Manifest hashes, JavaScript syntax and whitespace checks pass. Evidence:D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-boundary-20260908. Fixed H3情報チェック returnedPASS / local runtime deploy OK; its queued-identity note was applied by requiring the exact setup token and Run name beforeonExecutedmay release setup. Actual UI/GPU acceptance is deliberately not claimed. No workflow, Production/Sampling, commit, push or release change.
2026-09-08 — Live Review Each Chunk lifecycle accepted
- Replaced reliance on the unsupported frontend
beforeQueuePromptpath with the ComfyUI1.49.6widget lifecycle:beforeQueued->serializeValue->executed->execution_success. Added the exact lifecycle regression and deployed onlyweb/project_id.jsafter a one-file runtime snapshot. - Validation before deployment: supported-lifecycle test
1 passed; full guarded CPU suite1156 passedin50.60s;node --checkPASS. Final source/runtime JS SHA-256:78BA840BB4B57D21BC17AAB9F5CA3C89E3BC3C7AAE1A027BBD3FF54512717A53. - User-operated live 3 x 5s prompts completed as intended: Chunk 1 prompt
8a9a2275-7f31-4e3b-bf7c-95747feefc2bgenerated physical group 1 and opened Chunk 1 review; Chunk 2 prompt5427b850-0490-4b07-8243-77f737be7ba3reused one/generated one and opened Chunk 2 review; Chunk 3 promptef38c575-faa8-45ea-8485-6ced1500298ereused two/generated one and showed complete. Sampler node312was not cached in any run. - Final canonical run:
run_c05b0771405f40f78fbdd8055632e458, revision83cf80b0a4ff378b, active group revisionsf3e4cf7f740ce4c7ad469238,a49003f2b27ceb1088b3f776,fad8dff7f8641f6e4b5dcc78. Final preview was a continuous 15-second result. - No Python Production/Sampling, workflow, backend lifecycle, commit, push, tag, release, or Registry change. All GPU Queue actions were performed by the user. Record snapshot:
D:\Codex\_snapshots\ComfyUI-H3-Continuum\before-v38-live-review-record-20260908_160004(2 files).
Render History diagnosis after live acceptance
- Read-only live inspection captured the exact failure:
RunStorageError: selected Take provenance is invalid: group revision parent is missing. The displayed selections9db2a9287ef18796b770db00andf968bd1131c20912e207de0abelong to old sampling lineages. - Current
project.jsoncontains 76 revisions across 59 lineages. The accepted canonical lineage963eb7095602aef239570525bf999e77b71d8bfa2d5c5cf40f96c78a163fd72econtains onlyf3e4cf7f740ce4c7ad469238 -> a49003f2b27ceb1088b3f776 -> fad8dff7f8641f6e4b5dcc78. FrontendtakeCatalog()does not filter lineages, while backend_validated_take_selection()deliberately does. - Classification: feature exists and the backend rejection was safe; frontend catalog eligibility and rendering were defective. Fixed by filtering to canonical-lineage, fully resolvable ancestry; stale selections automatically return to a canonical Take with
Automaticaction. History output is compact and draw-height independent. - Focused Review/History regression
59 passed; full CPU regression1156 passedin52.09s. Runtime deployment changed onlyweb/project_id.jsafter Queue0/0; source/runtime SHA-256 match at549A8150BC926BF142758B1610E3730547D6ABCFBCDDB0B37ED8BAAAF86B38CB. Runtime snapshot:D:\Codex\_snapshots\ComfyUI-H3-Continuum\before-v38-render-history-runtime-deploy-20260908_162700. - No Take was adopted and no generation was started by Codex. Live browser confirmation is pending reload because the existing ComfyUI tab is marked unsaved; it was not forcibly reloaded.
Issue #20 workflow persistence hotfix (2026-09-10)
- User authorized repair and publication to main from pinned
6b453fe. The only runtime change isserialize: falseon the read-only Render History widget inweb/project_id.js; existingoptions.serialize: falseis retained. Backend widget order, Sampling, Audio, Run Storage, public workflow/ZIP bytes, and Easy Loader Bypass behavior are unchanged. - Added a regression using the real Continuum frontend script with modeled compact store persistence and forceInput migration. It covers ordinary serialization, store serialization, positional/named round trips, Advanced/Review/History/completion views, and API-input parity. It fails on the unmodified runtime; all 45 pre-existing fixture cases remain green. Full CPU pytest: 1154 passed, 2 skipped; frontend fixture: 46/46 passed.
- Validation ran on an isolated GitHub Actions CPU runner after byte-verified snapshots made with
tools/snapshot.ps1. This is not a real-browser or GPU validation. Windows source/runtime and saved Takes were not accessed or changed. Existing corrupted workflows need separate recovery; model/VAE/LoRA reset reports and the separate Bypass issue remain unverified. No Release/tag or Registry publication. Fixed H3情報チェック cross-task handoff was not sent from this session.
Prompt skill v1 guide update — GitHub (2026-09-10)
- Replaced repository-root
H3-Continuum-Skill-v1.zipwith the user-supplied updated prompt-guide package. Published ZIP SHA-256:DDE84DF88640B573031591F49F09A1BEB4DEE4F5199E7143F51BB650637F8D1D. - The existing English and Japanese README download links remain valid because the stable repository filename is unchanged. Package validation confirmed seven file entries and
write-continuum-h3-prompts/SKILL.md; runtime code and workflows were not changed.
Prompt skill explicit boundary opener update — GitHub (2026-09-11)
- Repacked repository-root
H3-Continuum-Skill-v1.zipfrom the validated installedwrite-continuum-h3-promptsskill after merging the preferred Chunk 2+ body openerContinuation of Chunk N.. UpdatedSKILL.md,references/continuity.md, andreferences/examples.md; the rule remains a prompt-level mitigation candidate rather than a sampler fix or quality guarantee. - Published package SHA-256:
7895D3E0C638D1D87E99F67EED13C04B85F5F8B76598C13EE9C48A01E707D808. ZIP signature504B0304, seven file entries, six skill files, requiredSKILL.md, and byte-level equality of all packaged skill files to the installed source were verified. Existing English and Japanese README links remain unchanged and valid; runtime code and workflows were not changed.
Prompt skill Timeline syntax hardening — GitHub (2026-09-12)
- Updated repository-root
H3-Continuum-Skill-v1.zipwithout changing runtime code or workflows. Promoted the standalone Timeline-header rule intoSKILL.md, refreshedreferences/continuum-contract.mdto the current V3.8 parser baselinebd12073c4322a88ce666fe062b7bb265f5de131a, added simple3x5,3x10, and2x15Timeline examples plus inline-input normalization toreferences/examples.md, and added the same critical syntax note toREADME_JA.md. - Validation:
tools/snapshot.ps1completed before modification; ZIP retained exactly seven entries and passed ZIP integrity validation; targetedtests/test_v2_prompts.pypassed 52/52. New ZIP SHA-256:144324850D4BA086938DA5F4ED7B02FF0668DEC554F092695328240E43FEC3EC. This is documentation/prompt-skill only; Sampling, Run Storage, public node schema, workflows, and runtime tensors are unchanged. H3情報チェック handoff remains pending because that cross-task destination is unavailable from this GitHub Actions job.
Loader persistence repair (2026-09-21)
- User approved Image Enable repair, Core Audio workflow replacement, and Core Video plus Enable-free 24fps adapter, including commit/push. Legacy Audio/Video IDs remain deprecated/loadable. Main Sampler, Sampling and stored generation contracts are unchanged.
- Source baseline
3d7c2fd692. Real source was snapshotted using tools/snapshot.ps1 before edits on an isolated Actions runner. Original-code regression exposed 3 failures; focused CPU/Node suite 46 PASS. See docs/LOADER_PERSISTENCE_REPAIR.md. - Windows worktree/runtime/browser and GPU were not accessed. Live frontend 1.53.6 tab switching/save/reopen and fixed H3情報チェック handoff remain PENDING. No Release/tag/Registry publication.
Validation result
- Original-code regression: 6 cases; 3 expected failures and 0 collection/runtime errors.
- Focused CPU/Node suite: 46 passed, zero failures/errors/skips.
- JavaScript module syntax, changed Python compilation, workflow JSON/ZIP parity, endpoint consistency and git diff whitespace checks passed.
- Official workflow structural guard preserved every non-loader node, each original link, and non-layout workflow metadata.
- Scope: isolated GitHub Actions CPU runner, not Windows deployment or a real-browser/GPU test.