Skip to content

[feat] video: add FastH3 dual-frame adapter and a writable Windows Codex sandbox - #292

Merged
kunkunblueberry merged 5 commits into
ThinkFlowLab:mainfrom
Probiotics5333:feat/fasth3-dual-frame
Nov 17, 2026
Merged

kunkunblueberry merged 5 commits into
ThinkFlowLab:mainfrom
Probiotics5333:feat/fasth3-dual-frame

Conversation

@Probiotics5333

@Probiotics5333 Probiotics5333 commented Nov 11, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Two scoped changes, requested together.

  • FastH3 first/last keyframes ([feat] providers: FastH3 支持首帧与尾帧双锚点生成 #290). New fasth3_dual video adapter. Its first/last-frame mode accepts at most one first and one last keyframe, uploads the anchors in first/last order, and submits fl2va with res_multistep at 20 steps. frame_indices marks the tail so it keeps the last frame of H3's native 17n+5 grid instead of being trimmed to the requested duration. Its reference mode is the existing 8-step omnimodal path. The modes are exclusive: first/last-frame mode rejects audio and video references and uses at most two images.
  • Windows Codex sandbox ([bug] agent: Windows 下 Codex 沙箱降为只读且命令发现失败 #291). Only on win32, append -c windows.sandbox="elevated" so --ignore-user-config cannot drop the Windows sandbox backend and downgrade --sandbox workspace-write to read-only. Approvals and isolation flags stay enabled, and no bypass switch is added. CLI discovery uses Get-Command opendirector in PowerShell and keeps command -v opendirector in POSIX shells.

The model-capability guide now documents the adapter, its two modes, the 20-step first/last path and the tail-frame trade-off.

Linked issue

  • Fixes #N included

Fixes #290
Fixes #291

Labels

  • Labels applied: type:feat, area:agent, area:provider, area:harness, modality:video

Test plan

  • Local:
    • pytest tests/test_fasth3_inputs.py tests/test_video_capabilities.py tests/test_creation_parameters.py — 90 passed.
    • pytest tests/test_creation_generation_service.py tests/test_openai_images.py — 28 passed.
    • pytest tests/test_native_runtime.py -k "codex or windows or proxy" — 8 passed, including the win32/linux/darwin parameter case.
    • python scripts/generate_contracts.py --check, npm run check:i18n, npm run check:browser-locale — passed.
    • npm test — 765 of 766 passed. The remaining browserHost case fails while creating a temporary virtualenv for a non-ASCII temp path (WinError 5); the same python -m venv call on that path fails independently of this branch.
  • Real service: three accepted FastH3 jobs, listed under Provider verification.
  • Left to CI: the production build, isolated browser workflows and installed-wheel checks in checks / checks.

Evidence that it works

The new adapter option in Model connections, captured against the isolated fixture API and Vite in both locales.

English (Video adapter → FastH3-Omni (first/last keyframe anchors)):

FastH3 dual adapter, en-US

Chinese (视频适配方式 → FastH3-Omni(首尾帧锚点)):

FastH3 dual adapter, zh-CN

Tail anchor preserved by a real run — order is last_frame anchor, generated first frame, generated last frame, first_frame anchor:

Anchor versus generated frames

Provider verification

  • Real-provider calls are described honestly, or none were made

Real vLLM-Omni FastH3 on this machine (FastH3-Omni-INT8, adapter fasth3_dual), not a double. Three jobs were accepted, succeeded and adopted:

Accepted Remote task Output
2026-11-06 video_gen_dual_26f7af99855a4be39f5c9128839a5622 4 s, 480P, 16:9, 832×480
2026-11-10 video_gen_dual_807e3ecf5eb4456798f213e8b1ac05c9 4 s, 480P, 16:9, 832×480
2026-11-10 video_gen_dual_cabeb72e2f6b45919031c47c6bd0239e 4 s, 480P, 1:1, 480×480

Each job recorded numInferenceSteps: 20, the fasth3_dual adapter, and first_frame / last_frame references in its input snapshot. Each output has 107 frames at 24 fps (4.458 s): the expected 17n+5 grid for a 4 s request, and the reason the dual-frame output is not trimmed back to 4 s. The stored technical check passed (H.264 + AAC, 832×480).

The tail anchor is preserved: the last generated frame matches the chosen last_frame.png (SSIM 0.972; first frame against first_frame.png 0.983), as shown by the comparison image above.

Bilingual UI

  • Both locales are updated

The new adapter label ships in zh-CN and en-US, and the guide update keeps the existing Chinese documentation set.

Scope note

The repository prefers one concern per pull request. These two changes were requested as a single delivery, so they stay together; an unrelated OPENDIRECTOR_TEMP_DIR override that had travelled with the FastH3 commit was removed, and the diff now contains only the two linked issues.

@github-actions github-actions Bot added area:agent Agent runtime, assistant chat and tools area:docs Documentation / 文档 area:provider Model providers and generation protocols labels Nov 11, 2026
@Probiotics5333 Probiotics5333 added area:harness Harness and skill integration (Default, OpenCode, Codex, Claude Code, Jiuwen, DSH) modality:video Video generation, editing or rendering type:feat New capability or behaviour labels Nov 11, 2026
@wuzhf9

wuzhf9 commented Nov 12, 2026

Copy link
Copy Markdown
Collaborator

REQUEST CHANGES — 9afdc3e67a3fd84e7f7dcf353f187180600b628d — declared Fixes #290, Fixes #291

The normal two-anchor path is connected, but the following boundary cases prevent closing #290 as implemented. I reproduced these through the shared generation service with both source=ui and source=agent, using isolated projects, real imported media and mocked provider HTTP.

  1. [P1] The same asset in both frame slots loses the first anchor. Create two image canvas items backed by the same asset, assign one as first_frame and the other as last_frame, then submit. The job is accepted and its input snapshot contains both anchors, but the actual multipart request uploads one image with frame_indices=[-1]. CreationGeneration.task() deduplicates reference IDs by asset, and generation_references() collapses roles into an assetId -> role dictionary, so the last role wins. Preserve reference/slot identity through request construction; asset-byte reads may still be deduplicated. Add a regression comparing the accepted snapshot with both actual uploads and their indices. See task construction and role lookup.

  2. [P2] Missing anchors are accepted, contrary to [feat] providers: FastH3 支持首帧与尾帧双锚点生成 #290. No anchors, first-only and last-only all create jobs; the no-anchor case submits t2va. The new tests explicitly expect these cases to succeed. [feat] providers: FastH3 支持首帧与尾帧双锚点生成 #290 instead requires preflight rejection when either anchor is missing. Enforce the agreed rule in shared business validation and test that rejection creates no job and makes no provider request. If single-anchor and text-to-video support is intentional, reconcile the issue, mode semantics and documentation before treating [feat] providers: FastH3 支持首帧与尾帧双锚点生成 #290 as closed. See new test expectations and text fallback.

  3. [P2] Audio/video references are silently ignored instead of rejected. Add an ordinary audio or video reference alongside the two anchors in first/last-frame mode. Submission succeeds; the extra reference becomes context_only, and only the images reach the provider. [feat] providers: FastH3 支持首帧与尾帧双锚点生成 #290 requires an explicit rejection with no job. Validate the original reference set before that filtering for this adapter/mode, preserving other adapters' existing semantics. Add acceptance-level rejection tests for both entry points. See reference filtering.

  4. [bug] agent: Windows 下 Codex 沙箱降为只读且命令发现失败 #291 still needs actual Windows acceptance evidence. The platform-specific argument change and PowerShell discovery guidance look reasonable, retain workspace-write, and add no sandbox-bypass flag. The new test detects the previous missing argument, but it mocks sys.platform and checks arguments only. Release checks runs on Ubuntu. Please provide a real Windows Codex run demonstrating shell execution, project reads/writes and opendirector invocation, plus an outside-workspace write denial under controlled test conditions. The evidence currently available does not establish that runtime acceptance criterion.

  5. Complete the bilingual user documentation. The new fasth3_dual instructions and tail-frame/duration trade-off are documented only in Chinese in docs/guides/video-model-capabilities.md. Add matching English user guidance as required by AGENTS.md and docs/development/i18n.md.

Ran: 128 existing relevant tests on this tip (118 media/capability/parameter/generation/image cases and 10 selected native-runtime cases); 16 supplementary boundary probes; generated-contract and i18n checks; diff whitespace check. The new Windows parameter test fails on the previous implementation for win32 and passes for Linux/macOS. Release checks is green on this exact tip; GitHub reports a clean merge against current main (af428f9c8ddbbd616106f1bdd4d189370175395c at review time).

Not run in this review: live provider calls, local browser workflows or real Windows Codex execution. The three real FastH3 jobs described by the author cover the successful generation path, not the boundary cases above.

@kunkunblueberry kunkunblueberry left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Requesting changes at 9afdc3e6. The adapter structure and the Windows sandbox half are sound — I checked the win32 flag, the i18n parity, the contracts, and the tests, and none of them is the problem. The problem is that the three boundary cases the linked issue calls out explicitly are not enforced, and one of them silently produces a wrong request.

1. Blocker — the same asset in both frame slots loses the first anchor

packages/core/src/opendirector/application/creation_generation.py:113 de-duplicates reference ids by asset:

reference_ids = list(dict.fromkeys(str(r.asset.assetId) for r in job.inputSnapshot.references if r.transmission != 'context_only'))

while :140 keeps both role entries for that asset in referenceRoles. packages/core/src/opendirector/integrations/reference_inputs.py:122 then collapses the roles into a per-asset dictionary:

role_by_asset = {str(value['assetId']): value['role'] for value in task.get('referenceRoles') or []}

so the last role wins, and references() (reference_inputs.py:32) de-duplicates the ids a second time. The result: two canvas image items backed by the same asset, assigned to first_frame and last_frame, pass validation (domain/creation/generation.py:448-451 only checks "more than one per role", and :380 only rejects duplicate (referenceId, assetId) pairs, which these are not), and the accepted job's snapshot really does contain two anchors — but the uploaded request contains one file and frame_indices=[-1].

The existing tests cannot catch this because they build the task dict directly and give each slot a distinct asset (tests/test_fasth3_inputs.py:353-357).

Minimal fix: carry slot identity instead of asset identity through request construction — iterate (referenceId, role, assetId) pairs and dedupe only the byte reads, so that images and frameIndices stay aligned ([0, -1]). A regression should compare the accepted snapshot (two anchors) against the actual multipart uploads and their indices.

2. Blocker — missing anchors are accepted, contrary to the issue

Issue #290's acceptance criteria say "首帧或尾帧缺失、槽位绑定多张图片时预检拒绝并给出可读原因", and "首尾帧模式下添加视频或音频参考时被明确拒绝,不产生任务".

Nothing enforces the minimum. domain/creation/generation.py:448-451 only caps each slot at one image and checks its kind; there is no rule requiring both slots. So with fasth3_dual in first_frame mode:

  • no anchors → task='t2va', 20 steps (integrations/providers.py:162-164);
  • first only → frame_indices=[0];
  • last only → frame_indices=[-1].

And the tests encode that as intended behaviour: tests/test_fasth3_inputs.py:353-357 parametrises single-slot success, and :406-430 asserts a no-anchor request is submitted as text-to-video.

Decide this explicitly rather than leaving it implicit. Either enforce the pair in validate_generation for videoAdapter == 'fasth3_dual' and videoMode == 'first_frame' (raise → no job, no provider call) and add the "both present" and "one missing" cases, or change issue #290's acceptance criteria and the mode description. As it stands the PR closes an issue whose stated rule it does not implement.

3. Major — video and audio references are silently dropped instead of rejected

packages/core/src/opendirector/application/creation_canvas.py:139-142 marks every non-text/image reference context_only in first_frame mode:

if spec.kind == 'video' and spec.parameters.videoMode == 'first_frame':
    ignored = kind not in {'text', 'image'} or kind == 'image' and (...)

Because domain/creation/generation.py:390 computes kinds from transmission != 'context_only', the guard at :446 (any(kind in {'audio', 'video'} for kind in kinds)) can never fire, and task() drops the reference at creation_generation.py:113. A user who connects a video or audio reference in first/last mode therefore gets a job that quietly ignores it.

The rule set cannot express the prohibition either: domain/input_rules.py:77-81 still advertises allowedKinds=['text','image','video','audio'] with maxVideos=3, maxAudios=3 for fasth3_dual, and only the UI precheck (src/backend/rules.ts:213-218) knows about the restriction. The HTTP, CLI, and agent entry points have no equivalent.

Minimal fix: validate the pre-policy reference set for this adapter and mode — or drop video/audio from the dual rules and enforce the per-mode prohibition in validate_generation — so that a rejection happens before job creation.

What I verified as correct, and would keep

  • The win32 sandbox argument: packages/core/src/opendirector/agent/native.py:123-127 appends -c windows.sandbox="elevated" only under sys.platform == 'win32', keeps --sandbox workspace-write and the approval flags, adds no bypass switch, and uses the same quoting idiom as the existing -c web_search="disabled". tests/test_native_runtime.py:283-294 asserts both halves.
  • Slot ordering and the tail anchor: reference_inputs.py:211-218 reorders uploads first-then-last regardless of canvas order and records symbolic 0/-1 indices, and integrations/fasth3.py:121-126 skips the duration trim when a last_frame role is present, which is what makes the tail anchor survive. There is no client-side off-by-one in frame_indices — they are symbolic, and num_frames is not sent.
  • Mode exclusivity for the adapter: input_rules.py:76-81 sets videoModes=['reference','first_frame'] and frameRoles=['first_frame','last_frame'] for fasth3_dual only, leaving plain fasth3 untouched.
  • Contracts and i18n: the videoAdapter enum is updated consistently in the Python literal, the TS type, generated.ts and creation.schema.json; models.service.adapterFastH3Dual exists in both locales and the zh/en key sets are identical at this head.
  • _profile_adapter returns None for the fasth3-omni profile (integrations/providers.py:47-52), so no ModelAdapter method is missing or unreachable for this adapter — it is driven through prepare_inputs and the generic same-origin result download.

Minor

integrations/fasth3.py:65 sends seconds as str(math.ceil(seconds)), and with the trim now skipped for tail-anchor jobs the delivered clip can exceed the requested duration (a 4 s request yields 107 frames = 4.458 s at 24 fps), while execution.durationSeconds still records the request. That is inherent to "put the tail on the 17n+5 grid", but the receipt and the probe then disagree; recording the effective duration would keep them consistent.

@FayeSpica

Copy link
Copy Markdown
Collaborator

Please resolve conflcts.

- New `fasth3_dual` video adapter. Its first/last-frame mode takes up to one
  first and one last keyframe and submits `fl2va` with `res_multistep` at 20
  steps; its reference mode keeps the existing 8-step omnimodal path. The two
  modes are not combined: first/last-frame mode rejects audio and video
  references.
- Anchors upload in first/last order with the slot in the filename, and
  `frame_indices` records the tail, so the tail keeps the last frame of H3's
  native 17n+5 grid instead of being trimmed away.
- Reuses the generic per-slot frame roles that landed for Ark/Seedance; the
  adapter only adds the frame-index upload and the plain prompt.
- Document the adapter, its two modes and the tail-frame trade-off in the
  video capability guide.

Contracts are regenerated for the new adapter value.

- `pytest tests/test_fasth3_inputs.py tests/test_video_capabilities.py
  tests/test_creation_parameters.py`: 90 passed.
- `python scripts/generate_contracts.py --check`: verified.
- `npm test`: 765 of 766 passed; the remaining browser-host case fails in
  this environment while creating a temporary virtualenv (WinError 5) and
  does not touch this change.
- `npm run check:i18n`: 0 problems.

Fixes: ThinkFlowLab#290
…#291)

## summary

- Add `windows.sandbox="elevated"` on win32 so `--ignore-user-config`
  cannot drop the Windows sandbox backend and downgrade
  `--sandbox workspace-write` to read-only. Approvals and the isolation
  flags stay enabled; no bypass switch is added.
- Locate the CLI with `Get-Command opendirector` in PowerShell while POSIX
  shells keep `command -v opendirector`, in the system prompt and in the
  execution context handed to the native runtime.
- Document the behavior in Chinese and English, and cover the win32, linux
  and darwin parameter differences with a regression test.

## validation

- `pytest tests/test_native_runtime.py -k "codex or windows or proxy"`:
  8 passed, including the new platform-parameterized case. The two
  remaining real-CLI parameterizations need a logged-in native CLI and
  fail in this environment for every runtime, not only Codex.

Fixes: ThinkFlowLab#291
…nkFlowLab#290)

## summary

- Emit one anchor per requested slot role, so one asset that fills both the
  first and the last slot reaches the wire twice instead of collapsing into
  the last role through an asset-keyed map.
- Require both anchors when submitting `fasth3_dual` first/last-frame mode,
  and reject audio or video references there instead of letting the
  reference policy turn them into `context_only` without a word.
- Keep configuration editing lenient: the stricter rules run only for a
  submission, through the shared validation both the UI and the Agent use.
- Cover the same-asset slots and the two submission rejections with tests.

## validation

- `pytest tests/test_fasth3_inputs.py tests/test_video_capabilities.py
  tests/test_creation_parameters.py tests/test_creation_generation_service.py`:
  103 passed.
- `pytest tests -q` on Windows: 2250 passed. The 77 failures and 20 errors
  are the same environment failures (owner-rights ACL, missing native CLIs,
  JiuwenSwarm module) seen on `origin/main` in a side-by-side run; the only
  set difference was one flaky baseline-only oplock case.

Fixes: ThinkFlowLab#290
…kFlowLab#290)

## summary

- Translate the video capability guide into English and keep the existing
  Chinese text in `video-model-capabilities.zh-CN.md`, matching the other
  paired guides under `docs/guides`.
- Document the `fasth3_dual` first/last-frame mode, its both-anchor
  requirement and the rejection of auxiliary audio or video references in
  both languages.
- Update the documentation index and the user-guide links to the new pair.

Refs: ThinkFlowLab#290
@Probiotics5333

Probiotics5333 commented Nov 17, 2026 •

Copy link
Copy Markdown
Collaborator Author

Rebased onto main (729b49bc) and addressed the boundary findings in 2c906d5d:

  1. Same asset in both slots. The asset-keyed role map is gone; each requested slot role now emits its own anchor, so one asset filling both the first and last slot uploads twice with frame_indices=[0,-1]. Regression: test_dual_h3_same_asset_fills_both_frame_slots.
  2. Missing anchors. Submission of fasth3_dual first/last-frame mode now requires both anchors in the shared validation (submitting=True, used by the UI and the Agent); the zero-anchor t2va path is unreachable and the provider no longer emits it. Editing a draft stays allowed. Tests: test_dual_h3_requires_both_anchors_at_submission (empty, first-only and last-only rejected; draft edit still allowed).
  3. Audio/video references. They are rejected explicitly for this adapter and mode before the reference policy can mark them context_only; reference mode keeps its existing omnimodal semantics. Test: test_dual_h3_rejects_auxiliary_media_at_submission.
  4. Real Windows Codex acceptance. OpenDirector launched standalone Codex CLI 0.169.0 with the elevated workspace-write sandbox. In a disposable project, PowerShell wrote and read back a workspace file, Get-Command opendirector found the CLI, opendirector project show read the project, and opendirector asset write-text persisted a text asset. The same write to a separate directory writable by the host was denied by the sandbox, and no file was created. The run used no skills or sandbox bypass. Details are in eb665947.
  5. Bilingual guide. video-model-capabilities.md is now English, the Chinese text moved to video-model-capabilities.zh-CN.md, and the index and user-guide links follow the pair. Both languages state the both-anchor and no-auxiliary-media rules.

Duration follow-up. eb665947 preserves execution.durationSeconds as the requested duration and shows the measured output duration separately in generation history when available. The regression covers a 4 s request and a 4.458 s output (107 frames at 24 fps).

Validation on 2c906d5d: pytest on the four related suites — 103 passed; full Python suite — 2250 passed with the same environment failures as origin/main (owner-rights ACL, missing native CLIs, JiuwenSwarm module); npm test — 831/832, the remaining case is the known temporary-venv environment failure.

Validation on eb665947: generation-history regression tests (4 passed), sandbox argument tests (3 passed), TypeScript and i18n checks, and the real Windows run above. The full frontend suite still has the known Windows temporary-venv permission failure.

## summary
Keep the requested duration in the execution snapshot and show the measured output duration alongside it in generation history.

## verification
The generation-history regression covers a 4 s request with a 107-frame, 24 fps output (4.458 s), and jobs without a measured duration.

On Windows, OpenDirector's real native runtime launched Codex CLI 0.169.0 with the elevated workspace-write sandbox. In a disposable project, Codex used PowerShell to write and read back a workspace file, resolved opendirector with Get-Command, read the project with opendirector project show, and persisted a text asset via opendirector asset write-text. A separate directory writable by the parent process rejected the same write from the sandbox and was left unchanged. This used the standalone Windows CLI and skipped skill discovery; no sandbox bypass was used.

Refs: ThinkFlowLab#290, ThinkFlowLab#291
@Probiotics5333

Copy link
Copy Markdown
Collaborator Author

Thanks for the reviews.

  • @wuzhf9: Item 4 is verified on Windows: PowerShell and project CLI reads/writes succeeded, and a controlled write outside the workspace was denied. The updated summary has the acceptance details and links the earlier fixes.
  • @kunkunblueberry: The three FastH3 boundary fixes are in 2c906d5d. For the duration note, eb665947 keeps the requested duration and displays the measured output separately (4 s requested, 4.458 s delivered in the regression).
  • @FayeSpica: The branch was rebased, and GitHub currently reports no merge conflicts.

@wuzhf9 wuzhf9 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@kunkunblueberry kunkunblueberry left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-reviewed at eb665947. All three findings from my 11-13 review are addressed, and I verified each by reverting it and watching the new test fail rather than by reading the diff alone. Approved.

The three findings

1. Same asset in both frame slots. reference_inputs.py:128-130 now collects a list per asset instead of a last-wins dict, and :185-187 emits one anchor per requested slot role. Mutation check: restoring the old last-role-wins form makes exactly test_dual_h3_same_asset_fills_both_frame_slots fail, and nothing else.

2. Missing anchors. domain/creation/generation.py:466-474 gates on submitting, requires slots == {'first_frame', 'last_frame'}, and rejects auxiliary media in the same block — before the reference policy can mark anything context_only. submitting=True reaches it from both entry points (application/creation_generation.py:379, domain/creation/canvas.py:246), so the UI and the Agent take the same path, and draft editing stays allowed. Mutation check: disabling that block fails both test_dual_h3_requires_both_anchors_at_submission and test_dual_h3_rejects_auxiliary_media_at_submission.

Worth noting the placement is the substantive part: the check reads references rather than transmitted, which is why it catches video/audio inputs that the first_frame policy would otherwise silently drop to context_only. That was the actual defect, and it is fixed at the right layer instead of in the UI.

3. Bilingual guide. docs/guides/video-model-capabilities.md (English) and video-model-capabilities.zh-CN.md now exist as a pair, and docs/README.md links both the way kling-video and user-guide already do. Both state the both-anchor and no-auxiliary-media rules.

Duration note

eb665947 keeps execution.durationSeconds as the requested value and shows the measured output separately (creationJobs.ts:58-65, GenerationHistoryPanel.tsx). That is the right resolution of what I raised as a minor: the receipt no longer contradicts the probe, it carries both facts. The regression pins 4 s requested against a 107-frame 4.458 s output.

Verification

  • tests/test_fasth3_inputs.py: 23 passed. I ran it with a pillow_heif stub because this machine's venv lacks that optional dependency — the stub replaces the import only, no test logic.
  • Mutations above, each restored afterwards; the worktree is clean.
  • Release checks on eb665947: all 7 jobs success, including static, python, and all three browser lanes. CodeQL also green.
  • Branch is 0 commits behind main and merge-tree is clean, so no rebase is needed behind this result.

What I did not verify

The Windows Codex acceptance (#291) is author-reported: you describe standalone Codex CLI 0.169.0, an elevated workspace-write sandbox, successful PowerShell and opendirector reads/writes, and a denied write outside the workspace. I have no Windows host here and did not reproduce it. Its commit message records enough detail for a reader to repeat, so I am taking it as recorded rather than testing it — and the argument change itself is covered by tests/test_native_runtime.py, which asserts the flag is added on win32 and that --sandbox workspace-write survives.

Merging.

@kunkunblueberry
kunkunblueberry merged commit ec205d1 into ThinkFlowLab:main Nov 17, 2026
12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

area:agent Agent runtime, assistant chat and tools area:docs Documentation / 文档 area:harness Harness and skill integration (Default, OpenCode, Codex, Claude Code, Jiuwen, DSH) area:provider Model providers and generation protocols modality:video Video generation, editing or rendering type:feat New capability or behaviour

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[bug] agent: Windows 下 Codex 沙箱降为只读且命令发现失败 [feat] providers: FastH3 支持首帧与尾帧双锚点生成

4 participants