Skip to content

[diagnostics] inspect CodeQL paths for upstream PR 371 - #1

Closed
wuli666 wants to merge 61 commits into
mainfrom
feat/issue-114-optional-plugins
Closed

wuli666 wants to merge 61 commits into
mainfrom
feat/issue-114-optional-plugins

Conversation

@wuli666

@wuli666 wuli666 commented Nov 16, 2026

Copy link
Copy Markdown
Owner

Temporary diagnostic PR for ThinkFlowLab#371 at de7fd6b.

Run the existing CodeQL workflow on the identical feature branch so its owner can inspect full SARIF data-flow paths. The upstream PR reports seven path-injection locations, but its code-scanning details are inaccessible to the fork author.

Do not merge. Close this diagnostic PR after the remaining data flow is understood and the upstream check is fixed. No extra product changes or scan suppressions are included.

wuli666 and others added 30 commits November 8, 2026 12:12
…wLab#221)

## summary

- Match chatgpt-image-* model ids in the openai-gpt-image profile.
  Image tools stop sending response_format, which OpenAI rejects
  for these models, and reference edits use image[] parts with the
  documented 16-image limit.
- Parametrize the profile and adapter tests over gpt-image-1 and
  chatgpt-image-latest.

## validation

- pytest tests/test_openai_images.py: 16 passed; the two new
  chatgpt-image-latest cases fail without the fix.
- pytest tests: 1098 passed, 6 skipped.
- lint-imports: 4 kept; generate_contracts.py --check: verified.
- npm test: 675 passed.

Fixes: ThinkFlowLab#219
…lowLab#217)

## summary

- Split creation project storage into commit documents, catalog scans,
  commit-log reads, and publication. Lifecycle commands stay on
  CreationProjects.
- Keep command routing in application.creation_commands. execute()
  still delegates there and does not import application services.
- Split canvas tool recipes into context, image edits, segmentation,
  and video. The catalog still exports TOOLS, Source, ToolContext,
  Output, and select_image_objects.
- Point storage tests at the module that now owns each patched helper.

## validation

- lint-imports kept all four layer contracts.
- pytest on creation, tool, agent, canvas-source, library, and moss
  modules: 679 passed, 3 skipped.
- test_rejects_removed_runtime_and_invalid_selection_before_acceptance
  also fails on main in this environment with 503 because the bundled
  agent CLI is not installed. The full tests tree was not run.

Fixes: ThinkFlowLab#177

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
…lowLab#223)

## summary

- gpt-image models no longer declare image.cutout. Their edits redraw
  and move the subject, so the returned alpha does not line up with
  the source pixels, and gpt-image-2 rejects transparent backgrounds.
  Cutout without a selection or segmentation model now reports the
  missing model instead of calling OpenAI.
- Add a regression test and update the gpt-image capability assertion.

## validation

- New test_gpt_image_is_never_used_as_a_cutout_matte fails without the
  fix and passes with it; tests/test_openai_images.py passes.
- pytest tests: 1097 passed, 6 skipped.
- lint-imports: 4 kept; generate_contracts.py --check: verified.
- npm test: 675 passed.

Fixes: ThinkFlowLab#222
## summary

Add a linked star demonstration to both READMEs with a reduced-motion
still image and a reproducible Pillow generator. The original repository
illustration contains no adoption counters or actual GitHub actions.

## validation

- Decode the GIF: 960x560, 5.6 seconds, infinite loop, matching seam.
- Inspect rendered still, pointer, click, and hold frames.
- Regenerate both assets and compare bytes; verify README asset links.
- npm run check:i18n and git diff --staged --check pass.
## summary

- After perf-outputs, run the ThinkFlowLab#180 storage micro-bench and reuse
  that run's drag and render sample.
- Publish a trend artifact and a one-line comparison against the
  previous run and a pinned baseline.
- Fail the advisory lane only for a regression past the configured
  ratio and absolute floor, or when the suite itself breaks.
- Keep the daily v0.1.x decision on gate parity.

## validation

- python -m pytest tests/test_daily_benchmark.py (13 passed)
- python scripts/nightly/daily_benchmark.py at 600 commits / 200
  nodes and 200 projects finished in 131s, within threshold, with
  the baseline still unseeded.

Fixes: ThinkFlowLab#213

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
…nt (ThinkFlowLab#232)

## summary

- A sample that parsed but lost a required measurement was recorded in
  `missing` and then ignored by the failure decision: an empty `drag`
  block under `--ui-required`, or a storage sample missing a tracked
  field, ended the summary in `within threshold` and left the lane
  green while the same summary printed `FAILED` for that clause.
- Required measurements now decide `notify`: storage metrics always,
  drag metrics only while the lane requires the UI sample, so an
  advisory run without `--ui-required` keeps its current behaviour.
- Name the gap explicitly as `INCOMPLETE <ids>` in the one-line
  summary and stop claiming `within threshold` when one exists. The
  artifact carries the same list as `incomplete`.

## validation

- python -m pytest tests/test_daily_benchmark.py -q --noconftest
  (16 passed, 3 new: empty-but-parsed UI sample, storage sample with a
  removed field, partial UI sample that stays advisory)
- build_report with `{'drag': {}, 'render': {}}` and ui_required=True:
  notify True, `INCOMPLETE ui.drag.node.frameP95,ui.drag.pan.frameP95`,
  exit 1
- build_report with `cold-open.data.open` removed: notify True,
  `INCOMPLETE storage.coldOpen.open`, exit 1
- Same inputs without ui_required still notify False

Refs: ThinkFlowLab#216
…SS pagination (ThinkFlowLab#218)

* [feat] audio: add a connection-scoped voice library with on-demand MOSS pagination

* [fix] audio: correct voice uploads and distinguish missing previews

* [fix] audio: enforce the upload connection and align voice validation errors

* [fix] audio: select compatible voices without previews from the whole card
…hinkFlowLab#210)

## summary

- Download a compatible image result from an unconfigured host only after
  pinning a checked public address: every DNS answer must be public, the
  connection uses that IP with the original Host header and TLS name, and
  every redirect hop is revalidated with a fixed hop limit.
- Keep provider credentials on the exact configured origin. An explicitly
  configured local service still works; a different local port does not.
- Persist stable `errorDetails.code`/params for matting and concept
  failures so both locales no longer depend on backend wording, and keep
  the text mappings only as a fallback for older records.
- Restore hidden placeholders before collision renaming, accept only ASCII
  numbering suffixes, and clarify reference ports for empty assets.

## validation

- `tests/test_result_download_security.py`: 18 rejected destinations,
  mixed and empty DNS answers, pinned IP with Host/sni/signed path,
  redirect and hop-limit cases, explicit local origin, generation recovery.
- `Release checks` on `b6625d6e`: success (Python, frontend, build,
  installed wheel, isolated browser, voice library, assistant panels).

Refs: ThinkFlowLab#208, ThinkFlowLab#209
…hinkFlowLab#229)

## summary

- Stop passing `--ephemeral` to codex so its rollout files are written and a
  later turn can resume the thread. The old flag existed precisely because
  turns did not resume.
- Before resuming a codex session, check that its rollout exists under
  `$CODEX_HOME/sessions` or `archived_sessions`. When it is gone, start a new
  native thread in the same execution directory with the full conversation
  context, then save the new thread ID so the next turn resumes it.
- Keep `work.mkdir(exist_ok=...)` tied to a saved session rather than to a
  usable session ID, so the recovery path does not raise `FileExistsError` on
  the existing work directory.

## validation

- `tests/test_native_runtime.py`: the fake CLI asserts the recovery prompt
  carries the previous turn's content and no `resume` argument, while
  opencode/claude/default still resume; the isolation-flag test asserts
  `--ephemeral` is never added even when the CLI supports it; a new test
  covers `codex_session_available()` before and after a rollout exists.
- `Release checks` on `2d849c64`: success.

Fixes: ThinkFlowLab#228
…kFlowLab#242)

## summary

- Audio settings now drop a stored voice selection when the resolved model
  publishes no voice library. `serviceAudioSettings` runs on the composer's read
  path (`getNodeGenerationSettings`), so a node reopened after its model
  configuration changed is no longer blocked.
- Before this, only a model *switch* cleared the selection. Editing the same
  configuration into a plain OpenAI-compatible TTS kept the voice, and
  `voiceIssue` reported it as belonging to another service. The composer offers
  the library only while that capability exists, so there was no control left to
  replace it and every later edit re-applied the same stale value.
- Dropping the voice is the same outcome the switch path already produces, and a
  model that still publishes a library keeps its selection untouched.

## validation

- `tests/voiceLibrary.test.ts`: 13 passed. Four cover this change — the rules
  level (blocked while stale, unblocked after servicing, unchanged for a model
  that keeps the library) and the read boundary the composer actually uses.
- Removing the `serviceAudioSettings` hunk makes the read-path test fail, so the
  regression is pinned.
- `node --test tests/*.test.ts`: 666 tests, 658 passed, 8 failed. The same 8 fail
  on a clean `main` checkout here because `node_modules` is not installed
  locally (`ERR_MODULE_NOT_FOUND: react`); the count is identical with and
  without this change.
- `npm run check:i18n`: 2935 keys, 0 problems. No new user-visible string.

Refs: ThinkFlowLab#218
## summary

- Set the npm package, lockfile, and Python distribution version to 0.1.2 for today's patch release.
- Report the same version from the local API, the DeepSeek ACP client, and the Claude plugin manifest.

## validation

- Parsed package.json, package-lock.json, and pyproject.toml and confirmed each project version is 0.1.2.
* [docs] agents: add opendirector-review skill

## summary

- Add the PR tip review recipe at
  .agents/skills/opendirector-review/SKILL.md.
- Record PASS, WITHHOLD, or REQUEST CHANGES from the diff,
  the linked issue, and Release checks on that tip SHA.
- Keep merge, force-push, and bypassing red CI outside this skill.

Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

* [docs] repo: add a pull request evidence template

## summary

- Add .github/pull_request_template.md asking for a demo or an
  explicit N/A, plus issue, label, test, provider, and locale notes.
- Point the CONTRIBUTING checklist at that evidence section.
- Have opendirector-review withhold or request changes when a
  user-visible tip has neither media nor a justified N/A.

Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

* [ci] repo: re-trigger Release checks

## summary

- Empty commit so Release checks run again on this docs-only tip.
- The previous run failed in Voice library workflows; this branch
  does not change that UI, tests, or workflows.

Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
…hinkFlowLab#239)

* [feat] providers: add official Kling 4.0 and Flash video generation

## summary

- Add optional Kling text, frame and image-reference video routes
  through the existing durable job, recovery and adoption workflow.
- Enforce model-specific output limits and preserve safe provider
  error codes and request IDs without changing default selections.
- Add bilingual setup, capability descriptions, documentation and
  protocol, recovery and browser regression coverage.

## validation

- Focused Python regression: 195 passed; frontend suite: 690 passed.
- Build, i18n, generated contracts and bilingual browser checks passed.
- One live Flash Omni request produced and adopted a 720P video
  lasting 3.04 seconds with audio. Other route checks use mocks.

* [bugfix] providers: secure Kling result downloads

## summary

Validate Kling result destinations and pin downloads to a checked
public IP while preserving the original host and TLS identity.
Apply the same protection to recovered downloads and cover private
addresses, DNS responses, and redirects with regression tests.

## validation

- Run result download security, Kling, OpenAI image, and proxy tests.

---------

Co-authored-by: Barkly1te <haochongliao@gmail.com>
…les (ThinkFlowLab#253)

* [bugfix] video: send Seedance first and last keyframes as distinct roles

Ark's video generation API keys each keyframe by its content role, so a
first_frame plus last_frame pair has to reach the request as two separate
image_url entries. The Seedance 2.0 profile never declared frameRoles, and
the shared reference pipeline labelled every image with a single role, so
the validator rejected last_frame before a job was created.

- declare frameRoles ['first_frame', 'last_frame'] on the seedance-2.0 profile
- carry the requested slot through generation_references as imageRoles,
  keeping one image per slot and rejecting a duplicated slot
- allow two keyframe images in first_frame mode when the model declares
  last_frame; single-frame models keep the limit of one
- send each image with its own role, while reference mode still uses
  reference_image for every image

* [bugfix] video: keep an unlabelled image a first frame instead of guessing

An image whose reference role is the default 'reference' now maps to
first_frame, matching the single-image behaviour that predates last_frame.
Previously the second such image was silently promoted to last_frame, which
would have turned a duplicate connection into a bogus keyframe pair instead
of surfacing it as a slot conflict.
…b#193)

* [bugfix] assistant: preserve chat when browsing assets and history

## summary

- Keep assistant visibility independent from asset search and generation
  history, with adjacent desktop panels and usable narrow layouts.
- Cover drafts, dismissal paths, both locales and active reply/generation
  preservation in isolated browser tests.
- Make related shortcut/navigation tests wait for readiness and select
  their expected locale, and document the new validation entry point.

## validation

- npm test: 585 passed
- npm run build
- npm run check:i18n: 2862 keys, 0 problems
- Chromium browser checks: assistant auxiliary panels, double Shift and
  running-reply navigation (isolated API and model fixtures)

Fixes: ThinkFlowLab#118

* [feat] timeline: add editable captions and text overlays

## summary

- Retain timed text, style, layer order, fonts and source assets in
  editable timelines through shared UI and Agent commands.
- Share Chinese text layout between preview and MP4 export and retain
  exact editor revisions for each published output.
- Exchange editable timeline packages with resource validation,
  dependency protection and recoverable imports.
- Add bilingual UI, regression coverage and acceptance documentation.

## validation

- npm test: 621 passed.
- Related Python checks across nine test files: 189 passed.
- Generated contracts, i18n catalogs and frontend build passed.
- Prior macOS x64 and native arm64 browser/wheel acceptance is recorded
  in docs/development/issue-87-validation.md.
- Windows acceptance is pending; the full Python suite has documented
  baseline failures and an interrupted native integration run.

Refs: ThinkFlowLab#87

* [bugfix] Fix Windows agent files and sandbox initialization

## summary

- Add handle-relative execution file I/O on Windows and pin POSIX roots,
  rejecting links, redirected execution roots and nonregular inputs.
- Initialize OWNER RIGHTS directories before file creation, guard DACL
  updates, coordinate leases across processes and clean up failures.
  Preserve denies and restricted-token capability boundaries.
- Fix native session resume, private skill snapshots, atomic JSON
  replacement, history cache verification and Windows launcher quoting.
- Keep active hidden-source references and structured prompts independent
  of display visibility, while preserving accepted input snapshots.
- Add platform and native sandbox regressions and record timeline text,
  editable-package, installed-wheel and POSIX validation limits.

## validation

- Staged code and tests match the frozen final validation snapshot.
- Full Python: 1149 passed, 5 existing skips; related tests: 95 passed.
- Fresh installed wheel: 403 passed, 5 existing skips, including real
  dsh and Jiuwen engine tests using local model fixtures.
- Node: 621 passed; seven browser suites passed; contracts and i18n passed.
- Real Linux: 45 passed, 49 Windows-only skips; 2000 rejected inputs and
  100 file round trips kept the descriptor count at four.
- Four Windows skips require POSIX change time; one file-symlink case
  remains limited by Windows error 1314. Existing skip gates are unchanged.

Refs: ThinkFlowLab#87

* [bugfix] agent: journal Windows ACL writes and support safe resume

## summary

- Journal owner ACL deltas at the native setter's successful write,
  preserving external denies through retries and exceptional cleanup.
- Resume validated DeepSeek workspaces without rewriting existing ACLs
  or retaining file probes that block ordinary PowerShell writers.
- Add tracked Windows browser process ownership, isolated environments
  and native memory sampling, with real host and WinAPI regressions.
- Retain existing business assertions and record native Windows, wheel
  and seven-group browser acceptance with explicit platform limits.

## validation

- Complete Python: 1,368 passed, 15 skipped; Node: 720 passed.
- Related native regressions: 81 passed; wheel: 448 passed, 7 skipped.
- Seven browser groups, Windows wheel build and fresh install passed.
- Contracts, i18n, browser checks, four import contracts and build passed.
- Existing POSIX gates and WinError 1314 file-symlink limit documented;
  no test deletion, broader skips or sandbox relaxation.

Refs: ThinkFlowLab#87, ThinkFlowLab#193

* [bugfix] preserve native cleanup ownership and live text edits

## summary

- Consume ACL removal deltas at successful native writes and refresh
  Claude skill aliases safely when continuing the same session.
- Preserve uncertain POSIX outputs instead of deleting a replacement
  pathname, while retaining exclusive creation and error propagation.
- Discover PATH Python in CI, render live text move/trim previews,
  and align hidden-reference browser checks with shared semantics.
- Add regressions, review evidence and a focused bilingual Windows
  follow-up; keep historical platform results tied to their revisions.

## validation

- Python: 1321 passed, 74 platform/capability skips; Node: 721 passed.
- Full browser gate: 19/19 passed on the second complete run; initial
  structured-editor timeout and successful focused rerun recorded.
- Real text editing, render, coverage, archive and hidden references
  passed; build, contracts, i18n and four import boundaries passed.
- Independent repair review found no further required code fix.
- New native Windows cleanup cases and rebuilt wheel remain pending.

Refs: ThinkFlowLab#87

* [bugfix] agent: recheck ACL guards before native writes

## summary

- Reject an already broken oplock at the native setter boundary, keeping
  successful-write receipts and one-time cleanup consumption intact.
- Strengthen the real concurrent cleanup regression and retain actual
  post-write fault injection after the added pre-write check.
- Record native Windows follow-up, fresh wheel and browser evidence,
  preserving prior results with their original revision boundaries.

## validation

- Complete Python: 1,378 passed, 17 skipped; Node: 721 passed.
- Related modules: 184 passed, 3 skipped; ACL modules: 77 passed.
- Fresh wheel: 195 passed, 3 skipped; three browser groups passed.
- Build, contracts, i18n, browser checks and four import contracts passed.
- Retain POSIX gates and the Windows file-symlink privilege limitation;
  no test deletion, expanded skips or sandbox relaxation.

Refs: ThinkFlowLab#87, ThinkFlowLab#193

* [docs] plan remaining PR 193 work and Windows handoff

## summary
- Assign remaining integration, native ACL validation, fixes, review and final acceptance to Windows.
- Define source freeze, targeted retesting, CI ownership and evidence requirements to avoid repeated cross-device handoffs.
- Publish aligned Chinese and English handoff instructions.

* [bugfix] Keep late project reads from reverting an editor save

## summary
- Recheck the editor transaction and captured snapshot when background
  hydration finishes, before installing content or its version basis.
- Exercise a delayed real projection during stage rendering and retain
  strict publication refusal, retry and immutable provenance assertions.
- Record the demonstrated asynchronous projection risk in development
  lessons without weakening publication version checks.

## validation
- The formal structured-editor regression fails without the fix and
  passes with it in native Windows Chrome.
- TypeScript check passed. The later source edit only adds explanatory
  comments; broad frontend and package validation follows this freeze.

Refs: ThinkFlowLab#87

* [docs] Record final Windows evidence and source boundaries

## summary
- Record main integration, real ACL write/release windows, scoped source
  freezes, failed attempts and final Windows/browser/wheel evidence.
- Preserve the earlier full Python result at its tested revision and
  document unchanged Python inputs after the frontend correction.
- Update historical plans and both handoffs; leave final delivery SHA,
  PR metadata and final-head CI to the subsequent delivery receipt.

## validation
- All 875 frozen runtime/test/build inputs remain byte-identical.
- Source freeze Release checks 42569337698 passed, including Python,
  19 browser workflows, the voice fixture and assistant panels.
- Latest clean Windows wheel regression: 324 passed, 3 existing skips.
- Documentation diff check passed; no Windows functionality rerun for
  these status and evidence changes.

Refs: ThinkFlowLab#87

* [bugfix] web: invalidate autosave before editor transactions

## summary

- Invalidate scheduled autosaves before editor saves, media
  publication and timeline imports, including hidden-page flushes.
- Preserve queued writes, explicit canvas draft saves and fixed
  publication versions; reject stale-project callbacks first.
- Exercise an expired real autosave at the editor acknowledgement
  boundary and isolate and clean up the held projection test.
- Record the controlled reproduction and focused validation without
  claiming the exact CI schedule or new Windows acceptance.

## validation

- Structured publication regression fails before the fix and passes
  twice afterward; all five focused browser workflows pass.
- All 734 Node tests, TypeScript and production build pass.
- i18n validates 2996 keys; browser locale and manifest checks pass.

* [docs] Record final Windows native acceptance for PR 193

## summary

- Close the focused Windows handoff at d0d8999 with 154 passing tests,
  zero skips and the complete eleven-install/twelve-cleanup matrix.
- Record the host ACL prerequisite failure, authorized restriction,
  immutable input hashes and exact native/CLI evidence boundaries.
- Update both handoffs and the current delivery plan while retaining
  historical full-suite, media and wheel revisions.

## validation

- Seven focused Python modules: 154 passed, 0 failed, 0 skipped.
- JSONL, JUnit, console and exit-code cross-check passed.
- Independent read-only review and 878-input byte comparison passed.
- git diff --staged --check passed.

Refs: ThinkFlowLab#87

* [test] Give voice reload and media reads separate deadlines

## summary

- Observe actual media responses before reload, then use a bounded
  navigation budget and the existing media deadline after readiness.
- Keep successful-response and persisted voice/result assertions, and
  verify the reopened player's asset identity, bytes and playback.
- Release the response listener and deadline timer on every exit.

## validation

- Bilingual voice workflow passed with Chromium 145 and Playwright 1.58.2.
- Controlled 17-second navigation failed before and passed after this
  change; the missing-media negative control still fails as intended.
- Exact staged bytes match the tested source; independent read-only
  review, node --check and git diff --staged --check passed.

Refs: ThinkFlowLab#87

* [docs] Record the voice CI repair and final source boundaries

## summary

- Preserve the failed delivery CI and all scoped voice probes, including
  unsuccessful injections and the intentional media-outage control.
- Record c19781a as the browser-test freeze with its exact input hashes,
  explicit navigation budget change and stronger media assertions.
- Retain native acceptance at d0d8999 and close both handoffs and the
  current plan without repeating unchanged Windows validation.

## validation

- Cold and normal bilingual voice workflows passed; missing media is
  still rejected by the fifteen-second media deadline.
- 878 frozen inputs match and all 877 native-protected inputs remain
  unchanged, including 242 Python files.
- Independent read-only review and git diff --staged --check passed.

Refs: ThinkFlowLab#87

* [docs] Record integrated Windows and wheel acceptance for PR 193

## summary

- Record the fixed-main source freeze, default-selection regression and
  preview-safe voice observer, with positive and negative evidence.
- Record 154 native passes, all 23 WinAPI windows and 308 installed-wheel
  passes, plus the bilingual wheel UI and actual fixture media.
- Preserve failed attempts, source/input boundaries, the file-symlink
  privilege limit and CI status at document creation. Close Windows
  handoff work while retaining historical evidence and Draft review.

Refs: ThinkFlowLab#87

* [docs] timeline: attach review screenshots and commit identity map

## summary

- Publish the existing Chinese and English timeline screenshots with
  their hashes, original capture context and visible draft notice.
- Map original acceptance revisions to equivalent email-normalized
  commits while retaining their historical test evidence.
- Record unchanged runtime inputs and no repeated Windows tests.

## validation

- Staged screenshots match the inspected original bytes and hashes.
- All 18 rewritten commits preserve their trees and merge topology.
- Staged documentation diff and relative image targets checked.

Refs: ThinkFlowLab#87
* [ci] request all reviewers mapped to PR labels

Route ready PRs to all matching label owners, preserve existing review requests and submitted reviews, and use fallback owners when no label matches. Add missing performance, docs and desktop labels on main. Include bilingual process documentation and mocked API regression coverage.

* [ci] type reviewer routing tests for strict builds

* [ci] fix action pin and skip fallback on label removal

* [ci] retrigger checks after media-tools timeout

---------

Co-authored-by: Tiance Wang <11400530+wangtiance@users.noreply.github.com>
Use explicit 45-second cold-load and reopen waits while retaining the 15-second interaction timeout. Surface failed media HTTP responses directly and provide a deterministic slow-response probe. Refs ThinkFlowLab#250.

Co-authored-by: Tiance Wang <11400530+wangtiance@users.noreply.github.com>
…Lab#202)

## summary

- Add a `deepseek-harness` plugin that exposes OpenDirector creation tools to
  DSH: project listing, snapshot reads, job polling, generation creation and
  cancellation, all through the same open/command/lease protocol the CLI uses.
- Serve canvas media over a DSH web route, forwarding `Range` and passing
  `Content-Range`/`Accept-Ranges`/MIME through, with a per-chunk idle deadline
  so a stalled upstream body aborts instead of hanging.
- Keep the tool schemas valid JSON Schema: `defineTool` moves the property-level
  `required` marker to each parent object's array, pinned by a test against the
  real loader.
- Run the plugin's own tests in a path-filtered workflow rather than the shared
  release lane.

## validation

- `DSH plugin checks` and `Release checks`: success on `33b0353e`.
- `test/media-route.test.js`: range forwarding, MIME mapping, `400` for an
  invalid asset id, and `AbortError` on a stalled body.
- `test/host.test.js`: required arrays, nested and output schemas verified
  through the real DSH loader.
ThinkFlowLab#231)

* [bugfix] providers: run Z-Image-Turbo text-to-image at 9 steps, no CFG

## summary

Z-Image-Turbo text-to-image requests sent no num_inference_steps or
guidance_scale, so vLLM-Omni ran 50 steps with CFG at 1.0 instead of
the 9 steps without CFG that its recipe uses. Edits keep the service
defaults: strength picks the starting step, so fewer steps would change
how much of the source survives, and edits without CFG have not been
checked yet.

* [bugfix] providers: freeze Z-Image-Turbo sampling at job acceptance

## summary

The sampling values were added when the request was sent, so the job
record did not show them, and a job resumed after an upgrade could run
with values it was not accepted with. They are now resolved in
CreationGeneration._parameters for the vllm-omni-image profile, kept in
adapterParameters, recorded on ImageExecution as numInferenceSteps and
guidanceScale, and sent from the job on execution and recovery. A job
accepted without them keeps the service defaults. The adapter no longer
adds sampling on its own.
…Lab#275)

## summary

- Frame loopback model responses with HTTP chunked encoding so the
  bundled runtime can finish reading each response.
- Keep interrupted streams and cancellation distinct from completion,
  preserve local error statuses, and record transport outcomes.
- Run bundled OpenCode gateway regressions in CI and cover streaming,
  cancellation, tool-result continuation and persisted asset identity.
- Keep delivery guidance in shared system instructions and allow
  replies without unnecessary project-tool discovery.

Fixes: ThinkFlowLab#274
* [feat] providers: add Zhipu BigModel GLM chat support

- Register the Zhipu BigModel provider preset with its canonical URL, the
  Z.ai brand logo and the GLM agent model catalog.
- Map the shared thinking levels to GLM reasoning_effort in the shared chat
  payload. GLM accepts only low, high and max, auto omits the parameter and
  unknown levels fall back as in the DeepSeek branch.
- Cover the preset shape, bilingual labels, saved connection identity and
  effort mapping with tests, and the provider form and assistant picker with
  an isolated browser case.

- .venv/bin/python -m pytest tests -q: 1083 passed, 6 skipped
- npm test: 678 passed
- npm run build
- npm run check:i18n
- CREATION_API_URL=http://127.0.0.1:8019 CREATION_WEB_URL=http://127.0.0.1:5179 node tests/browser/zhipu_provider.mjs
- Real credential acceptance: streaming conversation, tool call execution and
  both display languages verified against open.bigmodel.cn with screenshots.

Refs: ThinkFlowLab#236

* [test] web: cover Zhipu provider checks in both locales

## summary

- Register the Zhipu provider browser workflow in the extended lane.
- Reuse the assistant picker check for both Chinese and English labels.

* [test] web: isolate Zhipu picker browser scenarios

## summary

- Run Chinese and English picker checks on fresh pages and projects.
- Keep saved model configuration shared while avoiding assistant state and
  project lease interference.

## validation

- node tests/browser/zhipu_provider.mjs
- node --check tests/browser/zhipu_provider.mjs
- git diff --staged --check

Signed-off-by: MrDongsls <dhy1042542469@126.com>

---------

Signed-off-by: MrDongsls <dhy1042542469@126.com>
Co-authored-by: Weiming Liao <liaowm5@gmail.com>
* [ci] browser: harden nightly advisory lanes

- Dismiss the model-setup welcome before canvas clicks so a fresh
  server does not cover the extended and perf lanes.
- Point extendedFixture scripts at the assistant fixture web and API.
- Keep the activity harness off production preview, which cannot
  serve its Vite test module.
- Open default entry on a project that script owns, and follow the
  immediate project-create flow.
- Hold history flush, reference undo, and the JiuwenSwarm recheck
  on the requests those journeys actually make.

- `node scripts/check-browser-manifest.mjs`
- Isolated browser runs of the scripts that failed in nightly run
  42527176813, plus locale, tool drafts, and the drag profile.
- The full GitHub nightly was not reproduced.

Related to ThinkFlowLab#200

Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

* [ci] browser: dismiss welcome in new timeline text scripts

## summary

- Dismiss the model-setup welcome in the timeline text coverage and
  render scripts added on main. They are extended-lane pages, and the
  manifest check rejects a page that omits the dismissal.

## validation

- `node scripts/check-browser-manifest.mjs`

Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
…lowLab#281)

Co-authored-by: Tiance Wang <11400530+wangtiance@users.noreply.github.com>
Removed API key configuration details from the reviewer workflow section for security reasons.
* [docs] release: record SemVer tags and daily cuts

## summary

- Add the existing public tag and daily v0.1.x cut procedure to
  docs/development/release.md.
- Link that page from the documentation index and CONTRIBUTING.
- Point release readiness at issue ThinkFlowLab#7 without restating its checklist.

## validation

- Checked the new relative links and the version-field paths they name.
  Did not run the test suite; this change is documentation only.

Fixes: ThinkFlowLab#271
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

* [docs] release: cut a minor every 14 days

## summary

- Daily patches stay on the current minor, v0.N.x, under the same gate.
- Every 14 days the morning cut is the next minor, delayed until the
  gate is green rather than skipped or cut red.
- MAJOR stays v1.0.0 for a supported public product. No start date
  is set for v0.2.0.

## validation

- Rechecked the relative links and version-field paths in
  docs/development/release.md.
  Did not run the test suite; this change is documentation only.

Fixes: ThinkFlowLab#271
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
…hinkFlowLab#269)

## summary

MiniMax Music 3 renders with seed 0 when a request has none, and music
jobs sent none, so generating the same style, lyrics and length again
returned the same song. FastH3 video jobs always sent seed 1234. A
seed is now drawn when a job is accepted, kept in adapterParameters
and the execution record, and sent on execution and recovery. YuE2
already drew its own seed per request; it now gets the job's seed
too, so a recovered job renders the take it was accepted with.

## validation

- pytest tests: 1580 passed, 98 skipped
- 7 new or changed tests fail on main and pass here
- MiniMax-Music3 on vLLM-Omni, on an earlier revision with the same
  music request: two jobs on main returned identical audio; two jobs
  here differ, and resending a job's seed reproduces it

Fixes: ThinkFlowLab#267
## summary

- Set the npm package, lockfile, and Python distribution version to 0.1.3 for today's patch release.
- Report the same version from the local API, the DeepSeek ACP client, and the Claude plugin manifest.

## validation

- Parsed package.json, package-lock.json, and pyproject.toml and confirmed each project version is 0.1.3.

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Zhangbo-cn and others added 26 commits November 12, 2026 17:08
…inkFlowLab#238) (ThinkFlowLab#276)

* [feat] providers: add Zhipu BigModel GLM-TTS and voice management (ThinkFlowLab#238)

- Add `glm-tts` as an audio provider through the existing connection-scoped
  voice library: 7 documented preset voices plus the account's cloned voices,
  with clone (upload → `glm-tts-clone` → delete sample → read back), select,
  synthesize and delete available to both the UI and the agent tools.
- Request non-streaming 24 kHz WAV and accept only RIFF bodies; JSON errors
  returned with HTTP 200 fail the job. Map BigModel 1113 to quota and 1214
  "voice not found" to `voice_not_found`; record format and sample rate in the
  execution snapshot.
- Validate text length (1024), recording type (mp3/wav), 3–30 s duration and
  duplicate names before any job or remote call is created.
- Add Zhipu as a conversation-agent preset (models discovered via `/models`)
  and explain a disabled send button when no conversation model is configured.
- Add zh-CN/en-US messages and user-guide sections.

- `pytest`: 1611 passed, 98 skipped; `npm test`: 757 passed.
- `check:i18n`, `check:browser-manifest`, `build`, `tsc --noEmit`: clean.
- `tests/browser/zhipu_real.mjs` (manifest helper, paid) against the real
  BigModel API in zh-CN and en-US: preset narration and the full clone
  lifecycle passed; the account was left with no cloned voices or files.

* [docs] providers: describe the merged Zhipu preset for chat and GLM-TTS

* [test] providers: wait for the cloned GLM-TTS result before capturing evidence

* [bugfix] providers: share one pin-to-public-address guard for CDN downloads

The voice library had its own copy of the result-download guard without the
resolver deadline, the IPv6 zone-id check or address normalisation, so a slow
or hostile resolver could hold a worker thread for as long as it liked on a
path that runs during every voice listing.

Lift the guard into http_settings as public_target (async) and
public_target_sync (worker threads), parameterised by the error factory, and
call it from both volcengine and the voice library. Both now share the 15 s
DNS deadline, reject IPv6 zone ids and out-of-range ports instead of raising
a bare ValueError, and pin to a normalised address while keeping the TLS
identity. Sync lookups run on a bounded pool so a hung resolver cannot
exhaust the request workers.

Drive the optional transcript field from a new acceptsTranscript capability
instead of the provider name, and publish the GLM-TTS default voice as
inputRules.defaultVoice so the client no longer hardcodes it.

Refs ThinkFlowLab#238

* [bugfix] providers: record voice sample cleanup failures

The uploaded clone sample is deleted in a finally block that swallowed every
ProjectError, so a provider that refuses the delete left the file in the
user's account with nothing in the log to say so. Only the account holder can
remove what is left behind, and they had no way to learn it happened.

Keep the clone's own outcome authoritative, but log the failed cleanup as a
warning with the provider's error code and the original traceback attached.

Refs ThinkFlowLab#238

* [test] providers: name the invalid recording cases

A generated parametrize id carried the whole WAV sample, so
PYTEST_CURRENT_TEST grew past MAX_ARG_STRLEN (192KB vs 128KB) and every
ffprobe call in those cases died with Errno 7. pytest 9.2 truncates the
id and hid this locally; CI pins 9.1.1, where it fails.

---------

Co-authored-by: zhangbo <bo.zhang@videostrong.com>
Co-authored-by: dev <dev@example.com>
…nkFlowLab#305)

* [ci] establish shared static gates and Python test layers

## summary

- Run the local pre-commit configuration before PR/main test work,
  preserving the existing required job and full Python suite default.
- Add reusable pytest selection, register initial CI-layer marks and
  document incremental layout and selection rules in both languages.
- Limit PR checks to 15 minutes while retaining other 30-minute budgets.

Refs: ThinkFlowLab#273

* [ci] parallelize release checks behind the static gate

## summary

- Split Python, frontend/package and browser journeys into independent
  jobs after shared static checks, retaining every existing command.
- Preserve the required checks name with a fail-closed aggregate and
  test failed, cancelled, skipped and missing lane receipts.
- Restore 30-minute full-suite budgets and fetch history only for
  static range comparisons; update the bilingual validation policy.

Refs: ThinkFlowLab#273

* [ci] collect per-test timing and layer evidence

## summary

- Record opt-in per-test phase timings, outcomes, layer membership and
  run identity without changing pytest selection or exit status.
- Publish separate full-suite and bundled OpenCode JSON, Markdown and
  JUnit evidence for duration-based layer assessment.
- Verify partial runs, failures, skipped tests and marker conflicts;
  document comparable-sample rules before moving coverage.

Refs: ThinkFlowLab#273

* [ci] classify costly integration cases and protect PR coverage

## summary

- Use a successful Linux timing sample to mark 12 complete integration
  cases as merge, retaining all assertions and parameter scenarios.
- Protect 45 ready representatives across 15 critical capabilities and
  the browser export gate with a full-inventory audit.
- Reject partial inventories, overlapping marks and missing execution
  evidence; document the measured basis and cross-platform ID limits.
- Keep all existing callers on the full suite until queue activation.

Refs: ThinkFlowLab#273

---------

Co-authored-by: Tiance Wang <11400530+wangtiance@users.noreply.github.com>
…FlowLab#299) (ThinkFlowLab#306)

Upload the perf report and drag profiles from one nightly-perf/ directory.
Mixing the workspace report with absolute /tmp profiles rooted the artifact at
/, so daily-benchmark never found nightly-perf/nightly-perf.json and reported
the UI sample missing. Add a layout regression that models upload-artifact
rooting against the workflow.
… (ThinkFlowLab#296)

* [bugfix] assistant: reuse empty conversations on New (ThinkFlowLab#246)

Creating a session while the current one has no messages or unsent
content now returns the existing session instead of archiving it.

* ci: re-run release checks (voice library browser timeout unrelated to ThinkFlowLab#246; passes 3/3 locally)

* [bugfix] assistant: persist reused create ids and ignore blank draftContent (ThinkFlowLab#246)

Record create ids that resolved to an empty session so later replays stay
bound after drafts or a restart. Treat text-only whitespace draftContent as
empty, while reference and skill parts still start a new session.

* [bugfix] assistant: treat non-string draftContent text conservatively (ThinkFlowLab#246)

Accepted drafts can carry a non-string text part, which previously
crashed New with HTTP 500. Treat those parts as present content and
add HTTP and frontend regressions. Also keep citation projection
after merging main.
## summary

- Classify newly opened, edited and reopened issues using the existing
  label catalog, preserving manual priorities and skipping stale results.
- Share response validation with PR labeling while retaining its policy.
- Add area-only fallback rules, regression tests and bilingual guidance.

## validation

- 26 label and reviewer-routing tests passed.
- Changed-file pre-commit checks passed, including actionlint.
- i18n checks and production build passed during implementation.
- Model and GitHub writes were tested with doubles, not live calls.

Refs: ThinkFlowLab#145
* [feat] media: support HEIC and HEIF image assets

## summary

- Retain HEIC/HEIF originals and serve oriented browser previews.
- Preserve original bytes in downloads, clipboard transfers, media
  publication and canvas archive round trips.
- Enable mixed HEIF references for both MiniMax H3 image models with
  shared validation, bilingual errors and regression fixtures.

## validation

- 167 focused Python tests and 766 frontend tests passed.
- TypeScript/Vite build, i18n, browser manifest/locale, import boundaries
  and generated contract checks passed.
- Browser import, paste, archive and structured editor checks passed.
- Generation used mock providers; no live MiniMax requests were made.

Fixes: ThinkFlowLab#282

* [bugfix] media: preserve HEIF originals and honor preview caching

## summary

- Revalidate browser previews before decoding while checking the original
  asset's integrity and rejecting unsupported formats.
- Register standalone HEIF thumbnails, localize decode errors, composite
  transparent previews, and reject multi-image HEIF imports.
- Preserve original bytes and extensions in downloads, clipboard copies,
  canvas archives, and timeline/FCPXML packages. Reject orphaned previews.
- Add regression coverage and update both user guides and validation notes.

## validation

- 169 focused Python tests and 770 frontend tests passed.
- Build, i18n, browser manifest/locale, import boundaries and generated
  contract checks passed.
- Isolated Chrome verified original-file transfers, FCPXML ZIP bytes,
  bilingual rejection, orientation, reopening and fixture generation.
- Read-only DeepSeek review through pi found no blocking issues.
- MiniMax calls were mocked; live provider calls remain unverified.

Refs: ThinkFlowLab#282
Refs: ThinkFlowLab#287

* [bugfix] media: preserve HEIF uploads across save acknowledgements

## summary

- Reject failed HEIF previews before creating an empty canvas item,
  while preserving the registered original for recovery.
- Compare effective generation settings across save receipts without
  accepting actual content, model, frame-selection or lock changes.
- Add bilingual production upload/reload regression coverage and a
  local test screenshot showing three HEIC images on one canvas.

## validation

- 780 frontend tests and 76 focused Python tests passed.
- Build, bilingual catalogs, browser manifest, contracts, import
  boundaries and repository static checks passed.
- HEIF browser regressions passed, including production checks in
  both languages with three real camera samples.
- Read-only DeepSeek review via pi: no blocking findings.

Refs: ThinkFlowLab#282

* [bugfix] media: preserve ICC profiles in HEIF thumbnails

## summary

- Keep RGB color profiles in JPEG library and WebP asset previews.
- Drop incompatible channel profiles after RGB conversion and
  invalidate older browser PNG renditions while retaining 304 reads.
- Cover profiled, unprofiled and grayscale HEIC/HEIF outputs and
  document local canvas, asset preview and version history evidence.

## validation

- 125 focused Python tests passed across HEIF, assets and image tools.
- Verified the original P3 profile in PNG, WebP and JPEG renditions.
- Captured three visible Chrome views using the actual camera file.
- DeepSeek reviewed the fix and cache follow-up with no blockers.

Refs: ThinkFlowLab#282
## summary

Complete the documented test layout and CI stage contract for PR-1 in
roadmap ThinkFlowLab#273, distinguishing target policy from current execution.

- Define local, PR, merge-queue, main-evidence and nightly responsibilities
  and budgets while retaining the current bar until queue enforcement.
- Separate directories, CI layers, hardware platforms, capability skips
  and browser lanes, with incremental discovery-safe migration rules.
- Keep both languages together and preserve existing heading anchors.

Co-authored-by: Tiance Wang <11400530+wangtiance@users.noreply.github.com>
…wLab#318)

* [bugfix] timeline: accept durations rounded to the millisecond

## summary

- Compare a clip's out point against the probed duration with a shared
  half-millisecond tolerance instead of `1e-6`. Out points are seeded
  from the millisecond-rounded `durationMs` field, so a 121-frame clip
  at 24 fps measures 5.041667s but is stored as 5.042s; probing it again
  on import found an overflow of 0.000333s and rejected every clip whose
  frame count is not a whole millisecond.
- Move the tolerance into `rule_facts.py` as `TIMELINE_DURATION_EPSILON`
  and export it through `ruleFacts.generated.ts`, so the document
  boundary, the aggregate snapshot validation and both application gates
  share one value instead of three literal copies.
- Skip the comparison when an asset has no measured duration. Such a
  probe cannot prove an overflow, and the document boundary and the
  aggregate validation already read it that way; only the timeline gates
  treated it as fatal, so one document could pass a gate and fail the
  next. A clip that genuinely overflows is still refused.

## validation

- `pytest tests/` : 1933 passed, 27 skipped.
- The new rounding test fails with the previous `1e-6` tolerance and
  passes with the shared value.
- `tsc --noEmit`, import-boundary lint, i18n, browser locale and
  browser manifest checks, and `vite build` passed.

Fixes: ThinkFlowLab#256

* [bugfix] canvas: drop a stale assembly link on archive import

## summary

- Stop the document-to-project translation from sending a link whose clip
  kept the media it was assembled with after its source node was
  regenerated. The archive is not corrupt; the relation is stale, and a
  live project keeps it only because the server skips a connection whose
  value did not change.
- Keep a connection the project already recorded, so a live project is
  unaffected, and require the source node's current media only for fresh
  links.
- Keep the clip itself, because its media is the user's edit, and drop
  only the relation the source no longer proves.

## validation

- `node --test tests/*.test.ts` : 796 of 797 passed. The single failure
  (`browserHost` PATH discovery) also fails on unmodified `main`.
- Both directions are covered by new tests, and each one fails when its
  guard is removed: the stale link is dropped on import, and a recorded
  link survives its source being regenerated in place.
- `tsc --noEmit`, import-boundary lint, i18n, browser locale and
  browser manifest checks, and `vite build` passed.

Refs: ThinkFlowLab#256

* [docs] canvas: record the two import blockers found on real data

## summary

- Extend the archive-import write-up with the two blockers that only a
  real export exposed: millisecond rounding in clips whose frame count is
  not a whole millisecond, and assembly links an archive still carries
  after a source node is regenerated.
- Record why the existing tests missed both. They replace `client.batch`
  and `fetch` with doubles, so no server validation ran and a document
  the API refuses passed the suite.
- State the rule that follows: a document-to-project defect has to be
  verified through the real backend.

Refs: ThinkFlowLab#256
…lowLab#122) (ThinkFlowLab#308)

Run-before checks counted only follow-node inputs, so a loaded or
reopened image config with more pinned references than the model
allows still looked runnable. Include pinned assets in the shared
precheck, without counting the same asset twice or treating them as
first-last frame connections.
…wLab#117) (ThinkFlowLab#313)

## summary

- When an official editor.create or editor.save adds timeline clips
  without inputConnections, bind each new clip to a source node only
  when exactly one visible, editable, type-compatible canvas node
  currently holds that asset.
- Leave clips unconnected when the asset is imported with no canvas
  node, when several visible nodes share the same asset, or when the
  clip already has an explicit connection.
- Do not rewrite clip order, trim, or media references.

## validation

- New regressions failed on current main, then passed after the fix,
  covering unique two-node Agent creates, ambiguous same-asset nodes,
  external imports, mixed unique and external clips, explicit
  connection IDs, save of only newly added clips, hidden-only sources,
  and showing a unique source in the same batch.
- Full suite: 1866 passed, 100 skipped
  (PYTHONPATH=packages/core/src:apps/api/src pytest tests).
- lint-imports kept all four contracts; i18n, browser locale,
  browser manifest and npm run build (tsc --noEmit && vite build)
  passed.
- npm test: 793 passed. One existing host-env probe
  (tests/browserHost.test.ts) failed in this worktree without a
  local venv, seeing /usr/bin/node after PATH isolation; the same
  test passed in a checkout that has .venv. This change does not
  touch that code.
- Canvas line drawing was not re-run in a browser; official command
  persistence, reopen, and existing projection from inputConnections
  cover the reported path. No paid generation calls.

Fixes ThinkFlowLab#117
…kFlowLab#288)

* [feat] providers: add Beijing Model Studio Qwen chat preset

* Restore provider-neutral native runtime coverage
…#283)

* [feat] providers: add MiniMax speech with system voices

## summary

- Add mainland MiniMax synchronous TTS and a read-only system voice
  library through the shared generation and voice services.
- Preserve provider voice IDs, validate accepted inputs, classify
  HTTP-200 failures, and recover public MP3 downloads without reposting.
- Add bilingual presets, errors, docs, browser evidence and regression
  tests for UI/Agent use, persisted media and recovery.

## validation

- MiniMax speech: 58 tests passed; related generation, model, voice,
  media download security and contract suites passed.
- npm test: 768 passed. Build, import boundaries, generated contracts,
  i18n, browser locale and manifest checks passed.
- Bilingual voice-library browser workflow passed with mocked HTTP
  and local WAV playback. Python cases probe real MP3 fixture bytes.
- No live MiniMax calls or perceptual voice-quality checks were made.

Refs: ThinkFlowLab#240

* [bugfix] providers: preserve MiniMax recovery and model routing

## summary

- Recover downloads with the accepted model snapshot and no credentials,
  preserving public-host checks after configuration removal or replacement.
- Reject unsupported official speech models through shared input rules,
  with bilingual errors, atomic save coverage and updated contracts.
- Acknowledge session release before the persisted-audio browser check
  reloads, avoiding dependence on unload-time request delivery.

## validation

- 70 MiniMax tests, 239 related Python tests and 23 frontend tests passed.
- Bilingual voice workflow passed with Playwright 1.58.2.
- Build, i18n, browser manifests, import boundaries and contracts passed.
- Provider HTTP was mocked; live MiniMax synthesis was not checked.

* [bugfix] providers: keep accepted MiniMax results recoverable

## summary

- Match voice-library capabilities to supported MiniMax models.
- Project accepted speech jobs and download saved results without
  revalidating the current model whitelist; retain public-host checks
  and credential isolation through the existing downloader.
- Document complete native voice catalogues and pin self-hosted download
  behaviour after model configuration removal or replacement.

## validation

- 428 targeted Python tests and 22 frontend tests passed.
- Build, i18n, generated contracts and import boundaries passed.
- DeepSeek/pi read-only review found no blocking issues and independently
  verified the 144 focused provider, recovery and security tests.
- Provider HTTP was mocked; live MiniMax synthesis was not checked.

* [bugfix] providers: cap MiniMax speech at 3000 characters

## summary

- Limit non-streaming MiniMax speech to 3,000 characters through shared input rules.
- Align Chinese/English errors and guides; reject longer text before provider calls or job creation.
- Cover the 3,000/3,001 Unicode boundary and localized rejection receipts.

## validation

- 77 MiniMax speech tests passed.
- English and Chinese voice-library browser workflows passed with mocked provider HTTP.
- i18n, browser locale, generated contracts, diff and pre-commit checks passed.

---------

Co-authored-by: Weiming Liao <liaowm5@gmail.com>
…er-window conflict (ThinkFlowLab#277) (ThinkFlowLab#279)

## summary

- Keep the previous page's session ID in `sessionStorage`. Before its first
  request, a reloaded tab releases that session when no live page still holds
  its Web Lock, so a crashed or killed page no longer blocks the reload for
  the 60 s lease. A duplicated tab copies `sessionStorage` but finds the
  original's lock held, so it never releases a live window.
- Give "project open in another window" its own `project_open_elsewhere` code
  with bilingual text. `storage_busy` again means only a storage lock timeout;
  it keeps its retries and `Retry-After`. Only a reloaded tab waits briefly
  on its own lease; a fresh window is refused at once.

## validation

- `pytest`: 1581 passed, 98 skipped; `npm test`: 760 passed; `tsc`, `build`,
  `check:i18n`, `check:browser-manifest`: clean.
- New gate script `tests/browser/session_lease.mjs` (zh-CN, en-US): reload
  with a dropped release reopens in about 2 s with no 409, and a second window
  is refused with `project_open_elsewhere`. On `main` the reload times out.
- `scripts/run-browser-suite.mjs`: 20/20 passed on fresh fixture services.

Co-authored-by: zhangbo <bo.zhang@videostrong.com>
…owLab#315) (ThinkFlowLab#317)

A reload restores the last persisted viewport. When the view saved after
the generation step lands before the reload, the timeline created at
{x:700,y:100} sits above the canvas and the right-click times out with
"element is outside of the viewport".

Fit all content with Shift+1 after each reload that targets a node created
through the client, and wait until the node is inside the canvas and the
viewport is stable before clicking. Retries the fit up to three times and
reports the transform and bounds if the node never becomes reachable.

Fixes ThinkFlowLab#315
## summary

- Wait for the edited text binding, model, prompt and node positions
  before checking saved state in the selection/history browser test.
- Keep the AB asset-content, coalescing and command-retry assertions;
  bound the request gate and release it when an assertion fails.
- Document the stale-label race and bilingual validation guidance.

## validation

- Full canvas_selection_history.mjs passed on this isolated branch.
- Browser manifest, browser locale and i18n checks passed.
- Node syntax, Prettier, pre-commit and git diff checks passed.

Fixes: ThinkFlowLab#300
Align the npm package, lockfile, Python distribution, local API, DeepSeek ACP clientInfo, and Claude plugin manifest versions to 0.1.4.
* Add BigModel CogVideoX-3 video generation

* Explain BigModel watermark waiver above Generate

---------

Co-authored-by: unicorn1125 <192290080+unicorn1125@users.noreply.github.com>
…ds (ThinkFlowLab#278)

* feat(contract): stage/route capability mapping table

* fix(contract): key the table by (operation, profile) per upstream vocabulary

vllm-omni#10102 publishes six operations and twelve profiles; `stages[]` is
deployment topology (`id`/`type`/`depends_on`, numeric strings, the same full
list for every model) and a route cannot separate the three profiles that share
`/v1/audio/speech`.  The table is therefore keyed by (operation, profile), with
OPERATION_ROUTES kept as the ownership check that reports `route_mismatch`.

- OPERATION_PROFILES replaces STAGE_ROUTES, covering all twelve profiles
- capabilities are available only on a supported profile in an enabled
  operation; `missing` needs positive evidence, everything else stays `unknown`
- segmentations/depth keep a route but no operation, so image.segment,
  video.matting and video.motion_capture are `unknown` with `no_upstream_profile`
- `asr` reports `unmapped_stage`: upstream publishes no recognition operation
- fixtures follow the real response shape; five regressions pin the pairs,
  the shared speech path, topology-only stages and absent fields

* fix(contract): separate states, validate transport/method, make fixtures schema-valid

Round-2 fixes for the three independent reviews of ThinkFlowLab#278.

Product decisions (P1-P5):
- P1: image_response / audio_response no longer map to text.chat; they stay
  unknown with no_local_capability, since text.chat is served by a text-only
  adapter and an image-only deployment must not read as able to chat.
- P2: text_to_audio (audio_generation, /v1/audio/generate) maps to audio.music
  alongside lyrics_to_music; the StableAudio/LTX2-t2a shape no longer reads
  audio.music as missing.
- P3: support=unsupported and exposure=disabled stay "missing" but are told
  apart: resolve() now returns a `reasons` map (unsupported /
  endpoint_disabled / route_disabled) instead of dropping reason_code.
- P4: profiles[*].routes[*].exposure is read: all routes disabled -> missing +
  route_disabled, all unknown or routes: [] -> unknown, one enabled -> available.
- P5: one unrecognised route no longer voids a profile that answered; only when
  nothing usable remains does it fall to unknown.

Contract and implementation (I1-I3):
- I1: OPERATION_ROUTES is (transport, method, path); _check_route compares all
  three, so DELETE /v1/chat/completions and a websocket-declared http route are
  route_mismatch.
- I2: endpoint distinguishes absent (no answer: None, no models, models not a
  list) from empty (answered, no operation declared) from present.
- I3: the unreachable `item in UNROUTED` branch is deleted; unrouted ids never
  collect evidence, and the disjointness invariant is enforced by
  test_every_capability_id_is_routed_or_explained instead of dead runtime code.
  Malformed routes and supported profiles with no route are now reported
  (malformed_route / no_route_declared) instead of silently unknown.

Fixtures (F1-F3):
- All three fixtures are upstream-shaped: six operations per model, support and
  exposure on every operation, profile support present, profiles: [] instead of
  a missing key, image_response/audio_response outputs ["image"]/["audio"].
- Defensive shapes (asr key, misplaced profile, absent required support) moved
  out of the fixtures into the tests that defend against malformed input.
- test_fixtures_are_schema_valid asserts upstream's structural invariants only.
- exclude_none is described correctly (it drops None-valued optional fields such
  as `request`; default_factory lists serialize as []).

Tests (T1-T5):
- The tautological order assertion is now a literal vocabulary list, and the
  twelve pairs are literal too instead of an echo of OPERATION_PROFILES.
- test_a_profile_without_support_is_unknown_even_when_its_operation_is_supported
  covers profile-level absence (old N4 mutant now dies).
- test_a_profile_that_declares_no_route_is_unknown covers `supported + routes: []`
  (old M7/P3 mutant now dies).
- New cases: profile routes all disabled / all unknown, one unknown route does
  not void a profile, DELETE and websocket mismatches, image-only chat, text_to_audio
  -> audio.music, absent vs empty endpoints, reason separation, non-dict operation
  values, profile order independence.
- UNROUTED expectations are a literal table, not recomputed from the constant.

Docs (D1-D2):
- Corrected the exclude_none claim and the audio.score / /v1/audio/transcriptions
  statement (the capabilities body never publishes that path, so audio.score can
  only be unknown).
- Added the status/reason mapping table (CN + EN) covering every declaration
  shape, plus the profile/route/profile-output tables.

Tests: 60 passed (was 40); related subset 122 passed, 1 skipped. Nine mutations
of the new rules (including removing route_disabled) all fail the suite.

* fix(contract): keep audio_generation local-unknown, align required_inputs with upstream

* fix(contract): fixture construction invariants and unrecognised-route exposure cases

* fix(contract): fixtures obey the realtime-exposure rule, gate now asserts it

`/v1/realtime/video` is the one route whose exposure upstream rewrites:
`entrypoints/openai/serving_capabilities.py:342-347` forces `disabled` when
the profile's `required_inputs` intersect `{video, audio}` and otherwise keeps
the route's own value. Three fixture rows disagreed with that rule:

- full/`image_to_video` (`text`,`image`) carried `disabled` -> `enabled`
- partial/`mixed-operations`/`image_to_video` same
- partial/`unknown-enums`/`video_to_video` (`video`) carried `unknown` ->
  `disabled` (the override is unconditional on the route's own value)

`test_fixtures_obey_upstream_construction_invariants` claimed in its docstring
to pin this rule but only checked route-key sets, so the drift went unseen. It
now asserts the rule itself plus same-path exposure consistency inside one
operation; reverting any of the three rows turns it red (verified).

Resolution output is unchanged by these rows: no `expected` block moved.
Docs/PR body refreshed: upstream is now `43ff7b9` (vocabulary byte-identical to
`d1104f8`), test/doc counts and line references re-measured.

* fix(contract): map each video profile to the id its request shape serves

text_to_video / image_to_video / video_to_video are three different request
shapes in VllmOmniVideo (no reference + task=t2va, input_reference + task=fl2va,
source_video + masks). Folding all three onto video.generate left
video.first_frame and video.edit permanently unknown, so an image-to-video-only
deployment read as able to do text-to-video.

Add candidates_for() as the reverse index over the same table so a caller
starting from a local capability id does not need a second vocabulary, and say
explicitly that resolve() is catalogue-level: it never authorises a request.

Document that a top-level route is what upstream aggregates exposure from, so a
top-level disabled route closes the operation and no profile can reopen it.

* docs(contract): cite line numbers that exist on main

VllmOmniMusic / OpenAIChat moved with main, and application/model_availability.py
only exists on ThinkFlowLab#266, so evidence has to point at application/tool_tasks.py:362
for the capability a request selects, plus ThinkFlowLab#266 for its own request table.

* docs(contract): correct citations and the narrowing claim

An independent pass found stale line numbers (VllmOmniMusic, VllmOmniScore),
a count that mixed UNROUTED with no_upstream_profile, a branch-local path in
ThinkFlowLab#266's section, and a prescription ("keep the top level enabled, disable the
profile route") that this projection cannot emit: profile routes copy the
top-level exposure except on /v1/realtime/video. Say that instead.

* docs(contract): keep the contract doc in English only

The bilingual file carried the same tables twice and the two halves had already
drifted (a count wrong in one, a citation wrong in the other). Keep the English
half, promote its headings, and drop the duplicate from the doc-pin test: the
pinning is upstream literal table <-> doc table <-> fixtures, which does not
need a translation to be checked against.

* fix(contract): gate the narrow video ids on declared input roles

A modality-level profile does not prove a narrower tool: VACE hands the image
to reference_images, SeedVR2 declares V2V without latent-mask editing, and H3's
resolver marks V2V unsupported. REQUIRED_ROLES plus declared_roles() keep
video.first_frame / video.edit unknown unless one of the profile's own open
routes names the role with support=supported, reported as role_not_declared.

Two fixture cases pin the counterexamples, and the structural check now accepts
the optional route-level request instead of rejecting a body upstream emits.

* fix(contract): count roles only on this operation's open routes

A role declared on a closed route, or on a path the operation does not own, is
not evidence: filter by exposure and ownership, and read only the route
reference (upstream's projection already copies the profile-level constraints
into it, so reading the profile again would re-add what a route narrowed).

Fixtures now carry the request shape the projection emits (profile-level base
inputs, last_frame unsupported, the four transport parameters), the structural
guard asserts that shape, and a unit test proves the guard bites.

* docs(contract): name the per-route scope the gate does not decide

The gate answers whether a profile's own open routes name the role, not which
route a request will take; that stays ThinkFlowLab#266's per-request check.

* fix(contract): decide required roles per route and keep a refusal distinct

A required role gathered from a different route is not a request: an edit
call leaves on /v1/videos or on /v1/videos/sync, so source_video on one and
the two noise masks on the other used to read as available while neither
route would accept the call. Support is now decided per route and per
capability, so a gated id can no longer lend its verdict to an ungated one
in the same row.

A document that has decided every required role on every open route and
still offers no complete one is a positive no — one role refused everywhere,
or two routes each refusing a different role — and resolves to
missing/unsupported instead of unknown/role_not_declared. Silence, an
undecided support value, an owned route whose exposure cannot be read and an
unplaceable route left enabled all keep it unknown.

Signed-off-by: Xingkai98 <xingkai98@126.com>

* fix(contract): read the whole declaration inside the role gate

A three-argument role_verdict() used to judge one profile differently from
resolve(), which passed the unplaceable routes in separately: the same
declaration read missing when called directly and unknown inside the
report. The gate now reads those routes itself, so the public call and the
catalogue result cannot disagree, and the extra parameter is gone.

Also pins the remaining wording gaps the review found: the module rule and
the doc table now say at least one route has to be open, and the doc tells a
caller to ask role_verdict() instead of re-deriving the rule.

Signed-off-by: Xingkai98 <xingkai98@126.com>

* fix(contract): a malformed route reference is not a channel

A reference with no usable path was classified as an unplaceable route by
the role gate while _route_exposures() reports it as malformed_route and
keeps it out of the reachability decision, so one broken reference could
turn a refusal into unknown. The gate now skips it the same way.

Also pins the gate against resolve() with a parametrised consistency case
set, and splits the boundary assertions (bare role string, undecided route,
unplaceable route) out of one long test so each rule has its own guard.

Signed-off-by: Xingkai98 <xingkai98@126.com>

* fix(contract): honour count bounds, per-route refusals and the protocol envelope

- read `min_count`/`max_count` beside `support`: a window that cannot admit
  the single item the tool sends is a refusal (`missing` with `unsupported`),
  and a bound that is not a whole number leaves the entry undecided
- a route refusing one required role is out of the running whatever it leaves
  unmentioned, so every open route refusing at least one is `no` instead of
  `unknown`; undecided and unplaceable alternatives still keep `unknown`
- refuse a body whose `object`/`schema_version` names another contract:
  `endpoint="unreadable"`, neither `models` nor `stages` read, one
  `unsupported_format` note; `declarations()` applies the same gate
- `missing` now needs every model to say no, and a `models[]` element listing
  no operation is reported as `malformed_model` instead of being dropped
- docs: rule paragraphs, state table, reason list (eleven codes), endpoint
  states and the regression inventory

* fix(contract): validate count bounds and read only the postable route as evidence

Count bounds now validate domain as well as type: a negative whole number, a
fractional value, a bool and a numeric string are unreadable, and a window whose
`min_count` exceeds its `max_count` decides nothing. An unreadable bound can
neither grant a role nor turn it into a refusal, so the id follows the rest of
the declaration; `max_count: 0` and `min_count: 2` stay genuine refusals, and an
explicit `support: unsupported` still refuses whatever its counts say.

The same review found three places where an unreadable or unpostable declaration
was read in the permissive direction, so the table now:

- counts only the route this adapter actually posts as evidence
  (`POSTABLE_ROUTES`); `/v1/videos/sync`, the websocket route and another
  operation's paths are reported, never decisive
- reads `transport` as upstream's default only when the key is absent; a present
  but unreadable `transport`, `method` or `path` is `malformed_route`
- normalizes `support`/`exposure` at the route, profile and operation levels and
  normalizes published ids in all three readers

That last tightening needs three reasons: `exposure_unknown` for a postable
route whose exposure is absent or unreadable, `malformed_profile` for a profile
element that cannot be read, and `malformed_operation` for an operation value
that is not an object. Notes that explain a verdict are now emitted after the
cross-model aggregate, and only for a pair that still has an `unknown` id.

Verification: tests/test_capability_contract.py 147 passed; full local run 1761
passed / 1 failed / 100 skipped, the failure being
test_cancel_native_process_group[default], which fails identically on the parent
commit and needs a real agent CLI.

* docs(contract): state the postable-only rule everywhere and pin it

The status table, the role-gate exception and the "two boundaries to keep"
paragraph still prescribed `unknown` for a case the resolver decides as
`missing` with `route_disabled`: a profile whose postable route is disabled
beside an enabled route this table cannot place. The resolver moved to
`POSTABLE_ROUTES`; the prose did not follow.

- Rewrite the affected rows and paragraphs: an unplaceable route is a
  diagnostic only, and only an absent, unreadable or literal `unknown`
  exposure on a postable route keeps a verdict undecided.
- State the precondition every status row is read under (profile supported,
  operation supported + enabled) and scope `no_local_capability` to an open
  postable route, like every other profile-level note.
- Add a regression built from the documented table case: disabled
  `/v1/videos` beside an enabled unknown path, the sync and socket variants
  under every exposure, the open-route control, the role-gated neighbours
  and the `no_local_capability` pair.
- Pin the two rule sections and the `REQUIRED_ROLES` paragraph against
  `tests/fixtures/omni-capabilities/contract-rule.txt`, compared whole: an
  edit to either side turns the contract test red, so prose and resolver
  cannot drift apart unseen.

Module changes are docstrings only; resolver behaviour is unchanged.

* docs(contract): check the role vocabulary against vllm-omni 12d834c2

`vllm-omni` `b934b642` (head `12d834c2`) now publishes the four request roles
this table's `REQUIRED_ROLES` names, and the paragraph that introduced them
still said upstream published no such role — so it prescribed `unknown` for
declarations that now exist and are decidable.

- State what the tree declares: Wan I2V `first_frame` supported with
  `min_count: 0` / `max_count: 1` beside a required `image` whose counts are not
  additive; both VACE architectures `first_frame` `unsupported` with
  `max_count: 0`; the three editing roles supported `0..1` where the handler
  takes latent masks and `unsupported` `max_count: 0` where it does not.
- State the statuses that follow from the existing gate — `available` on Wan
  I2V, `missing` with `unsupported` on VACE and on a refusing V2V profile,
  `unknown` where the profile's own support is `unknown` — with no module
  change, because the gate reads the declaration rather than a role list.
- Add a test built from those four declarations, citing the upstream test
  anchors rather than prose, and regenerate the pinned copy of the sections in
  the same commit as the prose it pins.

The module change is a comment above `REQUIRED_ROLES`; behaviour is unchanged.

---------

Signed-off-by: Xingkai98 <xingkai98@126.com>
Co-authored-by: OpenDirector Bot <bot@opendirector.local>
…b#346)

## summary

Classify the complete PR diff after shared static checks and skip heavy
lanes only for an explicit nonempty documentation allowlist.

- Retain code, mixed, renamed-code and unknown comparisons in the full
  gate, including all main pushes, manual runs and nightly callers.
- Require successful static evidence and every skipped lane receipt in
  the stable aggregate; failures and cancellations remain blocking.
- Exercise real Git ranges and aggregate execution, document the policy
  in both languages, and align the PR budget target to 30 minutes.

Co-authored-by: Tiance Wang <11400530+wangtiance@users.noreply.github.com>
* [ci] Add CodeQL scanning for every pull request

## summary

- Analyze Python and JavaScript/TypeScript independently on every PR,
  with main baseline scans and pinned official Actions.
- Require successful analysis receipts in a uniquely named aggregate;
  document alert checks and pending administrator enforcement.
- Add workflow policy and aggregate-command regression coverage.

## validation

- 106 CI policy, scope, summary, profile and layer tests passed.
- Shared pre-commit checks passed against exact staged blobs.
- npm run check:i18n passed.

Refs: ThinkFlowLab#273

* [ci] Cancel outdated PR runs across check workflows

## summary

- Group CodeQL, Release checks and DSH plugin runs by workflow,
  event and PR number so updates cancel obsolete PR checks.
- Keep non-PR runs isolated by run ID without cancellation.
- Document concurrency boundaries in Chinese and English.

## validation

- 106 related CI tests passed.
- Shared static checks passed against the exact staged blobs.

---------

Co-authored-by: Tiance Wang <11400530+wangtiance@users.noreply.github.com>
…hinkFlowLab#344)

* [feat] canvas: resize Markdown text nodes and edit them fullscreen

## summary

- Add a selected-only text resize grip with independent dimensions,
  a usable minimum, lock checks and one undo step per gesture. Reuse
  resize-node and canvas.update, persisting after the drag ends.
- Render Markdown cards and add source/preview modes to the existing
  fullscreen window. Share text, selection, caret and scroll state;
  retain uncontrolled inline input and protect IME composition.
- Keep editing live: Done, closing and Escape finish the surface
  without a separate Cancel, and undo reverses the completed edit.
- Add canvas text create/show/update/resize commands using the existing
  asset and canvas services, lock checks and revision conflicts. Read
  complete UTF-8 source up to 1 MiB without widening agent read limits.
- Add bilingual labels and guidance, command and lock tests, and a
  browser gate for Markdown, editing handoff and resize persistence.

## validation

- npm run check:i18n: 3018 keys, zero problems.
- npx tsc --noEmit: passed without diagnostics.
- npm test: 820 passed, zero failed or skipped.
- PYTHONPATH=packages/core/src:apps/api/src .venv/bin/python -m pytest \
    tests/test_canvas_text_cli.py tests/test_cli.py \
    tests/test_agent_artifacts.py -q: 47 passed.
- PYTHONPATH=packages/core/src:apps/api/src .venv/bin/python \
    -m pytest -q -m ready: 90 passed, 2014 deselected.
- npm run check:browser-manifest: 76 entries, passed.
- npm run check:browser-locale: 71 pages, zero problems.
- node tests/browser/canvas_text_markdown.mjs: passed standalone
  against isolated API/Vite with Chrome, including both locales.
- npm run test:browser: 21/21 workflows passed with isolated services.
- /opt/anaconda3/bin/python -m pre_commit run: all applicable hooks
  passed; git diff --staged --check also passed.
- Browser model calls used a fixture double. Physical IME and
  multi-runtime evidence remain unverified.

Refs: ThinkFlowLab#336

* [docs] validation: record issue 336 canvas text node evidence

## summary

- Record verified resize zooms, Markdown rendering, Chromium CDP
  composition, fullscreen handoff, stale-save recovery and locks.
- Describe coordinator-verified Codex and DeepSeek CLI workflows,
  runtime limitations, the HTTP source read and lease renewal needs.
- Add four PNGs and a GIF unchanged from the supplied feature media,
  linked from the validation guide.

## validation

- Compare all five staged media files byte-for-byte with the supplied
  originals: identical.
- file: four PNG images and one GIF89a animation.
- du -kh: GIF 420K, evidence directory 672K; GIF is under 500 KB.
- git diff --staged --check: passed.
- /opt/anaconda3/bin/python -m pre_commit run: all applicable hooks
  passed without changing files.

Refs: ThinkFlowLab#336

* [bugfix] cli: require a listening API for complete text reads

## summary

- Explain source-read transport failures with an API URL, stable
  error code and English/Chinese guidance to start the API.
- Fall back to HTTP status and URL when an error response is blank,
  retaining structured API error codes.
- Test real CLI subprocesses through the agent socket bridge and a
  listening Uvicorn API, including complete BOM/CRLF source, revisions
  and a closed-port control; label TestClient coverage as bridge only.
- Preserve the 1 MiB CLI source bound and 64 KiB agent read budget.

## validation

- New HTTP/error tests: 7 passed; prior code: 6 failed, 1 passed.
- CLI text, CLI and agent artifact tests: 47 passed.
- pytest -q -m ready: 95 passed, 2016 deselected.
- Localization tests: 7 passed.
- npm run check:i18n: 3018 keys, 0 problems.
- npx tsc --noEmit: passed.
- npm test: 820 passed.
- git diff --staged --check: passed.

Refs: ThinkFlowLab#336

* [bugfix] canvas: hold the interaction guard through a text resize

## summary

- Reuse the node-drag interaction guard and presentation cancellation
  when text resizing starts, alongside the history transaction.
- Release the guard and history on completion or cancellation, keeping
  transient resize changes out of storage and preserving one undo step.
- Exercise immediate corner-drag selection and resize at zooms 0.8
  and 1.3, asserting a stable viewport and geometry within one canvas
  unit of the screen delta divided by zoom.
- Cover an active selection fit, Escape, pointer cancellation, outside
  release, restored controls, unchanged text and reopened dimensions.

## validation

- New viewport assertion failed on 90fee3b with the pointer held.
- Updated canvas_text_markdown workflow passed standalone.
- npm run test:browser: 21/21 workflows passed in 201.4 seconds.
- npm test: 820 passed.
- npm run check:i18n: 3018 keys, 0 problems.
- npx tsc --noEmit: passed.
- pytest -q -m ready: 95 passed, 2016 deselected.
- Browser manifest: 76 entries, gate=22; locale: 71 pages, 0 problems.
- git diff --staged --check: passed.

Refs: ThinkFlowLab#336

* [docs] validation: keep reusable text-node checks, move per-request evidence to the pull request

## summary

- Keep isolated API/Vite reproduction steps and the text-node browser
  regression's stable viewport, geometry, cancellation and save checks.
- Preserve the listening-API and 15-second lease requirements, linking
  to the existing CLI guide instead of duplicating command examples.
- Remove five PR media files and per-run results from the guide;
  pull-request evidence remains available at its pinned commit URLs.

## validation

- Confirm removed measurements and limitations appear in the published
  PR body before removing them from the validation guide.
- Verify the CLI guide is unchanged, the media directory is gone and
  the reusable section has no media embeds or duplicate CLI examples.
- git diff --staged --check: passed.

Refs: ThinkFlowLab#336
## summary

- Install validated local or public HTTPS Skill packages into a locked
  workspace registry, with explicit activation and exact dependencies.
- Filter native Skill discovery per harness and preserve accepted turn
  snapshots, conversation history and editable project outputs.
- Add bilingual plugin settings, an optional creative brief example,
  a custom package scaffold and installation and authoring guides.
- Cover archive and download boundaries, damaged-cache recovery,
  session invalidation and the bilingual browser workflow.

## validation

- Related Python regressions: 105 passed; final plugin suite: 59 passed.
- Frontend tests: 818 passed; production build and wheel inspection passed.
- Bilingual plugin browser flow, i18n, contracts and import checks passed.
- Antigravity independently closed all five initial review findings.
- No real paid-model calls or native Windows filesystem validation.

Refs: ThinkFlowLab#114
## summary

- Limit local sources to package folders inside the backend workspace,
  excluding project and internal storage; retain ZIP upload elsewhere.
- Read package files through pinned handles to reject link replacement
  and hard links, and validate control characters before path handling.
- Preserve empty plugin resources without relaxing execution-artifact
  defaults, with bilingual guidance and boundary regression coverage.

## validation

- Related Python checks: 208 passed; 15 Windows-only cases skipped.
- Production browser plugin flow, build, i18n and static checks passed.
- Antigravity independently rechecked 119 tests and 250 concurrent
  link-replacement attempts with no outside-file disclosure.
- CodeQL and native Windows verification remain for remote CI.

Refs: ThinkFlowLab#114
port = parsed.port
if port and (parsed.scheme, port) not in {('https', 443), ('http', 80)}: host += ':' + str(port)
endpoint = urlunsplit((parsed.scheme.lower(), host, parsed.path.rstrip('/'), '', ''))
account = hashlib.sha256((config.get('apiKey', '') if config.get('authMode') != 'none' else '').encode()).hexdigest()
assert exit.value.code == 1
payload = json.loads(capsys.readouterr().out)
assert payload['error'].strip(), payload
assert '502' in payload['error'] and 'http://api.test/' in payload['error']
# Only a complete, clean, successful transcript may yield a receipt.
pattern = (r'Command: [^\r\n]+\r?\n\s*Stdout: ([\s\S]*?)'
r'\r?\n\s*Stderr: \(empty\)\r?\n\s*Exit Code: 0\s*')
match = re.fullmatch(pattern, value)
})}
</p>
)}
{url && <audio key={url} src={url} controls preload="metadata" />}
if (typeof name !== 'string') fail('input.string', { field: 'name' })
const namespaceBytes = Buffer.from(namespace.replaceAll('-', ''), 'hex')
const digest = createHash('sha1')
.update(namespaceBytes)
}
if (!url.hostname) throw new ConfigError('config.urlInvalid', { field })
if (/[?#]/.test(raw)) throw new ConfigError('config.urlQuery', { field })
return raw.replace(/\/+$/, '')
const PROJECTS = '/api/creation/projects'

export function joinUrl(baseUrl, path) {
return `${baseUrl.replace(/\/+$/, '')}/${String(path).replace(/^\/+/, '')}`
evidence.projectId = projectId;
evidence.canvasUrl = afterAdopt.canvasUrl;
console.log(JSON.stringify(evidence, null, 2));
if (process.env.EVIDENCE_PATH) await writeFile(process.env.EVIDENCE_PATH, JSON.stringify(evidence, null, 2) + '\n');
return resilientRead(
async (signal) => {
const response = await fetch(
`/api${creationPath(this.id)}/assets/${assetId}/content`,
`/api${creationPath(this.id)}/assets/${assetId}/${representation}`,
## summary

- Require source strings to pass source resolution and authorization;
  reject direct Path and other unexpected input types.
- Route uploaded ZIP bytes through a separate archive entry while
  sharing the existing atomic, idempotent package registration.
- Cover wrong input types, upload provenance and URL parameter redaction,
  and record the mixed-entry boundary lesson.

## validation

- Plugin and localization tests: 90 passed.
- File-scoped pre-commit checks passed.
- Antigravity independently reviewed this exact patch and passed it.

Refs: ThinkFlowLab#114
@wuli666 wuli666 closed this Nov 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.