Repository navigation
Complete JiuwenSwarm knowledge acceptance and repair init evidence reuse - #327
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: bb74e1e4a4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| @@ -292,10 +292,14 @@ def from_env(cls, settings, *, state_dir: Path) -> "InitRuntime": | |||
| traces = TraceStore(state_dir / INIT_DIR / "traces") | |||
| clone = Path(os.environ.get(CLONE_ENV) or state_dir / INIT_DIR / "knowledge-repo") | |||
| ensure_knowledge_clone(clone) | |||
| from .model_dispatch import SharedModelDispatch | |||
There was a problem hiding this comment.
On Windows, InitRuntime.from_env() now unconditionally imports model_dispatch, whose module-level import fcntl is unavailable there. As a result, every adapter-backed kb init invocation fails before selecting a stage—including the documented no-model --suggest-seeds path—even though the repository exposes Windows installation and the imkbinit skill. Use a platform-aware locking implementation or avoid constructing the shared dispatcher on unsupported platforms.
Useful? React with 👍 / 👎.
The expanded JiuwenSwarm catalog has 173 features, but initialization left feature evidence undiscovered or attached to nearby implementation. This change locates pinned citation spans and related test assertions, prioritizes the requested feature during deepening, shares the existing 13 model slots across init paths, and restores navigation to retained audited depth pages. Successful Git reads are cached by immutable commit and path; moving refs are resolved first, errors remain uncached, and live PR metadata is rechecked.
At upstream
f0a69728c96b5961d993449f1a901cbd2f4dac5b, foundation knowledge increases from 927 to 1,005 of 1,038 cells (108 to 142 fully covered features). Depth knowledge increases from the published 552 to 1,178 of 1,211 cells, including 495 retained preview cells and 131 newly approved cells. All 173 features have accepted depth; each facet exceeds 90% (validation is 158/173), and production-file structural coverage remains 2,058/2,199. Foundation and depth each retain 33 unknown cells. Historical text and proofs are preserved; no unknown or missing-test claim is counted as approval.The Chinese report and 72 KB summary are in
eval/knowledge-depth/jiuwenswarm-init-completion-20261007.{md,json}. Complete native inputs, outputs, usage and tool traces remain outside Git with audit hashes. The report distinguishes lightweight acceptance, source navigation, actual injected context, and unexecuted upstream tests. The measured local source-read replay improves from 7.002 to 0.722 seconds with identical hashes; no new GLM PR-review accuracy or speed gain is claimed.Validation so far:
Merge only after the remaining acceptance checks and CI pass. The final gate allows truthful partial foundation knowledge and keeps all depth and production-file coverage targets unchanged.