Repository navigation
Run the release audit in each sweep and draft the adapter-baseline companion - #222
Conversation
…mpanion Sweeps already accepted audit hints, but nothing ran the repository's release auditor for them. The design also has the service draft the adapter-baseline update as a companion PR that runs in parallel with the sweep and never blocks it. - sweep_audit.audit_sweep runs the adapter's release.auditor plugin (generated-baseline mode) for the sweep's upstream SHA pair before the sweep; a fallback re-check with no upstream movement is skipped. It reads one fetched revision of the knowledge repository, materialized with git archive (never a checkout or snapshot that may lag main), and the sweep then runs on that same revision. Knowledge document issues become per-page hints for the sweep generator. An audit that fails goes to people once per sweep, and the sweep runs without hints. - Reconciliation findings (the committed adapter baseline lags the audited upstream) become one companion PR per sweep: the plugin names its baseline file (new baseline_file field, relative to the adapter; vllm-omni: release_baseline.yaml). Only the upstream: and inventories: blocks are replaced with the generated values, so everything else, comments included, is kept. A block runs to the next top-level key: a column-0 comment inside it does not end it, and comments just before the next key stay with that key. It is published as a draft through open_companion_pr and goes to people. The sweep never waits for it: the companion uses the sweep's revision (no extra fetch), and any failure drafting it is traced and sent to people once per sweep while the sweep still runs. The scheduler's recovery pass for staged companions is isolated per change set too, so one that keeps failing to publish never blocks later ticks. - The sweep summary reports the number of reconciliation findings. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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: e9dd619948
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".
| def run_sweep(rt, lifecycle, owner: str, sweep: dict, upstream: UpstreamRepo, | ||
| audit_hints: dict[str, list[dict]] | None = None) -> dict: | ||
| audit_hints: dict[str, list[dict]] | None = None, reconciliation: int | None = None, | ||
| base_sha: str | None = None) -> dict: |
There was a problem hiding this comment.
The new audit inputs remain optional here, so the documented playbooks/kb-sweep.yaml path still reaches engine/steps/knowledge.py::sweep, which calls run_sweep(...) without running audit_sweep or staging the baseline companion. Consequently, invoking kb run rather than the scheduler silently performs a release sweep with no audit hints, no reconciliation report, and no baseline PR; move the audit orchestration into the shared sweep path or update every entry point.
Useful? React with 👍 / 👎.
| committed = rt.knowledge.show(main, path) | ||
| if committed is None: | ||
| return None | ||
| updated = baseline_text(committed, result.generated_baseline) |
There was a problem hiding this comment.
When the audit reports any reconciliation, this always creates a PR that changes only upstream and inventories. The vLLM-Omni auditor also classifies unmatched_path, suspicious_path_route, stale_pin, and other routing/policy findings as reconciliation, none of which this edit resolves; in particular, a changed unmatched path can disappear from future audits once the sweep baseline advances. These findings need separate remediation or escalation rather than being treated as fixed by the baseline companion.
Useful? React with 👍 / 👎.
| hints: dict[str, list[dict]] = {} | ||
| for issue in result.issues: | ||
| page = str(issue.get("document") or "").removeprefix("knowledge/") | ||
| if page: | ||
| hints.setdefault(page, []).append(issue) |
There was a problem hiding this comment.
Audit issues are only converted to hints, but some valid enforced issues cannot be consumed by the sweep: owner_document_missing may omit document entirely, or name a document that is absent from the fetched tree. The former is dropped here and the latter is keyed to a path that rule_pages(base, ...) never iterates, so the sweep completes and advances its baseline without fixing or notifying people about the failed audit condition. Unmappable or nonexistent-page issues should be explicitly escalated.
Useful? React with 👍 / 👎.
Knowledge-service redesign follow-up (design §9.1). Each release sweep now runs the repo's release audit, and baseline drift becomes an adapter-baseline companion PR. Sweeps accepted audit hints before, but nothing ran the auditor.
sweep_audit.audit_sweepruns the adapter'srelease.auditorin generated-baseline mode for the sweep's upstream SHA pair (fallback re-checks with no upstream movement are skipped).git archive, and the sweep then runs on that same revision.baseline_filefield (vllm-omni:release_baseline.yaml).upstream:andinventories:blocks are replaced; other content and comments are kept, and block boundaries are line-aware.Commit hook approved; full suite and SPEC freshness pass locally.
🤖 Generated with Claude Code