[tzhouam]

Purpose

Owner rule request: when reviewing a PR, check the description's Test Plan / Test Result for (1) whether the CI guard cases covering the changed features/models were actually run, and (2) whether the commands match what CI really executes.

REV-3a in knowledge/general/review/rules.md states the gate. The comparison is answerable rather than a matter of taste, because CI declares both halves at the frozen head:

• a pipeline step's commands are the commands CI runs;
• its source_file_dependencies declare which changed paths that step guards;
• pyproject.toml owns the marker semantics those commands select on.

Equivalence is by selector -- test targets, -m expression, run-level -- never by literal text, so timeout, --cov* and container prefixes are not mismatches. Running a test file directly is explicitly not evidence that the lane collects it: the marker deselects it there.

Coverage: Direct consumers (including /imreview) get it via DIRECT_REVIEW_CHECKLIST; the Strict reviewer gets the repo-shaped form via the injected strict-review-checklist.md. The enforced version for the bot's Direct lane is omni-reviewbot#56.

The injection budget had no room

strict-review-checklist.md was 6,991 characters against a 7,000-character cap (engine/steps/review/steps.py). Anything appended would have been silently cut mid-sentence -- the gate would have looked present in the page while never reaching the model. The cap moves to 8,000 and the page now sits at 7,484.

Three places documented the cap as 4k or 7k while the code said 7,000; they now all say 8k (doc/knowledge/writing.md, doc/features/review-recall.md, the check_wiki_lint.py comment, the test docstring).

SPEC freshness

Touching direct_routing.py and engine/steps/review/ made their SPEC pages stale and failed the --strict gate on the first push. Both are re-verified against the change, and engine/steps/review.md now states the injection cap as a contract, since that is the thing this PR moves.

Test Plan

PYTHONPATH=$PWD/src python -m pytest -q
python knowledge/tools/check_knowledge_tree.py
python knowledge/tools/check_wiki_lint.py

test_review_checklist_resolves_from_adapter_knowledge now pins the budget from both sides: a marker past the old 7k cap must be injected, and a tail past the new 8k cap must not be.

Note for anyone running the suite here: the shared venv has infermatrix_copilot installed editable against another checkout, and this repo's pytest config does not set pythonpath. Without PYTHONPATH=$PWD/src the tests import the other tree and silently pass against unmodified code -- that is how the first run of the budget test looked green-adjacent while the injection was still 7k.

Test Result

Full suite on this branch: 1 failure, test_phase1_cluster.py::test_drift_guard_unimportable_checked_class_fails_guard, identical to origin/main at e40258d3b (verified by running the full suite on a pristine baseline worktree). A test_testing_substrate.py process-reaping test failed once under full-suite load and passes in isolation on both trees; it did not reproduce on rerun.

Knowledge validators: check_knowledge_tree.py 0 errors / 22 pre-existing size reminders; check_wiki_lint.py 0 errors, 0 warnings across 387 pages.

Two ratchets caught a first draft that named the target repo in a core comment (test_repo_neutral_core, test_parity_vocabulary_is_contained); the comment was neutralized rather than the ceilings bumped.

Not covered: no live review was run, so the rule's finding quality on real PRs is unmeasured.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JCk5xJGRyydouK7tuhs8kk

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.