Repository navigation
fix(review): filter self-negating findings the model labels not actionable - #84
Conversation
…nable The non-finding filter missed a class where the model self-labels a finding as "not actionable" or "not a defect" but the selection pipeline posts it anyway (observed on dev-site #63). Add SELF_NEGATING_TITLE regex matching "not actionable", "not a defect", "not a real issue/bug/problem" anywhere in the title, and extend CONFIRMATION_SCENARIO_SUFFIX with "not a defect" end-anchor. 553 tests (11 new). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
![umm-actually[bot]](https://avatars.githubusercontent.com/in/4233716?s=60&v=4){kind=small}
|
Document the new self-negating drop patterns in the Non-finding filter section
The 'A finding is dropped when:' enumeration documents the exact patterns that drop a finding, but it omits the two additions this PR ships: the title patterns 'not actionable' / 'not a defect' / 'not a real issue/bug/problem' (SELF_NEGATING_TITLE) and the failure_scenario ending 'not a defect'. Per the conventions file, README updates in the same change that alters behavior; an operator reading the doc to understand a dropped finding gets an incomplete picture. Failure scenario: An operator sees a finding titled '…— not a defect' dropped from a review and reads the Non-finding filter section to confirm why; the documented rules list only the prefix/end-anchor patterns, none of which apply to that title, so the drop is unexplained by the reference doc that claims to enumerate every drop condition. Suggested fixAdd a bullet to the Non-finding filter section covering the title self-negation patterns and extend the leaked-conclusion bullet to include '…not a defect'. Match the existing backtick-listing style.umm-actually · deepseek/deepseek-v4-flash-0731 |
|
umm-actually re-reviewed at No new findings (5 tracked finding(s) across all runs). Context notes
umm-actually · deepseek/deepseek-v4-flash-0731 |
…tle regex SELF_NEGATING_TITLE required article 'a' before 'issue', missing the grammatically correct 'not an issue' — a more natural model output than 'not a issue'. Changed `a` to `an?` so both forms are caught. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The comment said "start- and end-anchored patterns" but the new SELF_NEGATING_TITLE regex uses word boundaries, not position anchors. The individual regex constants already document their anchoring. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The regex alternation `(?:issue|bug|problem)` had tests for 'issue' and 'bug' but none for 'problem' — removing it from the alternation would have gone undetected. Adds a drop test for "not a problem" in a title. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Address bot review findings:
- SELF_NEGATING_TITLE now requires a separator or end-of-title after the
phrase, so qualified retractions survive ("not a bug on X but Y")
- CONFIRMATION_SCENARIO_SUFFIX extended with the full self-negating set
(not actionable, not a(n) (real) issue/bug/problem) — not just "not a
defect"
- README Non-finding filter section updated with the new patterns
- Test name cleaned of external PR reference
558 tests (3 new passthrough + scenario suffix tests).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
Fixed — README Non-finding filter section updated with the new self-negating title patterns and extended failure_scenario suffix entries. 🔍 ship-check · pr-monitor · claude-opus-4-6 |
…erns The summary sentence claimed all patterns are "anchored to the start or end of a field" but SELF_NEGATING_TITLE is separator-delimited mid-field. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1 participant
Summary
SELF_NEGATING_TITLEregex catches self-retractions anywhere in the title (word-bounded)CONFIRMATION_SCENARIO_SUFFIXwith "not a defect" end-anchor (alongside existing "no bug here" pattern)Observed on dev-site #63: finding titled "Edge-case navigation race — not actionable" with body "Not a defect." was posted, requiring the author to manually acknowledge the non-issue.
Test plan