feat: omniscient audit + tightened reviewer/structurizer/analyzer prompts
Major changes:
1. Audit (verifyIssues) rewrite — now THE final judge instead of a severity-recalibrator:
- Inputs structured issues + Task line with PR URL (audit fetches diff itself via
`gh pr diff` + Read/Grep/Glob). Does NOT consume reviewer chat or pre-stuffed diff.
- Output schema extended: verdict (keep/rewrite/drop/new), body, evidence, reason.
- Can DELETE false positives (not just downgrade), REWRITE weak descriptions, ADD
missed issues — especially cross-file pattern repetition.
- Optional .magpie-house-rules.md picked up from cwd as authoritative repo conventions.
- New config block `audit:` with claude-opus-4-7 + max effort by default.
2. Reviewer prompts (Round 1 + Round 2):
- Add severity vocabulary at reviewer stage (was only at structurizer before).
- Add reverse rubric: do NOT report build script polish, missing comments, forward-
compat hypotheticals, style preferences, theoretical-but-impossible cases.
- Require file:line + code quote + failure scenario for every issue.
- Drop "Review EVERY file / don't stop early" — brevity over completeness.
- Round 2: drop "find what others MISSED" anti-pattern; agreeing is fine.
3. Structurizer:
- line field now REQUIRED (drop issues that can't be anchored to a hunk line).
- Description must capture WHY + FAILURE scenario + FIX (so audit has basis to verify).
- Drop "STRICT — choose LOWER" severity bias.
4. Analyzer: add 6th "建议的 review 重点" section; parseFocusAreas now matches
English + Chinese headings, with-/no-space, bold variant; handles `-` `*` `•`
`·` ①-⑳ `1.`/`1、`/`1)` bullets.
5. Convergence judge: fix parse bug (verdict swallowed by trailing punctuation);
explicit one-word verdict format constraint.
Schema additions:
- MergedIssue: verdict, body, evidence, auditReason
- MagpieConfig: audit?: ReviewerConfig
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Li Liu
e3fd28c0f0