Branch: refs/heads/prompt-audit-fixes Home: https://github.com/xwiki/xwiki-dev-llm Commit: 99485c9518a517c3b849186366b458e9593ec782 https://github.com/xwiki/xwiki-dev-llm/commit/99485c9518a517c3b849186366b458... Author: Vincent Massol <[email protected]> Date: 2026-09-03 (Thu, 03 Sep 2026) Changed paths: M CLAUDE.md M README.md M xwiki/skills/xwiki-jira/SKILL.md M xwiki/skills/xwiki-review/SKILL.md M xwiki/skills/xwiki-translations/SKILL.md Log Message: ----------- [Misc] Fix the xwiki-review skeptic gate and stale prompt-surface text * Replace xwiki-review's 0-100 confidence score with the five named verdicts the rubric already described. The scale emitted only 0/25/50/75/100 while the gate dropped everything below 80, so the 75 tier — a finding resting on a rule stated explicitly in an OKF file or a CLAUDE.md — could never survive and only `100` ever reached the comment. * Run xwiki-review's eligibility gate and context pack in code. The PR state, the head-SHA comment check, the size ceiling, the bot-author check, the CLAUDE.md paths, module packaging and the jacoco ratio are all determined by their inputs; one model call is left in the gate, for the machine-generated / trivially-safe judgement. * Drop the per-angle "at most 5 findings" cap and the report's cap of 10. The reviewer contract already states the bar qualitatively, and the number truncated by arrival order, not by severity. * Record all three MCP servers and both launcher scripts in CLAUDE.md, which still described .mcp.json as holding two servers and omitted start-sonarqube-mcp.mjs. * State the xwiki-jira write guards once each, in plain register, instead of a `## NEVER` section repeating what the flow above it already says. * Drop the MUST/ALWAYS intensifiers in xwiki-translations, whose rules each carry their reason in the sentence that follows. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> To unsubscribe from these emails, change your notification settings at https://github.com/xwiki/xwiki-dev-llm/settings/notifications