lead-intelligence

Converge Loop

Drive a product to convergence: repeated rounds of gap discovery across agreed dimensions, fixing reversible findings in-round, parking the rest with recommendations, stopping when new-findings dry up. The executable instructions live in prompt.md; this body is the Codex-facing summary.

Key contracts:

• Every finding carries evidence (file:line, URL, or log excerpt) and a dimension × severity classification. Findings without evidence are dropped.
• Fixable = reversible AND unambiguous. Those are fixed in-round, verified by the project's real gate plus a live re-probe, and committed small.
• Everything else is parked with a recommendation (.claude/decisions.md / .claude/blockers.md), never silently dropped, never used to stop the loop.
• Convergence is explicit: default stop rule is two consecutive rounds with zero new P1/P2 findings (cap: 8 rounds). The final report is in Chinese, categorized by dimension and severity, listing fixed work with evidence, parked items with reasons, and mechanism gaps ("how would we know this succeeded" monitoring that does not exist).