convergence-looplisted

# Convergence Loop Verify claims about reality through multi-pass agent loops with composable behaviors. Implements tenet T20 (research_convergence_loops) and T25 (discovery phases SHOULD use convergence loops). ## Critical Rules 1. **Minimum 2 passes** — MUST never single-pass. First pass discovers, second verifies. One-pass work is not a convergence loop. 2. **T20 tally every pass** — MUST record Confirmed / Corrected / Extended / New for every pass. No pass without a tally. 3. **User gate before convergence declaration** — MUST present tally and recommend converged/not-converged. User decides. (CLAUDE.md guardrail #2) 4. **Save state after every pass** — MUST persist to state file. Convergence loops are long-running and WILL hit compaction. 5. **Graduated convergence** — per-claim, not all-or-nothing. Converged claims graduate; unconverged continue to next pass. 6. **Contradictions are findings** — disagreement between agents or sources is a discovery, not an error. Surface with evidence from both sides. ## When to Use - Verifying claims about codebase state (audit findings, diagnosis) - Discovery phases in any skill (T25 — deep-plan, skill-audit, ecosystem audits) - Validating plan completeness against requirements - Cross-referencing documentation against code reality - Any task where acting on wrong claims wastes significant effort ## When NOT to Use - Single-value lookups ("what version is X?") - Mechanical operations (git mv, file deletion) -