plan-execution-loop

Plan Execution Loop

Thesis: One plan slice at a time — implement, evidence-based audit (≥90/100, file:line proof required), fix, re-audit — then next slice. Parent agent orchestrates; subagents implement and audit.

Invoke when: User says "run the plan", "execute slice X", or points at docs/plans/*.md for implementation.

Per-slice loop (do not skip)

1. IMPLEMENT — Task subagent: read plan slice only; implement; run verification commands; update Progress row
2. AUDIT   — Task subagent (readonly): grade /100 vs slice acceptance with file:line evidence; PASS only if ≥90 and every criterion has proof
3. FIX     — if FAIL: Task subagent applies audit fix list; no scope creep
4. RE-AUDIT — repeat 2–3 until ≥90 with full evidence
5. NEXT SLICE — only after current slice PASS

Subagent prompts must include

• Exact slice ID and file path to plan section
• Acceptance criteria from plan (copy verbatim)
• Verification commands (npm run test:coverage, check_spec_drift.sh, etc.)
• Forbidden scope (other slices, unrelated refactors)
• Evidence rule: every deduction needs path:line citation; every PASS criterion needs path:line proof or automatic FAIL
• Fresh grade (audit dispatches only): grade the artifact as it exists now — start at 100, deduct only for current gaps with path:line proof. Do not reward "improvement since last pass", prior fix lists, or Progress table history. Parent must not anchor with prior scores (see Re-audit below).
• "Return: grade, PASS/FAIL, evidence table, fix list, files changed, blockers"

Evidence-based audit (mandatory)

Audit subagents grade evidence, not vibes. A numeric score without citations is invalid.

For every deduction — cite path:line (or path:line-line for a range) showing the gap. Example: -5 Scope: backend/internal/handler/foo.go:42 — touches bar.go, outside slice file list.

For every slice acceptance criterion — cite path:line (or command output) proving it is met. If the auditor cannot find proof after reading the diff and running verification commands, that criterion is automatic FAIL regardless of the running score.