Epic: mine loops.elorm.xyz closed-loop patterns into MAP

[azalio]

Context

Researched the full loops.elorm.xyz directory — 40 reusable "closed-loop" workflows for AI coding agents (Ralph-loop lineage). Each loop ships a uniform contract: goal + bounded max-iterations + a single deterministic check command + a binary exit condition + an anti-gaming guardrail block + a self-pacing kickoff prompt.

This epic captures the subset that is novel for MAP and in-scope for a language-agnostic autonomous-coding framework. The remaining ~32 loops were deliberately dropped:

• Already covered: test/build/lint-typecheck/format-until-green (existing gates), ship-pr / fix-ci-until-green (PR routine in CLAUDE.md), independent-verifier-pass (final-verifier), pre-commit / post-edit guards (existing hooks), autoloop-tdd (map-tdd), de-sloppify / dependency-upgrade-one-by-one (minimality Epic: integrate a minimality doctrine (Ponytail concept) across the MAP pipeline #180–Phase 4: gated decomposer pruning (deferred_yagni parking lot + pruneable classification) #184 + per-subtask atomic execution), docs-sync / changelog-sync (release + docs policy), spec-first-ship (decomposition-completeness / forward-coverage gate Evaluate STORM-style cross-agent reconciliation: ship deterministic Finding-schema + blind-spot/contradiction flags, discard LLM ceremony #249).
• Out of MAP core: a11y, bundle-size, visual-regression, openapi/api-contract, prisma migration, staging-smoke, deploy-verification, knip, npm/security-audit as a JS domain.

Cross-cutting observation worth keeping

Every loop hard-codes the same anti-gaming guardrail set: don't modify the check/exit to force success; don't skip/bypass checks; don't weaken/delete/skip tests; no trivial always-pass assertions; prefer fixing prod code over patching tests; (golden-file gates) bless baselines only for intentional changes. MAP enforces the spirit of this via Monitor + "fix every surfaced error", but it is worth auditing/consolidating as an explicit named doctrine (see slice on N-pass convergence + anti-gaming).

Slices

Tracking comment below enumerates the child issues. Each slice is independently shippable.

[azalio]

Slice tracking

Filtered 40 loops from loops.elorm.xyz → 8 novel, in-scope slices.

Tier 1 — correctness / safety (high value):

• Flaky-test triage: repetition-as-signal + Monitor 'deferred (nondeterministic)' disposition #252 — Flaky-test triage: repetition-as-signal + Monitor deferred (nondeterministic) disposition
• Intra-run failure memory: fail-twice→guardrail-sign + per-attempt reflexion fed back next iteration #253 — Intra-run failure memory: fail-twice→guardrail-sign + per-attempt reflexion fed back next iteration
• map-debug: require an executable minimal-repro that proves root cause before any fix #254 — map-debug: require an executable minimal-repro that proves root cause before any fix
• Bounded-effort retry + explicit escalation outcome ('act once, then escalate') #255 — Bounded-effort retry + explicit escalation outcome ("act once, then escalate")
• Merge-conflict resolution discipline: per-file, intent-preserving, test-after-each-batch #256 — Merge-conflict resolution discipline (directly addresses the recurring CONFLICTING-PR pain)

Tier 2 — convergence / gate primitives (medium):

• Monitor/self-review convergence: require N consecutive clean qualitative passes #257 — Monitor/self-review convergence: N consecutive clean qualitative passes
• Numeric-threshold convergence gate (coverage %, perf/size budget) + anti-metric-gaming #258 — Numeric-threshold convergence gate (coverage %, perf/size budget) + anti-metric-gaming

Tier 3 — exploratory (needs scoping decision):

• Exploratory: scheduled/watchdog trigger model + 'report-delivery' exit-condition shape #259 — Scheduled/watchdog trigger model + "report-delivery" exit-condition shape

Suggested order: #254, #253, #255 (tighten the debug/retry loop), then #252, #256, then #257/#258, then decide on #259.

[azalio]

✅ #254 landed via #260 (squash-merged to main).

Repro-probe root-cause gate is live in /map-debug: record_repro_probe (witness exit 42 before fix) + verify_repro_resolved (same frozen snapshot must flip 42→0 after fix), with an immutable locked snapshot, sentinel exit contract, repro_probe manifest stage, and a self-contained gitignored scratch dir. llm-council-reviewed (Option A). make check green (pyright 0, 2675 tests).

Updated slice tracking:

Tier 1 — correctness / safety:

• Flaky-test triage: repetition-as-signal + Monitor 'deferred (nondeterministic)' disposition #252 — Flaky-test triage
• Intra-run failure memory: fail-twice→guardrail-sign + per-attempt reflexion fed back next iteration #253 — Intra-run failure memory
• map-debug: require an executable minimal-repro that proves root cause before any fix #254 — map-debug repro gate (feat(map-debug): enforce repro-probe root-cause gate (#254) #260)
• Bounded-effort retry + explicit escalation outcome ('act once, then escalate') #255 — Bounded-effort retry + escalation
• Merge-conflict resolution discipline: per-file, intent-preserving, test-after-each-batch #256 — Merge-conflict resolution discipline

Tier 2: #257, #258 · Tier 3: #259

Suggested next: #253 (intra-run failure memory) or #255 (bounded-effort retry) to keep tightening the debug/retry loop.