Add 11 loops: model-upgrade-migration, PR-gate guardrails, stateful guardrails, long-horizon migrations

[DCCA]

Summary

Implements all remaining roadmap loops — the complex-loops ranking and the stateful / long-horizon backlog are now fully built out. 11 new loops across 4 commits, each following loopy's trigger → detect → act → output → guardrails contract with injected boundaries (fully unit-tested with fakes / a memory StateStore).

Model-upgrade migration (415f3e9)

• model-upgrade-migration — golden-set diff → human-gated model switch on runPlan (baseline → candidate → diff → approve → apply). Reuses prompt-eval-gate's scorecard/regression logic.

Wave 1 — PR-gate & freshness guardrails (5593467)

Deterministic, single-shot, report-PR loops:

• i18n-drift — flag missing/orphaned translation keys across locales
• perf-budget — flag performance regressions vs a stored baseline
• a11y-baseline — fail only on new accessibility violations vs a baseline
• runbook-freshness — flag runbooks past their review interval

Wave 2 — stateful guardrails (619e69b)

On the durable StateStore (+ a human Gate where approval is warranted):

• test-impact-budget — flag tests grown past a rolling EWMA runtime baseline; rolls the baseline forward
• data-contract-guard — additive schema changes auto-record; breaking changes block behind a human gate before a PR records the new baseline
• cost-guardrail — track per-resource idle streaks (grace period); gated remediation report once idle past minStreak — never auto-deletes
• eval-set-drift — surface production categories missing from the eval set; record surfaced categories so each reports once

Includes a real bug fix: cost-guardrail's idle streak only persisted in act, but act only runs once the streak crosses the threshold — a chicken-and-egg that meant it could never fire. The counter now advances+persists in detect (every run, as the grace period requires); act reads the already-advanced value to avoid double-counting.

Wave 3 — long-horizon migrations (954a0bb)

Export-only programmatic advance* loops on the resumable runPlan primitive (the model-upgrade-migration pattern; resumable via plan:<planId> in the StateStore):

• api-deprecation-rollout — announce → grace-period (long-horizon wait) → caller-drain → human-gated removal. Won't remove an API while callers remain.
• dep-major-migration — migrate one major bump (the kind dep-updates skips): verify the consumer build → human-gate a green result → emit the manifest bump. A red build is never gated or proposed; the failing log is surfaced.

Each wave-2 loop is wired into the catalog, loopy run (env-file boundaries + file StateStore), and package exports; the wave-3 programmatic loops add package exports. Every loop has a loop.yaml, README, playbook, and tests.

OpenSpec

Specs (source of truth) + archives added for all 11 loops under openspec/specs/ and openspec/archive/; openspec/roadmap.md and README.md updated.

Validation

• npm run typecheck / lint — clean
• npm test — 249 tests passing
• npm run build — clean
• CLI smoke: new runnable loops listed + wired (clear guidance without a data source)

Inventory: 26 loop directories, 21 runnable via the CLI catalog, 5 export-only programmatic long-horizon loops.

🤖 Generated with Claude Code

https://claude.ai/code/session_012CrMioorha3QtPUgWbtDbR

---

Generated by Claude Code

[DCCA]

Self-review

Reviewed the full diff (114 files, +5804 / −5, all four commits). Findings:

Consistency — all 11 loops follow the established contract and conventions: ESM NodeNext with .js relative imports, import type for type-only imports, *FromManifest factories with typeof config guards, deterministic renderers, and fail-safe detect/act. The wave-3 loops mirror model-upgrade-migration exactly (runPlan steps + an advance* export, resumable via plan:<planId>).

Boundaries — every external dependency is injected (timing/usage/schema/drift sources, verify, remainingCallers, readers) and faked in tests; no network in unit tests. CLI wiring reads env-file boundaries + a file StateStore, matching the wave-1 pattern.

Bug caught during integration — cost-guardrail persisted its idle streak only in act, but act only runs once the streak crosses the threshold, so the counter could never accumulate to fire. Fixed: the streak now advances+persists in detect (every run, which is what the grace period requires) and act reads the already-advanced value to avoid double-counting. Covered by the "builds a streak… then produces a PR once approved" and "never reaches the gate for a busy resource" tests.

Gates / safety — human-gated paths (data-contract-guard, cost-guardrail, api-deprecation-rollout, dep-major-migration) all hold until explicit approval and emit nothing destructive; cost-guardrail never auto-deletes, dep-major-migration never gates a red build, api-deprecation-rollout won't remove while callers remain.

Validation — typecheck, lint, 249 tests, and build all clean locally; CLI list/guidance smoke-tested. OpenSpec specs + archives present for all 11 loops.

LGTM — merging once the build check goes green.

---

Generated by Claude Code