Add codemod-campaign loop (cross-PR migration campaign)

[DCCA]

Summary

Ships the flagship complex loop — rank #1 from research deep-dive #3 — the codemod-campaign: a staged, throttled, ledger-tracked codemod campaign across the codebase, built on loopy's long-horizon primitives (durable StateStore + human Gate).

advanceCampaign(...) (called repeatedly, e.g. on a schedule) advances at most one batch per run:

• ledger in StateStore tracks migrated files, open batches, counts, pilot approval;
• reconciles against real PR state each run (merged → migrated, closed → files back to the pool);
• throttles at maxOpenPrs; pilot batch is human-gated (no PRs until approved);
• a failed batch (red tests) opens no PR; completes when nothing remains and no batches are open.

Idempotent and resumable. Injected boundaries (codemod, targets, test runner, prs) make it fully unit-tested with fakes + a memory store.

OpenSpec

Shipped via proposal → apply → archive; source of truth: openspec/specs/codemod-campaign.md.

Validation

• typecheck / lint — clean
• npm test — 166 tests (+5: full lifecycle — pilot gate → batches → throttle → reconcile/merge → completion → failure)
• npm run build — clean

Programmatic for now (like the experiment orchestrator); a GitHub-backed CampaignPrs adapter and a loopy campaign CLI verb are noted follow-ups. Next from the ranking: prompt-eval-gate, model-upgrade-migration.

🤖 Generated with Claude Code

https://claude.ai/code/session_012CrMioorha3QtPUgWbtDbR

---

Generated by Claude Code

[DCCA]

Code review

CI green (build ~23s); locally typecheck, lint, 166 tests, build all pass.

Strengths

• Best showcase of the long-horizon primitives together. The campaign ledger lives in the durable StateStore, the pilot batch is a real Gate, and advanceCampaign is idempotent + resumable — exactly the cross-PR campaign pattern the deep dive ranked #1.
• Correct staged-rollout invariants, each covered by a test: no PR until the pilot is approved (gate); throttle at maxOpenPrs; reconcile against real PR state every run (merged → migrated, closed → files back to the pool); files migrate only on merge; a red batch opens no PR and is recorded; completion only when nothing remains and no batches are open. This directly answers the anti-patterns from the research (runaway, stale state, partial-rollout corruption, gate bypass).
• Pure helpers (remainingTargets/selectBatch/reconcile/renderBurndown) are separately unit-tested; the whole engine is fakeable (codemod/targets/runner/prs + memory store) with no real network.

Notes (non-blocking, scoped)

• Programmatic for now — CampaignPrs is injected; a GitHub-backed adapter (open + state) and a loopy campaign CLI verb are the noted follow-ups (consistent with the experiment orchestrator).
• targets() returns files still matching the predicate, so campaignTotal is derived (migrated + in-flight + remaining); fine for burndown.

Recommend merge.

---

Generated by Claude Code