ci-agents

Agent / Team / Panel review infrastructure for GitHub Actions.

Three layers, each independently usable:

• Agent (sw2m/ci-agents/agent) — Run a single AI agent with schema validation and retry.
• Team (sw2m/ci-agents/team) — Run N agents on the same task, evaluate alignment via a user-provided evaluator.
• Panel (sw2m/ci-agents/panel) — Aggregate category results with a ruling algorithm (any-fail, majority, or custom script).

All actions are pure: structured data in (shared module), structured data out (shared + step outputs). No side effects — no comments, no reviews, no issues. The consumer decides what to do with the results.

Data flow

shared.set(ns, "context", data)
        ↓
  Agent / Team / Panel
        ↓
shared.get(ns, "result|alignment|ruling")

Budget tracking

The 7-5-3-1-0 decay schedule is applied at the highest layer used. Panel reads prior comments to determine the round, then includes round and budget in the ruling output. The consumer uses this to decide whether to continue iterating.

Dependencies

• sw2m/octoscript — Deno runtime, shared/output/serde/frontmatter globals