Pipeline

Architecture is pipes-and-filters. Heuristics catch obvious issues in seconds; full mode still runs AI so Ship Coach can merge local and model findings.

%%{init: {"flowchart": {"nodeSpacing": 20, "rankSpacing": 28, "padding": 8}, "themeVariables": {"fontSize": "13px"}}}%%
flowchart TD
  A[Collect context] --> B[Heuristics]
  B --> C{Full mode?}
  C -->|yes| D[Diff Analyst]
  D --> E[Risk Reviewer]
  E --> F[Ship Coach]
  F --> R[Report]
  C -->|no / --local| R
  R --> V[Verdict + draft PR]

Stages

1. Context — diff, changed files, optional test signals, Spec task counts, root ready.yml (git or --path)
2. Heuristics — secrets (.env, AKIA / ghp_ patterns), TODO/FIXME, debug logs, custom regex from ready.ymlready.yml
   Optional root `ready.yml` Definition of Ready — test globs, forbidden paths, max diff size, custom regex blockers, and path allowlists.
3. Diff Analyst — summarizes the change set
4. Risk Reviewer — scores risks and blockers
5. Ship Coach — merges findings into READY / READY WITH WARNINGS / NOT READYREADY / READY WITH WARNINGS / NOT READY
   Structured report outcome: READY, READY WITH WARNINGS, or NOT READY — with blockers, warnings, and checklist., checklist, draft PR title/body, top actions

Graceful degradation

On agent failure or missing Bedrock access, the product degrades to a heuristic-only report rather than failing the API hard — especially important for GitHub Actions (prefer HTTP 200 with a degraded report).

Tip

Heuristic blockers do not skip AI in full mode. Ship Coach merges local and model findings so the verdict stays consistent across surfaces.

Timeouts

Sequential Bedrock calls default to ~20s each. Lambda budget is ≥90s / 512 MB. Save-path Kiro hooks stay on heuristics (30s); full Bedrock is user-triggered (120s).