Agent Skills: review-program

review-program

Thin launcher. The canonical workflow — phases, roles, gates, severity rules, artifact contracts — lives in references/workflow.md. Read it completely before acting and follow it exactly; this file only adapts it to this surface. Do not redefine flags, phases, severity rules, file ownership, or completion gates here.

Use this skill for PolicyEngine-specific PR reviews even when the user starts from a generic built-in review command (such as Codex /review): the canonical workflow adds PolicyEngine source, reference, regulatory, PDF, and test-review rules a generic review lacks.

Treat the text after $review-program or /review as the raw workflow arguments:

$review-program [PR_NUMBER_OR_SEARCH] [PDF_URL] [--local] [--local-diff] [--full]
  [--skip-pdf] [--600dpi] [--resume] [--incremental REPORT] [--prefix NAME]
  [--sources MANIFEST] [--source-budget MINUTES]

With --local-diff, the PR argument may be omitted for unpublished local work; the canonical workflow resolves the base repository and reviews local HEAD without a PR.

Mandatory completion gate:

• Default to policy/source and code/test reviewers; consolidate their findings directly.
• Reviewers do their own investigation and never delegate helper or verification tasks. Use the surface adapter's reviewer profile; read-only model work still requires skills.
• Reviewers own finding validation. The coordinator resolves specific conflicts and assembles the result; calling workflows route repairs without reviewing it again.
• For country-model work, the coordinator and each reviewer load model-development in their own contexts before substantive work and supply the canonical loading evidence.
• The review is incomplete until {RUN_ROOT}/{PREFIX}-review-full-report.md and {RUN_ROOT}/{PREFIX}-review-summary.md exist, with {RUN_ROOT}/{PREFIX}-review-result.json for confirmed-critical repair dispatch.
• Phase 6 consolidation is required before displaying or posting findings.
• Report COMPLETE or PARTIAL separately from the critical count; missing material evidence cannot be reported as a clean review.

Surface adapters:

• Claude Code: also read references/claude-launcher.md — it maps canonical roles to this plugin's agent types and the workflow's abstract operations to Claude Code mechanics.
• Codex: use the delegation mapping below.

Delegation mapping (when subagent use is available and authorized):

• Map the two substantive reviewer roles to workers with skill access. Where supported, remove delegation tools from those workers. They perform read-only dependency tracing themselves rather than creating additional explorers or reviewers.
• Pass concrete RUN_ROOT, WORKTREE_ID, PREFIX, and (from Phase 1) SNAPSHOT values and the role's task spec from the canonical workflow to every subagent; name the role's skills explicitly in the subagent prompt — do not assume a worker infers them from parent context.
• For country-model work, each worker resolves model-development in its own runtime catalog, reads the full entrypoint with its catalog-specified access mechanism and supplies the canonical SKILLS_READY evidence. Missing skill availability/load is a blocker, not a reason to silently substitute repository excerpts.
• Every subagent appends progress lines to its canonical progress log while working and finishes by writing its assigned file and returning the one-line DONE message from the canonical completion contract. Dispatch reviewers before writing the scope brief when the runtime can message a running worker; otherwise include a brief capped at one minute.
• If subagents are unavailable or not authorized, execute the roles directly in phase order, preserving the same handoff files and read-only contract.