Agent Skills: encode-policy-v2

encode-policy-v2

Thin launcher. The canonical workflow—routing, arguments, phases, roles, checkpoints, artifact contracts, retry budgets, and completion gates—lives in references/workflow.md. Read it completely before acting and follow it exactly; do not redefine or compress its behavior here.

Treat the text after $encode-policy-v2 as raw workflow arguments:

$encode-policy-v2 STATE PROGRAM [--skip-review] [--research-only] [--600dpi]
  [--resume] [--from-phase N] [--full-validation] [--local]
  [--sources MANIFEST] [--source-budget MINUTES]

Mandatory gates that no surface may skip:

• Load model-development in each model worker's own context and verify its load evidence under the canonical loading contract before accepting work; missing skills block it.
• Route purely parametric changes to encode-reform unless the user chooses to continue.
• Make no GitHub write before the user-approved scope decision exists.
• Read all relevant rules and visually inspect tables, scans or ambiguous PDF evidence; share validated originals and extracts with the independent review.
• Do not push through a red structural or quick-audit gate, or silently push failing tests.
• Workers never delegate; brief both owners from one shared contract block, and treat the coordinator's single bounded test invocation as the run of record.
• In local mode, make no GitHub writes or pushes; preserve the implementation worktree.
• Keep new PRs draft and complete follow-up review after any review-fix round.
• Maximum two reviews: a second, incremental review follows repairs to confirmed criticals from the first. Stop after round 2 with any remaining blockers reported.
• Do not finish before the Phase 7 summary and canonical completion contract are satisfied.

Surface adapters:

• Claude Code: also read references/claude-launcher.md, which maps canonical roles and abstract operations to Claude Code mechanics.
• Codex: use the delegation mapping below.

Codex delegation mapping, when subagent use is available and authorized:

• The coordinator handles research, requirements, coverage/validation, Git and reporting; it may read code and full evidence directly. No summary-only read restriction.
• Use persistent workers for the two substantive roles: implementer (parameters and variables together) and test-author. Pass exact owned paths, concrete worktree/run values, approved scope, artifact contracts and the needed skill references.
• Each Codex worker resolves policyengine-model-development from its own available skills catalog and reads that entrypoint using its catalog-specified access mechanism, then its task references. Include the canonical loading contract in the prompt; a repository path alone does not demonstrate runtime skill availability. Do not assume the parent session's skill load supplies the worker's context.
• Test-source analysis can overlap implementation; final tests use the published variable contract. Reuse owners for repairs. Workers never stage, commit, push or independently sync/format the shared environment. They do their own work without helper or nested agents, run at most one bounded self-check of their owned tests, and report a CONTRACT MISMATCH instead of relocating cases; the coordinator runs the tests of record.
• Run review-program in a fresh review coordinator context with its canonical local or published arguments, raw source manifest, captured head and scope. Require that coordinator's own model-development load. Leave capacity for its reviewers and consume its completed decision directly for repair routing, without a second review by the encoder.
• Use only supported runtime capabilities. When delegation is unavailable, perform the implementation roles directly while preserving independent review context and gates.