Pre-Execution Review Policy (internal)

SkillDev tools

Internal owner of the pre-execution review cycle and the planning ledgers: independence, unioned findings, counter-evidence dismissal, no-progress, `CONVERGENCE-ANOMALY`, and the evidence/obligation/findings tables. Consumed by `review-spec`, `review-plan`, and the authoring skills. Not a menu entry.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Pre-Execution Review Policy (internal) skill

What this skill tells your AI

The instructions your AI receives, as published by gtrabanco/agentic-workflow in skills/pre-execution-review/SKILL.md and read by ahel’s review.

One owner for the rules that both pre-execution reviewers apply, so a Product review and a Plan review cannot drift into two different definitions of independence, union, or convergence. It states policy; it runs nothing, writes nothing, and emits no verdict of its own.

evidence-grounding  = how an author prepares and self-checks an artifact.
pre-execution-review = how any pre-execution reviewer judges it, and what the
                       frozen ledgers look like.
review-spec / review-plan = the only skills that emit a pre-execution verdict.

When to use

  • review-spec / review-plan — before the first check and again before any repeat, repair, or synthesis step.
  • design-feature, plan-feature, plan-feature-scaffold, plan-fix — when writing the ledgers and when a review comes back failed.
  • Nothing else. Candidate-source review (review-change), merge gating (audit-pr) and execution (execute-phase) keep their own contracts; this skill adds no authority over them.

Hard rule — no verdicts, no edits

This skill never prints SPEC-REVIEW-PASS, PLAN-REVIEW-PASS, SPEC-REVIEW-FAIL, PLAN-REVIEW-FAIL or NEEDS-DESIGN as a result of its own reading, and it never writes a unit artifact. Quoting a verdict shape here is a definition, not an issuance. A policy summary that reads like an approval is a contract violation — the verdict belongs to the reviewer turn that binds the snapshot.

The three references

Condition nowLOAD
Running or repairing a pre-execution reviewreferences/POLICY.md — independence, union, dismissal, diversity labels, author exclusion, untrusted content, critique/synthesis/arbitration bounds, no quorum, no-progress, batch repair, CONVERGENCE-ANOMALY, write-then-report
Building or re-checking a snapshot digestreferences/SNAPSHOT.md — the one executable recipe (scripts/pre-execution-snapshot.mjs), what each stage binds, how a consumer re-verifies a receipt, and why a snapshot digest is not a git blob id
Writing or validating a Plan-stage artifactreferences/LEDGERS.md — the planning-evidence table, the obligation ledger, the stage-aware planning-findings.md, the durable review mark, and who may write each

Guardrails

  • Single owner. A caller may restate a rule only as a one-line pointer plus the stage-specific detail it adds. Two copies of the union rule, the dismissal rule, or the convergence fields is the drift this skill exists to prevent.
  • Policy is not a receipt. Reading this file proves nothing about any artifact; only a reviewer turn that binds a snapshot does.
  • No weakening. The bounds here are ceilings and floors: a cycle may be shorter than the budget allows, never looser than the policy allows.
  • Vocabulary is closed. Severity, class, role, verdict and status words are exactly those named in the references. Inventing accepted, waived, mitigated, or a fourth verdict is a contract break.
  • Docs-language and commit conventions per the project's Workflow conventions.

Done when

  • Every pre-execution reviewer and authoring caller points here for the shared cycle and ledger rules instead of restating them, and each stage file keeps only the detail that is genuinely stage-specific.

Signals

GitHub stars
21
Forks
3
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
pre-execution-review
Source
github.com/gtrabanco/agentic-workflow