full-review: Non-Diff Audits, Evidence Triage & Direction Decisions¶
Use full-review when auditing an entire repository, subsystem, commit range, or non-diff evidence bundle (transcripts, logs, CI runs, benchmarks, prior ledgers), or when the user asks a Direction decision question (merge vs. rework, keep vs. close, option selection, sequencing).
- Source specification:
.agents/skills/full-review/SKILL.md - References:
evidence-intake.md,report-template.md,review-roles.md - Governing standard: PR Review Standard v1.5.0
- Canonical role map & closure gate: Roles, skill workflows, and runtime closure
When to use full-review¶
- Auditing repository state, a subsystem, or a multi-artifact evidence bundle beyond a single PR diff.
- Answering an architectural or release Direction decision with explicit constraints, options, falsifiers, and sequenced next actions.
- Keep diagnosis separate from unauthorized code edits. For standard PR diffs use
pr-review; for cold Day-1 launch audits useday-one-review.
Workflow¶
- Frame & evidence intake (
Responsible Reviewer): State the review or direction question in one sentence, pinSession Provenance, and build an explicit Evidence manifest (EID | Artifact | Locator | SHA-256 / size | Class | Access | Examined). Ingest large logs or transcripts in targeted batch windows rather than step-by-step polling. - Pin, trace & 11-category coverage (
Working Cohort->Responsible Reviewer): Trace contracts before runtime behavior and populate all eleven Standardv1.5.0categories (ARCH,DOC,SKILL,MCP,RUN,SVC,DB,SCRIPT,TEST,DEP,SEC), marking unreviewed or irrelevant buckets explicitly (Not assessedorNot applicablewith a reason; never silently skipped). Issue a provisional content verdict. - Direction decision (Section 7.2, when requested): Record the decision question, hard constraints, options table (including status quo and minimal option), load-bearing evidence, falsifiers, ordered immediate next actions, and explicit "What not to do" guardrails.
- Two-plane gate (
Responsible Reviewer<->Independent Integrity Auditor): When running as a team or/swarm(paired withswarm-governance), the flat Integrity Auditor checks tool traces and file hashes to emitCLEANorINTEGRITY VIOLATION(never contentPASS). OnCLEAN, the Responsible Reviewer re-checks final artifacts and issues the final verdict. - Delivery & quality gates: Format using
report-template.md(Sections 1–8, plus Section 9 only if/self-diagnosiscohorts ran), saving underdocs/reviews/<review-slug>.mdwhen adopted or explicitly requested, or delivering in chat or the Antigravity artifact directory.