Change Verdict

Use the smallest mode or action and scope that answer the request. The six modes, pending, and default deps read evidence and return results without changing workspace or review state. Explicit dependency-remediation authority permits only a bounded deps fix delegation to an available authorized writer. The parent retains its read-only command grants; neither action metadata nor the user's fix request adds direct shell, Write, or Edit capability.

Modes

Choose one mode. If the mode or requested scope is ambiguous or cannot be resolved, return INCONCLUSIVE and state the minimum evidence or clarification needed.

Actions

Natural-language pending-review and dependency requests select these actions; selectors are optional when the task is clear. Select one action or mode per observation and load only its local procedure.

When NOT to Use

• Implement or fix a source finding: use flow-drive. Explicit dependency remediation belongs to deps fix within its declared bound.
• Trace an unfamiliar code path: use code-trace.
• Specialized skill governance scoring or stocktake: the former audit capability is retired. For an existing skill document or change, select this skill's docs or risk mode within its declared scope.

Procedure

Follow references/shared/review-workflow.md for scope resolution, fixed-point handling, evidence collection, and aggregation. It owns the shared read-only sequence.

Load only the selected mode's references:

• code: use the branch prompt for a branch scope; otherwise use the fast or full prompt matching the selected depth. The shared code research instructions own Git reads; apply the code-only Standards and Spec status rules in references/shared/review-common.md.
• pr: read PR metadata and use bash ${CLAUDE_SKILL_DIR}/scripts/check-unrelated-changes.sh only as a read-only advisory scan when PR metadata is available.
• security: use references/security/codex-prompt-security.md and label findings with their OWASP category.
• tests: read source, tests, acceptance criteria, and available runtime evidence. --ac-trace maps criteria to evidence; --coverage loads references/tests/codex-prompt-test-review.md. Legacy --scope tests means --mode tests.
• docs: read the full target and enough source or configuration to check it. Load references/docs/review-loop-doc.md only when a prior snapshot is supplied.
• risk: run node ${CLAUDE_SKILL_DIR}/scripts/risk-analyze.js --json on the current read-only tree and use references/risk/ to interpret its score.

${CLAUDE_SKILL_DIR} is the directory containing this SKILL.md. Claude substitutes it in these helper commands and in allowed-tools. Hosts that do not substitute it, and references that write $SKILL_DIR, resolve that same directory. The working directory stays the consumer repository. risk-analyze.js and check-unrelated-changes.sh are the helper paths granted by allowed-tools; the grant is not unrestricted Bash. codex review and review-cli.sh stay explicit opt-in and do not start by themselves.

For every mode, report evidence gaps and use the response shape in templates/review_output.md. The code-only Standards and Spec section is omitted from all other modes. Severity and final verdict meanings are owned by references/shared/review-rubric.md.

Optional CLI second opinion

Only explicit --second-opinion=codex-exec may invoke bash ${CLAUDE_SKILL_DIR}/scripts/review-cli.sh --backend cli. Pass the selected scope and fixed point, not the primary conclusion. Keep the CLI result separate, redact it, and record its status. The primary review is complete without this optional check; an unavailable CLI is reported as degraded evidence.

Authority boundary

Return review results in the response. The six modes, pending, and default deps keep tracked/untracked files, index, and lockfiles unchanged. Their read-only scope excludes fixers, generators, staging, commits, artifact writers, gate/sentinel emitters, and writer dispatch. deps fix is the sole bounded writer-delegation exception described above; fix success cannot clear a gate or replace the independent security verdict. A re-review is a new observation of a caller-supplied snapshot. Missing required local resources fail closed before dependent action; obtain no replacement from sibling skills or an ambient source checkout.

References

• references/shared/review-workflow.md — shared sequence and fixed-point handling.
• references/shared/review-common.md — evidence, source labels, and finding normalization.
• references/shared/review-rubric.md — severity and final verdict definitions.
• references/shared/codex-research-instructions.md — research guidance used by code prompts.
• references/shared/cli-backend.md — explicit CLI transport contract.
• references/actions/pending.md, references/actions/deps.md — delegated pending review and dependency audit/remediation procedures.
• references/code/, references/security/, references/tests/, references/docs/, references/risk/ — mode-specific instructions.
• scripts/review-cli.sh, scripts/check-unrelated-changes.sh, scripts/risk-analyze.js — read-only helpers.

Completion check

• One mode or action and a readable, resolved scope are recorded.
• The reported fixed point and evidence anchors match the reviewed snapshot.
• Findings are normalized and ranked; missing evidence remains visible.
• Only an explicitly requested CLI opinion is reported, with its actual status.
• Read-only observations leave repository state unchanged; an explicit dependency fix records the authorized writer, bounded changes, and a separate independent security result. No review artifact, gate, or sentinel was written.