Nova review

Shared Execution Policy

This file is the supporting behavioral contract for /nova-plugin:review and the deprecated /nova-plugin:nova-review compatibility entrypoint. Prefer the direct command; the compatibility name remains only for the current major-version migration window.

• Resolve natural-language and explicit KEY=value inputs using ../_shared/parameter-resolution.md; explicit non-conflicting values take precedence.
• Apply ../_shared/safety-preflight.md before side effects. Never infer approval, destructive scope, credentials, or output destinations.
• Follow ../_shared/output-contracts.md and ../_shared/artifact-policy.md; report completed, skipped, and blocked validation truthfully.
• Respect the frontmatter tool boundary. Missing inputs, unavailable dependencies, overlapping user changes, or repository-policy conflicts are blockers rather than permission to broaden scope.

Execution

1. Parse $ARGUMENTS against the workflow-specific inputs below.
2. Read only the context required for the requested scope.
3. Apply the workflow contract and its strict output format.
4. Stop before unauthorized side effects; otherwise validate in proportion to risk and report residual risk.

Workflow Contract

<!-- BEGIN GENERATED BEHAVIOR CONTRACT -->

Generated from workflow-specs/behaviors.v2.json. This block is authoritative. Run node scripts/generate-behavior-surfaces.mjs --write after changing the IR; if explanatory text below conflicts, fail closed.

Generated Behavior Index

• Purpose: Perform evidence-grounded code or design review at the requested depth and output mode without implementation.
• Canonical inputs: REVIEW_SCOPE(required aliases=INPUT,SCOPE); LEVEL(optional aliases=DEPTH default="standard" exact="lite","standard","strict"); MODE(optional default="full" exact="full","findings-only"); REVIEW_PROFILE(optional default="general" exact="general","plan","codex-review-only","codex-verify-only")
• Resolved variant authority: {"REVIEW_PROFILE":"codex-review-only"} normalized={"LEVEL":"standard","MODE":"full","REVIEW_PROFILE":"codex-review-only"} -> runtime/contracts/codex-review-only.json; {"REVIEW_PROFILE":"codex-verify-only"} normalized={"LEVEL":"standard","MODE":"full","REVIEW_PROFILE":"codex-verify-only"} -> runtime/contracts/codex-verify-only.json; {"REVIEW_PROFILE":"plan"} normalized={"LEVEL":"standard","MODE":"full","REVIEW_PROFILE":"plan"} -> runtime/contracts/plan-review.json; {} normalized={"LEVEL":"standard","MODE":"full","REVIEW_PROFILE":"general"} -> runtime/contracts/review.json; {"LEVEL":"lite"} normalized={"LEVEL":"lite","MODE":"full","REVIEW_PROFILE":"general"} -> runtime/contracts/review-lite.json; {"LEVEL":"standard","MODE":"findings-only"} normalized={"LEVEL":"standard","MODE":"findings-only","REVIEW_PROFILE":"general"} -> runtime/contracts/review-only.json; {"LEVEL":"strict"} normalized={"LEVEL":"strict","MODE":"full","REVIEW_PROFILE":"general"} -> runtime/contracts/review-strict.json. Declared selector defaults are applied before matching. An exact normalized override wins; a non-exact combination that triggers an alias specialization stops as conflicting, and only a valid combination that triggers no specialization uses the canonical fallback. The complete resolved runtime contract is authoritative and no field falls back to canonical prose.
• Claude static-entrypoint gate: Native command and Skill frontmatter are static. A matching command wrapper may continue after it has verified that its invoked command id equals resolvedWorkflowId; this canonical Skill must not re-resolve or reject that validated wrapper. Only when this canonical Skill is itself the Claude native invoked entrypoint and no validated wrapper gate exists must resolvedWorkflowId equal review. Otherwise STOP before tools or side effects and invoke the exact direct command /nova-plugin:<resolved commandEntrypoint.directCommandId>; never execute the specialized contract under unmatched canonical frontmatter. Generic and Codex adapters may execute the resolved contract directly under adapter enforcement.
• Decision entries: 7; canonical routes and variants: review {"REVIEW_PROFILE":"plan"}, review {"REVIEW_PROFILE":"codex-review-only"}, review {"REVIEW_PROFILE":"codex-verify-only"}, review {"MODE":"findings-only"}, review {"LEVEL":"lite"}, review {"LEVEL":"standard"}, review {"LEVEL":"strict"}.
• Workflow steps: resolve-scope → route → inspect → emit
• Output: mode=chat; order=findings → impact rationale → directional guidance; severity=Critical, Major, Minor.
• Deviation/failure: mode=forbid; failure order=status → missing input → allowed values → safe next action.
• Full IR: runtime/contracts/review.json#behaviorContract embeds the complete decision table, invariants, stops, field definitions, validation, and failure contract from the same source. Detailed guidance below may not override it.

<!-- END GENERATED BEHAVIOR CONTRACT -->

Purpose

Provide structured review findings for code/design artifacts through one canonical surface, with depth and output shape selected by parameters.

Inputs

Resolve LEVEL, MODE, and REVIEW_PROFILE first, then use the matched runtime contract's required-input set:

Resolved profile	Selector	Required inputs
General review	{}	REVIEW_SCOPE
Lite review	{"LEVEL":"lite"}	REVIEW_SCOPE
Findings-only review	{"LEVEL":"standard","MODE":"findings-only"}	REVIEW_SCOPE
Strict review	{"LEVEL":"strict"}	REVIEW_SCOPE
Plan review	{"REVIEW_PROFILE":"plan"}	PLAN_INPUT_PATH
Codex review-only	{"REVIEW_PROFILE":"codex-review-only"}	REVIEW_SCOPE
Codex verify-only	{"REVIEW_PROFILE":"codex-verify-only"}	REVIEW_FILE

REVIEW_SCOPE is not a universal requirement: plan review uses PLAN_INPUT_PATH, while Codex verify-only uses REVIEW_FILE.

For the general, lite, findings-only, and strict profiles, REVIEW_SCOPE must materialize the evidence as supplied text or as an explicit readable file/path set. A label such as “current branch”, “this PR”, or “the diff against main” is not itself review evidence because this Skill has no Bash permission to resolve Git state. Stop and request patch text or an explicit readable file/path set. Mention the Codex review-only profile only as an optional alternative when the user explicitly asks for external Codex review; never switch profiles automatically.

Outputs

• MODE=full: complete review output at the selected LEVEL.
• MODE=findings-only: stop after prioritized findings at the selected LEVEL.
• standard and strict use severity buckets through Critical, Major, Minor.
• Directional suggestions only.

Workflow

1. Parse and materialize scope, LEVEL, MODE, and REVIEW_PROFILE; stop on semantic branch/diff labels that do not include patch text or readable paths.
2. Keep the selected identity as canonical nova-review; apply depth and output shape as structured parameters.
3. Emit findings with impact rationale, stopping after findings when MODE=findings-only.

Examples

• Natural trigger: Use review on this core module change.
• Explicit trigger: review LEVEL=standard INPUT="inventory service diff".

Safety

• No implementation patches.
• Clearly label facts vs assumptions.

General Review Guidance

Review only materialized evidence in the resolved scope. Stop on semantic Git labels without patch text or readable paths. Distinguish evidence, inference, and assumptions; do not modify code, redesign, broaden scope, or invoke an external Codex profile unless the user explicitly selected it.

Depth expands from high-signal correctness and test gaps (lite), through complexity, performance, concurrency, failure modes, and maintainability (standard), to security, integrity, boundaries, evolution, and operational resilience (strict). Full output groups Critical, Major, then Minor and states evidence, impact, and conceptual direction.

LEVEL=lite selects the bounded review. /nova-plugin:review-only remains the direct LEVEL=standard MODE=findings-only compatibility entrypoint; automatic routing must not select the alias. Codex profiles require their complete resolved contract plus separate shell, network, and authentication approval. Canonical read-only Skill frontmatter does not authorize that external runtime. All variants remain non-implementation workflows.