review-orchestrator (審査)

Unified skill for selecting review types, spawning multi-perspective and cognitive review agents, and managing quality gates. Consolidates 5 granular skills into a single review system.

Trigger: レビュー要求 (review requested)

Source skills: twin-review, cognitive-review, review-selector, staged-quality-gate, prompt-normalizer

Installation

openclaw install leegitw/review-orchestrator

Dependencies:

leegitw/failure-memory (for context)
leegitw/context-verifier (for file verification)

# Install with dependencies
openclaw install leegitw/context-verifier
openclaw install leegitw/failure-memory
openclaw install leegitw/review-orchestrator

Standalone usage: Review orchestration works independently for multi-perspective reviews. Integration with failure-memory enables automatic observation recording from review findings.

Data handling: This skill operates within your agent's trust boundary. When triggered, it uses your agent's configured model for multi-perspective review orchestration. No external APIs or third-party services are called. Review results are written to docs/reviews/ in your workspace.

What This Solves

One perspective has blind spots. This skill coordinates multiple review perspectives to catch what single-pass review misses:

Twin review — technical and creative perspectives for balance
Cognitive modes — analyzer ("what conflicts"), architect ("how to restructure"), implementer ("how to implement")

The insight: N=2 catches more than N=1. Different perspectives see different things. Coordinate them systematically.

Note: "Cognitive modes" are review perspectives with different analytical focus, not external API calls. Mode names (analyzer, architect, implementer) describe the review approach, not specific AI models or services.

Usage

/ro <sub-command> [arguments]

Sub-Commands

Command	CJK	Logic	Trigger
`/ro select`	選択	context×risk→type∈{twin,cognitive,code}	Explicit
`/ro twin`	双子	spawn(technical,creative)→findings[]	Explicit
`/ro cognitive`	認知	spawn(modes[])→analysis[]	Explicit
`/ro multi`	双視	alias for `/ro twin` (multi-perspective review)	Explicit
`/ro gate`	門番	staged_work→pass✓∨block✗	Explicit

Arguments

/ro select

Argument	Required	Description
context	Yes	Description of work to review
--risk	No	Risk level: `low`, `medium`, `high` (auto-detected if omitted)

/ro twin

Argument	Required	Description
target	Yes	File path(s) or topic to review
--technical-only	No	Skip creative perspective
--creative-only	No	Skip technical perspective

/ro cognitive

Argument	Required	Description
target	Yes	File path(s) or topic to review
--modes	No	Cognitive modes: `analyzer`, `architect`, `implementer` (default: all)

/ro multi

Alias for /ro twin. The name "twin" refers to the dual-perspective review pattern (technical + creative), not a specific team structure. /ro multi is provided for discoverability by users unfamiliar with the "twin" terminology.

/ro gate

Argument	Required	Description
stage	Yes	Stage name or number to gate
--strict	No	Require all checks pass (default: true)
--allow-minor	No	Allow minor issues to pass

Core Logic

Review Type Selection

Context	Risk	Recommended Review
Implementation	Low	`/ro twin --technical-only`
Implementation	Medium	`/ro twin` (both perspectives)
Implementation	High	`/ro twin` + `/ro cognitive`
Architecture	Any	`/ro cognitive`
Documentation	Any	`/ro twin --creative-only`
Security	Any	`/ro cognitive` + external review

Multi-Perspective Review

Perspective	Focus	CJK
Technical	Architecture, standards, patterns, security	技術
Creative	UX, communication, philosophy alignment	創造

Cognitive Modes

Cognitive modes provide different analytical perspectives. Modes are configurable; defaults shown below.

Mode	Perspective	Focus	CJK
analyzer	"Here's what conflicts"	Tensions, trade-offs, contradictions	審碼
architect	"Here's how to restructure"	Architecture, patterns, organization	審構
implementer	"Here's how to implement"	Concrete steps, complexity, path forward	審実

Note: Mode names are perspective-based, not model-specific. The underlying model used for each mode is configurable (see Configuration section below).

Configuration

Configuration is loaded from (in order of precedence):

.openclaw/review-orchestrator.yaml (OpenClaw standard)
.claude/review-orchestrator.yaml (Claude Code compatibility)
Defaults (built-in)

Cognitive Mode Interface

Each cognitive mode implements this interface:

Field	Type	Required	Description
id	string	Yes	Mode identifier (e.g., "analyzer", "architect")
perspective	string	Yes	Human-readable perspective description
prompt_prefix	string	Yes	Prompt prefix for this mode's analysis
model_hint	string	No	Optional model preference (not enforced)

Cognitive Mode Configuration

# .openclaw/review-orchestrator.yaml
cognitive_modes:
  - id: analyzer
    perspective: "Here's what conflicts"
    prompt_prefix: "Analyze tensions, trade-offs, and conflicts in..."
    model_hint: "prefer analytical model"
  - id: architect
    perspective: "Here's how to restructure"
    prompt_prefix: "Suggest architectural improvements for..."
    model_hint: "prefer architectural model"
  - id: implementer
    perspective: "Here's how to implement"
    prompt_prefix: "Provide implementation guidance for..."
    model_hint: "prefer practical model"

Quality Gate Configuration

# .openclaw/review-orchestrator.yaml
quality_gates:
  test_command: "npm test"    # Node.js (default)
  # test_command: "go test ./..."  # Go
  # test_command: "pytest"         # Python
  # test_command: "cargo test"     # Rust
  coverage_threshold: 5       # Max allowed coverage drop (%)
  require_docs: true          # Require documentation updates

Quality Gate Checks

Check	Condition	Severity
Tests pass	`{test_command}` exit 0	Critical
Coverage maintained	delta ≤ `{coverage_threshold}`%	Important
No critical findings	review.critical == 0	Critical
Docs updated	changed files have docs (if `require_docs`)	Minor

Checks use configured values from quality_gates section. Defaults: test_command=npm test, coverage_threshold=5, require_docs=true.

Output

/ro select output

[REVIEW SELECTION]
Context: "Refactoring authentication handler"
Risk: Medium (auto-detected: changes auth code)

Recommended: /ro twin
Rationale: Medium-risk implementation benefits from both technical and creative perspectives.

Alternative: /ro cognitive (for deeper architectural analysis)

/ro twin output

[MULTI-PERSPECTIVE REVIEW INITIATED]
Target: src/handlers/auth.go
Perspectives: technical, creative

--- Technical Perspective Findings ---
Severity: important
- I-1: Missing error handling on line 45
- I-2: Consider extracting validation logic

Severity: minor
- M-1: Inconsistent naming convention

--- Creative Perspective Findings ---
Severity: minor
- M-1: Error messages could be more user-friendly
- M-2: Consider adding debug logging for operators

Verdict: Approved with conditions

/ro cognitive output

[COGNITIVE REVIEW INITIATED]
Target: docs/architecture/auth-flow.md
Modes: analyzer, architect, implementer

--- Analyzer Perspective (Conflicts) ---
- Tension between security and usability in token refresh
- Trade-off: session duration vs re-auth frequency

--- Architect Perspective (Restructure) ---
- Suggested: Extract token service from handler
- Benefit: Cleaner separation of concerns

--- Implementer Perspective (Implement) ---
- Implementation path: 3 stages
- Estimated complexity: Medium

Verdict: Approved

/ro gate output (pass)

[QUALITY GATE: Stage 2]
Status: ✓ PASSED

Checks:
- ✓ Tests pass (exit 0)
- ✓ Coverage maintained (delta: -1.2%)
- ✓ No critical findings
- ✓ Documentation updated

Proceed to Stage 3.

/ro gate output (block)

[QUALITY GATE: Stage 2]
Status: ✗ BLOCKED

Checks:
- ✓ Tests pass
- ✗ Coverage dropped (delta: -8.3%, threshold: 5%)
- ✓ No critical findings
- ✗ Documentation not updated

Action required before proceeding:
1. Add tests to restore coverage
2. Update docs/handlers/README.md

Example: API Design Review

/ro cognitive api/openapi.yaml --modes analyzer,architect
[COGNITIVE REVIEW INITIATED]
Target: api/openapi.yaml
Modes: analyzer, architect

--- Analyzer Perspective (Conflicts) ---
- Tension: Versioning strategy (URL vs header) inconsistent across endpoints
- Trade-off: Pagination style (cursor vs offset) mixed usage

--- Architect Perspective (Restructure) ---
- Suggested: Standardize on cursor pagination for large collections
- Benefit: Consistent client SDK, better performance at scale

Verdict: Approved with conditions

Example: Performance Review

/ro twin src/handlers/search.go --technical-only
[MULTI-PERSPECTIVE REVIEW INITIATED]
Target: src/handlers/search.go
Perspectives: technical

--- Technical Perspective Findings ---
Severity: important
- I-1: N+1 query pattern on line 78 (database calls in loop)
- I-2: Missing index on search_terms table

Severity: minor
- M-1: Consider caching frequent searches (>100 req/min)

Verdict: Approved with conditions

Integration

Layer: Review
Depends on: failure-memory (for context), context-verifier (for file verification)
Used by: governance (for constraint reviews)

Failure Modes

Condition	Behavior
Invalid sub-command	List available sub-commands
Target not found	Error: "Target not found: {path}"
Agent spawn failed	Retry once, then report failure
Gate check failed	Block with specific remediation steps

Next Steps

After invoking this skill:

Condition	Action
Review complete	Write findings to `docs/reviews/`
Gate passed	Proceed to next stage
Gate blocked	Address findings, re-run gate
Critical findings	Create observation via `/fm record`

Workspace Files

This skill reads/writes:

docs/reviews/
├── [date]-[topic]-twin-technical.md
├── [date]-[topic]-twin-creative.md
├── [date]-[topic]-cognitive.md
└── [date]-[topic]-gate-results.md

Security Considerations

What this skill accesses:

Configuration files in .openclaw/review-orchestrator.yaml and .claude/review-orchestrator.yaml
Target files specified by user for review (read-only)
Its own output directory docs/reviews/

What this skill does NOT access:

Files outside declared workspace paths
System environment variables
Network resources or external APIs
Other tools' configuration

What this skill does NOT do:

Invoke AI models directly (instruction-only skill)
Call external services (Codex, Gemini, or any third-party API)
Send data to external services
Modify files outside its workspace (only writes to docs/reviews/)
Execute arbitrary code

Cognitive modes clarification: Mode names like "analyzer", "architect", and "implementer" describe review perspectives (analytical approaches), NOT external AI services. The skill provides prompts and structure; your agent executes the review using its configured model.

Configuration files: The config files (.openclaw/review-orchestrator.yaml and .claude/review-orchestrator.yaml) contain only local behavior settings (cognitive mode definitions, quality gate thresholds). They do NOT contain API keys, tokens, or external service endpoints by design.

Dependency clarification: References to failure-memory and context-verifier are skill-level dependencies installed via openclaw install. These are separate instruction-only skills with their own declared permissions — this skill does not grant them additional access or inherit their permissions.

Provenance note: This skill is developed by Live Neon (https://github.com/live-neon/skills) and published to ClawHub under the leegitw account. Both refer to the same maintainer.

Acceptance Criteria

Consolidated from 5 skills as part of agentic skills consolidation (2026-02-15).