unified-review

Unified Review Orchestration

Intelligently selects and executes appropriate review skills based on codebase analysis and context.

Quick Start

# Auto-detect and run appropriate reviews
/full-review

# Focus on specific areas
/full-review api          # API surface review
/full-review architecture # Architecture review
/full-review bugs         # Bug hunting
/full-review tests        # Test suite review
/full-review all          # Run all applicable skills

Verification: Run pytest -v to verify tests pass.

When To Use

• Starting a full code review
• Reviewing changes across multiple domains
• Need intelligent selection of review skills
• Want integrated reporting from multiple review types
• Before merging major feature branches

When NOT To Use

• Specific review type known
  • use bug-review
• Test-review
• Architecture-only focus - use architecture-review
• Specific review type known
  • use bug-review

Review Skill Selection Matrix

Codebase Pattern	Review Skills	Triggers
Rust files (*.rs, Cargo.toml)	rust-review, bug-review, api-review	Rust project detected
API changes (openapi.yaml, routes/)	api-review, architecture-review	Public API surfaces
Test files (test_*.py, *_test.go)	test-review, bug-review	Test infrastructure
Makefile/build system	makefile-review, architecture-review	Build complexity
Mathematical algorithms	math-review, bug-review	Numerical computation
Architecture docs/ADRs	architecture-review, api-review	System design
General code quality	bug-review, test-review	Default review
Post-implementation audit	imbue:justify	High add/delete ratio, test changes, new abstractions

Workflow

1. Analyze Repository Context

• Detect primary languages from extensions and manifests
• Analyze git status and diffs for change scope
• Identify project structure (monorepo, microservices, library)
• Detect build systems, testing frameworks, documentation

2. Select Review Skills

# Detection logic
if has_rust_files():
    schedule_skill("rust-review")
if has_api_changes():
    schedule_skill("api-review")
if has_test_files():
    schedule_skill("test-review")
if has_makefiles():
    schedule_skill("makefile-review")
if has_math_code():
    schedule_skill("math-review")
if has_architecture_changes():
    schedule_skill("architecture-review")
# Default
schedule_skill("bug-review")

Verification: Run pytest -v to verify tests pass.

3. Execute Reviews

Dispatch selected skills concurrently via the Agent tool. Use this mapping to resolve skill names to agent types:

Skill Name	Agent Type	Notes
bug-review	pensive:code-reviewer	Covers bugs, API, tests
api-review	pensive:code-reviewer	Same agent, API focus
test-review	pensive:code-reviewer	Same agent, test focus
architecture-review	pensive:architecture-reviewer	ADR compliance
rust-review	pensive:rust-auditor	Rust-specific
code-refinement	pensive:code-refiner	Duplication, quality
math-review	general-purpose	Prompt: invoke Skill(pensive:math-review)
makefile-review	general-purpose	Prompt: invoke Skill(pensive:makefile-review)
shell-review	general-purpose	Prompt: invoke Skill(pensive:shell-review)

Sub-agent isolation (required). One lens must not color how the next is read. Two ways to get that, and the second only runs when the user asks for it:

Path	How isolation holds
Agent tool	Dispatch ALL selected agents in a SINGLE parallel call, and read no output until every agent has returned. Reading the first result anchors synthesis toward it: each later result gets judged against the first rather than independently. Collect all, then synthesize once
/pensive:unified-review workflow (workflows/unified-review.js)	The script holds the results. It is not a reasoning entity, so it cannot be anchored by reading one stage before another, and findings pass between stages without entering anyone's context. Verification also starts per dimension instead of waiting for the slowest lens

Prefer the workflow when the dimension list is known before the work and each finding should face an adversarial check. Prefer the Agent tool when the roster has to adapt to what the first lens finds, or when no workflow was requested: a workflow never starts unasked.

Its subagents run in acceptEdits whatever the session's permission mode, so the shipped script scopes its prompts to reading. And it has no filesystem, so what it returns is a claim that survived refutation, not proof-of-work evidence. Reproduce before acting on a finding.

Rules:

• Never use skill names as agent types (e.g., pensive:math-review is NOT an agent)
• When pensive:code-reviewer covers multiple domains, dispatch once with combined scope
• For skills without dedicated agents, use general-purpose and instruct it to invoke the Skill tool
• Maintain consistent evidence logging across all agents
• Track progress via TodoWrite

4. Integrate Findings

• Consolidate findings across domains
• Identify cross-domain patterns
• Prioritize by impact and effort
• Generate unified action plan

Deferred capture for backlog findings: Findings that are triaged to the backlog (out-of-scope for the current review or deferred by the team) should be preserved so they are not lost between review cycles. For each finding assigned to the backlog, run:

python3 scripts/deferred_capture.py \
  --title "<finding title>" \
  --source review \
  --context "Review dimension: <dimension>. <finding description>"

The <dimension> value should match the review skill that surfaced the finding (e.g. bug-review, api-review, architecture-review). This runs automatically after the action plan is finalised, without prompting the user.

Review Modes

Auto-Detect (default)

Automatically selects skills based on codebase analysis.

Focused Mode

Run specific review domains:

• /full-review api → api-review only
• /full-review architecture → architecture-review only
• /full-review bugs → bug-review only
• /full-review tests → test-review only

Full Review Mode

Run all applicable review skills:

• /full-review all → Execute all detected skills

Quality Gates

Each review must:

1. Establish proper context
2. Execute all selected skills successfully
3. Document findings with evidence
4. Prioritize recommendations by impact
5. Create action plan with owners

Deliverables

Executive Summary

• Overall codebase health assessment
• Critical issues requiring immediate attention
• Review frequency recommendations

Domain-Specific Reports

• API surface analysis and consistency
• Architecture alignment with ADRs
• Test coverage gaps and improvements
• Bug analysis and security findings
• Performance and maintainability recommendations

Integrated Action Plan

• Prioritized remediation tasks
• Cross-domain dependencies
• Assigned owners and target dates
• Follow-up review schedule

Modular Architecture

All review skills use a hub-and-spoke architecture with progressive loading:

• Each skill has modules/: Domain-specific details loaded on demand
• Cross-plugin deps: imbue:proof-of-work, imbue:diff-analysis/modules/risk-assessment-framework

This reduces token usage by 50-70% for focused reviews while maintaining full capabilities.

Exit Criteria

• All selected review skills executed
• Findings consolidated and prioritized
• Action plan created with ownership
• Evidence logged per structured output format

Supporting Modules

• Review workflow core - standard 5-step workflow pattern for all pensive reviews
• Output format templates - finding entry, severity, action item templates
• Quality checklist patterns - pre-review, analysis, evidence, deliverable checklists

Troubleshooting

Common Issues

If the auto-detection fails to identify the correct review skills, explicitly specify the mode (e.g., /full-review rust instead of just /full-review). If integration fails, check that TodoWrite logs are accessible and that evidence files were correctly written by the individual skills.