Final Agent Review

Execution Context: This skill runs directly in the PAW session (not a subagent), preserving user interactivity for apply/skip/discuss decisions.

Automated review step that runs after all implementation phases complete, before Final PR creation. Examines the full implementation diff against specification to catch issues before external review.

Capabilities

• Review implementation against spec for correctness, patterns, and issues
• Multi-model parallel review with synthesis (CLI only)
• Society-of-thought review via paw-sot engine with specialist personas, parallel/debate execution, and confidence-weighted synthesis (CLI only)
• Single-model review (CLI and VS Code)
• Interactive, smart, or auto-apply resolution modes
• Generate review artifacts in .paw/work/<work-id>/reviews/

Procedure

Step 1: Read Configuration

Read WorkflowContext.md for:

• Work ID and target branch
• Final Review Mode: single-model | multi-model | society-of-thought
• Final Review Interactive: true | false | smart
• Final Review Models: comma-separated model names (for multi-model)

{{#cli}} If mode is multi-model, parse the models list. Default: latest GPT, latest Gemini, latest Claude Opus.

If mode is society-of-thought, also read:

• Final Review Specialists: all (default) | comma-separated specialist names | adaptive:<N>
• Final Review Interaction Mode: parallel (default) | debate
• Final Review Specialist Models: none (default) | model pool | pinned pairs | mixed {{/cli}} {{#vscode}} Note: VS Code only supports single-model mode. If multi-model is configured, proceed with single-model using the current session's model. {{/vscode}}

Step 2: Gather Review Context

Required context (review subagents gather this themselves via tools):

• Implementation diff: git diff <base-branch>...<target-branch>
• Spec.md, ImplementationPlan.md, CodeResearch.md in .paw/work/<work-id>/

Step 3: Create Reviews Directory

Create .paw/work/<work-id>/reviews/ if it doesn't exist. Create .paw/work/<work-id>/reviews/.gitignore with content * (if not already present).

Review Prompt (shared)

Provide this to all review subagents (single-model or each multi-model subagent). Subagents have full tool access — they gather context themselves rather than receiving it inline.

Review this implementation against the specification. Be critical and thorough.

## Context Locations
- **Diff**: Run `git diff <base-branch>...<target-branch>` to see all changes
- **Specification**: Read `.paw/work/<work-id>/Spec.md`
- **Implementation Plan**: Read `.paw/work/<work-id>/ImplementationPlan.md`
- **Codebase Patterns**: Read `.paw/work/<work-id>/CodeResearch.md`

Start by gathering the diff and reading the spec, then review against these criteria:

## Review Criteria
1. **Correctness**: Do changes implement all spec requirements? Any gaps?
2. **Pattern Consistency**: Does implementation follow established codebase patterns?
3. **Bugs and Issues**: Logic errors, edge cases, race conditions, error handling gaps
4. **Token Efficiency**: For prompts/skills, opportunities to reduce verbosity
5. **Documentation**: Missing or outdated documentation

For each finding, provide:
- Issue description
- Current code/text
- Proposed fix
- Severity: must-fix | should-fix | consider

Write findings in structured markdown.

{{#cli}}

Step 4: Execute Review (CLI)

If single-model mode:

• Execute review using the prompt above
• Save to REVIEW.md

If multi-model mode:

Read the resolved model names from WorkflowContext.md. Log the models being used, then start immediately — models were already confirmed during paw-init.

Then spawn parallel subagents using task tool with model parameter for each model. Each subagent receives the review prompt above. Save per-model reviews to REVIEW-{MODEL}.md.

After multi-model reviews complete, generate synthesis.

Important: If any findings involve interface changes, API modifications, or data flow updates, populate the Verification Checklist with specific components that need coordinated updates. This prevents half-fixes where only one side of an interface is updated.

REVIEW-SYNTHESIS.md structure:

# Review Synthesis

**Date**: [date]
**Reviewers**: [model list]
**Changes**: [branch/diff reference]

## Consensus Issues (All Models Agree)
[Highest priority - all models flagged these]

## Partial Agreement (2+ Models)
[High priority - multiple models flagged]

## Single-Model Insights
[Unique findings worth considering]

## Verification Checklist
[Populate with specific touchpoints for interface/data-flow changes]
- [ ] [Component A] updated
- [ ] [Component B] updated  
- [ ] Data flows end-to-end from [source] → [target]

## Priority Actions
### Must Fix
[Critical issues]

### Should Fix
[High-value improvements]

### Consider
[Nice-to-haves]

If society-of-thought mode:

Load the paw-sot skill and invoke it with a review context constructed from WorkflowContext.md fields and implementation artifacts:

After paw-sot completes orchestration and synthesis, proceed to Step 5 (Resolution) to process the REVIEW-SYNTHESIS.md findings.

{{/cli}}

{{#vscode}}

Step 4: Execute Review (VS Code)

Note: VS Code only supports single-model mode. If multi-model is configured, report to user: "Multi-model not available in VS Code; running single-model review."

If society-of-thought is configured, report to user: "Society-of-thought requires CLI for specialist persona loading (see issue #240). Running single-model review."

Execute single-model review using the shared review prompt above. Save to REVIEW.md. {{/vscode}}

Step 5: Resolution

If no findings: Report clean review and proceed.

If Interactive = true:

Present each finding to user:

## Finding #N: [Title]

**Severity**: [must-fix | should-fix | consider]
**Issue**: [Description]

**Current**:
[Show current code/text]

**Proposed**:
[Show proposed change]

**My Opinion**: [Rationale]

---

**Your call**: apply, skip, or discuss?

Track status for each finding:

• applied - Change made to codebase
• skipped - User chose not to apply
• discussed - Modified based on discussion, then applied or skipped

{{#cli}} For multi-model mode, process synthesis first (consensus → partial → single-model). Track cross-finding duplicates to avoid re-presenting already-addressed issues.

For society-of-thought mode, process REVIEW-SYNTHESIS.md findings by severity (must-fix → should-fix → consider). Present trade-offs from the "Trade-offs Requiring Decision" section for user decision. Track cross-finding duplicates. {{/cli}}

If Interactive = smart:

{{#cli}} If Final Review Mode is single-model, smart degrades to interactive behavior (no synthesis to classify). Follow the Interactive = true flow.

If Final Review Mode is multi-model, classify each synthesis finding, then resolve in phases:

Classification heuristic (applied per finding):

Consensus agreement implies models converged on the fix — no per-model cross-referencing needed.

If Final Review Mode is society-of-thought, classify each REVIEW-SYNTHESIS.md finding:

Phase 1 — Auto-apply: Apply all auto-apply findings without user interaction. Display batch summary:

## Auto-Applied Findings (N items)

1. **[Title]** (must-fix) — [one-line description]
2. **[Title]** (should-fix) — [one-line description]
...

Phase 2 — Interactive: Present each interactive finding using the same format as Interactive = true (apply, skip, or discuss).

Phase 3 — Summary: Display final summary of all dispositions:

## Resolution Summary

**Auto-applied**: N findings (consensus fixes)
**User-applied**: N findings
**User-skipped**: N findings
**Reported only**: N findings (consider-severity)

If all findings are auto-apply or report-only, skip Phase 2. If all findings are interactive, skip Phase 1. {{/cli}} {{#vscode}} Smart mode degrades to interactive behavior in VS Code (single-model has no agreement signal). Follow the Interactive = true flow above. {{/vscode}}

If Interactive = false:

Auto-apply all findings marked must-fix and should-fix. Skip consider items. Report what was applied.

{{#cli}}

Moderator Mode (society-of-thought only)

After finding resolution completes, if Final Review Mode is society-of-thought and (Final Review Interactive is true, or smart with significant findings remaining), invoke paw-sot a second time for moderator mode — pass the review context type (same as orchestration invocation, i.e., diff), output_dir (containing individual REVIEW-{SPECIALIST}.md files and REVIEW-SYNTHESIS.md), and review coordinates (diff range, artifact paths).

The paw-sot moderator mode handles specialist summoning, finding challenges, and deeper analysis. See paw-sot skill for interaction patterns.

Exit: User says "done", "continue", or "proceed" to exit moderator mode and continue to paw-pr.

Skip condition: If Final Review Interactive is false, or no findings exist, skip moderator mode entirely. {{/cli}}

Step 6: Completion

Report back:

• Total findings count
• Applied / skipped / discussed counts
• Summary of key changes made
• Review artifacts location
• Status: complete (ready for paw-pr)

Edge cases:

• Empty diff → Report "no implementation changes to review", proceed to paw-pr
• All findings skipped → Proceed to paw-pr (user's choice respected)

Review Artifacts

Location: .paw/work/<work-id>/reviews/ All files gitignored via .gitignore with * pattern.