Agent Skills: Athena PR Reviewer

Athena PR Reviewer

Instructions

Progress Tracking

Use TodoWrite throughout the review to show progress. Initialize at start:

TodoWrite([
  {content: "Detect PR target", status: "in_progress", activeForm: "Detecting PR target"},
  {content: "Gather context", status: "pending", activeForm: "Gathering context"},
  {content: "Select reviewers", status: "pending", activeForm: "Selecting reviewers"},
  {content: "Run reviews", status: "pending", activeForm: "Running reviews"},
  {content: "Aggregate findings", status: "pending", activeForm: "Aggregating findings"},
  {content: "Verify findings", status: "pending", activeForm: "Verifying findings"},
  {content: "Present summary", status: "pending", activeForm: "Presenting summary"}
])

QUICK MODE uses a shorter todo list instead:

TodoWrite([
  {content: "Detect PR target", status: "in_progress", activeForm: "Detecting PR target"},
  {content: "Gather context", status: "pending", activeForm: "Gathering context"},
  {content: "Run: code-reviewer", status: "pending", activeForm: "Running code-reviewer"},
  {content: "Verify findings", status: "pending", activeForm: "Verifying findings"},
  {content: "Present summary", status: "pending", activeForm: "Presenting summary"}
])

Update todos as you complete each step. Mark current step in_progress, previous step completed.

1. Detect Mode & PR Target

Mode Detection: Check if the user's message contains "quick" (case-insensitive). If yes → QUICK MODE (single code-reviewer). If no → FULL MODE (existing pipeline). The word "all" or "full" always forces FULL MODE even if "quick" also appears.

Parse user input to identify the PR:

• Direct PR reference (PR 123, #123): Extract number directly
• Jira ticket (PROJ-123): Run gh pr list --search "PROJ-123" --json number --jq '.[0].number'
• Current branch: Run gh pr view --json number --jq '.number'
• No PR found: Extract Jira from branch with git branch --show-current | grep -oE '[A-Z]+-[0-9]+'

2. Gather Data (Script)

Run the gather-context script which collects all data in parallel:

~/.claude/skills/athena-pr-reviewer/scripts/gather-context.sh ${PR_NUM} ${JIRA_TICKET}

This script:

• Creates work directory at /tmp/athena-review-${PR_NUM}/
• Fetches in parallel: PR metadata, diff, Jira ticket, epic, CLAUDE.md guidelines, git blame, prior PR comments
• Writes combined context to ${WORK_DIR}/context.md
• Writes diff to ${WORK_DIR}/diff.patch

Output files:

• context.md - Combined PR + Jira + guidelines + history data
• diff.patch - Full PR diff
• pr.json - Raw PR metadata
• jira.json - Raw Jira ticket data
• epic.json - Epic context (if linked)
• guidelines.md - All CLAUDE.md files from repo
• blame.md - Git blame for changed files (who wrote what, when)
• prior-comments.md - Comments from past PRs touching same files

3. Detect and Select Reviewers

[QUICK MODE: Skip this step entirely. Proceed to Step 4 with code-reviewer only.]

Before running reviews, detect available reviewer agents and let the user select which to use.

3.1 Detect Available Reviewers

Built-in reviewers (always available):

• 6 Claude specialists: comment-analyzer, test-analyzer, error-hunter, type-reviewer, code-reviewer, simplifier
• External LLMs: Gemini, Codex (auto-detected by run-reviews.sh)

Dynamic reviewers: Check available subagent_type values in your context for additional reviewers:

1. Pattern match - Find agents where name or description contains "reviewer" or "review"
2. Exclude: athena-pr-reviewer (this skill), data-gathering agents (hermes-pr-courier, heimdall-pr-guardian, etc.)

3.2 Check for "all" Flag

If the user's original message includes "all" or "all reviewers" or "don't ask" or "no questions" (case-insensitive), skip the selection prompt entirely and proceed to step 4 with all detected reviewers. Do NOT call AskUserQuestion.

Otherwise, proceed to 3.3.

3.3 Confirm or Customize Reviewers

Show all detected reviewers and ask for confirmation:

I'll run the review with these agents:

**Claude specialists (6):** (built-in, always available)
- comment-analyzer, test-analyzer, error-hunter
- type-reviewer, code-reviewer, simplifier

**External LLMs (2):** (run outside Claude via CLI)
- Gemini, Codex

**Installed agents (N):** (detected in your Claude setup)
- {agent-1}, {agent-2}, ...
- (or "None detected" if empty)

Proceed with all {total} reviewers?

Use AskUserQuestion with:

• Yes, run all (Recommended) - Proceed with all detected reviewers
• No, let me choose - Show detailed selection UI

3.4 Handle Response

If "Yes, run all": Proceed to step 4 with all detected reviewers.

If "No, let me choose": Show paginated multi-select UI:

• Max 4 options per question, max 4 questions per call
• Each batch shows "All in this batch" + 3 actual reviewers
• Use multiSelect: true
• Group by category (Built-in specialists, External agents, Dynamic agents)

Parse selection:

• If "All in this batch" selected → include all reviewers from that batch
• Otherwise include only individually selected reviewers

4. Run Reviews (Selected Reviewers in Parallel)

[QUICK MODE: Spawn only ONE Task agent using prompts/code-reviewer.md. No external LLMs (Gemini/Codex). No dynamic agents. Skip 4.1, 4.3. Only run 4.2 with code-reviewer. Then proceed to 4.4 to wait for it.]

Execute all selected reviews simultaneously in a SINGLE message.

First, update todos to show individual reviewers:

Replace the "Run reviews" todo with one todo per selected reviewer:

TodoWrite([
  {content: "Detect PR target", status: "completed", activeForm: "Detecting PR target"},
  {content: "Gather context", status: "completed", activeForm: "Gathering context"},
  {content: "Select reviewers", status: "completed", activeForm: "Selecting reviewers"},
  // One todo per selected reviewer:
  {content: "Review: comment-analyzer", status: "pending", activeForm: "Running comment-analyzer"},
  {content: "Review: test-analyzer", status: "pending", activeForm: "Running test-analyzer"},
  {content: "Review: error-hunter", status: "pending", activeForm: "Running error-hunter"},
  // ... etc for all selected reviewers
  {content: "Aggregate findings", status: "pending", activeForm: "Aggregating findings"},
  {content: "Verify findings", status: "pending", activeForm: "Verifying findings"},
  {content: "Present summary", status: "pending", activeForm: "Presenting summary"}
])

In ONE message, run all selected reviewers:

4.1 External LLMs (if selected)

If Gemini or Codex were selected, start with run_in_background: true:

~/.claude/skills/athena-pr-reviewer/scripts/run-reviews.sh ${WORK_DIR}

4.2 Built-in Claude Specialists (if selected)

For each selected built-in specialist, spawn a Task agent:

| Specialist | Prompt File | Output File | |------------|-------------|-------------| | comment-analyzer | prompts/comment-analyzer.md | claude-comments.md | | test-analyzer | prompts/test-analyzer.md | claude-tests.md | | error-hunter | prompts/error-hunter.md | claude-errors.md | | type-reviewer | prompts/type-reviewer.md | claude-types.md | | code-reviewer | prompts/code-reviewer.md | claude-general.md | | simplifier | prompts/simplifier.md | claude-simplify.md |

Task: general-purpose
Prompt: "Read ~/.claude/skills/athena-pr-reviewer/prompts/{SPECIALIST}.md for instructions.
Then read ${WORK_DIR}/context.md and ${WORK_DIR}/diff.patch.
Perform the review and use the Write tool to save your findings to: ${WORK_DIR}/reviews/{OUTPUT_FILE}
IMPORTANT: Use the Write tool directly, not Bash with cat/heredoc."

Note: Each Task call returns a task_id. Save these - you will need them in step 4.4.

4.3 Dynamic Reviewer Agents (if selected)

For each selected dynamic agent, spawn using its own agent type:

Task: {agent-name}
Prompt: "Review the PR for code quality issues.

Context file: ${WORK_DIR}/context.md (contains requirements, PR metadata, guidelines)
Diff file: ${WORK_DIR}/diff.patch (annotated with line numbers)

Use your expertise to identify issues. For each finding include:
- File path and line number (use format from diff annotations)
- Severity: Critical/High/Medium/Low
- Confidence: 0-100
- Description and suggested fix

Use the Write tool to save your review to: ${WORK_DIR}/reviews/{agent-name}.md
IMPORTANT: Use the Write tool directly, not Bash with cat/heredoc."

Note: Each Task call returns a task_id. Save these - you will need them in step 4.4.

4.4 Wait for ALL Agents to Complete

STOP. Do NOT proceed until you complete this step.

When you spawned Task agents in steps 4.1-4.3, each returned a task_id. You MUST now:

1. List all task IDs you received from spawning agents
2. For EACH task_id, call TaskOutput:

   TaskOutput(task_id: "abc123", block: true)
   TaskOutput(task_id: "def456", block: true)
   TaskOutput(task_id: "ghi789", block: true)
   ... one call per spawned agent
   

3. Do NOT skip any - even if you see some agent output appear automatically
4. Update todos - Mark each reviewer as completed when its TaskOutput returns

FORBIDDEN:

• ❌ Proceeding when "most" agents are done
• ❌ Relying on automatic output appearing
• ❌ Skipping TaskOutput for agents that seem fast
• ❌ Moving to Step 5 before ALL TaskOutput calls return

REQUIRED:

• ✓ Call TaskOutput for EVERY spawned agent
• ✓ Use block: true on each call
• ✓ Mark reviewer todo as completed when TaskOutput returns
• ✓ Wait for each call to return before proceeding

Only after ALL TaskOutput calls have returned, proceed to Step 5.

5. Aggregate Reviews

[QUICK MODE: Skip aggregation. Read ${WORK_DIR}/reviews/claude-general.md directly. Proceed to Step 5.5.]

Read ALL review files from ${WORK_DIR}/reviews/ directory and combine findings.

Possible reviewers (depending on selection):

• External: gemini.md, codex.md
• Built-in: claude-comments.md, claude-tests.md, claude-errors.md, claude-types.md, claude-general.md, claude-simplify.md
• Dynamic: {agent-name}.md (any additional detected agents)

Confidence Filtering:

• Drop findings with confidence < 80
• Keep findings 50-79 only if flagged by 2+ reviewers

Priority Boost Rule: Items flagged by 2+ reviewers get bumped up one severity level.

| Reviewers | Original | Final Severity | |-----------|----------|----------------| | 3+ | High | Critical | | 2 | High | Critical | | 3+ | Medium | High | | 2 | Medium | High | | 1 | Any | No boost |

Deduplicate similar findings, noting which reviewer(s) flagged each and average confidence.

Theme Clustering: Group findings by type or concept similarity so related issues can be tackled together. Let themes emerge naturally from the findings. Within each theme, sort by severity (Critical → High → Medium).

5.5 Verify Findings

For each aggregated finding, verify against actual code to filter hallucinations:

1. Read ${WORK_DIR}/diff.patch to get the actual code
2. For each finding with file:line reference:
   • Extract the actual code at that location from the diff
   • Compare the finding's description to what the code actually does
3. Use the verifier prompt (~/.claude/skills/athena-pr-reviewer/prompts/verifier.md) to validate each finding
4. Filter based on verdict:
   • ✓ VERIFIED → Keep in final output
   • ✗ REJECTED → Write to ${WORK_DIR}/rejected.md with reason
   • ⚠️ PARTIAL → Keep but move to "Suggestions" section

Output verified findings to ${WORK_DIR}/verified-findings.md

6. Synthesize Actionable Items

QUICK MODE uses a lighter template:

# Quick Review: PR #{PR_NUM} — {PR_TITLE}

## Findings (Verified)
- [ ] **Critical** file:line - issue - fix (92%) ✓
- [ ] **High** file:line - issue - fix (88%) ✓

### Suggestions
- partial/low-confidence items

## Recommendation: APPROVE / REQUEST_CHANGES

No requirements table, no reviewer attribution, no review sources section. Proceed to Step 7.

FULL MODE template:

# PR Review: {PR_TITLE} (#{PR_NUM})

## Requirements Status
| Requirement | Status | Notes |
|-------------|--------|-------|

## Action Items (Verified)

<!-- Group by theme (type/concept similarity), priority shown inline -->

### {Theme 1}
- [ ] **Critical** file:line - issue - fix [reviewer1 + reviewer2] (92%) ✓
- [ ] **High** file:line - issue - fix [reviewer1] (88%) ✓

### {Theme 2}
- [ ] **High** file:line - issue - fix [reviewer1 + reviewer2] (85%) ✓
- [ ] **Medium** file:line - issue - fix [reviewer1] (82%) ✓

### {Theme N}
- [ ] ...

### Suggestions
- improvements (including PARTIAL findings downgraded from higher severity)

## Rejected Findings
Findings that failed verification are saved to: `${WORK_DIR}/rejected.md`

## Review Sources
[List all .md files found in ${WORK_DIR}/reviews/]

## Recommendation: APPROVE / REQUEST_CHANGES

7. Offer to Address Findings

After presenting the summary, offer the user a choice using AskUserQuestion:

How would you like to address the findings?

Options:

• Walk through one by one - I'll present each item and you decide fix or skip
• Auto-fix all (Recommended) - I'll fix everything that doesn't require a product decision and report back
• I'm done - Just keep the review summary, no fixes

7a. Walk Through (one by one)

Replace todos with themes:

TodoWrite([
  {content: "Theme 1 (N issues)", status: "in_progress", activeForm: "Addressing Theme 1"},
  {content: "Theme 2 (N issues)", status: "pending", activeForm: "Addressing Theme 2"},
  // ... one todo per theme
])

Work through ONE THEME at a time:

For the current theme, go through each item:

1. Present the issue with code snippet
2. Explain the problem and share your opinion
3. Propose a solution
4. Wait for user decision: "fix" or "skip"
5. If fix: implement the change immediately, then commit
6. Move to next item in theme

Once all items in a theme are addressed (fixed or skipped), mark theme completed and present the next theme.

CRITICAL: Do NOT accumulate changes. Each item is either implemented+committed or discarded before moving on.

7b. Auto-fix All

Autonomously fix all findings that are purely technical (no product decisions needed).

First, classify each finding:

• Auto-fixable: Bug fixes, type fixes, error handling improvements, test additions, code simplification, comment corrections — anything where the correct fix is unambiguous from the code context
• Needs product decision: Changes that affect user-facing behavior, API contracts, feature scope, data models, or where multiple valid approaches exist and the "right" one depends on product intent

Then create a todo list with ALL findings:

TodoWrite([
  {content: "Fix: [finding description] (file:line)", status: "pending", activeForm: "Fixing [short desc]"},
  {content: "Fix: [finding description] (file:line)", status: "pending", activeForm: "Fixing [short desc]"},
  {content: "PRODUCT: [finding description] (file:line)", status: "pending", activeForm: "..."},
  // ... one todo per finding, prefixed with "Fix:" or "PRODUCT:"
])

Work through all auto-fixable items:

1. Mark item in_progress
2. Implement the fix
3. Mark item completed
4. Move to next item (do NOT wait for user input between items)

Skip all PRODUCT items — leave them as pending.

After all auto-fixable items are done, present a summary:

## Auto-fix Summary

### Completed ({N} items)
- [x] file:line - what was fixed
- [x] file:line - what was fixed
...

### Needs Product Decision ({M} items)
For each item, explain:
- **Issue**: what the finding is
- **Why it needs your input**: what decision is required
- **Options**: the possible approaches (if applicable)

Would you like to address the product decisions now, or commit what we have?

If the user wants to address product decisions: Walk through them one by one (same flow as 7a but only for the remaining PRODUCT items).

7c. I'm Done

Do nothing further. The review summary stands as-is.

Examples

User: "Review PR 456"

1. Detect PR 456, find linked Jira ticket
2. Gather context via script (parallel CLI calls)
3. Detect available reviewers → ask to confirm or customize
4. Run selected reviews in parallel
5. Aggregate findings, boost items flagged by 2+ reviewers
6. Verify findings against actual diff (filter hallucinations)
7. Present verified actionable summary
8. Offer: walk through, auto-fix, or done

User: "Review PR 456 with all reviewers"

1. Detect PR 456, find linked Jira ticket
2. Gather context via script
3. "all" keyword detected → skip reviewer selection, use all
4. Run all reviews in parallel
5. Full review workflow

User: "Review CSD-123"

1. Find PR linked to CSD-123
2. Gather context including acceptance criteria
3. Present reviewer selection (may include custom security-reviewer agent if installed)
4. Run selected reviews in parallel
5. Present findings with reviewer attribution
6. User chooses auto-fix → agent fixes technical items, reports product questions

User: "Review this branch"

1. Get PR from current branch
2. Extract Jira from branch name if needed
3. Detect all available reviewers
4. User selects reviewers via paginated UI
5. Full review workflow with selected reviewers

User: "Quick review PR 456"

1. "quick" detected → QUICK MODE
2. Detect PR 456, gather context via script
3. Skip reviewer selection
4. Run code-reviewer only (single agent)
5. Skip aggregation, verify findings
6. Present lighter summary template
7. Offer to address findings

User: "Quick check this branch"

1. "quick" detected → QUICK MODE
2. Get PR from current branch, gather context
3. Single code-reviewer pass
4. Lighter output

User: "Quick review PR 456 with all reviewers"

1. "all" overrides "quick" → FULL MODE
2. Full pipeline with all reviewers