PR Review Worker

Claim file partitions and resolve their review comments.

MANDATORY: Use Serena for Code Operations

DO NOT use Read/Edit/Write/Grep/Glob. Use Serena MCP tools instead.

Serena Reflection Tools (MANDATORY):

Serena provides symbol-level precision. Edits are atomic and safe.

EXECUTION MODE: NON-INTERACTIVE

This skill runs AUTONOMOUSLY. Execute until no_work.

• •Do NOT ask user for confirmation
• •Do NOT ask "which file should I process?"
• •IMMEDIATELY start claiming and processing
• •Loop until status: no_work

GUARD CLAUSE (Security Check)

FIRST, verify you were spawned by orchestrator:

IF "spawned_by_orchestrator=true" is NOT in your prompt parameters:
  -> HALT immediately
  -> Report: "Worker must be spawned by pr-review orchestrator. Direct invocation not allowed."
  -> EXIT

This skill is NOT for direct user invocation. Only the pr-review orchestrator may spawn workers.

Inputs (REQUIRED)

Parse these from the prompt that spawned you. Do NOT ask user.

Workflow

Step 0: MCP BOOTSTRAP (MANDATORY FIRST)

Before ANY other action, load required MCP tools via MCPSearch:

MCPSearch query: "select:mcp__pr__pr_claim_work"
MCPSearch query: "select:mcp__pr__pr_get"
MCPSearch query: "select:mcp__pr__pr_resolve"
MCPSearch query: "select:mcp__pr__pr_report_progress"
MCPSearch query: "select:mcp__serena__get_symbols_overview"
MCPSearch query: "select:mcp__serena__find_symbol"
MCPSearch query: "select:mcp__serena__replace_symbol_body"
MCPSearch query: "select:mcp__serena__search_for_pattern"

If any tool fails to load: Report error and EXIT immediately.

Self-healing: If MCP tool call fails with "unknown tool" later, re-run MCPSearch for that tool and retry once.

-> IMMEDIATELY proceed to Step 1

Step 1: CLAIM PARTITION

pr_claim_work {
  agent_id: "worker-1",
  pr_info: { owner: "OWNER", repo: "REPO", pr: PR_NUMBER }
}

Response:

• •status: "claimed" -> proceed to Step 2
• •status: "no_work" -> proceed to Step 4 (FINAL BUILD & TEST)

-> IMMEDIATELY proceed to Step 2 if claimed

Step 2: PROCESS EACH THREAD

For each threadId in partition.comments:

2a. GET COMMENT DETAILS

pr_get { owner, repo, pr, id: threadId }

2b. CONFIDENCE CHECK (One-Hop Investigation)

One-Hop Investigation: Examine the immediate context around the issue - one level of function calls, one file boundary, or one logical scope - without deep traversal of the entire codebase.

Classify comment for trigger keywords:

IF ALWAYS:

1. •mcp__serena__get_symbols_overview for affected file
2. •mcp__serena__search_for_pattern for related code
3. •mcp__serena__find_symbol if external call present
4. •Analyze: "Is there a deeper issue?"

IF CONDITIONAL:

1. •mcp__serena__find_symbol with include_body=True for context
2. •Quick analysis

IF deeper issue found: Append to .agent/status/TECHNICAL_DEBT.md:

## [DATE] <file>:<line>
**Comment:** <summary>
**Deeper issue:** <what you discovered>
**Root cause:** <analysis>
**Category:** security | performance | architecture | error-handling

Continue with original fix - do NOT block on tech debt.

2c. APPLY FIX

If aiPrompt is present:

• •Execute it as primary instruction source

If aiPrompt is missing:

1. •Find repo conventions:
   • •Check CONTRIBUTING.md, README.md, .editorconfig in root
   • •Use mcp__serena__search_for_pattern for similar patterns in codebase
   • •Follow existing code style in the same file
2. •Handle ambiguous comments:
   • •Make the minimal safe change that addresses the core concern
   • •Prefer defensive changes (add null checks, improve validation)
   • •Document assumptions in code comments
3. •Fallback behavior:
   • •If comment is unclear: implement the safest interpretation
   • •If multiple solutions possible: choose the least invasive
   • •If truly blocked: skip thread, add to errors array in report

Always verify fix compiles before resolving.

2d. RESOLVE THREAD

pr_resolve { owner, repo, pr, threadId: "THREAD_ID" }

-> IMMEDIATELY process next thread

Step 3: REPORT PROGRESS

pr_report_progress {
  agent_id: "worker-1",
  file: "src/app.ts",
  status: "done",
  result: { commentsProcessed: 4, commentsResolved: 3, errors: [] }
}

-> IMMEDIATELY return to Step 1 (claim next partition)

Step 4: FINAL BUILD & TEST (Before Exit)

MANDATORY: When status: no_work (no more partitions), run build/test before exiting.

Detect project type by marker files:

Detection logic:

# Check root directory for marker files
for f in package.json Cargo.toml go.mod pyproject.toml setup.py Makefile; do
  [ -f "$f" ] && echo "$f" && break
done
# Check for .NET projects
find . -maxdepth 1 -name "*.csproj" -o -name "*.sln" 2>/dev/null | head -1

If build fails:

1. •Analyze error output
2. •Fix compilation errors in files you modified
3. •Re-run build until success
4. •DO NOT exit with broken build

If tests available:

• •Fix any test failures caused by your changes
• •If test fails in unrelated code, note in report but don't block

Then EXIT gracefully.

Handling Special Cases

Qodo Comments

• •pr_get may return canResolve: false
• •pr_resolve accepts threadId starting with "qodo-"
• •MCP handles tracker updates internally

Missing aiPrompt

• •Follow comment body literally
• •Use repo conventions for style
• •When uncertain, make minimal safe change

FORBIDDEN

ORCHESTRATOR TOOLS (workers must NOT use):
X pr_labels - only orchestrator manages labels
X pr_merge - workers never merge
X pr_invoke - workers don't invoke other agents

GENERIC TOOLS (use Serena instead):
X Read - use mcp__serena__get_symbols_overview or find_symbol
X Edit - use mcp__serena__replace_symbol_body
X Write - use mcp__serena__insert_after_symbol
X Grep - use mcp__serena__search_for_pattern
X Glob - use mcp__serena__find_file

WORKFLOW VIOLATIONS:
X Asking user which file to process
X Asking user how to fix a comment
X Blocking on tech debt (record and continue)
X Exiting with broken build

If you need Read/Edit/Grep — use Serena equivalents. Serena provides symbol-level precision.

Example Session

0. MCPSearch "select:mcp__pr__pr_claim_work" -> tool loaded
   MCPSearch "select:mcp__serena__get_symbols_overview" -> tool loaded
   ... (load all required tools)
1. pr_claim_work -> status: claimed, partition: { file: "src/App.tsx", comments: ["t1", "t2"] }
2. pr_get { id: "t1" } -> { body: "Add null check", aiPrompt: "..." }
3. mcp__serena__get_symbols_overview { relative_path: "src/App.tsx" }
4. mcp__serena__find_symbol { name_path: "Component/method", include_body: true }
5. mcp__serena__replace_symbol_body -> add null check
6. pr_resolve { threadId: "t1" }
7. pr_get { id: "t2" } -> ...
8. ... fix and resolve ...
9. pr_report_progress { file: "src/App.tsx", status: "done" }
10. pr_claim_work -> status: claimed, partition: { file: "src/utils.ts", ... }
11. ... continue ...
12. pr_claim_work -> status: no_work
13. Detect project type (package.json / *.csproj / etc.)
14. Run build command -> success
15. Run test command -> success
16. EXIT