Skill Pr Review Implementation

PR Review Implementation Skill

Thin wrapper that validates inputs, determines whether this is a PR review response task (sources present) or a legacy PR description task (sources absent), and delegates to the appropriate agent. The critical difference from skill-pr-implementation is the status transition: this skill uses preflight pr_ready (not postflight pr_ready) to set the task to [PR READY], directing the user to manually post the response files.

Trigger Conditions

This skill activates when:

• Task type is "pr"
• /implement command targets a pr-type task
• The routing entry for implement.pr resolves to skill-pr-review-implementation

Dispatch Decision

After input validation, inspect the sources array in the task's state.json entry:

• Sources present and non-empty → PR review response workflow
  • Delegate to pr-review-implementation-agent
  • Agent composes pr-response.md and/or zulip-response.md
• Sources absent or empty → Legacy PR description workflow
  • Delegate to cslib-implementation-agent (same as skill-pr-implementation)
  • Agent composes pr-description.md from git diff

Execution Flow

Stage 1: Input Validation

Validate that:

1. task_number exists and resolves to an active project in state.json
2. task_type is "pr"

# Check task exists and get task_type
task_type=$(jq -r --argjson num "$task_number" \
  '.active_projects[] | select(.project_number == $num) | .task_type' \
  specs/state.json)

if [ "$task_type" != "pr" ]; then
  echo "Error: Task $task_number has type '$task_type', not 'pr'. skill-pr-review-implementation only handles pr tasks."
  exit 1
fi

Stage 1a: Dispatch Decision

Check for sources array in the task entry:

sources_count=$(jq --argjson num "$task_number" \
  '.active_projects[] | select(.project_number == $num) | .sources // [] | length' \
  specs/state.json)

if [ "$sources_count" -gt 0 ]; then
  workflow="review_response"
else
  workflow="pr_description"
fi

• workflow="review_response" → proceed with Stages 2-10 below
• workflow="pr_description" → delegate to cslib-implementation-agent (legacy path)

Legacy path: If workflow="pr_description", use the same delegation context and postflight as skill-pr-implementation. The agent should compose pr-description.md from the task description and git diff, and the skill should call postflight pr_ready at Stage 7.

Stage 2: Preflight Status Update

Update status to "implementing" BEFORE invoking subagent:

bash .claude/scripts/update-task-status.sh preflight "$task_number" implement "$session_id"

Stage 3: Create Postflight Marker

This skill previously wrote a bare, metadata-free marker via touch. Source skill-base.sh and derive padded_num/project_name (not previously derived anywhere in this skill), then follow @.claude/context/patterns/skill-preflight-flow.md's Stage 3 (marker creation) so this skill's marker carries the full Shape A payload, matching every other lifecycle skill:

source .claude/scripts/skill-base.sh
padded_num=$(printf "%03d" "$task_number")
project_name=$(jq -r --argjson num "$task_number" \
  '.active_projects[] | select(.project_number == $num) | .project_name' \
  specs/state.json)
skill_name="skill-pr-review-implementation"
operation="implement"
skill_create_postflight_marker "$padded_num" "$project_name" "$session_id" "$skill_name" "$operation"

Stage 3a: Read Artifact Number

Read next_artifact_number from state.json for naming summary files:

artifact_number=$(jq -r --argjson num "$task_number" \
  '.active_projects[] | select(.project_number == $num) | .next_artifact_number // 1' \
  specs/state.json)
artifact_number_padded=$(printf "%02d" "$artifact_number")

Stage 4: Prepare Delegation Context

Find the latest research report and plan files:

# Find latest report (pattern: reports/{NN}_pr-review-research.md or any .md report)
report_path=$(ls "specs/{NNN}_{SLUG}/reports/"*.md 2>/dev/null | sort -r | head -1)

# Find latest plan if it exists
plan_path=$(ls "specs/{NNN}_{SLUG}/plans/"*.md 2>/dev/null | sort -r | head -1)

Build the delegation JSON with PR-review-specific fields:

{
  "session_id": "{session_id}",
  "delegation_depth": 1,
  "delegation_path": ["orchestrator", "implement", "skill-pr-review-implementation"],
  "timeout": 3600,
  "task_context": {
    "task_number": N,
    "task_name": "{project_name}",
    "description": "{description}",
    "task_type": "pr"
  },
  "sources": [...],
  "report_path": "specs/{NNN}_{SLUG}/reports/{NN}_pr-review-research.md",
  "plan_path": "specs/{NNN}_{SLUG}/plans/{NN}_pr-review-impl-skill.md",
  "artifact_number": "01",
  "metadata_file_path": "specs/{NNN}_{SLUG}/.return-meta.json"
}

Extract sources from state.json:

sources=$(jq -c --argjson num "$task_number" \
  '.active_projects[] | select(.project_number == $num) | .sources // []' \
  specs/state.json)

Note: Use jq -c (compact output) for inline JSON in delegation context. Use the safe select(.project_number == $num) pattern (not !=) per jq-escaping-workarounds.md.

Stage 4a: Memory Retrieval (Optional)

If --clean flag is NOT set, retrieve relevant memories:

memory_context=$(bash .claude/scripts/memory-retrieve.sh "pr review response implementation" 2>/dev/null || echo "")

Include memory_context in the delegation context if non-empty.

Stage 5: Invoke Subagent

Use the Agent tool with:

• subagent_type: "pr-review-implementation-agent"
• Prompt: the prepared delegation JSON

The subagent will:

1. Write early metadata to .return-meta.json
2. Read the research report and extract reviewer feedback
3. Apply minor code changes if any are requested
4. Compose pr-response.md and/or zulip-response.md
5. Write final metadata to .return-meta.json
6. Return brief text summary

Stage 5b: Self-Execution Fallback

CRITICAL: If you performed the work above WITHOUT using the Agent tool (i.e., you read files, wrote artifacts, or updated metadata directly instead of spawning a subagent), you MUST write a .return-meta.json file now before proceeding to postflight. Use the schema from return-metadata-file.md with status: "implemented" and include the response file artifacts.

If you DID use the Agent tool, skip this stage -- the subagent already wrote the metadata.

Postflight (ALWAYS EXECUTE)

The following stages MUST execute after work is complete, whether the work was done by a subagent or inline (Stage 5b). Do NOT skip these stages for any reason.

Stage 6: Parse Subagent Return

Read the metadata file:

cat "specs/{NNN}_{SLUG}/.return-meta.json"

Check status field:

• "implemented" → success, proceed normally
• "partial" → partial success, note in status update
• "failed" → failure, update status accordingly

Stage 6a: Validate Artifact Content (Non-Blocking)

# Check which response files were created
if [ -f "specs/{NNN}_{SLUG}/pr-response.md" ] && [ -s "specs/{NNN}_{SLUG}/pr-response.md" ]; then
  echo "pr-response.md validated"
else
  echo "Warning: pr-response.md missing or empty"
fi

if [ -f "specs/{NNN}_{SLUG}/zulip-response.md" ] && [ -s "specs/{NNN}_{SLUG}/zulip-response.md" ]; then
  echo "zulip-response.md validated"
fi

Stage 7: Update Task Status (PR READY)

CRITICAL: Use preflight pr_ready to transition the task to [PR READY].

bash .claude/scripts/update-task-status.sh preflight "$task_number" pr_ready "$session_id"

Why preflight pr_ready and NOT postflight pr_ready:

• postflight pr_ready would mark the task [COMPLETED] (the postflight for pr_ready is "completed")
• preflight pr_ready sets the task to [PR READY] without advancing to completed
• The task stays [PR READY] until the user manually posts the response files and runs /pr

Do NOT use postflight implement (that would set [COMPLETED]), and do NOT use postflight pr_ready.

Stage 7a: Propagate Memory Candidates

If .return-meta.json contains memory_candidates array with entries, log them for later harvest by /learn --task N:

memory_candidates=$(jq -c '.memory_candidates // []' "specs/{NNN}_{SLUG}/.return-meta.json")
if [ "$memory_candidates" != "[]" ]; then
  echo "Memory candidates available: $memory_candidates"
fi

Stage 8: Link Artifacts in state.json

Add response file artifacts to state.json and regenerate TODO.md.

For each response file created, add an artifact entry. Use pr_response for pr-response.md and zulip_response for zulip-response.md:

# Read which artifacts were created from .return-meta.json
artifacts=$(jq -c '.artifacts // []' "specs/{NNN}_{SLUG}/.return-meta.json")

# Add each artifact to state.json
# Example for pr-response.md:
bash .claude/scripts/state-write.sh \
  '.active_projects |= map(if .project_number == $num then
     . + {"artifacts": ((.artifacts // []) + [{"type": "pr_response", "path": $path, "summary": $summary}]),
          "next_artifact_number": ((.next_artifact_number // 1) + 1)}
     else . end)' \
  --session-id "$session_id" \
  --argjson num "$task_number" \
  --arg path "specs/{NNN}_{SLUG}/pr-response.md" \
  --arg summary "GitHub PR comment response for review task" \
  --regen-todo

Stage 8a: TTS Lifecycle Notification

bash .claude/scripts/lifecycle-notify.sh "implement" "$task_number" "pr_ready" 2>/dev/null || true

Stage 9: Cleanup Marker Files

Follow @.claude/context/patterns/skill-postflight-flow.md's Stage 9 (cleanup), reusing the padded_num/project_name derived at Stage 3:

skill_cleanup "$padded_num" "$project_name"

Stage 10: Return Brief Text Summary

Return 3-6 bullet points summarizing:

• Research report used (if any)
• Code changes applied (count)
• Response files created (pr-response.md, zulip-response.md)
• Task status transition (now [PR READY])
• Guidance to manually post response files

Example:

- Read PR review research report (7 requested changes, 3 open questions)
- Applied 2 minor code changes (typo fix, import ordering)
- Created pr-response.md with responses to 3 reviewers (5 comments addressed)
- Created zulip-response.md for stream 'cslib' topic 'PR Review: GroupAlgebra'
- Task transitioned to [PR READY]
- Next: Post pr-response.md as GitHub PR comment, send zulip-response.md to Zulip

Do NOT return JSON.

MUST NOT (Postflight Boundary)

After the agent returns, this skill MUST NOT:

1. Edit .lean files - All CSLib proof work is done by agent
2. Run lake build/test/lint - Verification is done by agent
3. Use lean-lsp MCP tools - Domain tools are for agent use only
4. Grep for sorries - Debt analysis is agent work
5. Write pr-response.md or zulip-response.md - Artifact creation is agent work
6. Create feature branches or push to remote - Never in this skill
7. Post to GitHub or Zulip - User handles this manually
8. Call postflight pr_ready - This skill MUST use preflight pr_ready to set [PR READY]
9. Call postflight implement - That would incorrectly set [COMPLETED]

PROHIBITION: If the subagent returned partial or failed status, the lead skill MUST NOT attempt to continue, complete, or "fill in" the subagent's work. Report the partial/failed status and let the user re-run /implement to resume.

The postflight phase is LIMITED TO:

• Reading agent metadata file
• Updating state.json via jq (using preflight pr_ready status target)
• Linking artifacts in state.json
• Regenerating TODO.md
• Lifecycle notification
• Cleanup of temp/marker files

Reference: @.claude/context/standards/postflight-tool-restrictions.md

Return Format

Brief text summary (NOT JSON).