Pepe39/.claude/skills/run

/run - Single entry point for code review labeling tasks

Creates new tasks or resumes existing ones from where they left off.

Arguments

• $ARGUMENTS (positional): Task ID, optionally followed by auto. E.g.: /run 2937204136 or /run 2937204136 auto
• auto mode (bypass): When the second token is auto, the skill runs the full pipeline non-interactively: do NOT ask the user anything, do NOT wait for confirmations, do NOT pause between steps. This mode is used when the API / extension invokes the skill via claude -p. Manual console runs (no auto) keep the interactive prompts.

Ground Rule

Task isolation: During task execution, focus only on the task (PR, diff, comment, labeling). Do not modify project docs, guides, commands, or templates. If you notice a pipeline issue, finish the task first, then report it.

Instructions

1. Locate task

Check if task exists: find tasks/ -maxdepth 2 -type d -name "{id}" 2>/dev/null (use only the task id, not the full $ARGUMENTS, since it may include auto).

• If find returns a path -> task EXISTS (go to 1a)
• If find returns nothing -> task DOES NOT exist (go to 2a)

IMPORTANT: Do NOT use glob with ls tasks/*/ID/ as quotes around the path prevent * expansion and cause false negatives. Always use find.

1a. Idempotency check (BEFORE doing anything else)

If the task dir was found, IMMEDIATELY check whether the four deliverable files already exist on disk and are non-empty:

DIR=<path returned by find>
ALL=1
for f in quality.md severity.md context_scope.md advanced.md; do
  if [ ! -s "$DIR/deliverables/$f" ]; then ALL=0; break; fi
done
[ $ALL -eq 1 ] && echo ALREADY_DONE

• If ALREADY_DONE is printed -> the task is already labeled. Do NOT touch any file. Do NOT bootstrap progress.md. Do NOT run any step. Print: Task {id} already labeled (deliverables present on disk). Skipping. and STOP immediately. The API reads the deliverables from disk; nothing else is needed.
• If any file is missing -> continue to 2b or 2c as normal.

Why: the API can re-invoke /run {id} auto for tasks that were produced in earlier sessions. Rerunning the pipeline would waste work and may produce different labels for an already-validated task. Idempotency is mandatory in auto mode AND in interactive mode.

2a. If task does NOT exist -> Create task

1. Get current date in YYYY-MM-DD format
2. Create structure:

   mkdir -p tasks/{date}/{id}/deliverables
   mkdir -p tasks/{date}/{id}/work
   

3. Generate tasks/{date}/{id}/inputs.md:

# Task Inputs

Paste your task variables below. Fill in each field from the annotation platform.

## Task Variables

- **pull_request_url:** (paste here)
- **nwo:** (paste here)
- **head_sha:** (paste here)
- **comment_id:** (paste here)
- **body:** (paste here)
- **file_path:** (paste here)
- **diff_line:** (paste here)
- **discussion_url:** (paste here)
- **repo_url:** (paste here)
- **coding_language:** (paste here)

4. Generate tasks/{date}/{id}/progress.md:

# Progress: {id}

**Current Step:** 01 - Parse Inputs
**Status:** pending
**Last Updated:** {timestamp ISO 8601}

| # | Step | Status | Started | Completed |
|---|------|--------|---------|-----------|
| 01 | Parse Inputs | pending | | |
| 02 | Analyze PR | pending | | |
| 03 | Analyze Comment | pending | | |
| 04 | Label Quality | pending | | |
| 05 | Label Severity | pending | | |
| 06 | Label Context Scope | pending | | |
| 07 | Label Advanced | pending | | |
| 08 | Generate Output | pending | | |

5. Generate tasks/{date}/{id}/task_info.md:

# Task: {id}

## Status
Created: {timestamp ISO 8601}

## Input Data
(populated after step 01)

## Analysis
(populated after steps 02-03)

## Labels
(populated after steps 04-07)

## Output
(populated after step 08)

6. Confirm: "Task {id} created at tasks/{date}/{id}/"
7. Interactive mode only: Tell the user "Paste your task variables into inputs.md and confirm when ready." then wait for confirmation.
8. Auto mode: assume inputs.md is already populated by the API; continue immediately to step 3 without waiting.

2b. If task exists but has NO progress.md -> Bootstrap

1. Create progress.md with all steps as "pending"
2. Continue to step 3

2c. If task exists and HAS progress.md -> Read it

Read tasks/{date}/{id}/progress.md. Extract Current Step and Status.

3. Show summary

== Task {id} ==
Steps completed: {N}/8
Next step: {N} - {name} ({status})

If status is "in-progress": "NOTE: This step was interrupted. It will be resumed."

4. Determine step to execute

Dispatch logic:

1. If Current Step status = "in-progress" -> resume that step (interrupted session)
2. If Current Step status = "pending" -> start that step
3. If Current Step status = "done" -> find the next pending step

If ALL steps are "done": Show "All steps completed. Task ready for submission." and stop.

Step to command mapping:

| Step | Command | |---|---| | 01 | step-01-parse-inputs | | 02 | step-02-analyze-pr | | 03 | step-03-analyze-comment | | 04 | step-04-label-quality | | 05 | step-05-label-severity | | 06 | step-06-label-context-scope | | 07 | step-07-label-advanced | | 08 | step-08-generate-output |

5. Execute step

Read .claude/skills/{step-command}.md and follow ALL instructions from that file with id={id}.

6. After the step

1. Update progress.md (each step already does this when finishing)
2. Show: "Step {N} completed. Next: {N+1} - {name}"
3. Interactive mode: Ask "Continue with the next step? (yes/no)". If yes -> back to step 4. If no -> section 7 (Cleanup), then show "To resume: /run {id}".
4. Auto mode: Do NOT ask. Loop straight back to step 4 with the next pending step until all 8 steps are done, then proceed to section 7 (Cleanup) and stop.

7. Cleanup

After the pipeline completes (all steps done) or when the user declines to continue, clean up the cloned repository:

REPO_DIR="tasks/{date}/{id}/work/repo"
if [ -d "$REPO_DIR" ]; then
  rm -rf "$REPO_DIR"
  echo "Cleaned up repo clone at $REPO_DIR"
fi

This cleanup runs regardless of whether all steps completed or the user stopped early.