From autoworker
Sync tracking documents based on current conversation results. Updates subtask, progress, findings, task_plan, project CLAUDE.md. Use when finishing a task or reaching a milestone.
npx claudepluginhub phj128/autoworker --plugin autoworkerThis skill uses the workspace's default tool permissions.
Based on work results and discoveries in the current conversation, update all tracking documents.
Organizes and tracks complex multi-step tasks using persistent Markdown files: task_plan.md, findings.md, progress.md. Auto-recovers context after /clear, integrates with tools like git diff for session catch-up. Useful for project planning and progress monitoring.
Updates PROGRESS.md and NEXT_STEPS.md with changes, verifications, blockers, and exact next actions for resuming work sessions with minimal context loss.
Creates and archives Navigator task documentation with implementation plans. Use when starting new features, completing work, or documenting features.
Share bugs, ideas, or general feedback.
Based on work results and discoveries in the current conversation, update all tracking documents.
Optional argument: $ARGUMENTS
park: Sync + pause current subtask (set status: paused) + clear plan file. For "task still in progress (e.g., training) but need to switch to something else"archive: Sync + archive completed subtask (move to claude_docs/subtask/) + clear plan file. For "task fully complete"Find tracking files and document system in the project root:
Top-level tracking files:
subtask_*.md — current in-progress subtasksprogress.md — progress logfindings.md — core discoveriestask_plan.md — task planningCLAUDE.md — project rules and technical infoREADME.md — project intro (for humans)Progressive disclosure documents:
claude_docs/reference/*.md — detailed documents referenced from CLAUDE.mdCLAUDE.md — index files for each moduleRead each existing file to understand current recorded state.
Review work completed in this conversation, determine updates needed item by item:
| File | Update condition | Update content |
|---|---|---|
subtask_*.md | Steps were completed | Check off [x], record progress, add verification results |
progress.md | Task completed or progressed | Update task item status, add result summary |
findings.md | New conclusions or counter-intuitive behaviors discovered | Add new section or update existing section |
task_plan.md | Task completed/progressed, or plan changed | Read subtask's "task_plan positioning → corresponding Phase" field, locate and check off [x] in task_plan.md with completion timestamp; match by content when no positioning field; update future steps if plan changed |
CLAUDE.md | Reusable patterns or lessons learned discovered | Add to appropriate section (don't duplicate existing content) |
claude_docs/reference/*.md | Findings in this conversation belong to an existing reference topic | Update corresponding reference document |
Subdirectory CLAUDE.md | Changed code structure/files in a module | Update corresponding subdirectory file index |
README.md | Added feature or changed usage/installation | Sync update (user-perceivable changes) |
Dead link check: Scan project CLAUDE.md for all claude_docs/reference/ references, confirm files exist:
New reference identification: If this conversation produced new lessons or technical documentation (> 20 lines) and CLAUDE.md already exceeds 300 lines, suggest writing to claude_docs/reference/<topic>.md and adding a trigger heading in CLAUDE.md, rather than appending directly to CLAUDE.md.
For each file needing update:
Using claude_docs/ progressive disclosure structure, check if this conversation's changes affect subdirectory documents:
Flow:
Glob claude_docs/*/CLAUDE.md → get all subdirectory indexesPause the current active subtask, clean up workspace, facilitate switching to next task. Subtask file stays in workspace (not archived) and can be resumed later.
subtask_*.md → grep status: active → Edit to status: paused**Paused** — <infer reason from conversation, e.g., "waiting for training completion">progress.mdStatus filter:
subtask_*.md (exclude subtask_template.md) → grep status: to filterstatus: completedExecute archive (for each completed subtask):
claude_docs/subtask/CLAUDE.md to find max sequence + 1subtask_<name>.md → claude_docs/subtask/<sequence>_<name>_<date>.mdrm subtask_<name>.md (after archiving, don't keep in workspace — prevents dispatch from selecting it next time)claude_docs/subtask/CLAUDE.md archive list + Phase groupingprogress.md add archive link.claude/plans/*.md, find current project's plan file, clear contents (write empty string). Archiving = task done — stale plan would interfere with next autoworker:deep-plan.Check project-level CLAUDE.md size and structure (if exists):
Size check:
500 lines → bloated, list content that can be migrated
Bloat identification (when > 300 lines):
Scan CLAUDE.md sections, identify migratable content by these rules:
| Pattern | Should migrate to | How to detect |
|---|---|---|
| Experiment logs, hyperparameter comparison tables | progress.md | Contains epoch/loss/reward numeric tables |
| Bug fix history, lessons learned (> 20 lines) | findings.md or claude_docs/reference/ | Detailed content > 20 lines |
| Detailed data pipeline documentation | claude_docs/reference/data_pipeline.md | Data formats, field descriptions, transformation flows |
| Detailed API/config documentation | claude_docs/reference/<topic>.md | Detailed config item lists or code templates |
Output:
After park or archive completes, commit and push all changes to remote.
List all updated files and change highlights:
sync-docs complete:
- subtask_xxx.md: Steps 3-5 checked off, verification results added
- progress.md: Task X marked complete
- CLAUDE.md: Added "YYY" lesson
- [archive] → claude_docs/subtask/40_xxx_20260227.md
- CLAUDE.md health: N lines, healthy / getting large, suggest migrating X sections
status: completed: Only subtasks with completed status (after gate-check PASS) get archived