subtask-init

autoworker:subtask-init — Create Subtask First Half (Gate 1.3 Step A+B)

After Plan Mode discussion is complete and user has confirmed direction, invoke this skill to create the first half of subtask.md.

Execution Flow

1. Determine Sequence Number

Glob `subtask_*.md` (current directory) + Glob `claude_docs/subtask/*.md` (archive)
→ Extract all sequence numbers → take max + 1

2. Extract Confirmation Record from Conversation Context

Review the conversation and extract:

• Confirmation method: AskUserQuestion / confirmed in plan discussion / explicit in user instruction
• Key Q&A: scope, completion criteria, constraints, preferences
• Goal: WHY + WHAT (1-3 sentences)
• Success criteria: specific measurable items
• Acceptance criteria: Extract quantifiable metrics from plan (metric name + measurement method + expected value/range + tolerance). When no quantitative metrics exist, explicitly write "No quantitative metrics, reason: "

3. Identify Core Assumptions

Extract assumptions that depend on external systems from plan and conversation. Each assumption needs:

• Assumption description (I assume X because Y)
• Verification experiment (one-line command)

Pure internal logic changes (e.g., refactoring, documentation): Write one line No external dependencies | — | —.

4. Search for Reference Files

Glob/Grep search for related files → confirm paths and line numbers

Do not guess line numbers — must Read to confirm.

5. Extract Design Decisions and Plan

Extract from plan file / conversation:

• Design decisions (decision point + choice + reason)
• Implementation plan (Phase/Step structure)

5.5. Pause Old Active Subtask

Glob `subtask_*.md` (exclude subtask_template.md) →
  For each file grep `status: active` →
    Edit all found to `status: paused`

Purpose: Ensure at most 1 active subtask at any time. Creating a new subtask automatically pauses old ones.

6. Write Subtask First Half

Language rule: subtask.md is an internal work document. Write it in the user's conversation language (as configured in their CLAUDE.md or inferred from their messages). Do NOT use the task's target language — e.g., if the task is "translate to English", the subtask itself should still be in the user's language.

Reference the table structure in subtask template, Write file:

subtask_<sequence>_<name>_<date>.md

First line after title: write status: active (flush left, grep-friendly format).

Include these sections (result columns left empty):

• User confirmation record
• task_plan positioning (extract corresponding Phase from plan file, position in overall plan, impact after completion. Write "No task_plan, skip" when none exists)
• Goal + success criteria
• Acceptance criteria table (extract from plan, or write "No quantitative metrics, reason: ")
• Core assumptions table (result column left empty)
• Design decisions
• Plan
• Reference files
• Verification plan (left empty, filled by autoworker:subtask-plan)
• Test results (left empty, filled by autoworker:checkpoint)
• Confidence assessment (left empty, filled by autoworker:gate-check)
• Gate 3 self-check (left empty, filled by autoworker:gate-check)
• Progress log
• Findings / Conclusions

7. Run Assumption Verification (Critical Step)

Line by line, read the "verification experiment" column of the assumptions table:

1. Execute the command
2. Observe the output
3. Edit subtask to fill in actual results

Hard rules:

• Result column must come from command execution output, cannot be filled by reasoning
• "—" (no verification needed) must be because there truly are no external dependencies, not laziness
• Verification failure → stop, report to user, do not continue

8. Output

Subtask first half created: subtask_<sequence>_<name>.md
- Assumption verification: <N>/<N> passed
- Acceptance criteria: <N> (or "No quantitative metrics")
- Paused old subtask: <list paused filenames, or "None">
→ Automatically invoking autoworker:subtask-plan

9. Chain: Immediately Invoke autoworker:subtask-plan

After outputting the summary above, immediately invoke autoworker:subtask-plan. Do not wait for user instructions, do nothing else.

Important Notes

• Cannot Write entire file then Edit results: Write must leave result column empty; only Edit to fill after verification
• Assumption verification is a blocking step: Any verification failure → entire flow pauses
• Reference files must be confirmed: Glob/Read to confirm paths and line numbers exist — cannot write "approximately line XX"
• Chaining is mandatory: Must invoke autoworker:subtask-plan after completion, cannot skip or manually substitute