worker-dispatch（工人派遣）

Worker Dispatch

Overview

Spawns and manages worker Claude agents using the Task tool with run_in_background: true. Workers run as parallel background agents monitored via TaskOutput.

Core principle: Workers are isolated, scoped, and expendable. State lives in GitHub, not in workers.

Key optimization: Pre-extract issue context BEFORE spawning. Workers receive focused context, not raw issues.

Worktree Isolation (MANDATORY)

CRITICAL: Every worker MUST have its own git worktree. Workers NEVER operate in the main repository.

Main Repository (./)           ← Orchestrator only
    │
    └── Worktrees (isolated)
        ├── ../project-worker-123/    ← Worker for #123
        ├── ../project-worker-124/    ← Worker for #124
        └── ../project-worker-125/    ← Worker for #125

Orchestrator responsibility: Create worktree BEFORE spawning worker. Worker responsibility: Verify isolation BEFORE any work (see worker-protocol).

This prevents file clobbering between parallel workers.

Worker Types

Pre-Extraction (CRITICAL)

Extract issue context BEFORE spawning. Workers should not spend tokens re-reading issues.

extract_issue_context() {
  local issue=$1

  # Single API call to get everything
  ISSUE_JSON=$(gh issue view "$issue" --json title,body,labels,comments,assignees)

  TITLE=$(echo "$ISSUE_JSON" | jq -r '.title')
  BODY=$(echo "$ISSUE_JSON" | jq -r '.body')
  LABELS=$(echo "$ISSUE_JSON" | jq -r '[.labels[].name] | join(", ")')

  # Extract acceptance criteria if present
  ACCEPTANCE=$(echo "$BODY" | sed -n '/## Acceptance Criteria/,/^## /p' | head -20)
  [ -z "$ACCEPTANCE" ] && ACCEPTANCE=$(echo "$BODY" | sed -n '/- \[/p' | head -10)

  # Get latest handover if exists
  HANDOVER=$(echo "$ISSUE_JSON" | jq -r '
    [.comments[] | select(.body | contains("<!-- HANDOVER:START -->"))] | last | .body // ""
  ')

  # Get recent progress comments
  PROGRESS=$(echo "$ISSUE_JSON" | jq -r '
    [.comments[-3:][].body] | join("\n---\n")
  ' | head -50)
}

Spawning Implementation Workers

Step 1: Extract Context & Create Worktree

spawn_implementation_worker() {
  local issue=$1
  local attempt=${2:-1}

  # Pre-extract context
  extract_issue_context "$issue"

  worker_id="worker-$(date +%s)-$issue"
  issue_slug=$(echo "$TITLE" | tr '[:upper:]' '[:lower:]' | sed 's/[^a-z0-9]/-/g' | cut -c1-40)
  branch="feature/$issue-$issue_slug"
  worktree_path="../$(basename $PWD)-worker-$issue"

  # Create worktree
  git fetch origin main
  git branch "$branch" origin/main 2>/dev/null || true
  git worktree add "$worktree_path" "$branch"
}

Step 2: Register in GitHub

register_worker() {
  local issue=$1 worker_id=$2 worktree=$3

  gh issue comment "$issue" --body "<!-- WORKER:ASSIGNED -->
{\"worker_id\": \"$worker_id\", \"assigned_at\": \"$(date -u +%Y-%m-%dT%H:%M:%SZ)\"}
<!-- /WORKER:ASSIGNED -->

**Worker Assigned:** \`$worker_id\` at $(date -u +%H:%M:%S)Z"

  update_project_status "$issue" "In Progress"
}

Step 3: Construct Focused Prompt

~50 lines with pre-extracted context:

construct_worker_prompt() {
  local issue=$1 worker_id=$2 attempt=$3
  # Uses pre-extracted: TITLE, ACCEPTANCE, HANDOVER, PROGRESS

  cat <<PROMPT
Worker $worker_id | Issue #$issue | Attempt $attempt

## Task
$TITLE

## Requirements
$ACCEPTANCE

## Constraints
- Work ONLY on issue #$issue
- TDD: write tests first
- All tests must pass before PR
- Complete code review before PR (post artifact to issue)
- Max 100 turns - handover at 90+

## Exit Conditions
Return JSON when done:
\`\`\`json
{"status": "COMPLETED|BLOCKED|HANDOVER", "pr": null, "summary": "..."}
\`\`\`

| Status | Meaning |
|--------|---------|
| COMPLETED | PR created, tests pass |
| BLOCKED | Cannot proceed without external input |
| HANDOVER | Turn limit approaching, context posted |

## Progress Protocol
Post brief updates to issue. On handover, post full context with <!-- HANDOVER:START --> markers.

$([ -n "$HANDOVER" ] && echo "## Previous Handover
$HANDOVER")

$([ -n "$PROGRESS" ] && echo "## Recent Activity
$PROGRESS")

Begin implementation now.
PROMPT
}

Step 4: Spawn Worker

Task(
  description: "Issue #[ISSUE] worker",
  prompt: [FOCUSED_PROMPT],
  subagent_type: "general-purpose",
  run_in_background: true
)

Returns: task_id for monitoring.

Spawning Research Workers

Use Explore subagent for read-only investigation before implementation:

construct_research_prompt() {
  local issue=$1 question=$2

  cat <<PROMPT
Research for issue #$issue: $TITLE

## Question
$question

## Scope
Investigate codebase to answer the question. Return structured findings.

## Output Format
\`\`\`json
{
  "findings": ["key finding 1", "key finding 2"],
  "relevant_files": ["path/to/file.ts"],
  "patterns": ["existing pattern to follow"],
  "concerns": ["potential issue to address"],
  "recommendation": "summary recommendation"
}
\`\`\`

Do NOT modify any files. Research only.
PROMPT
}

Task(
  description: "Research for #[ISSUE]",
  prompt: [RESEARCH_PROMPT],
  subagent_type: "Explore",
  run_in_background: true
)

Use cases:

• Pre-implementation: "What patterns exist for similar features?"
• Debugging: "Where is this error originating?"
• Impact analysis: "What will this change affect?"

Parallel Dispatch

Spawn multiple workers in ONE message for concurrent execution:

## Dispatching 3 Workers

1. Extract context for each issue (sequential)
2. Invoke all Task tools in SAME message (parallel):

Task(description: "Issue #123 worker", prompt: [...], subagent_type: "general-purpose", run_in_background: true)
Task(description: "Issue #124 worker", prompt: [...], subagent_type: "general-purpose", run_in_background: true)
Task(description: "Issue #125 worker", prompt: [...], subagent_type: "general-purpose", run_in_background: true)

3. Store all returned task_ids

Monitoring Workers

TaskOutput(task_id: "[ID]", block: false, timeout: 1000)

Parse completion:

RESULT=$(echo "$OUTPUT" | grep -oP '\{.*\}' | jq -r '.status')
PR=$(echo "$OUTPUT" | grep -oP '\{.*\}' | jq -r '.pr // empty')

PR Workers

Resolve existing PRs (CI failures, missing reviews, merge):

construct_pr_worker_prompt() {
  local pr=$1 worker_id=$2

  cat <<PROMPT
PR Worker $worker_id | PR #$pr

## Mission
1. Check CI: \`gh pr checks $pr\`
2. If failing: investigate, fix, push
3. Verify review artifact exists on linked issue
4. If mergeable: \`gh pr merge $pr --squash --delete-branch\`

## Output
\`\`\`json
{"status": "MERGED|BLOCKED|HANDOVER", "summary": "..."}
\`\`\`

## Constraints
- Only fix CI-related issues
- Push to existing branch only
- Check for do-not-merge label

Begin with: \`gh pr checks $pr\`
PROMPT
}

Worker Cleanup

On completion (detected via TaskOutput):

cleanup_worker() {
  local issue=$1 task_id=$2 result=$3

  gh issue comment "$issue" --body "<!-- WORKER:ASSIGNED -->
{\"assigned\": false, \"cleared_at\": \"$(date -u +%Y-%m-%dT%H:%M:%SZ)\"}
<!-- /WORKER:ASSIGNED -->

**Worker Complete:** $result"

  case "$result" in
    COMPLETED) update_project_status "$issue" "In Review" ;;
    BLOCKED)   update_project_status "$issue" "Blocked" ;;
    HANDOVER)  ;; # Keep In Progress, spawn replacement
  esac
}

Replacement Workers

When worker returns HANDOVER:

1. Handover context is in issue comments (already posted by worker)
2. Extract via: extract_issue_context (includes HANDOVER variable)
3. Spawn replacement with attempt+1
4. New worker receives full context automatically

Checklist

Before spawning:

• Issue context pre-extracted
• Worktree created on feature branch
• Worker registered in GitHub
• Project board status: In Progress

Prompt includes:

• Pre-extracted title and acceptance criteria
• JSON output format requirement
• Previous handover (if any)
• Recent progress comments

On completion:

• JSON result parsed
• GitHub state updated
• Project board status updated
• Task_id removed from active list

Integration

Used by: autonomous-orchestration, claude-autonomous --pr

Uses: worker-protocol (behavior contract), worker-handover (context format), ci-monitoring (PR workers)