orchestrate（orchestrate 搜索）

Orchestrate

Coordinate a team of parallel agents to execute a phase from a task plan. Manages task assignment, monitoring, scope enforcement, and wave-transition quality gates.

Modes:

• Interactive (default): In-session agents via TeamCreate/Task/SendMessage — use for complex tasks needing inter-agent coordination
• Headless (--headless): Independent claude -p processes — use for parallelizable tasks with clear scope boundaries (~54% less coordination overhead)

Do NOT load when

• task is small enough for one agent to finish directly
• no written plan or no separable workstreams exist
• parallel execution would create more overlap than speedup

NEVER

• Never let parallel agents git add && git commit without flock — git's index is process-global. Concurrent writes silently mix staged files between commits. Every agent system prompt must include flock /tmp/orchestrate/{session-id}/git.lock. Symptom: commit A contains files from task B.
• Never skip file overlap check — two agents editing the same file produces merge conflicts neither can resolve (they have no knowledge of each other). Detect overlap at plan time, serialize conflicting tasks.
• Never trust is_error field alone for failure detection — budget exhaustion sets is_error: false with subtype: "error_max_budget_usd". Always check subtype.startsWith("error_"). Confirmed via live testing of claude -p --output-format json.
• Never reuse PID files across waves — OS recycles process IDs. Always clear /tmp/orchestrate/{session-id}/task-*.pid between waves or a stale PID matches an unrelated process, causing the monitor loop to wait indefinitely.
• Never spawn headless agents without --no-session-persistence — each claude -p writes a session file to ~/.claude/. 22 parallel tasks = 22 orphaned session files polluting session history.
• Never use --dangerously-skip-permissions without --allowedTools — the flag alone gives unrestricted tool access including TeamCreate/SendMessage/Task (recursive agents). Always pair with --allowedTools "Bash Edit Write Read Glob Grep".
• Never retry a failed agent by re-spawning the same one with appended errors — accumulated failed reasoning makes the second attempt worse. Always spawn a fresh fix-subagent with targeted instructions.
• Never run code quality review before spec compliance passes — quality review on spec-incorrect code wastes tokens and produces misleading results. Spec compliance first, always.
• Never consider a headless task done without running the full verification chain — even subtype: "success" doesn't guarantee: correct branch, in-scope files, passing build. Always: commit exists → scope check → verify command.

Mode selection decision tree

Is task count > 10 AND tasks have clear file boundaries?
  YES → Headless: ~54% less overhead, simpler monitoring
  NO → Interactive: better for coordination, cross-task decisions

Do tasks share many files (>50% overlap)?
  YES → Add git worktrees OR serialize into sequential waves
  NO → Standard parallel execution with flock

Resource loading (mandatory)

• Before spawning any agent: Read agent-roles.md for role assignments and model selection
• Headless mode, before each task prompt: Read prompt-templates/{role}.md for that task's role
• At every wave transition: Read wave-template.md for the pre/during/post checklist
• For claude -p flag details or error classification: headless-runner.md

Do NOT load prompt templates in interactive mode — interactive agents invoke /quality-commit and /tdd directly.

Scripts

bash scripts/setup-session-dir.sh
node scripts/check-file-overlap.js "apps/api/src/a.ts,apps/api/src/b.ts" "apps/api/src/b.ts,apps/web/src/c.ts"

Failure escalation (headless)

1. Retry with context (up to 2): Append error output, re-spawn same model
2. Model escalation (after 2 fails): haiku→sonnet, sonnet→opus. Budget ×1.5
3. User intervention: Print full failure history. Offer: Skip / Manual fix / Abort

Progressive stall escalation (interactive)

Integration map

Arguments

$ARGUMENTS accepts: Phase ID, plan file path, or inline task list.

Key flags: --headless, --interactive, --dry-run, --wave N, --max-agents N, --budget-per-task N, --timeout-multiplier N, --no-qa, --verbose

Examples

/orchestrate A                          # Phase A, interactive
/orchestrate A --headless               # Phase A, headless
/orchestrate A --headless --dry-run     # Preview prompts without spawning
/orchestrate A --wave 2                 # Resume from wave 2
/orchestrate docs/plans/custom.md       # Custom plan file
/orchestrate "add auth; write tests"    # Inline task list