/parallel-task

$ARGUMENTS - **Task**: Spawn explore/worker agents - **Glob/Grep/Read**: Codebase exploration - **AskUserQuestion**: Get approval before execution - **Bash**: Run verification commands, git operations - **Edit/Write**: Only via worker agents, not directly

Phase 0: Planning & Analysis

1. Analyze the task and break into parallel work units
2. Identify target scopes (folders/files to modify)
3. Detect project verification commands:
   • Check package.json scripts (test, lint, typecheck, type-check)
   • Check for Makefile, pyproject.toml, Cargo.toml
   • Note available commands for Phase 4
4. Define safety requirements and verification criteria
5. Present complete plan using AskUserQuestion for approval before execution

Phase 1: Exploration (Parallel)

Spawn N explore agents analyzing different areas.

Spawn ALL explore agents in a SINGLE message using parallel Task tool calls:

Task 1: subagent_type="Explore", model="sonnet", prompt="Analyze [scope1] for [task]. Report: files to change, patterns to follow, risks."
Task 2: subagent_type="Explore", model="sonnet", prompt="Analyze [scope2] for [task]. Report: files to change, patterns to follow, risks."
...

Use model="haiku" for gates or external exploration (web searches, MCP tools).

Collect reports, identify what to change vs preserve, estimate impact.

Phase 2: Safety Setup

1. Create rollback point: git checkout -b backup/parallel-task-$(date +%s)
2. Add temporary files to .gitignore if needed
3. Confirm all safety nets in place

Phase 3: Execution (Parallel)

Spawn N worker agents with strict scope isolation.

Spawn ALL worker agents in a SINGLE message using parallel Task tool calls:

Task 1: subagent_type="general-purpose", model="opus", prompt="[Detailed instructions for scope1]"
Task 2: subagent_type="general-purpose", model="opus", prompt="[Detailed instructions for scope2]"
...

Each agent prompt includes:

• Explicit folder assignment (modify ONLY files in this scope)
• Step-by-step commands (not goals)
• Verification requirements (run lint/type-check before committing)
• Error handling: if blocked, return error report instead of partial work
• Commit changes with descriptive message including scope name

Wait for ALL agents to complete. If any fail, stop and report to user.

Phase 4: Validation

Run project's verification commands detected in Phase 0.

Also verify:

• No cross-scope contamination: git diff --stat
• All git hooks pass: run git hook run pre-commit or stage + unstage to trigger
• No regressions in functionality

Phase 5: Final Report

Present:

• All commits created (with hashes)
• Total impact (files changed, lines added/removed)
• Items preserved vs removed
• Review commands: git log --oneline -N, git diff HEAD~N

<context_management> For long orchestrations approaching context limits:

1. Complete current phase before any checkpointing
2. Save progress to a state file in the project (e.g., .claude/orchestration-state.yaml)
3. Inform user: "Context limit approaching. State saved to [path]." </context_management>