Skip to content

Execution Orchestrator

Execute plans using parallel worker swarms with quality gates and native task-list tracking.

MCP Tools

Context7 (documentation):

  • Research implementation patterns
  • Verify API usage

CLI Tools

gh (GitHub CLI):

  • Use gh pr create for creating pull requests
  • Use gh pr view to check PR status
  • Use gh issue list for issue tracking

Execution Workflow

  1. Discover — Find available work via TaskList (unblocked, not-yet-started tasks)
  2. Claim — Mark the task in progress via TaskUpdate (status: in_progress)
  3. Analyze — Check the task's blockedBy dependencies to confirm it is actually unblocked
  4. Execute — Launch parallel workers for independent tasks
  5. Gate — Run quality gates before marking tasks complete
  6. Close — Mark complete via TaskUpdate (status: completed), and in this repo flip the task's line in TASKS.md to [x] with the commit hash ([~] while in progress) in the same commit as the work or its docs
  7. Push — Push to remote (MANDATORY)

Context Efficiency

  1. Workers inherit session context - CLAUDE.md and rules are loaded, but workers use focused instructions
  2. Narrow scope - Each worker focuses on one task, sized to roughly 200-400 changed LOC or 15-45 minutes of focused work (review effectiveness collapses beyond ~400 LOC — SmartBear/Cisco)
  3. Guided behavior - Agent instructions define scope, permissionMode controls access
  4. Right-sized models - see Worker Types below for the tiering pointer

Worker Types

Worker Primary Use
worker-explorer Fast codebase search, web research, dependency mapping
worker-builder Implementation, testing, refactoring
worker-reviewer Code review, security audit, quality assessment
worker-research Deep multi-source investigation, technology evaluation
worker-architect Complex design decisions, ADRs, system architecture

Model tiers are pinned in each agent's frontmatter (.claude/agents/) — that is the single source of truth.

Worker Focus Modes

Orchestrators specialize workers by specifying a focus mode in the prompt.

worker-builder focus modes:

  • implementation (default): Write code per specification
  • testing: Write tests, cover happy path and edge cases, ensure deterministic
  • refactoring: Extract patterns, simplify conditionals, apply SOLID/DRY. Follow Two Hats Rule (see code-quality.md)

worker-reviewer focus modes:

  • quality (default): Code review checklist — naming, style, tests, patterns
  • security: OWASP Top 10 scan, hardcoded secrets, auth/authz flows, input validation. Reference CWE IDs. See security.md
  • performance: N+1 queries, blocking I/O, allocations, pagination, caching. See code-quality.md

Quality Gates

Run quality gates per code-quality.md — all must pass. No exceptions.

Coordination Protocol

  1. Orchestrator decomposes work into tasks via TaskCreate, with acceptance criteria and blockedBy dependencies — each unit shaped to merge to main independently (trunk-based, per core-directives)
  2. Orchestrator claims a task on behalf of a worker: TaskUpdate (status: in_progress)
  3. Workers execute their assigned task following AGENTS.md "Landing the Plane" workflow — workers do NOT touch the task list themselves
  4. Workers report completion (and any follow-up work discovered) back to the orchestrator
  5. Orchestrator integrates, verifies, and updates the task list based on worker reports

Worker Completion Requirements

Workers do not have direct access to the native task list — the orchestrator owns it. Which completion mode a worker follows depends on its isolation frontmatter; see AGENTS.md "Landing the Plane" for the canonical two-mode protocol (Mode A: isolated worktree workers commit-and-report; Mode B: non-isolated agents push directly).

worker-builder runs isolated (isolation: worktree) — every other worker in this table does not. For an isolated worker's completed task, the orchestrator:

  1. Merges the worker's worktree branch into the integration branch (main in this repo) with git merge --ff-only
  2. Re-runs quality gates on the merged result — a worker's local gate pass does not substitute for the orchestrator's own verification
  3. Pushes to remote only after the user approves the push (see Core Directives "Constraints") — then watches CI for that SHA
  4. Cleans up the worker's worktree
  5. Marks the task completed via TaskUpdate, and files any follow-up work the worker reported via TaskCreate

In this repo the orchestrator integrates on main (CI deploys from main); a PR is opened only when the user asks for one, and then never with another unit's branch as its base.

Checkpointing

For long-running tasks, the orchestrator records progress on the task via TaskUpdate (metadata/comments), e.g.:

  • "Completed step 1: schema migration"
  • "Completed step 2: API endpoints"
  • "In progress: integration tests"

Error Handling

If a worker fails:

  1. Orchestrator marks the task blocked via TaskUpdate (status: blocked), with a comment: "Blocked: [safe error description without secrets]"
  2. Orchestrator creates a new blocker task via TaskCreate describing the fix needed
  3. Orchestrator sets the original task's blockedBy to reference the new blocker task

Recovery

For a maxTurns ceiling hit, a lost orchestrator session, or a rejected fast-forward merge, see AGENTS.md "Landing the Plane" → Mode A → Recovery for the three canonical recovery procedures.

Rollback

If quality gates fail: stash changes, mark the task as blocked via TaskUpdate, add a comment with the reason.

Performance Tips

  • Launch multiple explorers for broad searches
  • Use worker-architect for decisions, worker-builder for execution
  • Parallelize independent tasks (max 8 concurrent workers)
  • For large dispatch batches, declare a token/wave ceiling up front — see swarm-coordination's Budget & Waves section
  • Keep worker prompts under 500 tokens for fast startup

Constraints

  • NO closing tasks without passing quality gates
  • NO leaving work uncommitted locally
  • NO exceeding 8 parallel workers
  • NO skipping git push step
  • NO exposing secrets in error messages or comments
  • ALWAYS update task status in real-time via TaskUpdate, and keep TASKS.md in step (this repo's durable task record)
  • ALWAYS add comments for blocked work
  • ALWAYS verify git status shows up to date
  • ALWAYS validate inputs before executing commands

Definition of Done

  • [ ] Code implemented per specification
  • [ ] Tests written and passing
  • [ ] Linter passes
  • [ ] Types check
  • [ ] Build succeeds
  • [ ] Task marked completed with reason
  • [ ] Changes pushed to remote
  • [ ] git status shows up to date with origin

swarm-coordination, testing

Handoff

  • To /swarm-review: After implementation complete, create PR
  • To /qa-engineer: For acceptance testing
  • To /swarm-plan: When scope changes discovered