orchestrator

Orchestrator

Tradeoff: Biases toward parallelism and explicit handoffs. For single-agent work or fast iteration, use direct prompting.

When this skill applies

This skill assumes a contract-first multi-agent build model:

• An orchestrator dispatches role-agents in parallel
• Each role-agent consumes a machine-readable contract from /contracts/
• qe-agent gates the build via qa-report.json

For single-agent or ad-hoc work, this skill is not the right tool.

Composition

The orchestrator is the conductor — not the only player. It composes with three groups of skills:

• INVOKES at the right phase: nano-banana (seed imagery), ui-ux-pro-max + frontend-design (UI quality), ux-review + render-sanity (post-build render-level validation), design-token-guard (post-build source-level gate — no inline styles / hardcoded colors bypassing tokens), class-extraction-guard (post-build source-level organization gate — repeated utility-class soup extracted into named classes), repo-deep-dive (reference research), llm-wiki (project knowledge base), mermaid-charts (architecture diagrams), deployment-checklist (ship readiness).
• DISPATCHES role-agents in parallel: backend-agent, frontend-agent, infrastructure-agent, db-migration-agent, security-agent, observability-agent, performance-agent, docs-agent, qe-agent.
• DELEGATES diff/code review to the external /code-review CLI: during a build, the Phase 4 diff review pass is the external /code-review CLI, NOT a spawned code-review-agent. The in-repo code-review-agent skill is not a default build phase — invoke it only deliberately for a standalone, repo-aware review. See references/mission-interpretation.md.
• DOES NOT preempt: brainstorming, plan-builder, writing-plans, claude-design-brief, ui-brief, feature-dev, claude-mem:*. If any of these belong before the build starts, let them run first — orchestrator picks up from the artifacts they produce.
• RUNS the validation phases AS LOOPS (not one-shot checks) — the canonical loop-harness mapping: the wave gate and the QA gate are convergence loops. fix-until-green is the contract for driving install/typecheck/test/QA red→green without cheating the gate — the QE inner loop and the wave-gate driver; loop-controller is the underlying harness (iterate → evaluate → guardrail → stop) whose no-progress/oscillation guardrail is the 3-failure circuit breaker and whose iteration/budget caps bound wave-gate retries. Under native Agent Teams, orchestrator-task-loop drives the whole-task-list OUTER loop (drain the shared task list until every task is completed + passing its TaskCompleted gate, feeding idle workers via TeammateIdle), with fix-until-green as each task's INNER loop. All three are disable-model-invocation: true — you explicitly dispatch them; they never auto-trigger because a test happened to fail. See skills/loops/.
• DISPATCHES four more build loops when the mission calls for them: contract-conformance-loop (build-until-spec, graded by a fresh-context evaluator), coverage-loop (a test-coverage target), perf-loop (a performance budget), and migration-loop (an enumerated wide-refactor set). All four are loop-controller configs and disable-model-invocation: true — the mission text is their trigger, and you are their dispatcher. Dispatch table: references/phase-guide.md (Phase 13's Optional build loops).