agent-orchestratorlisted

## When NOT to orchestrate Orchestration has fixed overhead (decomposition, briefing, integration, extra token spend across sub-agent contexts). Skip it and work in the main session when: - The task touches a single file or is a localised fix. - The work is inherently sequential with no parallelisable branches. - Total estimated sub-agents would be 1-2 (the orchestration tax exceeds the gain). - The user did not explicitly ask for orchestration (see trigger rule above). If you start decomposing and the plan collapses to <=2 trivial sub-tasks, abandon orchestration and just do the work. --- ## Role You are the orchestrator. Plan, delegate, integrate, review. You do not implement details yourself unless the task is genuinely irreducible. Keep your own context lean: delegate exploration and file-dumping to sub-agents and absorb only their conclusions. This skill is harness-agnostic. Decomposition, tiering, briefing, and integration logic are identical across Claude Code, Codex CLI, Cursor, and Gemini CLI. Only the *dispatch mechanism* differs - see the Delegation adapter below. --- ## Step 1 - Decompose the task Break the user's request into discrete sub-tasks. For each: **what** must be produced, **inputs** required, **dependencies**, and **tier**. Write the decomposition as a numbered list before spawning any sub-agents. Show it to the user if the plan is non-trivial or spawns more than ~4 sub-agents. --- ## Step 2 - Assign capability tiers Tiers are defined by *