Orchestration Skills

The orch-* family coordinates multiple agents and skills to solve complex compound tasks. These skills operate at the meta-level — they manage other skills, not code directly.

Skills

Skill ID	Description	Model Class
`orch-agent-orchestrator`	Top-level orchestrator: decomposes a complex request into a skill execution plan, then coordinates execution	`strong`
`orch-delegation`	Delegates sub-tasks to appropriate agents or skill chains based on capability matching	`cheap`
`orch-multi-agent`	Manages concurrent execution of multiple agent instances, tracks state, handles partial failures	`strong`
`orch-result-synthesis`	Aggregates outputs from multiple parallel skill runs into a coherent, non-contradictory unified result	`strong`

When to Use

Situation	Skill(s)
Request spans multiple domains/skills	`orch-agent-orchestrator`
Need to fan out work to parallel agents	`orch-multi-agent` + `orch-delegation`
Combining results from 3+ parallel runs	`orch-result-synthesis`
Complex compound task	All four in sequence

Instructions That Invoke These Skills

orchestrate — primary consumer; all four coordinated
meta-routing — uses orch-agent-orchestrator to route compound requests
research — uses orch-result-synthesis to merge multi-model research outputs

Orchestration Flow

Complex Request
      ↓
orch-agent-orchestrator    → decompose into skill plan
      ↓
orch-delegation            → assign skills to agents/models
      ↓
orch-multi-agent           → execute in parallel, track state
      ↓
orch-result-synthesis      → merge, deduplicate, unify output

Synthesis Contract

orch-result-synthesis never discards contradictions — it surfaces them:

[SYNTHESIS NOTE] GPT-5.4 and Claude Haiku 4.5 disagree on X.
  - GPT-5.4 position: ...
  - Claude Haiku 4.5 position: ...
  - Resolution: deferred to strong-model judgment (Claude Sonnet 4.6)