About orchestrate
A skill for coordinating independent work through Codex collaboration tools. It helps an agent decide what to delegate, define ownership and completion criteria, and integrate results while retaining responsibility for the final response.
Key features
- 01Assess whether delegation is worthwhile
- 02Set task context and ownership boundaries
- 03Select configuration based on preferences
- 04Track and redirect ongoing work
- 05Validate and integrate returned findings
Use cases
- 01Delegate independent investigation
- 02Assign parallel work without overlapping edits
- 03Reconcile agent findings into one result
Install
Copy the command and run it in a terminal.
npx skills add https:/After installing, type/orchestratein Claude Code to confirm it loaded.
Bundled filesagents/openai.yaml
Details
From the project READMEOrchestrate
Use Codex collaboration tools to delegate independent outcomes. The root owns user communication, integration, and the final result. This skill does not override the session's delegation or action permissions.
Decide what to delegate
Delegate when another agent can make useful progress independently or provide a valuable second judgment. Stay single-agent when coordination, duplicated context, or shared mutable state would cost more than the delegation helps. A long-running command alone does not require another agent when the available tools can run it asynchronously.
Give each agent a concrete outcome, the relevant context and evidence, its ownership boundary, permitted actions, and a definition of completion. Share enough history to avoid rediscovery, without copying unrelated context. Avoid overlapping write ownership. Further delegation needs an explicitly assigned coordination role.
Choose the configuration
Honor the user's model and reasoning preferences. Otherwise use the current agent's configuration unless the assignment provides a concrete reason to choose differently. When selecting an alternative, weigh ambiguity, consequence, expected quality, latency, and cost using the models actually available. Do not equate a role with a fixed model or maximum reasoning effort.
Before spawning, briefly disclose the task name, role, requested model and reasoning effort, and whether configuration is inherited. Do not guess settings the runtime does not expose or present requested settings as verified runtime identity.
Integrate the work
Keep useful work moving locally while agents run. Reuse an existing agent when its context remains relevant. Redirect or stop work when assumptions diverge, scope expands, or ownership overlaps.
Treat returned findings as evidence to assess, not automatic proof of completion. Resolve material disagreements against authoritative sources, validate the combined result, and report remaining limitations. Return a coherent answer rather than forwarding separate agent reports.