| name | grok-subagents |
| description | Direct a compatible Codex root model to delegate bounded, independent reasoning work to the installed Grok 4.5 native subagent. Use when the user invokes `$grok-subagents`, or when implicit invocation is enabled and bounded Grok delegation improves independent review, debugging hypotheses, security or edge-case analysis, alternative designs, research, or other parallelizable read-heavy work. |
Grok Subagents
Use the installed native role grok_4_5_subagent, not a direct model="grok-4.5" override. The role pins the xAI provider and high reasoning effort.
Delegate Deliberately
- Keep the root agent as planner, integrator, and final quality gate.
- Delegate every genuinely independent, bounded workstream that materially helps. Do not impose a fixed worker count; use the available concurrency intentionally and avoid duplicate scopes.
- Give each worker a response contract proportional to its assignment. Specify an approximate maximum word or token budget and, when useful, a maximum number of findings or conclusions. Keep narrow checks compact and allow broader architecture work more room.
- Ask workers to prioritize the most severe, actionable, well-evidenced results within that response budget and briefly flag meaningful omitted areas when necessary.
- Control only the returned synthesis. Do not reduce Grok reasoning effort, useful tool access, or the number of useful workers merely to shorten responses.
- Do not delegate trivial work, sequential work whose output immediately blocks the next root action, or duplicate work already owned by the root.
- The installed Grok role is read-only. Have it return evidence and recommendations; make edits and final decisions in the root task.
Native Spawn Protocol
- Confirm the fresh parent task uses Multi-Agent V1. If it uses V2, do not spawn Grok; explain that a Codex restart and fresh supported-root task are required.
- For each bounded workstream, call the standard native spawn tool with
agent_type="grok_4_5_subagent" and fork_context=false. Do not pass model, provider, reasoning-effort, or service-tier overrides.
- Start all independent children before waiting. Give each a precise scope, expected evidence, return format, and dynamically chosen response budget or finding limit.
- Wait for the needed results, reconcile disagreements, verify important claims locally, and return one integrated answer.
Fail Clearly
unknown agent_type: the role is unavailable; verify the Grok setup, then restart Codex before retrying.- V2 or an encrypted
agent_message failure: do not retry through Grok; restart Codex and use a new V1 root task.
- 401, 403, or 429-style inference errors: report the exact error and check xAI OAuth status, loopback proxy health, SuperGrok entitlement, and xAI usage/rate limits. Do not silently fall back to another model.
- Any model/provider mismatch: fail the delegated result closed and tell the user what could not be verified.
Report
State how many Grok children ran, what each covered, their key evidence or disagreement, and how the root used the results. Surface an unavailable Grok route rather than concealing it.
