codex-subagent-router/skills/codex-subagent-router/SKILL.md

---
name: codex-subagent-router
description: Decide when Codex subagents help, route bounded work across Astra, Sol, Terra and Luna, and manage child evidence and lifecycle. Use for delegation requests, routing decisions, or subagent audits; inspecting this skill does not itself authorize spawning.
---

# Codex Subagent Router

Version 2.0.0. For new general engineering sessions recommend Sol / medium, with high when demonstrated reasoning needs justify it. Preserve the user's selected parent; installation does not switch it. Use Astra for bounded residual hard reasoning, rather than mandatory planning of every task.

Make delegation useful, observable and bounded. Preserve the selected parent model. The parent owns the critical path, shared decisions, integration and final acceptance. This skill neither changes configuration nor grants external permissions.

## Trigger before routing

Classify the request before using child tools:

- **Audit or advice:** inspect rules, configuration and available tools; report findings without spawning or changing settings.
- **Authorized execution:** the user requested delegation, or an applicable instruction explicitly authorizes it. Within that scope, actively delegate independent useful slices while the parent advances other work.
- **No delegation authorization:** work locally. Automatic skill discovery, model availability, task complexity and a request to edit this skill do not grant permission to spawn.

Honor higher-priority host restrictions even when a lower-priority rule permits delegation. Do not request authorization repeatedly after it has been granted. Never turn a routing recommendation into a new user-owned task.

Before dispatch classify each candidate:

- **P:** independent evidence or disjoint file AND semantic writes with its own acceptance.
- **C:** a bounded investigation/draft that needs latest-baseline parent integration.
- **S:** an immediate dependency, overlapping contract/state, permissions, production mutation, PR/merge/deploy or final acceptance; keep it in the parent.

Spawn only P/C work with a concrete output and useful parent work available. Do not spawn for one command, ceremonial probes, duplicate reviews or a model quota. Batch homogeneous small work.

## Select a supported route explicitly

| Work shape | Initial route |
| --- | --- |
| Known-source collection, fixed checks, logs | Luna low/medium |
| Bounded classification, conversion, settled patch with fixed checks | Luna high |
| Everyday implementation, debugging, locating/correlating artifacts | Terra medium; high when needed |
| Bounded complex analysis, design or financial/security evidence | Sol medium; high when needed |
| Hardest independent synthesis across code, tools and research | Astra medium; high when needed |
| Shared decisions, integration and external/final gates | Current parent, serial |

These are starting heuristics, not measured cost rankings. Pick sufficient capability directly; do not escalate through every model. Missing access/data and tool failures need diagnosis, not a stronger model. After two same-class failures, pause that slice and return the evidence to the parent.

Use the lowest adequate supported effort. Use xhigh/max for a concrete depth need or an explicit compatible role requirement; ultra only when exposed and justified by the workload. Model choice and working role are separate. A role named "explorer" is not proof of sandbox isolation.

At dispatch, specify model AND effort when the host permits selection. Otherwise omitted settings may inherit the parent or configured defaults. Prefer a self-contained contract and no history fork; use bounded history only when necessary. Respect full-history/override incompatibilities. Never claim prompt text changed a runtime parameter.

## Check capabilities once per unchanged host

Inspect the actual child-tool schema: models, per-model efforts, role bindings, defaults, history rules, slot counting, workspace isolation and controls. API catalogs, config files and task-creation tools cannot prove child availability. Recheck only after relevant changes.

Use the first authorized, low-risk real child task as the capability observation; for unfamiliar Luna multi-step work, start with a useful bounded read-only slice. A returned model name or marker is not identity proof. Record requested settings separately from host-observed identity; missing fields are unknown. Output can pass while identity stays unknown, unless the task explicitly requires verified identity.

Treat workspaces as shared unless isolation is confirmed. Tool schemas may lack per-child sandbox, timeout, role or close controls. Contract limits remain instructions, not enforced capabilities. No adequate child route: keep feasible work local and disclose the gap.

Read [runtime and configuration](references/platforms.md) only for configuration/CLI diagnosis. Read [routing boundaries](references/routing-matrix.md) for ambiguous choices or live artifact collection.

## Dispatch, observe, accept

Give a compact contract: goal/symptom; baseline and non-goals; read/write/forbidden paths and semantic owner; requested model/effort/working role and real isolation; permitted actions; acceptance; return evidence; escalation. Include a time/checkpoint budget when useful, but do not call it an enforced timeout unless the host supplies one.

Read [lifecycle](references/lifecycle.md) before a multi-round or cancellation-sensitive run. Keep child IDs, ownership, state and acceptance in a small ledger. Send running agents only incremental context; reuse idle agents for follow-up; interrupt obsolete work and verify its state. Never infer completion from a wait timeout or slot release from interruption. Only call close if that tool exists.

Ask for bounded facts, paths, changes, real check outcomes, gaps and escalation. Review artifacts/diffs and provenance; resolve conflicts from original evidence, not votes. Reuse checks only for the same relevant baseline, artifact and environment. Before delivery verify no unneeded child remains running or writing.

Use [evidence packets](references/evidence-packet.md) when structured evidence is required, a run is material, or identity is disputed. Ordinary bounded work can return concise prose with the same relevant facts; do not generate ceremonial JSON for every local decision. The validator checks structure, not truth or semantic correctness.

## Context and ownership discipline

Read [escalation and adjudication](references/escalation.md) when planning exposes uncertainty, a key check fails, shared contracts change or acceptance has unresolved counterexamples. Pause after two same-class failures; preserve issue history across agents. Repair missing inputs first, escalate capability only for reasoning/execution limits, and adjudicate by reproducible evidence rather than model rank. The parent must understand and verify decisive conclusions. Recover to a cheaper adequate route after resolution.

Load only needed references; keep long logs on disk. Avoid whole-history forks and repeating contracts, packets or prior findings. Account for parent context/reasoning, child usage, waiting, retries and integration when evaluating cost.

Production read-only collection is routable by work shape; mutation and final semantic acceptance stay parent-owned. Do not silently relax a project's literal model-owner requirement. Identify the actual rule and resolve it only through authorized changes.