Skip to content
Back to skills

Codex Agent Dispatch

ASecurity

Automatically use for every Codex implementation or independent-review dispatch. Select the neutral role, construct a minimal context packet, invoke a fresh native subagent with resolved model/reasoning, and persist adapter-owned evidence; degrade explicitly when subagents are unavailable.

  • 6 stars
  • 0 votes
  • 0 copies
  • 4 views
  • Added September 3, 2026
ai-agents

Security analysis

A100/100

Scanned September 3, 2026

npx -y skills add Eduardo-Salvador/Agent-Harness-Kit --skill codex-agent-dispatch --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Codex Agent Dispatch?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Codex Agent Dispatch
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/eduardo-salvador-codex-agent-dispatch/badge)](https://www.skillsdirectory.com/skills/eduardo-salvador-codex-agent-dispatch)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: codex-agent-dispatch
description: Automatically use for every Codex implementation or independent-review dispatch. Select the neutral role, construct a minimal context packet, invoke a fresh native subagent with resolved model/reasoning, and persist adapter-owned evidence; degrade explicitly when subagents are unavailable.
---

# Codex agent dispatch

Follow `../../../docs/contracts/CODEX-AGENT-DISPATCH.md` and `../../../adapters/codex.md`. This is the executable Codex bridge from neutral role/task artifacts to a live agent context.

1. Build a request from the pinned `TASK.md`, resolved `harness.model-dispatch/v1`, and current capability evidence. Include only the role, task SPEC, approved authority/rule references, `read_set`, and `impact_set`; never include the conversation, full plan, or broad repository context.
2. Run `agent-harness codex-dispatch <request.json>` and inspect the machine-readable plan. For implementation, retain the requested specialist identity and use `role:generic-specialist` when no dedicated neutral executor exists. For review, force `role:reviewer-integrator`, a distinct identity, and `fork_turns: none`.
3. When status is `ready-to-dispatch`, invoke the named native operation using exactly `native_call.arguments`, including model and reasoning effort. Do not claim dispatch before the host returns an agent/context reference.
4. Persist the sanitized host response, then run `agent-harness codex-dispatch <plan.json> --response <response.json>`. Link the resulting `harness.codex-agent-dispatch/v1` record from the task/graph transition before marking the child active.
5. An implementation request with no subagent capability records `sequential-fallback` and runs visibly in the orchestrator context. A review request records `manual-fresh-context-required`; same-context review is forbidden even under fallback.
6. Implementer and reviewer identities, contexts, packets, and dispatch evidence remain distinct. A user-visible Codex task is created only when the user explicitly requests one; the normal route is an internal subagent.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.