# Codex Workflow Recommender

> Audit a repository and current Codex capabilities, then recommend the smallest evidence-backed read-only improvement or no change. Use for Codex setup optimization, surface selection, unapplied MCP/plugin/subagent plans, 优化 Codex 工作流, 审阅 Codex 配置, Codex 能力推荐. Exclude direct AGENTS/code-map edits, docs questions, skill audits, dynamic workflow implementation, code review, and any install/config/write.

- Skill: `bahayonghang/codex-workflow-recommender` (Agent Skill, multi-file: 22 files)
- Install (CLI): `npx skillmds@latest add bahayonghang/codex-workflow-recommender`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bahayonghang/codex-workflow-recommender/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: bahayonghang (https://skillmd.com/u/bahayonghang)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/bahayonghang/codex-workflow-recommender

---


# Codex Workflow Recommender

Read-only: do not create, edit, install, remove, configure, or write externally;
reports never authorize implementation.

## Workflow

1. Confirm outcome, repo root, surface, and scope; read applicable
   `AGENTS.md`/`code_map.md`.
2. Inventory relevant gates, roots, trusted config, plugins, and MCP. Prefer
   structured tools or `rg` and current help.
3. Summarize minimum fields, never raw doctor/config/auth/provider/env values.
   Preserve provenance: `built-in`, `user-config`, `project-config`,
   `plugin-provided`, `installed-enabled`, `installed-disabled`,
   `available-uninstalled`, `unsupported`, `missing evidence`.
4. Decide whether persistence is justified. `No change recommended` is valid.
   Smallest owner: one-off -> prompt; repo rule -> AGENTS; learned context ->
   memory; repeated flow -> skill; team bundle -> plugin; live external need ->
   MCP; independent role -> subagent; runtime default -> config/rule; lifecycle
   event -> hook; schedule -> automation.
5. Reuse suitable native/installed capability first. Technology detection alone
   is only a signal. Rank by impact, evidence, dependency, effort,
   reversibility, and Permission/data risk.
6. Stop at supported decisions; keep OMX conditional and omit irrelevant surfaces.

## Output Contract

Return `Outcome`; `Evidence and Unknowns`; supported `Prioritized Recommendations`;
dependency/risk `Implementation Sequence`; `Verification and Rollback`; and
`Approval Options` split by local versus persistent/external action. Each item:
Observed evidence; Existing capability/provenance; scope; prerequisites;
Permission/data risk; confidence or `missing evidence`; Verification;
Rollback/defer reason. Omit empty categories.

## References

- [Surface map](references/codex-surface-map.md)
- [Skills](references/skills-reference.md)
- [Subagents](references/subagent-templates.md)
- [MCP](references/mcp-servers.md)
- [Plugins](references/plugins-reference.md)
- [Config/hooks](references/hooks-patterns.md)

Load only relevant references. CLI help overrides examples and dated facts
override memory; otherwise use `missing evidence`.
Route direct AGENTS/code-map edits to `agents-md-improver`, fact questions to
`openai-docs`, skill reviews to `agent-skill-review`/`yao-meta-skill`, and
dynamic workflow builds to `codex-dynamic-workflows`.