Codex Workflow Recommender
Read-only: do not create, edit, install, remove, configure, or write externally;
reports never authorize implementation.
Workflow
- Confirm outcome, repo root, surface, and scope; read applicable
AGENTS.md/code_map.md.
- Inventory relevant gates, roots, trusted config, plugins, and MCP. Prefer
structured tools or
rg and current help.
- 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.
- 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.
- Reuse suitable native/installed capability first. Technology detection alone
is only a signal. Rank by impact, evidence, dependency, effort,
reversibility, and Permission/data risk.
- 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
- Skills
- Subagents
- MCP
- Plugins
- Config/hooks
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.
1---2name: codex-workflow-recommender3description: >-4 Audit a repository and current Codex capabilities, then recommend the smallest5 evidence-backed read-only improvement or no change. Use for Codex setup6 optimization, surface selection, unapplied MCP/plugin/subagent plans, 优化 Codex7 工作流, 审阅 Codex 配置, Codex 能力推荐. Exclude direct AGENTS/code-map edits,8 docs questions, skill audits, dynamic workflow implementation, code review,9 and any install/config/write.10version: 1.1.011category: developer-tools-integrations12tags:13 - codex14 - skills15 - mcp16allowed-tools: Read, Glob, Grep, Bash(codex --version), Bash(codex * --help), Bash(codex mcp list *), Bash(codex plugin list *), Bash(git status *), Bash(rg *)17---1819# Codex Workflow Recommender2021Read-only: do not create, edit, install, remove, configure, or write externally;22reports never authorize implementation.2324## Workflow25261. Confirm outcome, repo root, surface, and scope; read applicable27 `AGENTS.md`/`code_map.md`.282. Inventory relevant gates, roots, trusted config, plugins, and MCP. Prefer29 structured tools or `rg` and current help.303. Summarize minimum fields, never raw doctor/config/auth/provider/env values.31 Preserve provenance: `built-in`, `user-config`, `project-config`,32 `plugin-provided`, `installed-enabled`, `installed-disabled`,33 `available-uninstalled`, `unsupported`, `missing evidence`.344. Decide whether persistence is justified. `No change recommended` is valid.35 Smallest owner: one-off -> prompt; repo rule -> AGENTS; learned context ->36 memory; repeated flow -> skill; team bundle -> plugin; live external need ->37 MCP; independent role -> subagent; runtime default -> config/rule; lifecycle38 event -> hook; schedule -> automation.395. Reuse suitable native/installed capability first. Technology detection alone40 is only a signal. Rank by impact, evidence, dependency, effort,41 reversibility, and Permission/data risk.426. Stop at supported decisions; keep OMX conditional and omit irrelevant surfaces.4344## Output Contract4546Return `Outcome`; `Evidence and Unknowns`; supported `Prioritized Recommendations`;47dependency/risk `Implementation Sequence`; `Verification and Rollback`; and48`Approval Options` split by local versus persistent/external action. Each item:49Observed evidence; Existing capability/provenance; scope; prerequisites;50Permission/data risk; confidence or `missing evidence`; Verification;51Rollback/defer reason. Omit empty categories.5253## References5455- [Surface map](references/codex-surface-map.md)56- [Skills](references/skills-reference.md)57- [Subagents](references/subagent-templates.md)58- [MCP](references/mcp-servers.md)59- [Plugins](references/plugins-reference.md)60- [Config/hooks](references/hooks-patterns.md)6162Load only relevant references. CLI help overrides examples and dated facts63override memory; otherwise use `missing evidence`.64Route direct AGENTS/code-map edits to `agents-md-improver`, fact questions to65`openai-docs`, skill reviews to `agent-skill-review`/`yao-meta-skill`, and66dynamic workflow builds to `codex-dynamic-workflows`.
bahayonghang