Skip to content
Back to search

self-improving-for-codex

A Codex-native self-improving skill based on AGENTS.md, memories, and nightly review automation.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/godgod126--self-improving-for-codex/badge.svg)](/listings/godgod126--self-improving-for-codex)

HTML

<a href="/listings/godgod126--self-improving-for-codex"><img src="/api/listings/godgod126--self-improving-for-codex/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

将仓库克隆到 Codex 的全局 Skill 目录:

$skillsRoot = Join-Path $env:USERPROFILE '.codex\skills'
New-Item -ItemType Directory -Force -Path $skillsRoot | Out-Null
git clone https://github.com/GODGOD126/self-improving-for-codex.git `
  (Join-Path $skillsRoot 'self-improving-for-codex')

重新打开 Codex,确认技能列表中出现 `self-improving-for-codex`。

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 111 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: self-improving-for-codex
  3. description: Build or maintain a Codex-native self-improving memory loop using global `AGENTS.md`, a persistent memories directory, and optional nightly refinement automation. Use when Codex needs to adapt OpenClaw-style self-improvement ideas to Codex, set up long-term user/profile memory, create `PROFILE.md` / `ACTIVE.md` / `LEARNINGS.md` / `ERRORS.md` / `FEATURE_REQUESTS.md`, add promotion rules from raw learnings into active guidance, or create a recurring memory-refinement automation.
  4. ---
  5. # Self-improving for Codex
  6. ## Overview
  7. Use this skill to give Codex a durable, Codex-native self-improving loop without depending on OpenClaw-only primitives such as `SOUL` or `HEARTBEAT.md`.
  8. This skill assumes one stable rule-entry file and one stable memory directory:
  9. - Global rule entry: `~/.codex/AGENTS.md`
  10. - Global memory directory: prefer `~/.codex/memories/`
  11. ## Workflow
  12. ### 1. Audit the current state
  13. Inspect these locations first:
  14. - global `AGENTS.md`
  15. - the candidate memory directory
  16. - any existing `PROFILE.md`, `ACTIVE.md`, `LEARNINGS.md`, `ERRORS.md`, `FEATURE_REQUESTS.md`
  17. - any existing automation related to nightly review or memory maintenance
  18. If the environment already contains a partial setup, preserve it and extend it instead of replacing it blindly.
  19. ### 2. Establish the memory layout
  20. Create or normalize these files in the global memory directory:
  21. - `PROFILE.md`
  22. - `ACTIVE.md`
  23. - `LEARNINGS.md`
  24. - `ERRORS.md`
  25. - `FEATURE_REQUESTS.md`
  26. Read [references/memory-files.md](references/memory-files.md) when creating or repairing these files.
  27. Use this separation consistently:
  28. - `PROFILE.md`: long-term stable user profile and communication preferences
  29. - `ACTIVE.md`: compact high-priority rules worth reading at the start of future tasks
  30. - `LEARNINGS.md`: reusable learnings and corrections not yet promoted to top-level rules
  31. - `ERRORS.md`: reusable debugging and environment failure knowledge
  32. - `FEATURE_REQUESTS.md`: missing capabilities worth tracking across sessions
  33. ### 3. Wire the loop through `AGENTS.md`
  34. Use `AGENTS.md` as the single Codex-native entry point.
  35. Its job is to tell Codex:
  36. - which memory files to read before starting work
  37. - when to log new entries
  38. - how to classify entries by file
  39. - when to promote content from raw logs into `ACTIVE.md`
  40. - that `AGENTS.md` itself must not be edited automatically unless the user explicitly asks
  41. Read [references/agents-snippet.md](references/agents-snippet.md) before proposing or updating the `AGENTS.md` text.
  42. Unless the user explicitly asks for direct edits, propose the exact `AGENTS.md` snippet in chat and let the user apply it manually.
  43. ### 4. Add an optional nightly review loop
  44. When the user wants recurring maintenance, create a nightly automation that:
  45. - reviews the current memory files
  46. - primarily refines `LEARNINGS.md`, `ERRORS.md`, and `FEATURE_REQUESTS.md`
  47. - proposes or applies safe updates to the memory files
  48. - never edits `AGENTS.md` automatically
  49. Read [references/nightly-review.md](references/nightly-review.md) before designing the automation.
  50. ### 5. Validate the loop
  51. Before finishing, confirm the setup actually forms a loop:
  52. 1. `AGENTS.md` points Codex to `PROFILE.md` and `ACTIVE.md`
  53. 2. the five memory files exist and have sane content
  54. 3. promotion rules are explicit
  55. 4. if automation was requested, the automation prompt clearly explains the refinement-only role and promotion rules
  56. ## Promotion Rules
  57. Apply these promotion rules consistently:
  58. - Promote to `ACTIVE.md` only when the content is stable, cross-task useful, and likely to improve future execution or communication
  59. - Keep `PROFILE.md` limited to durable user identity, style, and preference facts
  60. - Keep temporary context out of `PROFILE.md`
  61. - Keep one-off noise out of all memory files
  62. - If a candidate item is ambiguous, keep it in a raw log or leave it as a proposal instead of promoting it
  63. ## Safety Rules
  64. - Do not assume Codex automatically reads arbitrary memory files; route the loop through `AGENTS.md`
  65. - Do not describe OpenClaw-only mechanisms as if they exist natively in Codex
  66. - Do not edit `AGENTS.md` automatically unless the user explicitly asks
  67. - Prefer updating `ACTIVE.md` over bloating `AGENTS.md`
  68. - Prefer compact, maintainable rules over long narrative summaries
  69. ## Deliverables
  70. When using this skill, aim to produce some or all of these:
  71. - a memory directory with the five core files
  72. - a proposed `AGENTS.md` snippet
  73. - an optional nightly automation prompt
  74. - a short explanation of what was created, what was not changed, and how the loop works
README.md · README · 113 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # Self-improving for Codex
  2. 让 Codex 通过 `AGENTS.md` 与本地 Markdown 记忆文件形成可审计、可维护的长期改进闭环。
  3. > A Codex-native, file-based memory loop. No background service, telemetry, or OpenClaw-only runtime required.
  4. ## 为什么需要它
  5. Codex 不会自动把任意文件当作长期记忆。这个 Skill 使用 Codex 原生的 `AGENTS.md` 作为稳定入口,把长期偏好、可复用经验、环境报错和能力缺口分别写入职责清晰的文件,避免把所有内容堆进一个不断膨胀的提示词。
  6. ```text
  7. AGENTS.md 启动规则
  8. │
  9. ├── 读取 PROFILE.md + ACTIVE.md
  10. │
  11. ▼
  12. 执行日常任务
  13. │
  14. ├── LEARNINGS.md
  15. ├── ERRORS.md
  16. └── FEATURE_REQUESTS.md
  17. │
  18. ▼
  19. 定期精炼、去重与晋升
  20. │
  21. ├── ACTIVE.md
  22. └── PROFILE.md
  23. ```
  24. ## 核心能力
  25. - 审计已有的全局 `AGENTS.md`、记忆目录和相关自动化,优先保留现有配置。
  26. - 建立并维护五类职责分离的记忆文件。
  27. - 生成适合 Codex 的 `AGENTS.md` 集成片段。
  28. - 定义从原始记录晋升到 `ACTIVE.md` / `PROFILE.md` 的保守规则。
  29. - 可选设计夜间精炼任务,用于去重、合并和清理过时内容。
  30. - 明确区分 Codex 原生能力与 OpenClaw 的 `SOUL`、`HEARTBEAT.md` 等专用机制。
  31. ## 安装
  32. 将仓库克隆到 Codex 的全局 Skill 目录:
  33. ```powershell
  34. $skillsRoot = Join-Path $env:USERPROFILE '.codex\skills'
  35. New-Item -ItemType Directory -Force -Path $skillsRoot | Out-Null
  36. git clone https://github.com/GODGOD126/self-improving-for-codex.git `
  37. (Join-Path $skillsRoot 'self-improving-for-codex')
  38. ```
  39. 重新打开 Codex,确认技能列表中出现 `self-improving-for-codex`。
  40. ## 使用
  41. 在 Codex 中直接说明要使用这个 Skill,例如:
  42. ```text
  43. Use $self-improving-for-codex to audit my current global memory setup and propose the safest Codex-native configuration.
  44. ```
  45. 也可以用中文:
  46. ```text
  47. 使用 self-improving-for-codex 检查我现有的 AGENTS.md 和 memories,保留已有内容并补齐缺失的记忆闭环。
  48. ```
  49. 建议先让 Codex 审计并给出变更清单,再授权修改本地文件。默认情况下,Skill 只应提出 `AGENTS.md` 的准确片段;除非用户明确授权,不应自动修改 `AGENTS.md`。
  50. ## 记忆文件分工
  51. | 文件 | 用途 | 不应该存放 |
  52. | --- | --- | --- |
  53. | `PROFILE.md` | 长期稳定的用户画像、沟通和工作偏好 | 临时任务、一次性要求 |
  54. | `ACTIVE.md` | 每次任务开始前都值得读取的高优先级规则 | 冗长过程、未经确认的推断 |
  55. | `LEARNINGS.md` | 可复用经验、用户纠正、知识更新 | 显而易见的小失误和噪音 |
  56. | `ERRORS.md` | 环境故障、工具异常和可复用排障方法 | 没有复用价值的偶发输出 |
  57. | `FEATURE_REQUESTS.md` | 当前尚未闭环、值得长期跟踪的能力需求 | 已完成的普通待办 |
  58. 默认推荐的全局位置是:
  59. ```text
  60. ~/.codex/AGENTS.md
  61. ~/.codex/memories/
  62. ```
  63. 实际路径应根据操作系统和 Codex 配置调整。
  64. ## 仓库结构
  65. ```text
  66. .
  67. ├── SKILL.md # Skill 主流程、边界和交付标准
  68. ├── agents/openai.yaml # Codex 展示名称与默认提示词
  69. └── references/
  70. ├── agents-snippet.md # AGENTS.md 推荐集成片段
  71. ├── memory-files.md # 五类记忆文件模板与职责
  72. └── nightly-review.md # 可选夜间精炼规则和提示词
  73. ```
  74. ## 安全与隐私
  75. - 这个仓库本身只包含 Markdown 指令和 Skill 元数据,没有后台服务或遥测代码。
  76. - 记忆文件可能包含个人偏好、环境路径或项目经验,不要把真实的 `~/.codex/memories` 直接提交到公开仓库。
  77. - 不应把密码、令牌、私钥或完整 `.env` 内容写入记忆文件。
  78. - 夜间精炼只维护已有记忆,不应默认回放全部历史对话。
  79. - 自动化不得自行修改 `AGENTS.md`;顶层规则应由用户保持最终控制。
  80. ## 贡献
  81. 欢迎提交 Issue 或范围清晰的 PR。请说明使用场景、预期行为和安全边界;如果修改了记忆分类或晋升规则,请同步更新相关 reference 文件。
  82. ## License
  83. [MIT License](LICENSE) © 2026 GODGOD126

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.