Skip to content

SkillOpt-Sleep Codex 集成反馈 #57

New issue
New issue
Closed
Closed
SkillOpt-Sleep Codex 集成反馈#57

Description

@bjvgukv25842-cmyk
LITMUS (bjvgukv25842-cmyk)
opened on Jun 14, 2026

我在 macOS + Codex Desktop / Codex CLI 0.140.0-alpha.2 上尝试安装并运行 plugins/codex。整体思路可行，但当前 Codex 集成有一些明显边界和功能问题。

  1. Codex integration 仍强依赖 Claude 路径
    当前 skillopt_sleep/config.py 默认使用：
    ~/.claude/projects
    ~/.claude/history.jsonl
    ~/.claude/skills
    cycle.py 也默认写项目下的：
    CLAUDE.md
    这导致 plugins/codex 虽然文档说支持 Codex，但实际不会开箱即用地读取 Codex Desktop/CLI 的历史记录。我的 Codex 历史在：
    ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl
    ~/.codex/history.jsonl
    建议：增加真正的 Codex transcript adapter，识别 session_meta、turn_context、response_item.message、response_item.function_call 等结构，并把 Codex memory 目标改为项目级 AGENTS.md，skill 目标改为 ~/.codex/skills/.../SKILL.md。
  2. Codex skill 安装路径疑似不对
    plugins/codex/install.sh 把 skill 安装到：
    ~/.agents/skills/skillopt-sleep/SKILL.md
    但我当前 Codex Desktop 实际加载的是：
    ~/.codex/skills//SKILL.md
    导致安装脚本可能显示成功，但 Codex Desktop 不一定能看到这个 skill。
    建议：检测当前 Codex 版本支持的 skill path，至少同时支持 ~/.codex/skills，或在 README 中明确版本差异。
  3. 文档说 custom prompts 稳定，但 Codex 当前更偏 skills / plugins
    plugins/codex/README.md 说使用 ~/.codex/prompts/.md 作为 /sleep。但当前 Codex 有 codex plugin 命令，也有 skills 机制；custom prompt 的长期稳定性需要再确认。
    建议：提供一个正式 Codex plugin manifest，或明确说明这是“临时 prompt + skill 适配”，不是 Codex 原生插件包。
  4. run-sleep.sh 无法指定 Python 路径
    我的系统 Python 是 3.9.6，但 Codex Desktop 自带 Python 3.12。原始 runner 只在 PATH 里找 python3.12/python3.11/python3.10/python3，无法通过环境变量直接指定。
    建议：支持：
    SKILLOPT_SLEEP_PYTHON=/path/to/python3
    否则很多 macOS 用户会因为系统 Python 版本低而安装失败。
  5. 真实 Codex backend 运行成本和耗时不可控
    我用：
    --backend codex --max-tasks 1 --edit-budget 1
    跑真实学习。一次成功 run 大约花了两分钟，并生成 accepted proposal；另一次 run 卡在 candidate memory validation 的 codex exec 子进程里，数分钟无输出，需要手动中断。
    问题点：
    每个 replay/judge/reflect 都 spawn codex exec
    缺少实时进度日志
    缺少 per-call timeout 的 CLI 参数
    max_tokens_per_night 配置存在，但未看到真实强制预算控制
    建议：增加 --max-tasks、--codex-timeout、--max-codex-calls、实时阶段日志，并真正执行 token/time budget。
  6. Accepted 后会 stage 未变化文档
    我遇到一次 accepted memory edit，但 staging 同时生成了 proposed_SKILL.md。当 live skill 文件里已有测试残留/旧内容时，它会被一起 stage，哪怕本轮 accepted edit 只针对 memory。
    建议：stage 前比较 result.new_skill != original_skill、result.new_memory != original_memory，只有实际变化的文件才写入 manifest。
  7. 测试或 mock run 容易污染真实用户状态
    在本地测试过程中，测试用的 mock persona 规则曾写入真实：
    ~/.codex/skills/skillopt-sleep-learned/SKILL.md
    ~/.skillopt-sleep/state.json
    这会导致后续真实 run 的 staged proposal 混入 mock 示例规则，例如：
    Always wrap the final answer in ... tags.
    Report arXiv ids in the exact form arXiv:XXXX.XXXXX.
    建议：测试默认必须使用 isolated temp state_dir / agent_home，不应触碰真实 home；CLI demo 也应显式提醒 state 写入位置。
  8. Staging 文件命名仍是 proposed_CLAUDE.md
    即使目标是 Codex 的 AGENTS.md，staging 里仍写：
    proposed_CLAUDE.md
    这会让用户误解它要改 Claude 配置。
    建议：根据 platform/memory filename 生成：
    proposed_AGENTS.md
    proposed_CLAUDE.md
    或统一叫 proposed_memory.md。
  9. Secret redaction 配置存在但实现不明显
    config.py 有：
    "redact_secrets": True
    但我没有看到 harvest 阶段对 Codex/Claude transcript 做明确 secret redaction。由于历史会话可能包含 API key、路径、私有代码、业务信息，这个边界很重要。
    建议：在 harvest 后、LLM miner 前做强制 redaction，并在 report 里说明 redaction 是否启用。
  10. lookback_hours 似乎没有实际参与 harvest
    配置里有：
    lookback_hours: 72
    但当前 cycle 主要依赖 state.last_harvest_for(project)；首次运行时没有看到按 lookback window 限制历史读取。
    建议：首次运行也应按 lookback_hours 限制，避免第一次就扫描大量历史并触发大量 LLM mining。
  11. README 中“Codex 支持的路径”与现实有偏差
    README 写 Codex skills 在：
    ~/.agents/skills//SKILL.md
    但我的 Codex Desktop 环境中可见 skills 是：
    ~/.codex/skills
    建议更新 README，并注明 Codex Desktop、CLI、旧版/新版路径差异。
  12. 缺少清晰的“不会随 Codex 启动自动运行”的说明
    用户容易误以为安装后会后台自动学习、同步启动、持续消耗 token。实际不是，只有手动 run --backend codex 才会消耗。
    建议在 README 明确写：
    Installation does not start a daemon.
    No learning happens on Codex startup.
    Only /sleep run or skillopt_sleep run triggers replay and model calls.

Activity

  1. Yif-Yang commented on Jun 17, 2026

    @Yif-Yang
    Yifan Yang (Yif-Yang)
    on Jun 17, 2026
    Contributor

    Thanks for the detailed feedback! Several of the issues you raised have been addressed in recent PRs:

    ✅ Resolved: Codex transcript support

    PR #48 added --source codex to harvest directly from Codex Desktop's ~/.codex/archived_sessions/*.jsonl:

    # Harvest from Codex Desktop archived sessions
bash plugins/run-sleep.sh run --project "$(pwd)" --source codex

# Non-default Codex home
bash plugins/run-sleep.sh run --project "$(pwd)" --source codex --codex-home /path/to/.codex

# Auto-detect (tries Codex first, falls back to Claude)
bash plugins/run-sleep.sh run --project "$(pwd)" --source auto

    The new harvest_codex.py module correctly parses Codex's session_meta, response_item.message, and related structures, while filtering out tool arguments, raw tool outputs, and sensitive content.

    ✅ Resolved: Codex skill-first

    PR #49 migrated the Codex integration from the deprecated custom prompt to a SKILL.md-based entrypoint. The /sleep slash command is no longer used — the skill is the sole entrypoint.

    🔜 To be improved: default paths and adoption target

    Your point about cycle.py defaulting to CLAUDE.md instead of AGENTS.md, and the skill target path issue, are valid and will be addressed in a follow-up update.

    Thanks for testing and reporting!

  2. Yifan Yang (Yif-Yang)
    added a commit that references this issue on Jun 20, 2026

    fix: add SKILLOPT_SLEEP_PYTHON override + lookback_hours first-run fa…

    889238b

  3. Yif-Yang commented on Jun 20, 2026

    @Yif-Yang
    Yifan Yang (Yif-Yang)
    on Jun 20, 2026
    Contributor

    Hi, following up — several more points from your feedback have now been addressed:

    ✅ Newly fixed (PR #74)

    • Point 4 (Python path): run-sleep.sh now supports SKILLOPT_SLEEP_PYTHON=/path/to/python3.12 to explicitly set the interpreter. This solves the macOS system Python 3.9 issue.
    • Point 10 (lookback_hours): First run now applies the lookback_hours config (default 72h) as a time cutoff, so it won't scan your entire transcript history and trigger massive LLM mining. Use --lookback-hours 0 to explicitly opt out and scan everything.

    ✅ Previously fixed (PRs #48, #49, #60, #72, #73)

    • Point 1: --source codex reads Codex Desktop transcripts ✅
    • Point 3: Custom prompts removed, skill-first ✅
    • Point 5: --max-tasks, --max-sessions, --progress, per-call timeout all added ✅
    • Point 9: Secret redaction implemented in harvest_codex.py ✅
    • Point 12: Docs clarify no daemon, no auto-spend ✅

    🔜 Remaining (lower priority, planned for follow-up)

    • Point 2/11 (install path ~/.agents/skills vs ~/.codex/skills): Codex Desktop is still in alpha and the skill path may change — we'll update once it stabilizes.
    • Point 8 (proposed_CLAUDE.md naming): Planned to make platform-aware (proposed_AGENTS.md for Codex).
    • Point 6 (staging unchanged files): Will add a diff check before staging.
    • Point 7 (test isolation): Will document state directory location more prominently.

    Thanks again for the thorough feedback — it's been very helpful!

  4. Yifan Yang (Yif-Yang)
    closed this as completedon Jun 25, 2026

  5. Yif-Yang commented on Sep 30, 2026

    @Yif-Yang
    Yifan Yang (Yif-Yang)
    on Sep 30, 2026
    Contributor

    A concrete follow-up on point 6: #277 has now merged after official CI passed. Accepted cycles stage only skill/memory documents whose text actually changed. The regression covers skill-only, memory-only, both/neither changed, and whitespace-only changes, including manifest flags and artifact presence.

    This update concerns unchanged-document staging only. It does not reopen this issue or claim that every historical integration concern is resolved.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions