SkillOpt Sleep for Codex runs the skillopt_sleep engine against local Codex sessions. It mines recurring tasks, replays them under your current skill, and proposes bounded edits to a Codex skill you name with a target path.
With the default gate on, only edits that raise a held-out score are kept. Proposals are staged for review and live files change only on explicit adoption, with backups.
What You Get
- Run actions for status, harvest, dry-run, run, adopt, schedule and unschedule
- Runner scripts for macOS and Linux (shell) and for Windows (CMD and PowerShell)
- A mock backend for a first smoke check, plus claude, codex, copilot, handoff and azure_openai backends
- Staged proposals with a report, so the agent can show baseline and candidate scores and the exact edits
Common Use Cases
- Teach a Codex skill from the tasks you repeat across sessions
- Schedule a nightly run for a Codex-targeted skill after setting the transcript source and target path in the config file
- Dry-run with the mock backend to check the setup without spending budget
- Run on a Windows machine through the PowerShell or CMD runner
What Makes It Unique
This variant is written around Codex limits. It requires an explicit target skill path, warns that the shared engine would otherwise write a Claude-managed skill, and notes that the scheduler does not remember the source or target path, so they go into the config file first. It cites one recorded benchmark run and states it is evidence for that setup only.
Under The Hood
The skill is prose that tells the agent to run a bundled runner script (run-sleep.sh, or the Windows CMD and PowerShell equivalents) located through the SKILLOPT_SLEEP_REPO variable, which starts the Python skillopt_sleep engine. It reads archived Codex sessions from the Codex home folder, read-only, and removes secret-shaped strings before use, though the skill says this is not a guarantee. A real backend sends truncated task content to its provider. Runs write staging files in the project, adopt overwrites the target skill and project CLAUDE.md after a backup, and schedule installs a nightly cron entry. The validation example also runs a benchmark script that needs an external eval data folder.