Extract reusable lessons from completed Codex sessions and turn them into future skills.
continuous-learning runs at the end of a session and looks for patterns worth preserving:
- error resolutions
- user corrections
- repeatable workarounds
- debugging techniques
- project-specific conventions
The goal is to make useful behavior durable instead of rediscovering it in every new session.
- A Stop hook runs after a session.
- The session is checked against configurable thresholds.
- Useful patterns are extracted.
- Learned material is saved under the configured learned-skills directory.
Edit config.json to adjust session length, extraction threshold, output path, detected pattern types, and ignored pattern types.
The included evaluate-session.sh script is intended to run as a Stop hook from your agent configuration.
Example target:
~/.Codex/skills/continuous-learning/evaluate-session.sh
This is the simpler first-generation version. For hook-based observation, confidence scoring, instinct export/import, and evolution into skills or commands, see continuous-learning-v2.
MIT. See LICENSE.