reflexion

Closed-loop learning for AI coding agents. Auto-captures errors and corrections, recalls relevant past solutions when similar situations arise, and promotes...

This skill appears to do what it says: it captures command outputs and user corrections and stores learnings under .reflexion/, recalls relevant entries before prompts, and can auto-append promoted rules to CLAUDE.md. Before installing, consider: (1) .reflexion/ contains captured command outputs and commands (the scripts attempt to redact tokens but redaction is heuristic and may miss things) — add .reflexion/ to your .gitignore and review entries before committing; (2) installing the hooks globally enables cross-project capture (privacy risk for multi-project/multi-user machines); (3) promoted rules will modify project files (CLAUDE.md) automatically when criteria are met — if you prefer manual control, run promote.sh manually rather than enabling background promotion; (4) the code requires common tools (git, python3) but no secrets; review the scripts yourself if you want to confirm redaction/behavior. Overall the skill is internally consistent and not requesting unexpected privileges or credentials.

Type: OpenClaw Skill Name: reflexion Version: 1.0.1 The reflexion skill bundle implements a local 'Verbal Reinforcement Learning' loop for AI agents, allowing them to capture command errors and recall solutions in future sessions. The scripts (capture.sh, recall.sh, promote.sh) operate entirely on the local filesystem within the project's .reflexion directory and include explicit logic to redact sensitive information like API keys and Bearer tokens before storage. The behavior is transparently documented and strictly aligned with the stated purpose of improving agent performance through episodic memory.