Skip to content

Latest commit

 

History

History
25 lines (18 loc) · 2.44 KB

File metadata and controls

25 lines (18 loc) · 2.44 KB

Changelog

[0.1.5] - 2026-03-04

Fixed

  • Doc discovery scans wrong directory — Reflex was using cwd (the agent's current working directory) instead of CLAUDE_PROJECT_DIR (the project root). When the agent cds into subdirectories, docs and skills at the project root were invisible. Now uses CLAUDE_PROJECT_DIR first, falling back to cwd.

[0.1.1] - 2026-02-20

Added

  • OpenClaw plugin (hooks/openclaw/): Uses the before_agent_start plugin hook to inject context via prependContext, which prepends to the user's prompt string. The agent sees the injection as part of the current message rather than buried in the system prompt. Discovers skills from both .openclaw/skills/ and .claude/skills/ for cross-tool compatibility. Reads conversation history directly from event.messages — no session file parsing needed.

Fixed

  • Binary discovery (hooks/claude-code/reflex-hook.py): Subprocess hooks don't inherit the interactive shell's PATH, so go install binaries in ~/go/bin/ were invisible. Now checks REFLEX_BIN env var, then shutil.which(), then common install paths (~/go/bin, ~/.local/bin, /opt/homebrew/bin, /usr/local/bin).
  • Injection wording: Replaced [Reflex]-prefixed tags with a direct imperative instruction. Docs get a bulleted list under "read these files now"; skills get a plain "use this skill" instruction.
  • CLI example in README: Updated registry format from old flat []RegistryItem array to current grouped {docs, skills} object.

[0.1.0] - 2026-02-19

Added

  • reflex route CLI: Framework-agnostic context router. Reads JSON from stdin (messages, registry, session state), calls an OpenAI-compatible LLM, returns {"docs": [...], "skills": [...]}.
  • Provider-agnostic architecture: Any OpenAI-compatible API works via ~/.config/reflex/config.yaml. Default: Kimi K2.5 (kimi-k2.5-preview at api.moonshot.ai).
  • Session state filtering: Items already read/used this session are excluded from routing decisions — no duplicate injections.
  • Fast path: If registry is empty after filtering, returns immediately without an API call.
  • Claude Code hook (hooks/claude-code/reflex-hook.py): UserPromptSubmit hook that auto-discovers skills from .claude/skills/*/SKILL.md and docs from any *.md file with summary + read_when frontmatter. No registry file to maintain. Calls reflex route and injects additionalContext with relevant docs and skills.