Ensure that only executable files are copied from LOCAL_TOOLS_DIR to avoid copying non-executable or invalid files. Added check for executable permission (-x) alongside file existence (-f).
- Augmente MAX_TOOL_ITERATIONS à 20 et timeout HTTP à 300s pour tolérer les erreurs/temps d'attente
- Ajoute des directives de gestion d'erreurs dans CLAUDE.md pour les appels MCP
- Introduit un mode bootstrap (`CRAZYCLAUDE_BOOTSTRAP=1`) pour éviter les git pull en boucle
- Découpe les tâches longues et préconise des corrections ciblées pour les artefacts mineurs
- Rename MCP tool `qwen3_task` → `task` in server.py and all references
- Enforce direct use of `mcp__qwen3__task` (no sub-agents) in docs
- Add bootstrap logic to install.sh/update_crazyclaude.sh for repo self-installation
- Introduce tools-local/ directory for local tool overrides
- Add test suite (test_qwen3.sh, test_qwen3.md) to validate Qwen3 file generation
- Update .gitignore: ignore test_qwen3_output/, add missing .claude/
- Add system prompt to server.py for fs_write enforcement
- Update agent docs (qwen3-worker, code-reviewer, delegation-rules, task-planner) to reflect new tool name and workflow
- Add system instruction: always use fs_write/fs_patch, never display content in response
Introduit un script bash (.claude/update_crazyclaude.sh) pour automatiser la mise à jour de CrazyClaude et llm-functions depuis leurs dépôts upstream, puis reconstruire les outils argc. Le script supporte le mode --dry-run et vérifie la présence des dépendances (git, argc, jq).
Replace LM Studio direct integration with aichat-based MCP server, enabling Qwen3-Coder to autonomously use filesystem, shell and web tools via llm-functions. Key changes:
- Update server.py to launch aichat on a dynamic port and handle tool_calls loop
- Add llm-functions directory with prebuilt tools (fs_cat, fs_ls, execute_command, etc.)
- Revise delegation rules to remove file count limit and emphasize autonomous tool usage
- Update agents and skills docs to reflect new architecture and workflow
- Add setup instructions for aichat, argc, jq and llm-functions build step
- Bump MCP server version to 3.0.0
Added comprehensive documentation for agentic AI assistants, including build/test/lint commands, code style guidelines per language (Rust, Go, Python, R, Bash), naming conventions, error handling strategies, documentation practices, and integration rules for Cursor/Copilot. Also includes hooks configuration and MCP server setup details.
Update configuration in .claude/CLAUDE.md with expanded stack (Python, R, bash), detailed linting/test commands, and stricter bash safety rules (no uncontrolled eval, prefer shellcheck). Adjust MCP servers to use configurable port 8888 instead of hardcoded 1248, and improve code formatting in agent_lm.py and server.py. Update README.md with concrete repository URL for cloning/subtree operations, and add missing tool requirements (aichat, jq, argc).
Add comprehensive Claude Code configuration for Rust/Go projects, including:
- CLAUDE.md with delegation heuristics and guardrails
- Subagents: qwen3-worker (LM Studio), code-reviewer, task-planner
- Skills for delegation rules and Rust/Go conventions
- Hooks: post-write-lint, pre-bash-guard, subagent-stop-log
- MCP server (stdio) to interface with LM Studio
- README and technical report documenting the architecture
