# Copy to machine.yaml (never commit it) and fill in. Read by twinlab/settings.py. Every path must exist. claude: shell: /usr/bin/zsh # SHELL for Claude Code's Bash tool; must print nothing on stderr when started bin: ~/.local/bin/claude # tested with Claude Code 2.1.319 install_dir: ~/.local/share/claude config_dir: ~/.claude # Claude Code config dir; for a Claude actor sign in with CLAUDE_CONFIG_DIR=~/.claude claude launch_flags: ["--enable-auto-mode", "--dangerously-skip-permissions"] proxy: "" # HTTP(S)_PROXY for every Claude Code process; empty for none no_proxy: "localhost,127.0.0.0/8" codex: # the values are used only with --actor-agent codex; the keys must exist bin: ~/.local/bin/codex # tested with Codex 0.166 model_catalog: ~/.codex/model_catalog.json # Codex model_catalog_json, for model names Codex does not know provider: # a [model_providers.] table of Codex's config.toml name: openai base_url: https://api.openai.com/v1 env_key: OPENAI_API_KEY wire_api: responses api_key: sk-... web_search: live # live | cached | disabled proxy: null # null: the provider is reached directly model_backends: # non-Claude models inside Claude Code; one entry per model id deepseek-flash: endpoint: deepseek # URL and key come from DEEPSEEK_API_KEY and DEEPSEEK_BASE_URL_* (README.md); # without this line, write base_url and auth_token here llm_proxy: # used only when DEEPSEEK_BASE_URL_OPENAI is set (scripts/llm_proxy.py) url: http://localhost:4001 master_key: sk-local-twinlab # the proxy's own token (--master-key), not the DeepSeek key proxy: null # null: reached directly config_dir: ~/.config/twinlab/claude-ds # an empty directory for the TwinLab and judge sessions cli_env: CLAUDE_CODE_MAX_RETRIES: "110" CLAUDE_CODE_MAX_CONTEXT_TOKENS: "1000000" # the model's real window; unknown models default to 200k CLAUDE_CODE_MAX_OUTPUT_TOKENS: "128000" # thinking counts as output; the default 32k cuts judge answers actor_default_effort: {} # used when --actor-effort is not given, e.g. claude-opus-5-5: high sandbox: bwrap: /usr/bin/bwrap ro_binds: [~/.local/bin, ~/.local/share/claude] # agent binaries; the Python running the benchmark is added itself # own_proc: true # optional; unset = detected at start (false inside Docker, where bwrap cannot mount proc)