| 1 | # ╔══════════════════════════════════════════════════════════════════════════════╗ |
| 2 | # ║ Codewhale Configuration ║ |
| 3 | # ║ ║ |
| 4 | # ║ Terminal coding agent for any model — open models first. ║ |
| 5 | # ╚══════════════════════════════════════════════════════════════════════════════╝ |
| 6 | |
| 7 | # See `docs/CONFIGURATION.md` for how config is loaded (profiles, env overrides, etc.). |
| 8 | |
| 9 | # ───────────────────────────────────────────────────────────────────────────────── |
| 10 | # Active provider + DeepSeek defaults |
| 11 | # ───────────────────────────────────────────────────────────────────────────────── |
| 12 | # Choose which provider to use by default. Per-provider credentials live in the |
| 13 | # `[providers.*]` sections near the bottom of |
| 14 | # this file — keeping both stored at once means `/provider deepseek` and |
| 15 | # `/provider nvidia-nim` (or `--provider openai`, `--provider wanjie-ark`, |
| 16 | # `--provider volcengine`, `--provider openrouter`, `--provider xiaomi-mimo`, |
| 17 | # `--provider fireworks`, `--provider siliconflow`, `--provider siliconflow-CN`, |
| 18 | # `/provider arcee`, `/provider moonshot`, `/provider qianfan`, `/provider sglang`, `/provider vllm`, |
| 19 | # `/provider ollama`, `/provider huggingface`, `/provider stepfun`, `/provider openmodel`, |
| 20 | # `/provider opencode-go`, `/provider opencode-zen`, `/provider meta`, `/provider xai`) toggle without having to re-enter keys. Top-level |
| 21 | # `api_key` / `base_url` are |
| 22 | # still read as DeepSeek defaults when `[providers.deepseek]` is absent |
| 23 | # (backward compatibility). |
| 24 | provider = "deepseek" # deepseek | deepseek-cn | deepseek-anthropic | nvidia-nim | openai | atlascloud | wanjie-ark | volcengine | openrouter | orcarouter | xiaomi-mimo | novita | fireworks | siliconflow | siliconflow-CN | arcee | moonshot | zai | stepfun | minimax | sglang | vllm | ollama | ollama-cloud | huggingface | modelscope | together | qianfan | openai-codex | anthropic | openmodel | deepinfra | sakana | longcat | opencode-go | opencode-zen | meta | xai | mistral | telecomjs | google | edenai | zenmux | csdn | concentrate | modelstudio-token-plan | modelstudio-coding-plan | codewhale |
| 25 | api_key = "YOUR_DEEPSEEK_API_KEY" # must be non-empty |
| 26 | base_url = "https://api.deepseek.com/beta" |
| 27 | # provider = "deepseek-cn" # legacy alias (official host is still https://api.deepseek.com) |
| 28 | # base_url = "https://api.deepseek.com" # opt out of DeepSeek beta features |
| 29 | # Optional custom model request headers for OpenAI-compatible gateways. |
| 30 | # Authorization and Content-Type are managed by the client and cannot be overridden here. |
| 31 | # http_headers = { "X-Model-Provider-Id" = "your-model-provider" } |
| 32 | |
| 33 | # ───────────────────────────────────────────────────────────────────────────────── |
| 34 | # Default Models |
| 35 | # ───────────────────────────────────────────────────────────────────────────────── |
| 36 | # DeepSeek V4 family: |
| 37 | # deepseek-v4-pro — flagship reasoning model on DeepSeek Platform |
| 38 | # deepseek-v4-flash — fast, cost-efficient (legacy aliases: deepseek-chat, deepseek-reasoner) |
| 39 | # deepseek-ai/deepseek-v4-pro — NVIDIA NIM-hosted Pro model ID |
| 40 | # deepseek-ai/deepseek-v4-flash — NVIDIA NIM-hosted Flash model ID |
| 41 | # deepseek/deepseek-v4-pro — default OpenRouter DeepSeek model ID |
| 42 | # arcee-ai/trinity-large-thinking — OpenRouter Arcee Trinity Large Thinking |
| 43 | # xiaomi/mimo-v2.5-pro — OpenRouter Xiaomi MiMo 2.5 Pro |
| 44 | # xiaomi/mimo-v2.5 — OpenRouter Xiaomi MiMo 2.5 |
| 45 | # z-ai/glm-5.1 — OpenRouter Z.AI GLM 5.1 |
| 46 | # z-ai/glm-5.2 — OpenRouter Z.AI GLM 5.2 |
| 47 | # z-ai/glm-5.3 — OpenRouter Z.AI GLM 5.3 (live on Z.ai since 2026-08-13; |
| 48 | # metadata inherited from 5.2, unpriced) |
| 49 | # z-ai/glm-5.3-flash — OpenRouter Z.AI GLM 5.3 Flash (1M multimodal; $0.15/$0.50 list) |
| 50 | # z-ai/glm-5-turbo — OpenRouter Z.AI GLM 5 Turbo (scout fast sibling of 5.2) |
| 51 | # GLM-5.3 — default direct Z.AI Coding Plan model (live since 2026-08-13; |
| 52 | # metadata inherited from 5.2, unpriced) |
| 53 | # GLM-5.3-Flash — direct Z.AI GLM 5.3 Flash (faster/explore sibling of 5.3) |
| 54 | # GLM-5.2 — direct Z.AI GLM 5.2 (previous default; explicit selections keep it) |
| 55 | # GLM-5.1 — direct Z.AI smaller model |
| 56 | # GLM-5-Turbo — direct Z.AI fast model (scout fast sibling of 5.2) |
| 57 | # step-3.7-flash — default direct StepFun / StepFlash model ID |
| 58 | # kimi-k3 — direct Moonshot K3 model ID (1M context) |
| 59 | # kimi-k2.7-code — default direct Moonshot/Kimi K2.7 model ID |
| 60 | # k3 — Kimi Code membership K3 API model ID |
| 61 | # kimi-for-coding — Kimi Code membership K2.7 compatibility ID |
| 62 | # kimi-for-coding-highspeed — Kimi Code membership high-speed roster ID |
| 63 | # gpt-4.1 — default generic OpenAI-compatible model ID |
| 64 | # deepseek-ai/deepseek-v4-flash — default AtlasCloud model ID |
| 65 | # deepseek-reasoner — default Wanjie Ark model ID |
| 66 | # mimo-v2.5-pro — default Xiaomi MiMo model ID |
| 67 | # mimo-v2.5-pro-ultraspeed — Xiaomi MiMo V2.5 Pro UltraSpeed chat model ID |
| 68 | # mimo-v2.5 — Xiaomi MiMo V2.5 Omni model ID |
| 69 | # mimo-v2.5-tts — Xiaomi MiMo speech/TTS model ID |
| 70 | # mimo-v2.5-tts-voicedesign — Xiaomi MiMo voice-design TTS model ID |
| 71 | # mimo-v2.5-tts-voiceclone — Xiaomi MiMo voice-clone TTS model ID |
| 72 | # accounts/fireworks/models/deepseek-v4-pro — Fireworks AI Pro model ID |
| 73 | # deepseek-ai/DeepSeek-V4-Pro — SiliconFlow hosted Pro model ID |
| 74 | # deepseek-ai/DeepSeek-V4-Flash — SiliconFlow hosted Flash model ID |
| 75 | # trinity-large-thinking — default direct Arcee AI API model ID |
| 76 | # trinity-large-preview — direct Arcee AI API model ID |
| 77 | # deepseek-ai/DeepSeek-V4-Pro — SGLang self-hosted Pro model ID |
| 78 | # deepseek-ai/DeepSeek-V4-Flash — SGLang self-hosted Flash model ID |
| 79 | # auto — auto-select between flash and pro based on task complexity. |
| 80 | # Complex tasks (debugging, refactoring, architecture) → pro; |
| 81 | # simple tasks (lookups, formatting, Q&A) → flash. |
| 82 | default_text_model = "deepseek-flash" |
| 83 | |
| 84 | # ───────────────────────────────────────────────────────────────────────────────── |
| 85 | # Thinking Mode (DeepSeek V4 reasoning effort) |
| 86 | # ───────────────────────────────────────────────────────────────────────────────── |
| 87 | # "off" — disables chain-of-thought (thinking.type = disabled) |
| 88 | # "low" — compat-maps to "high" server-side |
| 89 | # "medium" — compat-maps to "high" server-side |
| 90 | # "high" — reasoning_effort = high (DeepSeek default) |
| 91 | # "max" — reasoning_effort = max (deepest reasoning) |
| 92 | # |
| 93 | # Ctrl+T in the TUI cycles the reasoning tier. The header shows the current |
| 94 | # tier as a ⚡ chip. (Shift+Tab cycles the permission posture — Ask / |
| 95 | # Auto-Review / Full Access — not the reasoning tier.) |
| 96 | reasoning_effort = "max" |
| 97 | |
| 98 | # NOTE: `show_thinking`, `thinking_default_expanded`, `thinking_preview_lines`, |
| 99 | # `help_expand_groups`, `pin_last_prompt`, and `cost_currency` live in |
| 100 | # `~/.codewhale/settings.toml`, not here — `Config` has no such fields and |
| 101 | # unknown keys are ignored. See crates/tui/src/settings.rs. |
| 102 | # |
| 103 | # Density (Grok-like compact defaults; turn these up if you want more shown): |
| 104 | # thinking_preview_lines = 2 # 0 header-only, 10 older dump |
| 105 | # thinking_default_expanded = false |
| 106 | # help_expand_groups = false # true = F1 starts fully expanded |
| 107 | # pin_last_prompt = true |
| 108 | # show_tool_details = false |
| 109 | |
| 110 | # Signed catalog overlays are optional and inactive without approved trust keys. |
| 111 | # Explicit route/model selections and provider-owned rosters keep priority. |
| 112 | # CODEWHALE_DISABLE_CLOUD_FACTS=1 disables cache, local files and network too. |
| 113 | [cloud_facts] |
| 114 | enabled = false |
| 115 | channel = "stable" |
| 116 | ttl_hours = 6 |
| 117 | # url = "https://codewhale.net/api/facts/v1/{channel}" |
| 118 | |
| 119 | # ───────────────────────────────────────────────────────────────────────────────── |
| 120 | # Startup update check |
| 121 | # ───────────────────────────────────────────────────────────────────────────────── |
| 122 | # The TUI checks for newer Codewhale releases in the background at startup. |
| 123 | # Set check_for_updates = false in managed or air-gapped environments. |
| 124 | # The result is cached in ~/.codewhale/update-check.json, so the network is |
| 125 | # touched at most once per check_interval_hours while the notice still shows |
| 126 | # on every launch. Set 0 to check on every launch. |
| 127 | # Checks are skipped entirely in CI, and when CODEWHALE_NO_UPDATE_CHECK or |
| 128 | # NO_UPDATE_NOTIFIER is set. |
| 129 | # update_uri may point at a GitHub-compatible latest-release JSON endpoint. |
| 130 | [update] |
| 131 | check_for_updates = true |
| 132 | check_interval_hours = 24 |
| 133 | # update_uri = "https://internal.mirror.example/codewhale/releases/latest" |
| 134 | |
| 135 | # ───────────────────────────────────────────────────────────────────────────────── |
| 136 | # Hotbar slots (#2061 / #2064) |
| 137 | # ───────────────────────────────────────────────────────────────────────────────── |
| 138 | # Optional 1-8 workbar hotbar bindings. Since #3807 a missing `hotbar` key |
| 139 | # renders NO bar — the Hotbar is hidden until you configure [[hotbar]] |
| 140 | # tables here (or via `/hotbar`). An explicit `hotbar = []` also disables it. |
| 141 | # |
| 142 | # Invalid slots are skipped with a warning, duplicate slots use the last entry, |
| 143 | # and unknown actions are preserved so the UI can show a disabled entry. |
| 144 | # Slash commands can be bound as slash.<name>, for example slash.workflow. |
| 145 | # `/hotbar on` writes the default slots: slash.workflow, slash.goal, slash.auto, |
| 146 | # then Plan/Work/Operate, palette, and workbar toggle. Commands that require arguments |
| 147 | # pre-fill the composer instead of running incomplete. |
| 148 | # |
| 149 | # [[hotbar]] |
| 150 | # slot = 1 |
| 151 | # label = "voice" |
| 152 | # action = "voice.toggle" |
| 153 | # |
| 154 | # [[hotbar]] |
| 155 | # slot = 2 |
| 156 | # action = "session.compact" |
| 157 | # |
| 158 | # [[hotbar]] |
| 159 | # slot = 3 |
| 160 | # label = "mode" |
| 161 | # action = "slash.mode" |
| 162 | |
| 163 | # ───────────────────────────────────────────────────────────────────────────────── |
| 164 | # Paths |
| 165 | # ───────────────────────────────────────────────────────────────────────────────── |
| 166 | # New installs write product state under ~/.codewhale/. Existing ~/.deepseek/ |
| 167 | # files are still read as compatibility fallbacks when the .codewhale file is |
| 168 | # absent. |
| 169 | skills_dir = "~/.codewhale/skills" |
| 170 | mcp_config_path = "~/.codewhale/mcp.json" |
| 171 | notes_path = "~/.codewhale/notes.txt" |
| 172 | |
| 173 | # Anchors the native memory store. The filename itself is not written: under |
| 174 | # the Native backend (the only backend) the store is re-rooted to |
| 175 | # `<parent-of-this-path>/memory/global/MEMORY.md`. With the default below that |
| 176 | # resolves to `~/.codewhale/memory/global/MEMORY.md` (plus workspace-scoped |
| 177 | # files and a rebuildable SQLite FTS5 index). See docs/MEMORY.md. |
| 178 | memory_path = "~/.codewhale/memory.md" |
| 179 | |
| 180 | # instructions = ["./AGENTS.md", "~/.codewhale/global.md"] |
| 181 | # |
| 182 | # Optional list of additional instruction files concatenated into the |
| 183 | # system prompt in declared order (#454). Useful for layering |
| 184 | # repo-specific rules on top of a global preferences file. Each entry |
| 185 | # is expanded so `~` and env vars work; missing files are skipped with |
| 186 | # a tracing warning. Files are capped at 100 KiB per entry. |
| 187 | # |
| 188 | # Project-level config (.codewhale/config.toml in the workspace) replaces |
| 189 | # the user-level array wholesale rather than merging — list `~/global.md` |
| 190 | # inside the project array if you want both. An explicit empty array |
| 191 | # (`instructions = []`) clears the user list for the current repo. |
| 192 | |
| 193 | # ───────────────────────────────────────────────────────────────────────────────── |
| 194 | # User memory (#489) — opt-in. When enabled, the TUI loads the native store |
| 195 | # derived from memory_path (see above), injects a bounded recall block into |
| 196 | # the system prompt, intercepts `# foo` in the composer, and registers the |
| 197 | # `remember` / `memory_search` / `memory_get` tools. |
| 198 | # ───────────────────────────────────────────────────────────────────────────────── |
| 199 | [memory] |
| 200 | # enabled = true # turn the feature on (default: false) |
| 201 | # Override the env-var equivalent: `DEEPSEEK_MEMORY=on` |
| 202 | |
| 203 | # Xiaomi MiMo speech/TTS defaults. Also configurable with |
| 204 | # XIAOMI_MIMO_SPEECH_OUTPUT_DIR / MIMO_SPEECH_OUTPUT_DIR. |
| 205 | [speech] |
| 206 | # output_dir = "./speech" |
| 207 | |
| 208 | # ───────────────────────────────────────────────────────────────────────────────── |
| 209 | # Reasoning-only recovery |
| 210 | # ───────────────────────────────────────────────────────────────────────────────── |
| 211 | # When a reasoning model returns only hidden thinking without any answer text |
| 212 | # or tool call, the engine can re-request the answer automatically. |
| 213 | # Set max_reprompts = 0 to disable automatic recovery entirely. |
| 214 | # [reasoning_only] |
| 215 | # max_reprompts = 2 |
| 216 | # reprompt_message = "So, what's up ? Keep running !" |
| 217 | |
| 218 | # ─────────────────────────────────────────────────────────────────────────── |
| 219 | # Model-bound key redaction ([redaction]) |
| 220 | # ─────────────────────────────────────────────────────────────────────────── |
| 221 | # Codewhale masks credential-looking values in tool output before it reaches |
| 222 | # the model (the "model boundary"), so a file that contains a configured API |
| 223 | # key, a bare provider token, or a credential-shaped opaque string never leaks |
| 224 | # those bytes to the model. Leave this enabled unless the model must read and |
| 225 | # edit files that contain real credentials. |
| 226 | # |
| 227 | # Disabling is a security decision, so it is never a plain flag: |
| 228 | # * Set model_bound = "disabled" here, restart Codewhale, and the startup |
| 229 | # gate asks twice - a first confirmation, then a red "are you really |
| 230 | # sure?" stage. Only the second confirmation takes effect, and it applies |
| 231 | # on later launches while model_bound stays "disabled". |
| 232 | # * Going back to "enabled" - or rewriting config.toml after the |
| 233 | # confirmation - invalidates it: requesting "disabled" again always |
| 234 | # asks for a fresh confirmation. |
| 235 | # * Until a confirmation exists - including in non-interactive/headless |
| 236 | # runs, which never confirm anything - masking stays on regardless of |
| 237 | # this key. Choosing "keep masking on" on the gate leaves the key |
| 238 | # untouched, so the next launch asks again. |
| 239 | # * The value is forgiving: false/"off" mean "disabled"; true/"on" mean |
| 240 | # "enabled". |
| 241 | # [redaction] |
| 242 | # model_bound = "enabled" # mask keys before they reach the model (default) |
| 243 | # model_bound = "disabled" # request the opt-out (restart + confirm required) |
| 244 | |
| 245 | # Native tool catalog controls (#2076). By default only the core tool surface |
| 246 | # is loaded into the model context; less common native tools are discoverable |
| 247 | # through ToolSearch and loaded on first use. |
| 248 | # [tools] |
| 249 | # always_load = ["git_show", "notify"] |
| 250 | # |
| 251 | # `request_user_input` payload ceilings (#5949). The model may ask up to |
| 252 | # `user_input_max_questions` questions per call, each offering up to |
| 253 | # `user_input_max_options` options. Out-of-range values are clamped to the |
| 254 | # supported ranges (1..=10 and 2..=10) with a warning. |
| 255 | # user_input_max_questions = 6 |
| 256 | # user_input_max_options = 4 |
| 257 | # |
| 258 | # Seconds a question or an approval decision waits before it cancels (#6003). |
| 259 | # 0 disables the timeout entirely; values above 86400 (24h) are clamped. |
| 260 | # user_input_timeout_seconds = 300 |
| 261 | # |
| 262 | # Bound an interactive approval card's wait (#6101). When the window elapses |
| 263 | # the card resolves to deny (fail-closed) and the transcript says the bound |
| 264 | # denied the call; omitted or 0 waits indefinitely. |
| 265 | # [approval] |
| 266 | # timeout_seconds = 300 |
| 267 | |
| 268 | # ───────────────────────────────────────────────────────────────────────────────── |
| 269 | # Product telemetry — optional, enabled by default with a startup disclosure |
| 270 | # ───────────────────────────────────────────────────────────────────────────────── |
| 271 | # Anonymous usage analytics are enabled when this key is omitted, unless an |
| 272 | # existing opt-out or a run-scoped kill switch disables them. No first-run |
| 273 | # affirmative opt-in is required. See docs/TELEMETRY.md for the complete schema. |
| 274 | # Local session history and usage diagnostics work with telemetry disabled; |
| 275 | # they require no account or hosted analytics service. |
| 276 | # |
| 277 | # Setting it to `false` here is an answer, not just a flag: it deletes the |
| 278 | # random install id, truncates every buffered event, and leaves a tombstone |
| 279 | # that a session already running re-checks before it sends anything. Every |
| 280 | # later run re-reads this key and re-asserts that tombstone, so it stands for |
| 281 | # as long as the `false` does — and nothing outranks it, not `--telemetry true` |
| 282 | # and not `CODEWHALE_TELEMETRY=1`. Turning telemetry back on means writing |
| 283 | # `telemetry = true` here. |
| 284 | # |
| 285 | # The environment variable and the flag are different: they stop the run and |
| 286 | # erase nothing, so a harness that disables telemetry for one command does not |
| 287 | # discard the machine owner's install id and dry-run records. |
| 288 | # |
| 289 | # codewhale config set telemetry false # opt out: stops it and erases state |
| 290 | # CODEWHALE_TELEMETRY=0 codewhale # kill switch: stops it, erases nothing |
| 291 | # |
| 292 | # What is never collected: prompts, completions, tool arguments, diffs, file |
| 293 | # contents, filenames, paths, git remotes, repo or branch names, memory |
| 294 | # entries, chat history, API keys or tokens (not even a boolean saying one |
| 295 | # exists), model ids, custom provider table names, MCP server names, error or |
| 296 | # panic message bodies, per-event timestamps, keystrokes, clipboard, |
| 297 | # screenshots, or location. The complete schema is `docs/TELEMETRY.md`, and a |
| 298 | # test asserts this file and the serializer agree. |
| 299 | # |
| 300 | # A repo-local `.codewhale/config.toml` can set neither key: someone else's |
| 301 | # repository cannot turn your telemetry on or aim it at a host of their choosing. |
| 302 | # telemetry = false |
| 303 | |
| 304 | # Where batches are POSTed. Leaving this unset selects the shipped default, |
| 305 | # the first-party ingest service: |
| 306 | # |
| 307 | # https://telemetry.codewhale.net/v1/telemetry |
| 308 | # |
| 309 | # That default is only ever consulted for a session that is already enabled — |
| 310 | # it decides where a batch goes, never whether one exists. Nothing is sent |
| 311 | # until `telemetry` is on AND the first-run notice was answered with Enable. |
| 312 | # |
| 313 | # Two overrides, both of which beat the default: |
| 314 | # |
| 315 | # telemetry_endpoint = "https://collector.internal/v1/batch" # your own sink |
| 316 | # telemetry_endpoint = "" # contact nobody |
| 317 | # |
| 318 | # The empty string is the local dry-run sink: batches are serialized exactly as |
| 319 | # a real endpoint would see them, appended to |
| 320 | # `$CODEWHALE_HOME/telemetry/dryrun.jsonl`, and no HTTP client is ever |
| 321 | # constructed. Read that file to see precisely what would have been sent. |
| 322 | # `CODEWHALE_TELEMETRY_ENDPOINT` overrides this file, and setting it to the |
| 323 | # empty string means the same "contact nobody". |
| 324 | # |
| 325 | # `https://` is required; plain `http://` is accepted only for loopback, and no |
| 326 | # environment variable overrides that refusal. |
| 327 | # telemetry_endpoint = "https://telemetry.codewhale.net/v1/telemetry" |
| 328 | |
| 329 | # ───────────────────────────────────────────────────────────────────────────────── |
| 330 | # Security |
| 331 | # ───────────────────────────────────────────────────────────────────────────────── |
| 332 | allow_shell = true |
| 333 | approval_policy = "on-request" # on-request | untrusted | never |
| 334 | sandbox_mode = "workspace-write" # read-only | workspace-write | danger-full-access | external-sandbox |
| 335 | |
| 336 | # Whether a workspace-write sandbox also lets shell commands reach the |
| 337 | # network. Default false: being allowed to edit this repository is not a |
| 338 | # reason to be allowed to open outbound connections, so `curl`, package |
| 339 | # installs, and `git fetch` are denied by the OS sandbox unless you opt in. |
| 340 | # When a command is denied, Codewhale offers an elevation prompt that grants |
| 341 | # network for that call only; set this to `true` to grant it for the whole |
| 342 | # session instead. `danger-full-access` and `yolo` are unsandboxed and |
| 343 | # unaffected by this key. Note that on platforms with no OS sandbox backend |
| 344 | # (default Linux without bubblewrap, and Windows) nothing is enforced either |
| 345 | # way -- `/status` and `doctor` both say so. |
| 346 | # sandbox_network_access = false |
| 347 | |
| 348 | # Which other agents' instruction files to import as project instructions. |
| 349 | # Empty by default. Codewhale reads AGENTS.md (the cross-agent standard) and |
| 350 | # its own .codewhale/instructions.md without being asked; a CLAUDE.md, |
| 351 | # .cursorrules, or .github/copilot-instructions.md written as law for a |
| 352 | # different tool is not treated as law here until you say so. Codewhale's own |
| 353 | # files always outrank anything imported. |
| 354 | # Accepts: "claude", "cursor", "cline", "windsurf", "gemini", "copilot", |
| 355 | # "muse", or "all". Env: CODEWHALE_PROJECT_INSTRUCTION_IMPORTS (comma-separated). |
| 356 | # project_instruction_imports = ["claude"] |
| 357 | # |
| 358 | # Everything that reaches the model as standing project instruction authority — |
| 359 | # the repository-root -> workspace AGENTS.md chain, the global fallback layer, |
| 360 | # .codewhale/rules/*.md, and any imported foreign files — shares one 48 KiB |
| 361 | # aggregate budget. Instructions claim it first and are trimmed from the |
| 362 | # broadest scope inward, so the nearest-scope file is the last thing dropped. |
| 363 | # prompt_suggestion = true # opt-in: show ghost-text follow-up question in composer after each turn |
| 364 | |
| 365 | # Optional tab/window title shown as `[title] …` in front of the terminal |
| 366 | # window title (the `Codewhale` / `reasoning…` / `using tool…` / `done` |
| 367 | # states). Multi-window setups can give each workspace its own `--config` |
| 368 | # file (or profile) so alt-tabbed sessions are identifiable at a glance. |
| 369 | # The `/title` command overrides this per session; `/config title … --save` |
| 370 | # persists a new default here. Run `/title off` to drop a session override. |
| 371 | # title = "workspace-x" |
| 372 | |
| 373 | # Typed permission rules live in a sibling `permissions.toml` file, not in |
| 374 | # config.toml. Each `[[rules]]` entry accepts `tool`, optional `command` |
| 375 | # or `path`, optional absolute `workspace`, optional `command_exact = true`, |
| 376 | # and an `action` field: `"deny"`, `"ask"` (default), or `"allow"`. |
| 377 | # Deny always wins over ask, which wins over allow. The active user's sibling |
| 378 | # file is the only rule source today; project config overlays do not load a |
| 379 | # project-local permissions.toml. `workspace` is a repo scope, not a source. |
| 380 | # Globs and broad directory rules remain future work. |
| 381 | # |
| 382 | # In supported approval cards, press `S` to allow once and save an exact ask |
| 383 | # rule. Eligible safe calls also offer `P` / "Always allow this exact rule in |
| 384 | # this repo", which saves an exact `allow` rule scoped to the current absolute |
| 385 | # workspace. Dangerous/critical calls and repo-law prompts cannot save allow |
| 386 | # grants. The approval UI still does not save deny rules: |
| 387 | # exec_shell -> exact approved command string |
| 388 | # write_file -> exact workspace-relative target path |
| 389 | # edit_file -> exact workspace-relative target path |
| 390 | # apply_patch -> one exact workspace-relative path per validated touched file |
| 391 | # `read_file` rules can be written manually, but the approval UI does not save |
| 392 | # them. |
| 393 | # |
| 394 | # `/permissions list` shows the active source, effective matcher, repo/global |
| 395 | # scope, and whether each rule applies in the current workspace. |
| 396 | # `/permissions remove <number>` previews a deletion and prints a confirmation |
| 397 | # command bound to the current file snapshot. If the file changes first, the |
| 398 | # confirmation fails safely. Removal preserves unrelated comments/formatting, |
| 399 | # uses the same lock as approval-card appends, and atomically replaces the file. |
| 400 | # `/config ask-rules` remains a compatibility entry to the same list. |
| 401 | # |
| 402 | # Example ~/.codewhale/permissions.toml: |
| 403 | # |
| 404 | # [[rules]] |
| 405 | # tool = "exec_shell" |
| 406 | # command = "cargo test" |
| 407 | # action = "ask" |
| 408 | # |
| 409 | # # Block dangerous commands |
| 410 | # [[rules]] |
| 411 | # tool = "exec_shell" |
| 412 | # command = "sed" |
| 413 | # action = "deny" |
| 414 | # |
| 415 | # [[rules]] |
| 416 | # tool = "exec_shell" |
| 417 | # command = "awk" |
| 418 | # action = "deny" |
| 419 | # |
| 420 | # # Allow trusted commands without asking |
| 421 | # [[rules]] |
| 422 | # tool = "exec_shell" |
| 423 | # command = "git status" |
| 424 | # command_exact = true |
| 425 | # workspace = "/absolute/path/to/project" |
| 426 | # action = "allow" |
| 427 | # |
| 428 | # # Path-based deny |
| 429 | # [[rules]] |
| 430 | # tool = "write_file" |
| 431 | # path = "src/main.rs" |
| 432 | # action = "deny" |
| 433 | # |
| 434 | # [[rules]] |
| 435 | # tool = "edit_file" |
| 436 | # path = "src/lib.rs" |
| 437 | # action = "ask" |
| 438 | # |
| 439 | # [[rules]] |
| 440 | # tool = "apply_patch" |
| 441 | # path = "src/patch-target.rs" |
| 442 | # |
| 443 | # [[rules]] |
| 444 | # tool = "read_file" |
| 445 | # path = "secrets/api_key.txt" |
| 446 | # action = "deny" |
| 447 | # ───────────────────────────────────────────────────────────────────────────────── |
| 448 | # External Sandbox Backend (pluggable remote execution) |
| 449 | # ───────────────────────────────────────────────────────────────────────────────── |
| 450 | # When sandbox_backend is set to "opensandbox", all exec_shell calls are |
| 451 | # routed through an external OpenSandbox-compatible HTTP API instead of |
| 452 | # spawning a local process. The backend sends `POST {sandbox_url}/v1/sandbox/run` |
| 453 | # with `{"cmd": "...", "env": {...}}` and expects |
| 454 | # `{"stdout": "...", "stderr": "...", "exit_code": 0}`. |
| 455 | # |
| 456 | # sandbox_backend = "none" # "none" (default) or "opensandbox" |
| 457 | # sandbox_url = "http://localhost:8080" # OpenSandbox-compatible API base URL |
| 458 | # sandbox_api_key = "YOUR_API_KEY" # Optional Bearer token sent with requests |
| 459 | # |
| 460 | # Env-var overrides: |
| 461 | # DEEPSEEK_SANDBOX_BACKEND → sandbox_backend |
| 462 | # DEEPSEEK_SANDBOX_URL → sandbox_url |
| 463 | # DEEPSEEK_SANDBOX_API_KEY → sandbox_api_key |
| 464 | # |
| 465 | # Example OpenSandbox setup: |
| 466 | # |
| 467 | # sandbox_backend = "opensandbox" |
| 468 | # sandbox_url = "http://localhost:8080" |
| 469 | # sandbox_api_key = "sk-opensandbox-secret" |
| 470 | # |
| 471 | # The backend uses a 30-second HTTP timeout. Background, interactive, and |
| 472 | # TTY modes are not supported with external backends — all commands run |
| 473 | # synchronously via HTTP. |
| 474 | # ───────────────────────────────────────────────────────────────────────────────── |
| 475 | # Bubblewrap (Linux only, additional filesystem isolation) |
| 476 | # ───────────────────────────────────────────────────────────────────────────────── |
| 477 | # When set to true and `/usr/bin/bwrap` is present, exec_shell commands are |
| 478 | # routed through bubblewrap instead of relying solely on Landlock. Bubblewrap |
| 479 | # creates a read-only view of the root filesystem with write access limited to |
| 480 | # the working directory. Install separately: |
| 481 | # |
| 482 | # Ubuntu/Debian: apt install bubblewrap |
| 483 | # Fedora: dnf install bubblewrap |
| 484 | # Arch: pacman -S bubblewrap |
| 485 | # |
| 486 | # prefer_bwrap = false # default — use Landlock only |
| 487 | # |
| 488 | # Env override: CODEWHALE_PREFER_BWRAP=true |
| 489 | # Legacy alias (deprecated until 0.10.0): DEEPSEEK_PREFER_BWRAP=true |
| 490 | # |
| 491 | # With prefer_bwrap = true, the sandbox gets a private /dev and /proc plus a |
| 492 | # writable isolated /tmp by default, so toolchains work without widening the |
| 493 | # filesystem policy (#5410). Two optional escape hatches: |
| 494 | # |
| 495 | # bwrap_ro_roots = [] # extra host paths bind-mounted read-only, e.g. |
| 496 | # # ["/usr/lib", "/usr/lib64"] for system-library |
| 497 | # # linking when the root bind is narrowed |
| 498 | # bwrap_dev_roots = [] # host device nodes bind-mounted read-write, e.g. |
| 499 | # # ["/dev/null"] for redirection against the host |
| 500 | # # node — character/block devices only, never |
| 501 | # # directories, so this cannot widen file writes |
| 502 | |
| 503 | # auto_allow entries match by command prefix, not raw string. |
| 504 | # See command_safety.rs for the prefix dictionary. |
| 505 | # |
| 506 | # Examples: |
| 507 | # auto_allow = ["git status"] # auto-approves: git status, git status -s, git status --porcelain |
| 508 | # # does NOT auto-approve: git push, git checkout |
| 509 | # auto_allow = ["cargo check", "npm run"] |
| 510 | # |
| 511 | # auto_allow = [] |
| 512 | max_subagents = 10 # optional (default 64, clamped to 1-128) |
| 513 | |
| 514 | # Optional sub-agent tuning. max_concurrent overrides top-level max_subagents. |
| 515 | # [subagents] |
| 516 | # max_concurrent = 10 |
| 517 | # api_timeout_secs = 600 # per-step API timeout, clamped to 1..=3600 |
| 518 | # |
| 519 | # How many levels of nested sub-agents the `agent` tool may spawn: |
| 520 | # max_depth = 0 # opt out completely — the agent never spawns sub-agents |
| 521 | # max_depth = 1 # the agent may spawn sub-agents, but those may not spawn more |
| 522 | # max_depth = 2 # one more level of nesting, etc. |
| 523 | # Unset defaults to 3; any value is clamped to the hard ceiling (8). The depth |
| 524 | # limit is enforced in code, not requested of the model — a sub-agent past the |
| 525 | # limit cannot be spawned regardless of what the model decides. |
| 526 | # max_depth = 3 |
| 527 | |
| 528 | # Optional managed policy paths (defaults to /etc/deepseek/*.toml on unix): |
| 529 | # managed_config_path = "/etc/deepseek/managed_config.toml" |
| 530 | # requirements_path = "/etc/deepseek/requirements.toml" |
| 531 | |
| 532 | # ───────────────────────────────────────────────────────────────────────────────── |
| 533 | # Per-provider credentials (peer providers — NIM is first-class, not a flag) |
| 534 | # ───────────────────────────────────────────────────────────────────────────────── |
| 535 | # Providers can be stored at once; `provider = "..."` (top of file) or |
| 536 | # `/provider deepseek` / `/provider nvidia-nim` / `--provider openai` / |
| 537 | # `--provider wanjie-ark` / `/provider volcengine` / `/provider fireworks` / |
| 538 | # `--provider siliconflow` / `/provider arcee` / `/provider moonshot` |
| 539 | # switches between them without having to re-enter keys. Env vars override anything set here: |
| 540 | # DeepSeek: DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL, DEEPSEEK_MODEL |
| 541 | # DeepSeek Anthropic-compatible: DEEPSEEK_API_KEY, DEEPSEEK_ANTHROPIC_BASE_URL |
| 542 | # NIM: NVIDIA_API_KEY (or NVIDIA_NIM_API_KEY), NIM_BASE_URL |
| 543 | # (or NVIDIA_NIM_BASE_URL / NVIDIA_BASE_URL), NVIDIA_NIM_MODEL |
| 544 | # OpenAI-compatible: OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL |
| 545 | # Wanjie Ark: WANJIE_ARK_API_KEY (or WANJIE_API_KEY), WANJIE_ARK_BASE_URL, WANJIE_ARK_MODEL |
| 546 | # Volcengine Ark: VOLCENGINE_API_KEY (or VOLCENGINE_ARK_API_KEY / ARK_API_KEY), VOLCENGINE_BASE_URL, VOLCENGINE_MODEL |
| 547 | # OpenRouter: OPENROUTER_API_KEY, OPENROUTER_BASE_URL, OPENROUTER_MODEL |
| 548 | # Xiaomi MiMo: XIAOMI_MIMO_API_KEY (or XIAOMI_API_KEY / MIMO_API_KEY), XIAOMI_MIMO_BASE_URL, XIAOMI_MIMO_MODEL |
| 549 | # Token Plan: XIAOMI_MIMO_TOKEN_PLAN_API_KEY (or MIMO_TOKEN_PLAN_API_KEY), XIAOMI_MIMO_MODE/MIMO_MODE |
| 550 | # Novita: NOVITA_API_KEY, NOVITA_BASE_URL, NOVITA_MODEL |
| 551 | # Fireworks: FIREWORKS_API_KEY, FIREWORKS_BASE_URL |
| 552 | # SiliconFlow: SILICONFLOW_API_KEY, SILICONFLOW_BASE_URL, SILICONFLOW_MODEL |
| 553 | # Arcee: ARCEE_API_KEY, ARCEE_BASE_URL, ARCEE_MODEL |
| 554 | # Moonshot/Kimi: MOONSHOT_API_KEY (or KIMI_API_KEY), MOONSHOT_BASE_URL, MOONSHOT_MODEL |
| 555 | # SGLang: SGLANG_BASE_URL, SGLANG_MODEL, optional SGLANG_API_KEY |
| 556 | # vLLM: VLLM_BASE_URL, VLLM_MODEL, optional VLLM_API_KEY |
| 557 | # Ollama: OLLAMA_BASE_URL, OLLAMA_MODEL, optional OLLAMA_API_KEY |
| 558 | # Hugging Face: HUGGINGFACE_API_KEY (or HF_TOKEN), HUGGINGFACE_BASE_URL (or HF_BASE_URL), |
| 559 | # HUGGINGFACE_MODEL (or HF_MODEL) |
| 560 | # Meta Model API: META_MODEL_API_KEY (or MODEL_API_KEY), META_MODEL_API_BASE_URL |
| 561 | # (or MODEL_API_BASE_URL), META_MODEL_API_MODEL (or MODEL_API_MODEL) |
| 562 | # |
| 563 | # Custom DeepSeek-compatible APIs usually do not need a new provider table: |
| 564 | # set `provider = "deepseek"` and override [providers.deepseek].base_url/model. |
| 565 | # For generic OpenAI-compatible gateways, use `provider = "openai"` and the |
| 566 | # [providers.openai] table below. Keep provider/api_key/base_url in user config |
| 567 | # or environment variables; project overlays are not allowed to set them. |
| 568 | # |
| 569 | # Provider is the route/account/endpoint; model is the ID on that route. |
| 570 | # Common DeepSeek routes: |
| 571 | # provider = "deepseek" model = "deepseek-v4-pro" |
| 572 | # provider = "nvidia-nim" model = "deepseek-ai/deepseek-v4-pro" |
| 573 | # provider = "openrouter" model = "deepseek/deepseek-v4-pro" |
| 574 | # provider = "fireworks" model = "accounts/fireworks/models/deepseek-v4-pro" |
| 575 | # provider = "siliconflow" model = "deepseek-ai/DeepSeek-V4-Pro" |
| 576 | |
| 577 | # DeepSeek Platform (https://platform.deepseek.com) |
| 578 | [providers.deepseek] |
| 579 | # api_key = "YOUR_DEEPSEEK_API_KEY" |
| 580 | # base_url = "https://api.deepseek.com/beta" |
| 581 | # model = "deepseek-v4-pro" |
| 582 | # Custom DeepSeek-compatible example: |
| 583 | # base_url = "https://your-provider.example/v1" |
| 584 | # model = "deepseek-ai/DeepSeek-V4-Pro" |
| 585 | # http_headers = { "X-Model-Provider-Id" = "your-model-provider" } # optional custom request headers |
| 586 | # path_suffix = "/chat/completions" # override the API path; skips /v1 versioning when set |
| 587 | # reasoning_stream_style = "inline_tags" # route <think>...</think> content into Thinking cells |
| 588 | |
| 589 | # DeepSeek Anthropic-compatible Messages route (opt-in) |
| 590 | # [providers.deepseek_anthropic] |
| 591 | # api_key = "YOUR_DEEPSEEK_API_KEY" |
| 592 | # base_url = "https://api.deepseek.com/anthropic" |
| 593 | # model = "deepseek-v4-pro" |
| 594 | # [providers.deepseek.auth] # provider-scoped auth source metadata; command execution lands in a follow-up slice |
| 595 | # source = "command" |
| 596 | # command = ["secret-tool", "lookup", "service", "codewhale-deepseek"] |
| 597 | # timeout_ms = 2000 |
| 598 | # insecure_skip_tls_verify = true # last resort for private gateways; prefer SSL_CERT_FILE |
| 599 | |
| 600 | # NVIDIA NIM-hosted DeepSeek V4 (https://build.nvidia.com) |
| 601 | [providers.nvidia_nim] |
| 602 | # api_key = "YOUR_NVIDIA_API_KEY" |
| 603 | # base_url = "https://integrate.api.nvidia.com/v1" |
| 604 | # model = "deepseek-ai/deepseek-v4-pro" # or deepseek-ai/deepseek-v4-flash |
| 605 | |
| 606 | # Generic OpenAI-compatible endpoint. Use the built-in `openai` provider for |
| 607 | # third-party gateways; do not invent a custom provider name. For non-local |
| 608 | # http:// gateways, launch with DEEPSEEK_ALLOW_INSECURE_HTTP=1 only on a |
| 609 | # trusted network. |
| 610 | [providers.openai] |
| 611 | # api_key = "YOUR_OPENAI_COMPATIBLE_API_KEY" |
| 612 | # base_url = "https://api.openai.com/v1" |
| 613 | # model = "gpt-4.1" |
| 614 | # Gateway example: |
| 615 | # base_url = "https://gateway.example/v1" |
| 616 | # model = "your-deepseek-compatible-model" |
| 617 | # Alibaba Bailian / Model Studio DashScope OpenAI-compatible example: |
| 618 | # base_url = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1" |
| 619 | # model = "qwen-plus" |
| 620 | # context_window = 1000000 # set the gateway/model's real total context window |
| 621 | # insecure_skip_tls_verify = true # last resort for private gateways; prefer SSL_CERT_FILE |
| 622 | |
| 623 | # AtlasCloud OpenAI-compatible endpoint (https://www.atlascloud.ai/docs/models/llm) |
| 624 | [providers.atlascloud] |
| 625 | # api_key = "YOUR_ATLASCLOUD_API_KEY" |
| 626 | # base_url = "https://api.atlascloud.ai/v1" |
| 627 | # model = "deepseek-ai/deepseek-v4-flash" |
| 628 | |
| 629 | # Wanjie Ark / 万界方舟 OpenAI-compatible endpoint |
| 630 | [providers.wanjie_ark] |
| 631 | # api_key = "YOUR_WANJIE_API_KEY" |
| 632 | # base_url = "https://maas-openapi.wanjiedata.com/api/v1" |
| 633 | # model = "deepseek-reasoner" # or the exact model ID enabled on your Wanjie account |
| 634 | |
| 635 | # Volcengine / Volcano Engine Ark Coding API |
| 636 | [providers.volcengine] |
| 637 | # api_key = "YOUR_VOLCENGINE_API_KEY" |
| 638 | # base_url = "https://ark.cn-beijing.volces.com/api/coding/v3" |
| 639 | # model = "DeepSeek-V4-Pro" # or DeepSeek-V4-Flash |
| 640 | |
| 641 | # OpenRouter — multi-provider gateway (https://openrouter.ai) |
| 642 | [providers.openrouter] |
| 643 | # vendor = "deepinfra" # exact upstream slug; disables OpenRouter fallbacks |
| 644 | # api_key = "YOUR_OPENROUTER_API_KEY" |
| 645 | # base_url = "https://openrouter.ai/api/v1" |
| 646 | # model = "deepseek/deepseek-v4-pro" |
| 647 | # OpenRouter-compatible gateways can reuse this provider so reasoning/cache |
| 648 | # parsing stays on the OpenRouter-compatible path instead of generic OpenAI: |
| 649 | # base_url = "https://openrouter-compatible.example/v1" |
| 650 | # model = "deepseek/deepseek-v4-pro" |
| 651 | # Recent large model IDs also accepted here include arcee-ai/trinity-large-thinking, |
| 652 | # minimax/minimax-m3, minimax/minimax-m2.7, xiaomi/mimo-v2.5-pro, qwen/qwen3.6-flash, |
| 653 | # qwen/qwen3.6-35b-a3b, qwen/qwen3.6-max-preview, qwen/qwen3.6-27b, qwen/qwen3.6-plus, |
| 654 | # qwen/qwen3.7-max, google/gemma-4-31b-it, z-ai/glm-5.1, z-ai/glm-5.2, |
| 655 | # moonshotai/kimi-k2.6, |
| 656 | # nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free, and nvidia/nemotron-3-ultra. |
| 657 | |
| 658 | # OrcaRouter — OpenAI-compatible aggregation gateway (https://www.orcarouter.ai) |
| 659 | [providers.orcarouter] |
| 660 | # api_key = "YOUR_ORCAROUTER_API_KEY" |
| 661 | # base_url = "https://api.orcarouter.ai/v1" |
| 662 | # model = "deepseek/deepseek-v4-pro" |
| 663 | # Namespaced wire models pass through verbatim (deepseek/deepseek-v4-pro, |
| 664 | # deepseek/deepseek-v4-flash); OrcaRouter's own auto-routing model is |
| 665 | # selectable as "orcarouter/auto". |
| 666 | |
| 667 | # Xiaomi MiMo OpenAI-compatible endpoint (https://platform.xiaomimimo.com) |
| 668 | [providers.xiaomi_mimo] |
| 669 | # api_key = "YOUR_XIAOMI_KEY" |
| 670 | # base_url = "https://token-plan-sgp.xiaomimimo.com/v1" # Token Plan / tp- keys |
| 671 | # # base_url = "https://token-plan-ams.xiaomimimo.com/v1" # Token Plan Europe / Amsterdam |
| 672 | # # base_url = "https://api.xiaomimimo.com/v1" # Pay-as-you-go / sk- keys |
| 673 | # model = "mimo-v2.5-pro" # chat/reasoning |
| 674 | # Chat model IDs: mimo-v2.5-pro, mimo-v2.5-pro-ultraspeed, mimo-v2.5 |
| 675 | # Token Plan subscriptions use separate tp-* API keys plus api-key auth. |
| 676 | # mode = "token-plan-sgp" # default Token Plan endpoint |
| 677 | # mode = "token-plan-cn" # China cluster |
| 678 | # mode = "token-plan-ams" # Europe cluster |
| 679 | # mode = "pay-as-you-go" # standard API / sk- keys |
| 680 | # TTS aliases are also accepted by `codewhale speech`: tts, voice-design, voice-clone |
| 681 | # TTS model IDs: mimo-v2.5-tts, mimo-v2.5-tts-voicedesign, mimo-v2.5-tts-voiceclone, mimo-v2-tts |
| 682 | |
| 683 | # Novita AI-hosted inference (https://novita.ai) |
| 684 | [providers.novita] |
| 685 | # api_key = "YOUR_NOVITA_API_KEY" |
| 686 | # base_url = "https://api.novita.ai/openai/v1" |
| 687 | # model = "deepseek/deepseek-v4-pro" # or deepseek/deepseek-v4-flash |
| 688 | |
| 689 | # Fireworks AI-hosted DeepSeek V4 (https://fireworks.ai) |
| 690 | [providers.fireworks] |
| 691 | # api_key = "YOUR_FIREWORKS_API_KEY" |
| 692 | # base_url = "https://api.fireworks.ai/inference/v1" |
| 693 | # model = "accounts/fireworks/models/deepseek-v4-pro" |
| 694 | |
| 695 | # SiliconFlow-hosted DeepSeek V4 (https://siliconflow.com) |
| 696 | [providers.siliconflow] |
| 697 | # api_key = "YOUR_SILICONFLOW_API_KEY" |
| 698 | # base_url = "https://api.siliconflow.com/v1" |
| 699 | # model = "deepseek-ai/DeepSeek-V4-Pro" # or deepseek-ai/DeepSeek-V4-Flash |
| 700 | |
| 701 | # SiliconFlow China-hosted DeepSeek V4 (https://siliconflow.cn) |
| 702 | # Falls back to [providers.siliconflow] for api_key / base_url / model when unset. |
| 703 | [providers.siliconflow-CN] |
| 704 | # api_key = "YOUR_SILICONFLOW_API_KEY" |
| 705 | # base_url = "https://api.siliconflow.cn/v1" |
| 706 | # model = "deepseek-ai/DeepSeek-V4-Pro" |
| 707 | |
| 708 | # Arcee AI direct OpenAI-compatible endpoint (https://docs.arcee.ai) |
| 709 | [providers.arcee] |
| 710 | # api_key = "YOUR_ARCEE_API_KEY" |
| 711 | # base_url = "https://api.arcee.ai/api/v1" |
| 712 | # model = "trinity-large-thinking" # or trinity-large-preview |
| 713 | |
| 714 | # Moonshot/Kimi OpenAI-compatible endpoint (https://platform.kimi.ai) |
| 715 | [providers.moonshot] |
| 716 | # api_key = "YOUR_MOONSHOT_API_KEY" # or KIMI_API_KEY |
| 717 | # base_url = "https://api.moonshot.ai/v1" # or KIMI_BASE_URL |
| 718 | # model = "kimi-k3" # direct Moonshot K3 wire ID |
| 719 | # Direct K3 is always-thinking: off -> low, medium -> high, and max remains max. |
| 720 | # The exact route sends top-level reasoning_effort, max_completion_tokens, and |
| 721 | # omits temperature/top_p per https://platform.kimi.ai/docs/guide/kimi-k3-quickstart. |
| 722 | # Kimi Code membership path (key: https://www.kimi.com/code/console): |
| 723 | # api_key = "YOUR_KIMI_CODE_API_KEY" |
| 724 | # base_url = "https://api.kimi.com/coding/v1" |
| 725 | # model = "k3" # Kimi Code K3 wire ID |
| 726 | # K3 membership off -> enabled/low; dispatched auto selects a concrete tier. |
| 727 | # Only an omitted reasoning setting leaves the provider default in control. |
| 728 | # Moderato plans are capped at 262144; Allegretto and above unlock up to 1048576. |
| 729 | # context_window = 262144 # manually cap k3 to the Moderato/256K window |
| 730 | # context_window = 1048576 # Allegretto+ only; do not claim an unavailable entitlement |
| 731 | # Alternatively use model = "k3-256k" with context_window = 262144 for the fixed 256K route. |
| 732 | # Omit context_window to keep Codewhale's safe 262144-token bare-k3 baseline. |
| 733 | # `k3[1m]` is a Claude Code-only convention, not an API model ID; Codewhale rejects it. |
| 734 | # Kimi Code K2.7 remains available to all members as model = "kimi-for-coding". |
| 735 | # Kimi OAuth is not supported. Legacy auth_mode = "kimi_oauth" fails closed |
| 736 | # to the API-key guidance above without probing Kimi CLI credential files. |
| 737 | |
| 738 | # Z.AI GLM Coding Plan endpoint (https://docs.z.ai) |
| 739 | [providers.zai] |
| 740 | # api_key = "YOUR_ZAI_API_KEY" # or Z_AI_API_KEY |
| 741 | # base_url = "https://api.z.ai/api/coding/paas/v4" |
| 742 | # # General API endpoint, if you are not using the Coding Plan: |
| 743 | # # base_url = "https://api.z.ai/api/paas/v4" |
| 744 | # model = "GLM-5.3" # default; GLM-5.3-Flash is the fast sibling, GLM-5.2 the previous default, GLM-5.1 the smaller model, GLM-5-Turbo the 5.2 fast sibling |
| 745 | # # GLM-5.3 is live on the Z.ai Coding Plan (2026-08-13). Its catalog metadata |
| 746 | # # (limits, reasoning options) is inherited from GLM-5.2 until Z.ai publishes |
| 747 | # # distinct 5.3 numbers, and it carries no price. GLM-5.3-Flash (2026-08-26) |
| 748 | # # is the 1M multimodal picker row (`model = "GLM-5.3-Flash"`). An explicit |
| 749 | # # model = "GLM-5.2" keeps sending GLM-5.2; only the default moved. Accounts |
| 750 | # # not provisioned for 5.3 can still see a 429 with entitlement code 1311. |
| 751 | |
| 752 | # StepFun / StepFlash direct OpenAI-compatible endpoint (https://platform.stepfun.ai) |
| 753 | [providers.stepfun] |
| 754 | # api_key = "YOUR_STEPFUN_API_KEY" # or STEP_API_KEY |
| 755 | # base_url = "https://api.stepfun.ai/v1" # or STEP_BASE_URL |
| 756 | # # Coding Plan endpoint: |
| 757 | # # base_url = "https://api.stepfun.ai/step_plan/v1" |
| 758 | # model = "step-3.7-flash" # or STEPFUN_MODEL / STEP_MODEL |
| 759 | |
| 760 | # MiniMax direct OpenAI-compatible endpoint (https://platform.minimax.io) |
| 761 | [providers.minimax] |
| 762 | # api_key = "YOUR_MINIMAX_API_KEY" |
| 763 | # base_url = "https://api.minimax.io/v1" |
| 764 | # model = "MiniMax-M3" # or MiniMax-M2.7, MiniMax-M2.7-highspeed |
| 765 | # # MiniMax also publishes Anthropic-compatible endpoints: |
| 766 | # # global https://api.minimax.io/anthropic, China https://api.minimaxi.com/anthropic. |
| 767 | |
| 768 | # Self-hosted SGLang OpenAI-compatible server |
| 769 | [providers.sglang] |
| 770 | # api_key = "OPTIONAL_SGLANG_TOKEN" |
| 771 | # base_url = "http://localhost:30000/v1" |
| 772 | # model = "deepseek-ai/DeepSeek-V4-Pro" # or deepseek-ai/DeepSeek-V4-Flash |
| 773 | |
| 774 | # Self-hosted vLLM OpenAI-compatible server |
| 775 | [providers.vllm] |
| 776 | # api_key = "OPTIONAL_VLLM_TOKEN" |
| 777 | # base_url = "http://localhost:8000/v1" |
| 778 | # model = "deepseek-ai/DeepSeek-V4-Pro" # or deepseek-ai/DeepSeek-V4-Flash |
| 779 | |
| 780 | # Self-hosted Ollama OpenAI-compatible server |
| 781 | [providers.ollama] |
| 782 | # api_key = "OPTIONAL_OLLAMA_TOKEN" |
| 783 | # base_url = "http://localhost:11434/v1" |
| 784 | # model = "deepseek-v4-flash" # or any local Ollama tag |
| 785 | |
| 786 | # Hugging Face Inference Providers (https://huggingface.co/docs/api-inference) |
| 787 | # Provider aliases: huggingface, hugging-face, hugging_face, hf |
| 788 | # Env var aliases: HUGGINGFACE_API_KEY / HF_TOKEN, HUGGINGFACE_BASE_URL / HF_BASE_URL, |
| 789 | # HUGGINGFACE_MODEL / HF_MODEL |
| 790 | [providers.huggingface] |
| 791 | # api_key = "YOUR_HF_TOKEN" |
| 792 | # base_url = "https://router.huggingface.co/v1" |
| 793 | # model = "deepseek-ai/DeepSeek-V4-Pro" # or deepseek-ai/DeepSeek-V4-Flash |
| 794 | |
| 795 | # DeepInfra — AI inference cloud (https://deepinfra.com) |
| 796 | [providers.deepinfra] |
| 797 | # api_key = "YOUR_DEEPINFRA_TOKEN" |
| 798 | # base_url = "https://api.deepinfra.com/v1/openai" |
| 799 | # model = "deepseek-ai/DeepSeek-V4-Pro" # or deepseek-ai/DeepSeek-V4-Flash |
| 800 | |
| 801 | # ───────────────────────────────────────────────────────────────────────────────── |
| 802 | # Sakana AI Fugu Provider (https://api.sakana.ai) |
| 803 | # Provider aliases: sakana, sakana-ai, sakana_ai, fugu |
| 804 | # Env var aliases: FUGU_API_KEY, SAKANA_API_KEY |
| 805 | [providers.sakana] |
| 806 | # api_key = "YOUR_FUGU_API_KEY" |
| 807 | # base_url = "https://api.sakana.ai/v1" |
| 808 | # model = "fugu" # or fugu-ultra-20260615 |
| 809 | |
| 810 | # Meituan LongCat Provider (https://longcat.chat/platform) |
| 811 | # OpenAI-compatible curated gateway for Meituan's LongCat models. |
| 812 | # Provider aliases: longcat, long-cat, meituan-longcat, meituan |
| 813 | # Env var aliases: LONGCAT_API_KEY |
| 814 | [providers.longcat] |
| 815 | # api_key = "YOUR_LONGCAT_API_KEY" |
| 816 | # base_url = "https://api.longcat.chat/openai/v1" |
| 817 | # model = "LongCat-2.0" |
| 818 | |
| 819 | # OpenCode Go (https://opencode.ai/docs/go/) |
| 820 | # Subscription-backed OpenAI-compatible Chat Completions route. |
| 821 | # Env vars: OPENCODE_GO_API_KEY, OPENCODE_GO_BASE_URL, OPENCODE_GO_MODEL |
| 822 | # Chat Completions models: deepseek-v4-pro, grok-4.5, glm-5.2, glm-5.1, |
| 823 | # kimi-k3, kimi-k2.7-code, kimi-k2.6, deepseek-v4-flash, mimo-v2.5, |
| 824 | # mimo-v2.5-pro. |
| 825 | # Models documented only on OpenCode Go's Anthropic `/messages` endpoint are |
| 826 | # intentionally not advertised by this provider yet. |
| 827 | [providers.opencode_go] |
| 828 | # api_key = "YOUR_OPENCODE_GO_API_KEY" |
| 829 | # base_url = "https://opencode.ai/zen/go/v1" |
| 830 | # model = "deepseek-v4-pro" |
| 831 | |
| 832 | # OpenCode Zen (https://opencode.ai/docs/zen/) |
| 833 | # Model-aware gateway: GPT/Muse Spark use Responses, Claude/Qwen use Anthropic |
| 834 | # Messages, and DeepSeek/MiniMax/GLM/Kimi/Grok/free models use Chat Completions. |
| 835 | # Gemini uses a Google-specific protocol that Codewhale does not implement and |
| 836 | # therefore fails closed instead of being sent with the wrong request shape. |
| 837 | # Env vars: OPENCODE_ZEN_API_KEY (preferred), OPENCODE_API_KEY, |
| 838 | # OPENCODE_ZEN_BASE_URL, OPENCODE_ZEN_MODEL |
| 839 | [providers.opencode_zen] |
| 840 | # api_key = "YOUR_OPENCODE_ZEN_API_KEY" |
| 841 | # base_url = "https://opencode.ai/zen/v1" |
| 842 | # model = "gpt-5.5" # Responses |
| 843 | # model = "muse-spark-1.2-contributor-free" # Responses (free tier, auto-routed to Responses — no wire needed) |
| 844 | # model = "claude-sonnet-4-6" # Anthropic Messages example |
| 845 | # model = "deepseek-v4-pro" # Chat Completions example |
| 846 | # Custom gateway equivalent (when not using the opencode_zen provider): |
| 847 | # [providers.my_opencode] |
| 848 | # kind = "openai-compatible" |
| 849 | # base_url = "https://opencode.ai/zen/v1" |
| 850 | # model = "muse-spark-1.2-contributor-free" |
| 851 | # wire = "responses" |
| 852 | # api_key_env = "OPENCODE_ZEN_API_KEY" |
| 853 | |
| 854 | # Meta Model API / Muse Spark (https://developer.meta.com/ai/) |
| 855 | # OpenAI-compatible Chat Completions route. |
| 856 | # Provider aliases: meta, meta-ai, meta-model-api, muse, muse-spark |
| 857 | # Env var aliases: META_MODEL_API_KEY / MODEL_API_KEY, |
| 858 | # META_MODEL_API_BASE_URL / MODEL_API_BASE_URL, |
| 859 | # META_MODEL_API_MODEL / MODEL_API_MODEL |
| 860 | [providers.meta] |
| 861 | # api_key = "YOUR_META_MODEL_API_KEY" |
| 862 | # base_url = "https://api.meta.ai/v1" |
| 863 | # model = "muse-spark-1.1" |
| 864 | |
| 865 | # xAI / Grok Provider (https://console.x.ai/) |
| 866 | # OpenAI-compatible Chat Completions route. |
| 867 | # Provider aliases: xai, x-ai, x_ai, grok |
| 868 | # Env var aliases: XAI_API_KEY, XAI_BASE_URL, XAI_MODEL |
| 869 | # |
| 870 | # Auth modes: |
| 871 | # api_key (default) — console.x.ai pay-per-use key via api_key / XAI_API_KEY / keyring |
| 872 | # oauth — `codewhale auth xai-device` uses Codewhale-owned storage. |
| 873 | # Reading an existing Grok CLI file requires explicit |
| 874 | # `codewhale auth external-consent --provider xai --mode read-only`. |
| 875 | [providers.xai] |
| 876 | # api_key = "YOUR_XAI_API_KEY" |
| 877 | # auth_mode = "oauth" # or "device_code" / "grok_cli" |
| 878 | # base_url = "https://api.x.ai/v1" |
| 879 | # model = "grok-4.6" # or grok-4.5, grok-4.3, grok-build |
| 880 | |
| 881 | # Mistral AI — la Plateforme (https://console.mistral.ai/) |
| 882 | # OpenAI-compatible Chat Completions route. |
| 883 | # Provider aliases: mistral, mistral-ai, mistralai, la-plateforme |
| 884 | # Env var aliases: MISTRAL_API_KEY, MISTRAL_BASE_URL, MISTRAL_MODEL |
| 885 | [providers.mistral] |
| 886 | # api_key = "YOUR_MISTRAL_API_KEY" |
| 887 | # base_url = "https://api.mistral.ai/v1" |
| 888 | # model = "mistral-code-latest" # or mistral-medium-latest, mistral-small-latest, mistral-large-latest |
| 889 | |
| 890 | # Google Gemini — Google AI Studio (https://aistudio.google.com/apikey) |
| 891 | # OpenAI-compatible Chat Completions route on the official Gemini endpoint; |
| 892 | # this is the supported Gemini path (see docs/PROVIDERS.md). |
| 893 | # Provider aliases: google, gemini, google-gemini, ai-studio |
| 894 | # Env var aliases: GEMINI_API_KEY, GOOGLE_API_KEY, GEMINI_BASE_URL, GOOGLE_BASE_URL |
| 895 | [providers.google] |
| 896 | # api_key = "YOUR_GEMINI_API_KEY" |
| 897 | # base_url = "https://generativelanguage.googleapis.com/v1beta/openai/" |
| 898 | # model = "gemini-3.1-pro-preview" |
| 899 | |
| 900 | # ───────────────────────────────────────────────────────────────────────────────── |
| 901 | # Alibaba Cloud Model Studio — Token Plan |
| 902 | # (https://bailian.console.aliyun.com/) |
| 903 | # |
| 904 | # Token Plan Personal and Team share the same AP-Southeast (Singapore) endpoint. |
| 905 | # Available text/coding models: qwen3.8-max, qwen3.8-max-preview, qwen3.7-plus, |
| 906 | # qwen3.7-max, qwen3.6-flash, deepseek-v4-pro, deepseek-v4-flash-0731, glm-5.2 |
| 907 | # |
| 908 | # Provider aliases: modelstudio-token-plan, modelstudio_token_plan, |
| 909 | # alibaba-token-plan, dashscope-token-plan |
| 910 | # Env var aliases: MODELSTUDIO_API_KEY (preferred), DASHSCOPE_API_KEY, |
| 911 | # MODELSTUDIO_TOKEN_PLAN_BASE_URL, MODELSTUDIO_TOKEN_PLAN_MODEL |
| 912 | [providers.modelstudio_token_plan] |
| 913 | # api_key = "YOUR_MODELSTUDIO_API_KEY" |
| 914 | # base_url = "https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1" |
| 915 | # model = "qwen3.8-max" |
| 916 | # # Anthropic-compatible dialect (same key, /apps/anthropic path): |
| 917 | # # provider = "modelstudio-token-plan-anthropic" |
| 918 | # # base_url = "https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic" |
| 919 | |
| 920 | # ───────────────────────────────────────────────────────────────────────────────── |
| 921 | # Alibaba Cloud Model Studio — Coding Plan |
| 922 | # (https://bailian.console.aliyun.com/) |
| 923 | # |
| 924 | # Coding Plan uses a separate international endpoint (coding-intl.dashscope). |
| 925 | # Same model catalog as the Token Plan. |
| 926 | # |
| 927 | # Provider aliases: modelstudio-coding-plan, modelstudio_coding_plan, |
| 928 | # alibaba-coding-plan, dashscope-coding-plan |
| 929 | # Env var aliases: MODELSTUDIO_API_KEY (preferred), DASHSCOPE_API_KEY, |
| 930 | # MODELSTUDIO_CODING_PLAN_BASE_URL, MODELSTUDIO_CODING_PLAN_MODEL |
| 931 | [providers.modelstudio_coding_plan] |
| 932 | # api_key = "YOUR_MODELSTUDIO_API_KEY" |
| 933 | # base_url = "https://coding-intl.dashscope.aliyuncs.com/v1" |
| 934 | # model = "qwen3.8-max" |
| 935 | # # Anthropic-compatible dialect (same key, /apps/anthropic path): |
| 936 | # # provider = "modelstudio-coding-plan-anthropic" |
| 937 | # # base_url = "https://coding-intl.dashscope.aliyuncs.com/apps/anthropic" |
| 938 | |
| 939 | # ───────────────────────────────────────────────────────────────────────────────── |
| 940 | # Together AI Provider (https://www.together.ai/) |
| 941 | # Env var aliases: TOGETHER_API_KEY, TOGETHER_BASE_URL, TOGETHER_MODEL |
| 942 | [providers.together] |
| 943 | # api_key = "YOUR_TOGETHER_API_KEY" |
| 944 | # base_url = "https://api.together.xyz/v1" |
| 945 | # model = "deepseek-ai/DeepSeek-V4-Pro" # or deepseek-ai/DeepSeek-V4-Flash |
| 946 | |
| 947 | # ───────────────────────────────────────────────────────────────────────────────── |
| 948 | # Baidu Qianfan Provider (https://intl.cloud.baidu.com/product/qianfan.html) |
| 949 | # Provider aliases: qianfan, baidu-qianfan, baidu_qianfan, baidu |
| 950 | # Env var aliases: QIANFAN_API_KEY / BAIDU_QIANFAN_API_KEY, |
| 951 | # QIANFAN_BASE_URL / BAIDU_QIANFAN_BASE_URL, |
| 952 | # QIANFAN_MODEL / BAIDU_QIANFAN_MODEL |
| 953 | [providers.qianfan] |
| 954 | # api_key = "YOUR_QIANFAN_API_KEY" |
| 955 | # base_url = "https://api.baiduqianfan.ai/v1" |
| 956 | # model = "ernie-4.0-turbo-8k" # or your Qianfan service/model id |
| 957 | |
| 958 | # ───────────────────────────────────────────────────────────────────────────────── |
| 959 | # OpenAI Codex (ChatGPT) Provider — EXPERIMENTAL |
| 960 | # Run `codex login`, then explicitly grant read-only access to that exact file: |
| 961 | # codewhale auth external-consent --provider openai-codex --mode read-only |
| 962 | # Codewhale never refreshes or rewrites the Codex CLI file. No API key is |
| 963 | # stored here. Talks to the OpenAI Responses API at /codex/responses. |
| 964 | # Env var aliases: OPENAI_CODEX_ACCESS_TOKEN / CODEX_ACCESS_TOKEN (token override), |
| 965 | # OPENAI_CODEX_BASE_URL / CODEX_BASE_URL, OPENAI_CODEX_MODEL / CODEX_MODEL, |
| 966 | # OPENAI_CODEX_ACCOUNT_ID / CODEX_ACCOUNT_ID, OPENAI_CODEX_AUTH_FILE, CODEX_HOME |
| 967 | [providers.openai_codex] |
| 968 | # base_url = "https://chatgpt.com/backend-api" |
| 969 | # model = "gpt-5.5" |
| 970 | # The CLI writes this table after informed consent; do not copy it between |
| 971 | # providers or machines. Absence means disabled. |
| 972 | # [providers.openai_codex.external_credentials] |
| 973 | # access = "read_only" |
| 974 | # provider = "openai-codex" |
| 975 | # source = "codex_cli" |
| 976 | # path = "/absolute/path/to/.codex/auth.json" |
| 977 | # consent_version = 1 |
| 978 | |
| 979 | # ───────────────────────────────────────────────────────────────────────────────── |
| 980 | # Anthropic Provider (native Messages API) |
| 981 | # Talks to https://api.anthropic.com/v1/messages with x-api-key auth — not an |
| 982 | # OpenAI-compatible route. Models: claude-opus-4-8, claude-sonnet-4-6 (default), |
| 983 | # claude-haiku-4-5. Env vars: ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL, |
| 984 | # ANTHROPIC_MODEL. |
| 985 | [providers.anthropic] |
| 986 | # api_key = "sk-ant-..." |
| 987 | # base_url = "https://api.anthropic.com" |
| 988 | # model = "claude-sonnet-4-6" |
| 989 | |
| 990 | # OpenModel Provider (Anthropic-compatible Messages API) |
| 991 | # Talks to https://api.openmodel.ai/v1/messages with Bearer auth. OpenModel |
| 992 | # routes DeepSeek, DashScope, Xiaomi, Claude, and other models by model id. |
| 993 | # Env vars: OPENMODEL_API_KEY, OPENMODEL_BASE_URL, OPENMODEL_MODEL. |
| 994 | [providers.openmodel] |
| 995 | # api_key = "om-..." |
| 996 | # base_url = "https://api.openmodel.ai" |
| 997 | # model = "deepseek-v4-flash" |
| 998 | |
| 999 | # Codewhale API (https://app.codewhale.net/settings?section=api) |
| 1000 | # Account-backed model access over the provider keys connected to the |
| 1001 | # customer's Codewhale account. The account's authenticated `GET /v1/models` |
| 1002 | # listing is the catalog authority: each `provider/model` row states its wire |
| 1003 | # (`chat-completions` → `/v1/chat/completions`, `anthropic-messages` → |
| 1004 | # `/v1/messages`, `responses` → `/v1/responses`). Every protocol authenticates |
| 1005 | # the account key with `Authorization: Bearer`, never `x-api-key`. |
| 1006 | # Provider aliases: codewhale-api, cw-api, codewhale-cloud |
| 1007 | # Env vars: CODEWHALE_API_KEY, CODEWHALE_API_BASE (HTTPS except loopback); |
| 1008 | # the generic CODEWHALE_MODEL override also applies. |
| 1009 | [providers.codewhale] |
| 1010 | # api_key = "cwc_key_..." # or `codewhale account api-keys create --name <name> --use` |
| 1011 | # base_url = "https://api.codewhale.net/v1" |
| 1012 | # model = "deepseek/deepseek-v4-pro" # provider/model exactly as the account catalog returns it |
| 1013 | |
| 1014 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1015 | # Web Search Provider |
| 1016 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1017 | # Choose which backend the Web tool's search uses. Default is keyless Firecrawl |
| 1018 | # (bounded per-IP quota; `[search] api_key` or FIRECRAWL_API_KEY raises it). |
| 1019 | # Bing and DuckDuckGo HTML scraping need no API key. Tavily, Bocha, Metaso, |
| 1020 | # Baidu, Volcengine, Sofya and Serply need an api_key; SearXNG needs base_url. |
| 1021 | # API runtime failures and empty responses visibly degrade through DuckDuckGo |
| 1022 | # then Bing. Missing configuration and network-policy denials fail closed. |
| 1023 | # |
| 1024 | # [search] |
| 1025 | # provider = "firecrawl" # firecrawl | bing | duckduckgo | tavily | bocha | metaso | searxng | baidu | volcengine | sofya | serply |
| 1026 | # # firecrawl: Firecrawl Cloud search, keyless with a bounded per-IP quota |
| 1027 | # # duckduckgo: HTML scrape with Bing fallback |
| 1028 | # # bing: HTML scrape, no API key |
| 1029 | # # tavily: https://tavily.com — AI search, needs api_key |
| 1030 | # # TAVILY_API_KEY set + provider unset auto-selects tavily |
| 1031 | # # at runtime; it is never written to config.toml |
| 1032 | # # bocha: https://bochaai.com — 博查AI搜索，国内友好，需api_key |
| 1033 | # # metaso: https://metaso.cn — 秘塔AI搜索，需 api_key |
| 1034 | # # 设置 METASO_API_KEY 或 [search] api_key |
| 1035 | # # searxng: https://docs.searxng.org — trusted/self-hosted JSON API, |
| 1036 | # # set base_url; no public instance is used by default |
| 1037 | # # instance must enable JSON (settings.yml search.formats); HTML-only instances 403 |
| 1038 | # # baidu: 百度 AI Search via qianfan.baidubce.com，需 api_key |
| 1039 | # # volcengine: 火山引擎 Ark web_search (免费 2 万次/月), 需 api_key |
| 1040 | # # 也回退到 VOLCENGINE_API_KEY / VOLCENGINE_ARK_API_KEY / ARK_API_KEY 环境变量 |
| 1041 | # # sofya: https://sofya.co — AI search returning full page |
| 1042 | # # content (not snippets), needs api_key (ay_live_...); |
| 1043 | # # also falls back to the SOFYA_API_KEY env var |
| 1044 | # # serply: https://serply.io Google organic results with |
| 1045 | # # snippets, needs api_key; |
| 1046 | # # also falls back to the SERPLY_API_KEY env var |
| 1047 | # base_url = "https://search.example/" # optional DuckDuckGo-compatible HTML endpoint; |
| 1048 | # # required SearXNG root or /search endpoint |
| 1049 | # api_key = "YOUR_SEARCH_KEY" # required for tavily, bocha, metaso, baidu, volcengine, sofya, and serply; optional for firecrawl (raises the keyless quota); unused by searxng |
| 1050 | # # WARNING: treat config.toml like a secret file when |
| 1051 | # # storing API keys. Prefer env vars for local smoke tests. |
| 1052 | # |
| 1053 | # Env-var overrides: |
| 1054 | # CODEWHALE_SEARCH_PROVIDER → search.provider |
| 1055 | # DEEPSEEK_SEARCH_PROVIDER → search.provider (legacy alias) |
| 1056 | # CODEWHALE_SEARCH_API_KEY → search.api_key |
| 1057 | # DEEPSEEK_SEARCH_API_KEY → search.api_key (legacy alias) |
| 1058 | # CODEWHALE_SEARCH_BASE_URL → search.base_url |
| 1059 | # DEEPSEEK_SEARCH_BASE_URL → search.base_url (legacy alias) |
| 1060 | # FIRECRAWL_API_KEY → firecrawl key fallback |
| 1061 | # METASO_API_KEY → metaso key fallback |
| 1062 | # BAIDU_SEARCH_API_KEY → baidu key fallback |
| 1063 | # VOLCENGINE_API_KEY / VOLCENGINE_ARK_API_KEY / ARK_API_KEY → volcengine key fallback |
| 1064 | # SOFYA_API_KEY → sofya key fallback |
| 1065 | # SERPLY_API_KEY → serply key fallback |
| 1066 | # TAVILY_API_KEY → tavily autodetect when provider is unset (explicit |
| 1067 | # provider = "..." or CODEWHALE_SEARCH_PROVIDER still |
| 1068 | # wins; the autodetect is never persisted) |
| 1069 | |
| 1070 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1071 | # Network Policy (#135) |
| 1072 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1073 | # Per-domain allow/deny rules for outbound network calls made by the TUI's |
| 1074 | # tools (`fetch_url`, `web_search`) and the MCP HTTP transport. Stdio MCP |
| 1075 | # servers and direct LLM API calls are unaffected. |
| 1076 | # |
| 1077 | # Precedence: deny wins. A host listed in both `allow` and `deny` is denied. |
| 1078 | # |
| 1079 | # Host-matching rules: |
| 1080 | # - Exact match: `api.deepseek.com` matches only `api.deepseek.com`. |
| 1081 | # - Subdomain wildcard: an entry starting with `.` (e.g. `.example.com`) |
| 1082 | # matches `api.example.com` and `a.b.example.com` but not the apex |
| 1083 | # `example.com`. To cover both, list both. `*.example.com` is also accepted. |
| 1084 | # |
| 1085 | # Defaults are intentionally conservative: when this section is absent, no |
| 1086 | # policy is enforced (mirrors pre-v0.7.0 behavior). To opt in: |
| 1087 | # |
| 1088 | # [network] |
| 1089 | # default = "prompt" # allow | deny | prompt |
| 1090 | # allow = ["api.deepseek.com", "github.com", ".githubusercontent.com"] |
| 1091 | # deny = [] |
| 1092 | # proxy = ["github.com", ".githubusercontent.com"] |
| 1093 | # proxy_fake_ip_cidrs = ["198.18.0.0/15"] # requires both matching host and address |
| 1094 | # audit = true # one line per call to ~/.codewhale/audit.log |
| 1095 | |
| 1096 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1097 | # Verifier preview (#2093) |
| 1098 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1099 | # Enables automatic claim-of-done verifier preview once the runtime trigger is |
| 1100 | # active. Manual `run_verifiers` remains available even when this is false. |
| 1101 | # |
| 1102 | # [verifier] |
| 1103 | # enabled = false |
| 1104 | |
| 1105 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1106 | # Advisor / Watcher (#3982) |
| 1107 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1108 | # Optional background watcher that fires after each turn that contains tool |
| 1109 | # calls. It reads a bounded slice of recent tool interactions, makes a concise |
| 1110 | # LLM advisory call, and emits a brief note into the status area. |
| 1111 | # |
| 1112 | # Off by default. Enable with `[advisor] enabled = true` or toggle per-session |
| 1113 | # with `/advisor on` / `/advisor off`. |
| 1114 | # |
| 1115 | # Options (the full set — `AdvisorConfigToml` in crates/config/src/lib.rs): |
| 1116 | # enabled — master switch (default: false) |
| 1117 | # max_tool_calls — number of recent tool call/result pairs to read |
| 1118 | # (default: 10, clamped to 1–50) |
| 1119 | # rate_limit_secs — minimum seconds between successive advisor notes |
| 1120 | # (default: 60, clamped to 5–3600) |
| 1121 | # dedup_window_secs — notes whose content hash matches the previous note |
| 1122 | # within this window are dropped (default: 300) |
| 1123 | # model — model override for the advisory call; when absent the |
| 1124 | # advisor reuses the session's current model |
| 1125 | # |
| 1126 | # Unknown keys under [advisor] are silently ignored, not rejected — a typo |
| 1127 | # leaves the default in place with no error. |
| 1128 | # |
| 1129 | # [advisor] |
| 1130 | # enabled = false |
| 1131 | # max_tool_calls = 10 |
| 1132 | # rate_limit_secs = 60 |
| 1133 | # dedup_window_secs = 300 |
| 1134 | # model = "deepseek-v4-flash" |
| 1135 | |
| 1136 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1137 | # Skills (#140) |
| 1138 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1139 | # Settings for the `/skill install <spec>` community-skill installer. |
| 1140 | # * registry_url — curated index.json that resolves bare names to |
| 1141 | # `github:owner/repo` specs. Override to point at |
| 1142 | # a private fork or internal mirror. |
| 1143 | # * max_install_size_bytes — per-skill uncompressed size cap. Tarballs that |
| 1144 | # exceed this limit are rejected during validation. |
| 1145 | # Default: 5 MiB. |
| 1146 | # |
| 1147 | # `/skill install` is gated by `[network]`. Make sure `github.com` and |
| 1148 | # `raw.githubusercontent.com` are reachable (default `prompt` is fine — you'll |
| 1149 | # be asked once and can persist) before running it. |
| 1150 | # |
| 1151 | # [skills] |
| 1152 | # registry_url = "https://raw.githubusercontent.com/Hmbown/deepseek-skills/main/index.json" |
| 1153 | # max_install_size_bytes = 5_242_880 |
| 1154 | # scan_codewhale_only = false # true: ignore Claude/OpenCode/Cursor/agentskills.io skill dirs |
| 1155 | |
| 1156 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1157 | # TUI |
| 1158 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1159 | [tui] |
| 1160 | alternate_screen = "auto" # auto/always start on the alternate screen; never starts inline (keeps terminal scrollback). /fullscreen and /inline switch at runtime |
| 1161 | mouse_capture = true # true: TUI-owned mouse selection (payload set by selection_copy_markdown); false: raw terminal selection/copy |
| 1162 | selection_copy_markdown = true # true: copy transcript drag selections as Markdown source; false: rendered text |
| 1163 | terminal_probe_timeout_ms = 500 # optional startup terminal-mode timeout (100-5000ms) |
| 1164 | stream_chunk_timeout_secs = 900 # optional SSE idle timeout per chunk (0 = default, 1-3600) |
| 1165 | # Model steps are uncapped by default. Uncomment to install an explicit |
| 1166 | # ceiling; omission or 0 means no step limit, positive values clamp to 1-100000. |
| 1167 | # max_model_steps = 1000 |
| 1168 | # Wall-clock and stream budgets remain finite; 0 selects their default. |
| 1169 | turn_wall_clock_secs = 3600 # cumulative wall clock for one turn, excluding time |
| 1170 | # blocked on a human approval (0 = default, 30-86400) |
| 1171 | stream_max_content_mb = 10 # per-step cap on streamed content (0 = default, 1-512) |
| 1172 | stream_max_duration_secs = 1800 # per-step cap on one stream's duration (0 = default, 10-86400) |
| 1173 | osc8_links = true # emit OSC 8 escapes around URLs (Cmd+click in iTerm2/Ghostty/Kitty/WezTerm/Terminal.app 13+); set false for terminals that misrender |
| 1174 | # What the bottom chrome shows. Each key is one thing on screen: `mode` is the |
| 1175 | # posture bar's plan/act/operate chip, everything else is a segment of the |
| 1176 | # metrics line under it. Omit the key to use the built-in default; set [] to |
| 1177 | # strip the metrics line down to the help hint. You can also edit this |
| 1178 | # interactively with `/statusline`. |
| 1179 | # Supported keys: mode, model, context_percent, cost, |
| 1180 | # balance (prepaid providers only: DeepSeek, DeepSeekCN, OpenRouter, |
| 1181 | # SiliconFlow), cache, tokens, ttft, output_rate, workspace, git_branch. |
| 1182 | # Legacy session_metrics enables both ttft and output_rate. |
| 1183 | # Retired in 0.9.13: status, agents, reasoning_replay, prefix_stability, |
| 1184 | # last_tool_elapsed, rate_limit — they drove nothing. Old files |
| 1185 | # keep loading; the retired keys are ignored. |
| 1186 | # status_items = ["mode", "model", "context_percent", "cost", "tokens"] |
| 1187 | # Size presets for the two rows themselves (#5950) — composition stays in |
| 1188 | # status_items; these only decide how much of a row paints: |
| 1189 | # posture_bar = "full" # full | compact | hidden (default full) |
| 1190 | # # compact keeps the posture chips (and the cap |
| 1191 | # # warning) and drops the clocks, counts and hint; |
| 1192 | # # hidden gives the row to the transcript. |
| 1193 | # metrics_line = "compact" # full | compact | hidden (default compact) |
| 1194 | # # compact keeps the route, context reading, cost |
| 1195 | # # and balance plus selected TTFT/rate when they fit. |
| 1196 | # # It drops secondary counts and help; hidden gives the row to the transcript. |
| 1197 | # # Also settable at runtime: /config posture_bar compact |
| 1198 | # notification_condition = "unfocused" # unfocused | always | never |
| 1199 | # "unfocused" = notify only after this terminal has been |
| 1200 | # in the background for two seconds (default); |
| 1201 | # "always" = allow configured notifications while focused; |
| 1202 | # successful turns skip the threshold; |
| 1203 | # "never" = suppress all operator notifications. |
| 1204 | # locale = "auto" # UI chrome language: auto | en | ja | zh-Hans | zh-Hant | pt-BR | es-419 |
| 1205 | # # | vi | ko | ca | de | fr | id | hi | ru | uk |
| 1206 | # # "auto" reads LC_ALL → LC_MESSAGES → LANG; falls back to English. |
| 1207 | # # Override: `locale = "zh-Hans"` for Simplified Chinese regardless of OS locale. |
| 1208 | # # Also settable at runtime: /config locale zh-Hans |
| 1209 | # # Note: this only affects TUI labels/chrome — it does NOT change model output language. |
| 1210 | # mention_menu_behavior = "fuzzy" # fuzzy | browser; browser lists immediate directory children for @-mentions. |
| 1211 | |
| 1212 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1213 | # Transcript |
| 1214 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1215 | # Prose — user messages, assistant answers, and reasoning/thinking — fills the |
| 1216 | # full content width, consistent with tool/status cells and the wide-frame |
| 1217 | # decision in #5322 (#5436). Owners who want a bounded reading measure on |
| 1218 | # ultrawide terminals can cap it in columns. |
| 1219 | [transcript] |
| 1220 | # prose_measure = 120 # positive integer: cap prose wrap at N columns. |
| 1221 | # 0 or absent = full content width (default). |
| 1222 | # Must be a positive whole number; tool, diff, and |
| 1223 | # status cells always keep the full content width. |
| 1224 | |
| 1225 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1226 | # Feature Flags |
| 1227 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1228 | [features] |
| 1229 | shell_tool = true |
| 1230 | subagents = true |
| 1231 | web_search = true # enables canonical web.run plus the compatibility web_search alias |
| 1232 | apply_patch = true |
| 1233 | mcp = true |
| 1234 | exec_policy = true |
| 1235 | # vision_model = false # enable vision model for image_analyze tool |
| 1236 | # verify_tool = false # disable the agent-callable `verify` self-critique tool |
| 1237 | # (#4196). On by default; the agent decides when to spend |
| 1238 | # the extra reasoning, so cost is only incurred on demand. |
| 1239 | # Set false to remove it from the model's tool catalog. |
| 1240 | |
| 1241 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1242 | # Vision Model Configuration (optional) |
| 1243 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1244 | # Uses an OpenAI-compatible vision model API for the `image_analyze` tool. |
| 1245 | # api_key inherits from the main config if not specified. |
| 1246 | # |
| 1247 | # [vision_model] |
| 1248 | # model = "gemini-3.1-flash-lite-preview" # Required: vision-capable model ID |
| 1249 | # api_key = "YOUR_API_KEY" # Optional: defaults to main api_key |
| 1250 | # base_url = "https://generativelanguage.googleapis.com/v1beta/openai/" # Optional |
| 1251 | # |
| 1252 | # Xiaomi MiMo image understanding can be configured through the same tool: |
| 1253 | # model = "mimo-v2.5" |
| 1254 | # api_key = "YOUR_XIAOMI_KEY" |
| 1255 | # base_url = "https://token-plan-sgp.xiaomimimo.com/v1" # Token Plan / tp- keys |
| 1256 | |
| 1257 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1258 | # Retry Configuration |
| 1259 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1260 | [retry] |
| 1261 | enabled = true |
| 1262 | max_retries = 3 |
| 1263 | initial_delay = 1.0 |
| 1264 | max_delay = 60.0 |
| 1265 | exponential_base = 2.0 |
| 1266 | |
| 1267 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1268 | # Goal loop (`[goal]`) — operate-mode persistent goals |
| 1269 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1270 | # Operate-mode goals run to their completion gate with no default token, time, |
| 1271 | # or continuation ceiling. Token/time budgets, when supplied, are telemetry |
| 1272 | # only and do not stop a goal. The keys below are the opt-in circuit breakers. |
| 1273 | # [goal] |
| 1274 | # Optional safety backstop on automatic goal continuation passes. |
| 1275 | # Default: 0 (unlimited). Set a positive value to opt into a ceiling. |
| 1276 | # max_continuations = 100 |
| 1277 | # Optional cancellable quiet period between successful turns, useful for |
| 1278 | # coordinator goals that poll on a cadence instead of keeping one provider |
| 1279 | # turn open. Default: 0 (continue immediately). Cap: 86400 (24h). |
| 1280 | # continuation_delay_seconds = 300 |
| 1281 | # Per-turn step allowance while a goal is active (#5994): larger but still |
| 1282 | # finite. Default: 1000 (0/absent resolves to 1000). Range: 1..=100,000. |
| 1283 | # Bounds each turn, never the number of continuation passes; explicit |
| 1284 | # per-invocation ceilings (exec --max-turns, worker caps) always win. |
| 1285 | # max_steps = 1000 |
| 1286 | |
| 1287 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1288 | # Context Compaction |
| 1289 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1290 | # Auto-compaction is a saved UI setting edited with `/config` (`auto_compact`). |
| 1291 | # The optional saved threshold setting is `auto_compact_threshold_percent` |
| 1292 | # (default 80). There is no config-file |
| 1293 | # `[compaction]` table yet; runtime compaction budgets are chosen by the TUI |
| 1294 | # from the active model/context window. |
| 1295 | |
| 1296 | # [context] supports `enabled` (default false) and `project_pack` (#4781). |
| 1297 | # The old seam-manager keys (verbatim_window_turns, l1/l2/l3_threshold, |
| 1298 | # seam_model) are parsed but ignored — seam_manager.rs no longer exists. |
| 1299 | |
| 1300 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1301 | # Workshop / Large-Output Routing (#548) |
| 1302 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1303 | # Tool outputs exceeding `large_output_threshold_tokens` are routed through a |
| 1304 | # V4-Flash synthesis sub-agent. Only the synthesis reaches the parent context; |
| 1305 | # the raw text is stored in the workshop variable `last_tool_result` so the |
| 1306 | # parent can call `promote_to_context` later if it needs the full content. |
| 1307 | # |
| 1308 | # Per-tool overrides let high-volume tools (e.g. Bash) use tighter |
| 1309 | # thresholds without changing the global default. Name tools by the identifiers |
| 1310 | # the model actually sees — `exec_shell`, `grep_files`, and `web_search` were |
| 1311 | # retired and an override keyed to them matches nothing. |
| 1312 | # |
| 1313 | # There is no per-call `raw = true` escape. It was documented here, but |
| 1314 | # `raw = true` on a tool call does NOT bypass routing: the adaptive router |
| 1315 | # ignores it, and it is honoured only under the legacy |
| 1316 | # `CODEWHALE_CLASSIC_OUTPUT_ROUTING` switch. No tool advertises it. |
| 1317 | # |
| 1318 | # [workshop] |
| 1319 | # large_output_threshold_tokens = 4096 |
| 1320 | # # Optional model-visible byte ceilings (#5367). Absent keeps the |
| 1321 | # # compile-time defaults (read 50KiB / read_file 16KiB, then the |
| 1322 | # # compact 12K-char floor). Values raise the floor; they never lower |
| 1323 | # # it. Hard cap is 2MiB. |
| 1324 | # # read_result_max_bytes = 102400 |
| 1325 | # # tool_result_max_bytes = 102400 |
| 1326 | # [workshop.per_tool_thresholds] |
| 1327 | # Bash = 2048 # shell output synthesised aggressively |
| 1328 | # Web = 8192 # web results can be large; give them more room |
| 1329 | |
| 1330 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1331 | # Profile Example (for multiple environments) |
| 1332 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1333 | # Select a profile with `deepseek --profile <name>` or `DEEPSEEK_PROFILE=<name>`. |
| 1334 | [profiles.work] |
| 1335 | api_key = "WORK_DEEPSEEK_API_KEY" |
| 1336 | base_url = "https://api.deepseek.com/beta" |
| 1337 | |
| 1338 | [profiles.dev] |
| 1339 | api_key = "DEV_DEEPSEEK_API_KEY" |
| 1340 | allow_shell = true |
| 1341 | |
| 1342 | [profiles.nvidia-nim] |
| 1343 | provider = "nvidia-nim" |
| 1344 | api_key = "YOUR_NVIDIA_API_KEY" |
| 1345 | base_url = "https://integrate.api.nvidia.com/v1" |
| 1346 | default_text_model = "deepseek-ai/deepseek-v4-pro" |
| 1347 | |
| 1348 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1349 | # Desktop Notifications (OSC 9 / BEL on long agent-turn completion) |
| 1350 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1351 | # Emits an escape sequence to the terminal when a turn **completes successfully** |
| 1352 | # and took longer than `threshold_secs`. Failed or cancelled turns are |
| 1353 | # intentionally silent. Useful when you tab away from the TUI and want an alert |
| 1354 | # for "your task is ready". |
| 1355 | # |
| 1356 | # method = "auto" # auto | osc9 | kitty | ghostty | bel | off |
| 1357 | # auto: native/terminal banner for known terminals. |
| 1358 | # On macOS, the native banner is silent; on Linux, |
| 1359 | # unknown terminals stay silent instead of ringing BEL. |
| 1360 | # osc9: \x1b]9;<msg>\x07 (iTerm2-style; shows macOS notification) |
| 1361 | # bel: explicit \x07 beep (MessageBeep on Windows) |
| 1362 | # off: disable banners; an explicit completion sound is independent |
| 1363 | # threshold_secs = 30 # only notify when the turn took >= this many seconds |
| 1364 | # include_summary = false # include elapsed time + cost in the notification body |
| 1365 | # subagent_completion = "final-only" # always | final-only | off — per-subagent |
| 1366 | # notifications during fleet/workflow runs. final-only |
| 1367 | # (default) stays quiet mid-run and fires once when the |
| 1368 | # batch drains; off silences them entirely. |
| 1369 | # completion_sound = "off" # off | beep | bell | file — opt-in turn-completion sound |
| 1370 | # sound_file = "E:\\google\\downloads\\notify.wav" # WAV used when completion_sound = "file" (Windows) |
| 1371 | [notifications] |
| 1372 | # method = "auto" |
| 1373 | # threshold_secs = 30 |
| 1374 | # include_summary = false |
| 1375 | # subagent_completion = "final-only" |
| 1376 | # completion_sound = "off" |
| 1377 | # sound_file = "E:\\google\\downloads\\notify.wav" |
| 1378 | |
| 1379 | # Lifecycle event outbox: opt-in JSONL stream of session/turn/subagent |
| 1380 | # lifecycle events for supervisors and automation harnesses. One JSON line |
| 1381 | # per event (RuntimeEventEnvelope schema), appended and flushed on every |
| 1382 | # emit; seq is monotonic per file and recovers from the last line on open. |
| 1383 | # UNCOMMENT `path` TO ENABLE — unset/empty = OFF = behavior unchanged. |
| 1384 | # Fires for interactive TUI sessions AND headless `codewhale exec` runs. |
| 1385 | # See docs/CONFIGURATION.md → Lifecycle Outbox for the file contract. |
| 1386 | # [lifecycle_outbox] |
| 1387 | # path = "~/.codewhale/notifications/outbox.jsonl" |
| 1388 | # webhook_url = "https://example.com/hooks/codewhale" # optional: POST {"at", "event"} JSON per event |
| 1389 | # webhook_token = "" # optional: sent as `Authorization: Bearer <token>` |
| 1390 | # # delivery is best-effort: failures are logged and dropped |
| 1391 | |
| 1392 | # Opt-in per-event sound cues (#4817): deterministic, terminal-bell level |
| 1393 | # (BEL bytes only — functional signals, platform-safe no-op when the terminal |
| 1394 | # ignores BEL). Off by default. When completion_sound is active, turn-complete |
| 1395 | # is left to that channel so the two never double-ding. |
| 1396 | # [notifications.event_sound] |
| 1397 | # enabled = false # master switch, default off |
| 1398 | # events = ["turn-complete", "approval-needed"] # allow-list; unknown names ignored |
| 1399 | # min_interval_ms = 2000 # per-event rate limit |
| 1400 | # quiet = false # true silences all event sounds |
| 1401 | |
| 1402 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1403 | # Workspace Snapshots (#137) |
| 1404 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1405 | # Each turn the TUI takes a `pre-turn:<seq>` and `post-turn:<seq>` snapshot of |
| 1406 | # your workspace into a side-git repo at: |
| 1407 | # |
| 1408 | # ~/.codewhale/snapshots/<project_hash>/<worktree_hash>/.git |
| 1409 | # |
| 1410 | # Your own `.git` is never touched — `--git-dir` and `--work-tree` are always |
| 1411 | # set together when shelling out to git. Use `/restore N` (slash command) or |
| 1412 | # the `revert_turn` tool to roll the working tree back. Conversation history |
| 1413 | # is unaffected. |
| 1414 | # |
| 1415 | # Disk footprint: ~1-2 GB worst case for a 100 MB workspace × 12 turns/day, |
| 1416 | # typically far less thanks to git's content-addressed storage. The session |
| 1417 | # boot prunes anything older than `max_age_days` (default 7). |
| 1418 | # |
| 1419 | # [snapshots] |
| 1420 | # enabled = true # Snapshot workspace pre/post each turn for /restore |
| 1421 | # max_age_days = 7 # Older snapshots pruned at session start |
| 1422 | # max_workspace_gb = 2 # Snapshots self-disable on first init when the |
| 1423 | # # non-excluded workspace exceeds this size in GB |
| 1424 | # # (v0.8.32). Default 2 GB protects against running |
| 1425 | # # codewhale in directories with hundreds of GB |
| 1426 | # # of datasets / model weights / docker dumps where |
| 1427 | # # `git add -A` would hang the TUI for hours. Set |
| 1428 | # # to 0 to disable the cap (v0.8.31 behaviour); |
| 1429 | # # raise to a higher number for legitimate large |
| 1430 | # # monorepos. |
| 1431 | |
| 1432 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1433 | # LSP Diagnostics (post-edit) (#136) |
| 1434 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1435 | # After every successful file edit (`edit_file`, `apply_patch`, `write_file`), |
| 1436 | # the engine asks an LSP server for diagnostics on the file and injects them |
| 1437 | # as a synthetic system message before the next API call. This lets the agent |
| 1438 | # see compile breaks immediately without round-tripping through the user. |
| 1439 | # |
| 1440 | # Enabled by default. Failure modes are non-blocking: a missing LSP binary, |
| 1441 | # a crashed server, or a timeout simply skips the post-edit hook for that |
| 1442 | # turn — the agent's work is never blocked. |
| 1443 | # |
| 1444 | # Built-in language → server defaults: |
| 1445 | # rust → rust-analyzer |
| 1446 | # go → gopls serve |
| 1447 | # python → pyright-langserver --stdio |
| 1448 | # typescript → typescript-language-server --stdio |
| 1449 | # java → jdtls |
| 1450 | # php → intelephense --stdio |
| 1451 | # vue → vue-language-server --stdio |
| 1452 | # c, cpp → clangd |
| 1453 | # |
| 1454 | # Java support uses Eclipse JDT LS via the `jdtls` command. IntelliJ IDEA is |
| 1455 | # not required, and installing IntelliJ IDEA alone does not install `jdtls`. |
| 1456 | # |
| 1457 | # Override the defaults via the `servers` table below. |
| 1458 | # |
| 1459 | # For languages not in the built-in list (Ruby, C#, Swift, etc.), use |
| 1460 | # `[lsp.custom.<ext>]` to register a language server: |
| 1461 | # |
| 1462 | # [lsp.custom.rb] |
| 1463 | # command = "ruby-lsp" |
| 1464 | # args = ["--stdio"] |
| 1465 | # language_id = "ruby" |
| 1466 | # |
| 1467 | # [lsp.custom.cs] |
| 1468 | # command = "csharp-ls" |
| 1469 | # args = [] |
| 1470 | # language_id = "csharp" |
| 1471 | # |
| 1472 | # [lsp.custom.swift] |
| 1473 | # command = "sourcekit-lsp" |
| 1474 | # language_id = "swift" |
| 1475 | [lsp] |
| 1476 | # enabled = true |
| 1477 | # poll_after_edit_ms = 5000 |
| 1478 | # max_diagnostics_per_file = 20 |
| 1479 | # include_warnings = false |
| 1480 | # [lsp.servers] |
| 1481 | # rust = ["rust-analyzer"] |
| 1482 | # go = ["gopls", "serve"] |
| 1483 | # java = ["jdtls"] |
| 1484 | # php = ["intelephense", "--stdio"] |
| 1485 | # vue = ["vue-language-server", "--stdio"] |
| 1486 | |
| 1487 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1488 | # Hooks (optional) |
| 1489 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1490 | # Hooks run shell commands on lifecycle events (session start/end, tool calls, etc.). |
| 1491 | # Configure as `[[hooks.hooks]]` under a `[hooks]` table. |
| 1492 | # |
| 1493 | # SCOPE: hooks are a TUI runtime feature. They fire from the interactive TUI |
| 1494 | # and the engine turn loop it drives. `codewhale exec`, the CLI subcommands, |
| 1495 | # the app-server / ACP surfaces, and the `workflow` tool do NOT fire them. |
| 1496 | # |
| 1497 | # Available events (all 11): session_start, session_end, message_submit, |
| 1498 | # tool_call_before, tool_call_after, mode_change, on_error, turn_end, |
| 1499 | # subagent_spawn, subagent_complete, shell_env. |
| 1500 | # See docs/HOOKS.md for the per-event payload, env var, and steering contract. |
| 1501 | # |
| 1502 | # `message_submit`, `tool_call_before`, and `shell_env` are the only events |
| 1503 | # whose result can change what Codewhale does. The rest are observer-only — |
| 1504 | # which means their RESULT is ignored, not that the command is harmless. Every |
| 1505 | # hook is an arbitrary shell command running with your credentials. |
| 1506 | # |
| 1507 | # Note: `default_timeout_secs` below OVERRIDES each hook's own `timeout_secs`. |
| 1508 | # Leave it unset if you want per-hook timeouts to apply. `/hooks list` shows the |
| 1509 | # effective value and names the override. `default_timeout_secs = 0` is |
| 1510 | # REJECTED at load — it would expire every hook immediately — and per-hook |
| 1511 | # `timeout_secs` applies instead. The timeout applies to background hooks too: |
| 1512 | # on expiry the whole process group is killed and then reaped, best-effort, |
| 1513 | # with a bounded reap wait (see docs/HOOKS.md → Timeouts). |
| 1514 | # |
| 1515 | # `background = true` means submitted and never awaited. The hook still gets |
| 1516 | # the documented stdin payload, environment, and timeout — it just has no exit |
| 1517 | # code, so it cannot steer. `shell_env` ignores the flag and always runs in the |
| 1518 | # foreground because its stdout is the contract. |
| 1519 | # |
| 1520 | # A condition that references context its event never carries is REJECTED at |
| 1521 | # load, logged, and shown by `/hooks list` — for example an `exit_code` |
| 1522 | # condition outside `tool_call_after` / `on_error`, or a `mode` condition on |
| 1523 | # `shell_env`. Rejection is per entry: a broken hook never drops another one |
| 1524 | # that happens to share its name. `on_error` fires for tool failures with the |
| 1525 | # tool name, call id, and reported exit code attached, so tool-scoped and |
| 1526 | # exit-code-scoped `on_error` hooks are supported. |
| 1527 | # |
| 1528 | # `shell_env` (#456) is special: the hook runs immediately before each |
| 1529 | # `exec_shell` invocation and its stdout is parsed as `KEY=VALUE\n` lines. |
| 1530 | # Those vars are applied on top of `exec_shell`'s environment. For LOCAL |
| 1531 | # execution that environment is built from a fixed allowlist of parent |
| 1532 | # variables (PATH, HOME, LANG, TERM, …) — an ambient secret exported in your |
| 1533 | # terminal is NOT forwarded to `exec_shell` by itself, so this hook is the |
| 1534 | # supported way to supply one. If an external sandbox backend is configured, |
| 1535 | # that allowlist does NOT apply: the backend owns its base environment and your |
| 1536 | # `shell_env` values are TRANSMITTED to it. Later hooks override |
| 1537 | # earlier ones. Use this for ephemeral credentials, per-skill PATH adjustments, |
| 1538 | # or short-lived tokens. The resolved KEY names (NEVER values) are written to |
| 1539 | # `~/.codewhale/audit.log` so each session can be reconciled later. Hook |
| 1540 | # failure / timeout simply contributes no vars — it does not abort the shell |
| 1541 | # call. |
| 1542 | # |
| 1543 | # [hooks] |
| 1544 | # enabled = true |
| 1545 | # default_timeout_secs = 30 |
| 1546 | # |
| 1547 | # [[hooks.hooks]] |
| 1548 | # event = "session_start" |
| 1549 | # command = "echo 'Codewhale session started'" |
| 1550 | # |
| 1551 | # # Inject ephemeral creds into every shell call. Output one |
| 1552 | # # KEY=VALUE per line on stdout (export prefix optional). |
| 1553 | # [[hooks.hooks]] |
| 1554 | # name = "aws-creds" |
| 1555 | # event = "shell_env" |
| 1556 | # command = "aws-vault export my-profile --format=env" |
| 1557 | # # Optionally limit to specific tool names / categories: |
| 1558 | # # condition = { type = "tool_category", category = "shell" } |
| 1559 | # |
| 1560 | # # Observe sub-agent lifecycle events. These hooks receive bounded JSON |
| 1561 | # # metadata on stdin and are warn-only: failures do not affect sub-agent |
| 1562 | # # scheduling, prompts, or results. continue_on_error has no effect for |
| 1563 | # # these observer events; later matching hooks always continue. |
| 1564 | # [[hooks.hooks]] |
| 1565 | # name = "subagent-audit" |
| 1566 | # event = "subagent_complete" |
| 1567 | # command = "~/.codewhale/hooks/subagent-audit.sh" |
| 1568 | |
| 1569 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1570 | # Runtime API (`deepseek serve --http`) (#561) |
| 1571 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1572 | # Tuning knobs for the local HTTP/SSE daemon. The server binds to 127.0.0.1 |
| 1573 | # by default and is intended for local UIs (whalescale-desktop, dashboards, |
| 1574 | # automation scripts). Today this section only controls the CORS allow-list; |
| 1575 | # host/port/workers stay on `--host`, `--port`, and `--workers` flags. |
| 1576 | # |
| 1577 | # Built-in defaults always include: |
| 1578 | # http://localhost:3000 http://127.0.0.1:3000 |
| 1579 | # http://localhost:1420 http://127.0.0.1:1420 |
| 1580 | # tauri://localhost |
| 1581 | # |
| 1582 | # Use `cors_origins` to add extra dev origins (e.g. Vite's default `:5173`). |
| 1583 | # User entries STACK on top of the defaults — they do not replace them. The |
| 1584 | # CLI flag `--cors-origin URL` (repeatable) and env var |
| 1585 | # `DEEPSEEK_CORS_ORIGINS=url1,url2` resolve to the same merged list. |
| 1586 | # |
| 1587 | # [runtime_api] |
| 1588 | # cors_origins = ["http://localhost:5173", "http://127.0.0.1:5173"] |
| 1589 | |
| 1590 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1591 | # Tool Overrides & Plugins ([tools]) |
| 1592 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1593 | # The `[tools]` table lets you replace any built-in tool with a custom |
| 1594 | # implementation (script or command) or disable it entirely — without |
| 1595 | # forking or recompiling the binary. |
| 1596 | # |
| 1597 | # Plugin scripts dropped in the plugin directory are auto-discovered and |
| 1598 | # registered as model-visible tools alongside the built-in ones. |
| 1599 | # |
| 1600 | # Scripts receive the tool's JSON input on **stdin** and must return a |
| 1601 | # JSON `ToolResult` (`{"content": "...", "success": true}`) on **stdout**. |
| 1602 | # |
| 1603 | # [tools] |
| 1604 | # # Custom plugin directory (defaults to `~/.codewhale/tools/`) |
| 1605 | # plugin_dir = "~/.codewhale/tools" |
| 1606 | # |
| 1607 | # [tools.overrides] |
| 1608 | # # Disable a tool entirely — removes it from the model-visible catalog. |
| 1609 | # "code_execution" = { type = "disabled" } |
| 1610 | # |
| 1611 | # # Replace a tool with a script. Relative paths resolve against plugin_dir. |
| 1612 | # "exec_shell" = { type = "script", path = "audit-exec-shell.sh" } |
| 1613 | # |
| 1614 | # # Replace a tool with a command (binary on PATH or absolute path). |
| 1615 | # "read_file" = { type = "command", command = "bat", args = ["--paging=never"] } |
| 1616 | # |
| 1617 | # # Scripts can also accept static arguments before the JSON input: |
| 1618 | # "fetch_url" = { type = "script", path = "cached-fetch.sh", args = ["--ttl", "300"] } |
| 1619 | |
| 1620 | # ──────────── Enterprise example: audit-logging exec_shell wrapper ────────────── |
| 1621 | # Drop `audit-exec-shell.sh` in `~/.codewhale/tools/` and enable with: |
| 1622 | # |
| 1623 | # [tools.overrides] |
| 1624 | # "exec_shell" = { type = "script", path = "audit-exec-shell.sh" } |
| 1625 | # |
| 1626 | # The wrapper logs every request to `~/.codewhale/audit/exec_shell.log`, then |
| 1627 | # delegates to your own approved shell executor. Do not pipe the raw JSON |
| 1628 | # request into `sh -s`; parse the command field and enforce your policy first. |
| 1629 | # |
| 1630 | # ```sh |
| 1631 | # #!/usr/bin/env sh |
| 1632 | # # name: exec_shell |
| 1633 | # # description: Audit-logging wrapper for exec_shell |
| 1634 | # # approval: required |
| 1635 | # LOGDIR="${HOME}/.codewhale/audit" |
| 1636 | # mkdir -p "$LOGDIR" |
| 1637 | # LOGFILE="$LOGDIR/exec_shell.log" |
| 1638 | # input=$(cat) |
| 1639 | # echo "[$(date -Iseconds)] $input" >> "$LOGFILE" |
| 1640 | # printf '%s\n' '{"content":"audit wrapper dry run: configure an executor","success":false}' |
| 1641 | # ``` |
| 1642 | |
| 1643 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1644 | # Workflow automatic launch, approval, isolation, and activity (#4128) |
| 1645 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1646 | # First-class knobs for automatic Workflow orchestration. When the table is |
| 1647 | # omitted entirely, the runtime uses these product defaults. Later launch, |
| 1648 | # approval, and activity-persistence paths all read through this one model. |
| 1649 | # [workflow] |
| 1650 | # # Allow the parent agent to auto-launch Workflow for multi-agent work. |
| 1651 | # # Set false to require an explicit `/workflow` opt-in. |
| 1652 | # automatic = true |
| 1653 | # # Auto-start read-only plans without an approval card when automatic is on. |
| 1654 | # auto_start_read_only = true |
| 1655 | # # Require an approval card before write/shell/network/high-budget launches. |
| 1656 | # require_approval_for_writes = true |
| 1657 | # # Hard ceiling on agents in one Workflow run (matches VM lifetime cap). |
| 1658 | # max_children = 1000 |
| 1659 | # # Maximum concurrently live agents inside one run (others wait for a slot). |
| 1660 | # max_concurrent = 16 |
| 1661 | # # Maximum structural nesting depth accepted for Workflow IR (default 5). |
| 1662 | # # This is separate from Runtime child delegation below. |
| 1663 | # max_depth = 5 |
| 1664 | # # Default shared token budget for a Workflow run and its children. |
| 1665 | # default_token_budget = 120000 |
| 1666 | |
| 1667 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1668 | # Agent Fleet roster, role registry, and execution requests (#3165, #3167) |
| 1669 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1670 | # [fleet] |
| 1671 | # # Fleet stores member identity and route intent. Project trust, host identity, |
| 1672 | # # secrets, approvals, sandboxing, filesystem/network reach, and tool authority |
| 1673 | # # are Runtime policy and are deliberately not configured here. Pre-0.9.11 |
| 1674 | # # trust keys are accepted only as ignored migration input. |
| 1675 | # |
| 1676 | # # Headless worker execution hardening (#3027) |
| 1677 | # [fleet.exec] |
| 1678 | # # Tools always allowed regardless of role |
| 1679 | # allowed_tools = [] |
| 1680 | # # Tools always disallowed (overrides role and task spec) |
| 1681 | # disallowed_tools = ["exec_shell"] |
| 1682 | # # Optional hard ceiling on worker steps (tool calls + model turns). |
| 1683 | # # Omit or set 0 for the unbounded default; use a positive value to opt in. |
| 1684 | # max_turns = 500 |
| 1685 | # # Runtime child-agent depth for fleet workers. Shares ONE recursion axis with |
| 1686 | # # standalone sub-agents (a fleet worker IS a headless sub-agent). This is an |
| 1687 | # # execution request, not Fleet member identity. 0 blocks child agents (the |
| 1688 | # # root worker still runs); 3 is the default and 8 is the opt-in hard ceiling. |
| 1689 | # max_spawn_depth = 3 |
| 1690 | # # Extra system prompt injected into every headless worker |
| 1691 | # append_system_prompt = "Never modify .git/config or change remotes." |
| 1692 | # # Output format: "text" (default) or "stream-json" for ndjson events |
| 1693 | # output_format = "text" |
| 1694 | # |
| 1695 | # # Fleet profiles define named agent configurations the roster can dispatch. |
| 1696 | # # Built-in profiles are always available: manager, operator, scout, builder, |
| 1697 | # # reviewer, verifier, synthesizer, general. User-defined profiles under |
| 1698 | # # [fleet.profiles] override or extend the built-in set by id. Precedence |
| 1699 | # # is Workspace (.codewhale/agents/*.toml) > Config ([fleet.profiles]) > BuiltIn. |
| 1700 | # # See /fleet setup for an in-app profile-authoring wizard. |
| 1701 | # [fleet.profiles.ci-linter] |
| 1702 | # slot = "verifier" |
| 1703 | # loadout = "fast" |
| 1704 | # model = "deepseek-v4-pro" |
| 1705 | # |
| 1706 | # [fleet.profiles.ci-linter.role] |
| 1707 | # name = "CI Linter" |
| 1708 | # description = "Runs linters and formatters" |
| 1709 | # instructions = "Run cargo fmt --check and cargo clippy; never apply fixes." |
| 1710 | # |
| 1711 | # [fleet.profiles.pr-reviewer] |
| 1712 | # slot = "reviewer" |
| 1713 | # loadout = "inherit" |
| 1714 | # |
| 1715 | # [fleet.profiles.pr-reviewer.role] |
| 1716 | # name = "PR Reviewer" |
| 1717 | # description = "Reviews PRs with GitHub access" |
| 1718 | # instructions = "Review diffs for correctness, regressions, and missing tests." |
| 1719 | |
| 1720 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1721 | # Named fleets live in files, not inline tables (0.9.14). |
| 1722 | # |
| 1723 | # Inline [fleets.<name>] tables were removed: define named fleets as |
| 1724 | # `fleets/<name>.toml` under $CODEWHALE_HOME or the workspace instead |
| 1725 | # (see docs/FLEET.md). Old inline tables still parse but are ignored. |
| 1726 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1727 | |
| 1728 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1729 | # Requirements (admin constraints) example file |
| 1730 | # ───────────────────────────────────────────────────────────────────────────────── |
| 1731 | # allowed_approval_policies = ["on-request", "untrusted", "never"] |
| 1732 | # allowed_sandbox_modes = ["read-only", "workspace-write"] |
| 1733 |