codewhale/config.example.toml

# ╔══════════════════════════════════════════════════════════════════════════════╗
# ║                         CodeWhale Configuration                               ║
# ║                                                                              ║
# ║  Terminal coding agent for DeepSeek.                                          ║
# ╚══════════════════════════════════════════════════════════════════════════════╝

# See `docs/CONFIGURATION.md` for how config is loaded (profiles, env overrides, etc.).

# ─────────────────────────────────────────────────────────────────────────────────
# Active provider + DeepSeek defaults
# ─────────────────────────────────────────────────────────────────────────────────
# Choose which provider to use by default. Per-provider credentials live in the
# `[providers.*]` sections near the bottom of
# this file — keeping both stored at once means `/provider deepseek` and
# `/provider nvidia-nim` (or `--provider openai`, `--provider wanjie-ark`,
# `--provider volcengine`, `--provider xiaomi-mimo`, `--provider fireworks`,
# `--provider siliconflow`, `/provider moonshot`, `/provider sglang`,
# `/provider vllm`, `/provider ollama`) toggle without having to re-enter keys. Top-level
# `api_key` / `base_url` are
# still read as DeepSeek defaults when `[providers.deepseek]` is absent
# (backward compatibility).
provider = "deepseek" # deepseek | deepseek-cn | nvidia-nim | openai | atlascloud | wanjie-ark | volcengine | openrouter | xiaomi-mimo | novita | fireworks | siliconflow | moonshot | sglang | vllm | ollama
api_key = "YOUR_DEEPSEEK_API_KEY" # must be non-empty
base_url = "https://api.deepseek.com/beta"
# provider = "deepseek-cn"                       # legacy alias (official host is still https://api.deepseek.com)
# base_url = "https://api.deepseek.com"         # opt out of DeepSeek beta features
# Optional custom model request headers for OpenAI-compatible gateways.
# Authorization and Content-Type are managed by the client and cannot be overridden here.
# http_headers = { "X-Model-Provider-Id" = "your-model-provider" }

# ─────────────────────────────────────────────────────────────────────────────────
# Default Models
# ─────────────────────────────────────────────────────────────────────────────────
# DeepSeek V4 family:
#   deepseek-v4-pro             — flagship reasoning model on DeepSeek Platform
#   deepseek-v4-flash           — fast, cost-efficient (legacy aliases: deepseek-chat, deepseek-reasoner)
#   deepseek-ai/deepseek-v4-pro   — NVIDIA NIM-hosted Pro model ID
#   deepseek-ai/deepseek-v4-flash — NVIDIA NIM-hosted Flash model ID
#   deepseek/deepseek-v4-pro      — default OpenRouter DeepSeek model ID
#   arcee-ai/trinity-large-thinking — OpenRouter Arcee Trinity Large Thinking
#   qwen/qwen3.7-max              — OpenRouter Qwen 3.7 Max
#   xiaomi/mimo-v2.5-pro          — OpenRouter Xiaomi MiMo 2.5 Pro
#   xiaomi/mimo-v2.5              — OpenRouter Xiaomi MiMo 2.5
#   gpt-4.1                         — default generic OpenAI-compatible model ID
#   deepseek-ai/deepseek-v4-flash   — default AtlasCloud model ID
#   deepseek-reasoner               — default Wanjie Ark model ID
#   mimo-v2.5-pro                   — default Xiaomi MiMo model ID
#   accounts/fireworks/models/deepseek-v4-pro — Fireworks AI Pro model ID
#   deepseek-ai/DeepSeek-V4-Pro    — SiliconFlow hosted Pro model ID
#   deepseek-ai/DeepSeek-V4-Flash  — SiliconFlow hosted Flash model ID
#   deepseek-ai/DeepSeek-V4-Pro    — SGLang self-hosted Pro model ID
#   deepseek-ai/DeepSeek-V4-Flash  — SGLang self-hosted Flash model ID
default_text_model = "deepseek-v4-pro"

# ─────────────────────────────────────────────────────────────────────────────────
# Thinking Mode (DeepSeek V4 reasoning effort)
# ─────────────────────────────────────────────────────────────────────────────────
# "off"    — disables chain-of-thought (thinking.type = disabled)
# "low"    — compat-maps to "high" server-side
# "medium" — compat-maps to "high" server-side
# "high"   — reasoning_effort = high (DeepSeek default)
# "max"    — reasoning_effort = max (deepest reasoning)
#
# Shift+Tab in the TUI cycles between off / high / max. The header shows the
# current tier as a ⚡ chip.
reasoning_effort = "max"

# ─────────────────────────────────────────────────────────────────────────────────
# Cost Display
# ─────────────────────────────────────────────────────────────────────────────────
# Display estimated usage in USD or CNY. Aliases `yuan` and `rmb` normalize to `cny`.
cost_currency = "usd" # usd | cny

# ─────────────────────────────────────────────────────────────────────────────────
# Startup update check
# ─────────────────────────────────────────────────────────────────────────────────
# The TUI checks for newer CodeWhale releases in the background at startup.
# Set check_for_updates = false in managed or air-gapped environments.
# update_uri may point at a GitHub-compatible latest-release JSON endpoint.
[update]
check_for_updates = true
# update_uri = "https://internal.mirror.example/codewhale/releases/latest"

# ─────────────────────────────────────────────────────────────────────────────────
# Paths
# ─────────────────────────────────────────────────────────────────────────────────
# New installs write product state under ~/.codewhale/. Existing ~/.deepseek/
# files are still read as compatibility fallbacks when the .codewhale file is
# absent.
skills_dir = "~/.codewhale/skills"
mcp_config_path = "~/.codewhale/mcp.json"
notes_path = "~/.codewhale/notes.txt"

memory_path = "~/.codewhale/memory.md"

# instructions = ["./AGENTS.md", "~/.codewhale/global.md"]
#
# Optional list of additional instruction files concatenated into the
# system prompt in declared order (#454). Useful for layering
# repo-specific rules on top of a global preferences file. Each entry
# is expanded so `~` and env vars work; missing files are skipped with
# a tracing warning. Files are capped at 100 KiB per entry.
#
# Project-level config (.codewhale/config.toml in the workspace) replaces
# the user-level array wholesale rather than merging — list `~/global.md`
# inside the project array if you want both. An explicit empty array
# (`instructions = []`) clears the user list for the current repo.

# ─────────────────────────────────────────────────────────────────────────────────
# User memory (#489) — opt-in. When enabled, the TUI reads memory_path on
# startup and injects its contents into the system prompt as a
# <user_memory> block, intercepts `# foo` typed in the composer to append
# the line as a timestamped bullet, and registers a `remember` tool the
# model can call to add durable notes itself.
# ─────────────────────────────────────────────────────────────────────────────────
[memory]
# enabled = true            # turn the feature on (default: false)
# Override the env-var equivalent: `DEEPSEEK_MEMORY=on`

# Parsed but currently unused (reserved for future versions):
# tools_file = "./tools.json"

# Native tool catalog controls (#2076). By default only the core tool surface
# is loaded into the model context; less common native tools are discoverable
# through ToolSearch and loaded on first use.
# [tools]
# always_load = ["git_show", "notify"]

# ─────────────────────────────────────────────────────────────────────────────────
# Security
# ─────────────────────────────────────────────────────────────────────────────────
allow_shell = true
approval_policy = "on-request" # on-request | untrusted | never
sandbox_mode = "workspace-write" # read-only | workspace-write | danger-full-access | external-sandbox

# ─────────────────────────────────────────────────────────────────────────────────
# External Sandbox Backend (pluggable remote execution)
# ─────────────────────────────────────────────────────────────────────────────────
# When sandbox_backend is set to "opensandbox", all exec_shell calls are
# routed through an external OpenSandbox-compatible HTTP API instead of
# spawning a local process. The backend sends `POST {sandbox_url}/v1/sandbox/run`
# with `{"cmd": "...", "env": {...}}` and expects
# `{"stdout": "...", "stderr": "...", "exit_code": 0}`.
#
# sandbox_backend = "none"          # "none" (default) or "opensandbox"
# sandbox_url = "http://localhost:8080"  # OpenSandbox-compatible API base URL
# sandbox_api_key = "YOUR_API_KEY"  # Optional Bearer token sent with requests
#
# Env-var overrides:
#   DEEPSEEK_SANDBOX_BACKEND   → sandbox_backend
#   DEEPSEEK_SANDBOX_URL       → sandbox_url
#   DEEPSEEK_SANDBOX_API_KEY   → sandbox_api_key
#
# Example OpenSandbox setup:
#
#   sandbox_backend = "opensandbox"
#   sandbox_url = "http://localhost:8080"
#   sandbox_api_key = "sk-opensandbox-secret"
#
# The backend uses a 30-second HTTP timeout. Background, interactive, and
# TTY modes are not supported with external backends — all commands run
# synchronously via HTTP.
# ─────────────────────────────────────────────────────────────────────────────────
# Bubblewrap (Linux only, additional filesystem isolation)
# ─────────────────────────────────────────────────────────────────────────────────
# When set to true and `/usr/bin/bwrap` is present, exec_shell commands are
# routed through bubblewrap instead of relying solely on Landlock. Bubblewrap
# creates a read-only view of the root filesystem with write access limited to
# the working directory. Install separately:
#
#   Ubuntu/Debian:  apt install bubblewrap
#   Fedora:         dnf install bubblewrap
#   Arch:           pacman -S bubblewrap
#
# prefer_bwrap = false  # default — use Landlock only
#
# Env override: DEEPSEEK_PREFER_BWRAP=true

# auto_allow entries match by command prefix, not raw string.
# See command_safety.rs for the prefix dictionary.
#
# Examples:
#   auto_allow = ["git status"]   # auto-approves: git status, git status -s, git status --porcelain
#                                 # does NOT auto-approve: git push, git checkout
#   auto_allow = ["cargo check", "npm run"]
#
# auto_allow = []
max_subagents = 10 # optional (1-20)

# Optional sub-agent tuning. max_concurrent overrides top-level max_subagents.
# [subagents]
# max_concurrent = 10
# api_timeout_secs = 120 # per-step API timeout, clamped to 1..=1800

# Optional managed policy paths (defaults to /etc/deepseek/*.toml on unix):
# managed_config_path = "/etc/deepseek/managed_config.toml"
# requirements_path = "/etc/deepseek/requirements.toml"

# ─────────────────────────────────────────────────────────────────────────────────
# Per-provider credentials (peer providers — NIM is first-class, not a flag)
# ─────────────────────────────────────────────────────────────────────────────────
# Providers can be stored at once; `provider = "..."` (top of file) or
# `/provider deepseek` / `/provider nvidia-nim` / `--provider openai` /
# `--provider wanjie-ark` / `/provider volcengine` / `/provider fireworks` /
# `--provider siliconflow` / `/provider moonshot`
# switches between them without having to re-enter keys. Env vars override anything set here:
#   DeepSeek: DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL, DEEPSEEK_MODEL
#   NIM:      NVIDIA_API_KEY (or NVIDIA_NIM_API_KEY), NIM_BASE_URL
#             (or NVIDIA_NIM_BASE_URL / NVIDIA_BASE_URL), NVIDIA_NIM_MODEL
#   OpenAI-compatible: OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL
#   Wanjie Ark: WANJIE_ARK_API_KEY (or WANJIE_API_KEY), WANJIE_ARK_BASE_URL, WANJIE_ARK_MODEL
#   Volcengine Ark: VOLCENGINE_API_KEY (or VOLCENGINE_ARK_API_KEY / ARK_API_KEY), VOLCENGINE_BASE_URL, VOLCENGINE_MODEL
#   OpenRouter: OPENROUTER_API_KEY, OPENROUTER_BASE_URL, OPENROUTER_MODEL
#   Xiaomi MiMo: XIAOMI_MIMO_API_KEY (or XIAOMI_API_KEY / MIMO_API_KEY), XIAOMI_MIMO_BASE_URL, XIAOMI_MIMO_MODEL
#   Novita:     NOVITA_API_KEY, NOVITA_BASE_URL, NOVITA_MODEL
#   Fireworks:  FIREWORKS_API_KEY, FIREWORKS_BASE_URL
#   SiliconFlow: SILICONFLOW_API_KEY, SILICONFLOW_BASE_URL, SILICONFLOW_MODEL
#   Moonshot/Kimi: MOONSHOT_API_KEY (or KIMI_API_KEY), MOONSHOT_BASE_URL, MOONSHOT_MODEL
#   SGLang:    SGLANG_BASE_URL, SGLANG_MODEL, optional SGLANG_API_KEY
#   vLLM:      VLLM_BASE_URL, VLLM_MODEL, optional VLLM_API_KEY
#   Ollama:    OLLAMA_BASE_URL, OLLAMA_MODEL, optional OLLAMA_API_KEY
#
# Custom DeepSeek-compatible APIs usually do not need a new provider table:
# set `provider = "deepseek"` and override [providers.deepseek].base_url/model.
# For generic OpenAI-compatible gateways, use `provider = "openai"` and the
# [providers.openai] table below. Keep provider/api_key/base_url in user config
# or environment variables; project overlays are not allowed to set them.

# DeepSeek Platform (https://platform.deepseek.com)
[providers.deepseek]
# api_key = "YOUR_DEEPSEEK_API_KEY"
# base_url = "https://api.deepseek.com/beta"
# model = "deepseek-v4-pro"
# Custom DeepSeek-compatible example:
# base_url = "https://your-provider.example/v1"
# model = "deepseek-ai/DeepSeek-V4-Pro"
# http_headers = { "X-Model-Provider-Id" = "your-model-provider" } # optional custom request headers

# NVIDIA NIM-hosted DeepSeek V4 (https://build.nvidia.com)
[providers.nvidia_nim]
# api_key = "YOUR_NVIDIA_API_KEY"
# base_url = "https://integrate.api.nvidia.com/v1"
# model = "deepseek-ai/deepseek-v4-pro"     # or deepseek-ai/deepseek-v4-flash

# Generic OpenAI-compatible endpoint. Use the built-in `openai` provider for
# third-party gateways; do not invent a custom provider name. For non-local
# http:// gateways, launch with DEEPSEEK_ALLOW_INSECURE_HTTP=1 only on a
# trusted network.
[providers.openai]
# api_key = "YOUR_OPENAI_COMPATIBLE_API_KEY"
# base_url = "https://api.openai.com/v1"
# model = "gpt-4.1"
# Gateway example:
# base_url = "https://gateway.example/v1"
# model = "your-deepseek-compatible-model"

# AtlasCloud OpenAI-compatible endpoint (https://www.atlascloud.ai/docs/models/llm)
[providers.atlascloud]
# api_key = "YOUR_ATLASCLOUD_API_KEY"
# base_url = "https://api.atlascloud.ai/v1"
# model = "deepseek-ai/deepseek-v4-flash"

# Wanjie Ark / 万界方舟 OpenAI-compatible endpoint
[providers.wanjie_ark]
# api_key = "YOUR_WANJIE_API_KEY"
# base_url = "https://maas-openapi.wanjiedata.com/api/v1"
# model = "deepseek-reasoner"                # or the exact model ID enabled on your Wanjie account

# Volcengine / Volcano Engine Ark Coding API
[providers.volcengine]
# api_key = "YOUR_VOLCENGINE_API_KEY"
# base_url = "https://ark.cn-beijing.volces.com/api/coding/v3"
# model = "DeepSeek-V4-Pro"                  # or DeepSeek-V4-Flash

# OpenRouter — multi-provider gateway (https://openrouter.ai)
[providers.openrouter]
# api_key = "YOUR_OPENROUTER_API_KEY"
# base_url = "https://openrouter.ai/api/v1"
# model = "deepseek/deepseek-v4-pro"
# Recent large model IDs also accepted here include arcee-ai/trinity-large-thinking,
# qwen/qwen3.7-max, xiaomi/mimo-v2.5-pro, qwen/qwen3.6-35b-a3b,
# google/gemma-4-31b-it, z-ai/glm-5.1, moonshotai/kimi-k2.6, and
# nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free.

# Xiaomi MiMo OpenAI-compatible endpoint (https://platform.xiaomimimo.com)
[providers.xiaomi_mimo]
# api_key = "YOUR_XIAOMI_KEY"
# base_url = "https://api.xiaomimimo.com/v1"
# model = "mimo-v2.5-pro"

# Novita AI-hosted inference (https://novita.ai)
[providers.novita]
# api_key = "YOUR_NOVITA_API_KEY"
# base_url = "https://api.novita.ai/v1"
# model = "deepseek/deepseek-v4-pro"         # or deepseek/deepseek-v4-flash

# Fireworks AI-hosted DeepSeek V4 (https://fireworks.ai)
[providers.fireworks]
# api_key = "YOUR_FIREWORKS_API_KEY"
# base_url = "https://api.fireworks.ai/inference/v1"
# model = "accounts/fireworks/models/deepseek-v4-pro"

# SiliconFlow-hosted DeepSeek V4 (https://siliconflow.com)
[providers.siliconflow]
# api_key = "YOUR_SILICONFLOW_API_KEY"
# base_url = "https://api.siliconflow.com/v1"
# model = "deepseek-ai/DeepSeek-V4-Pro"     # or deepseek-ai/DeepSeek-V4-Flash

# Self-hosted SGLang OpenAI-compatible server
[providers.sglang]
# api_key = "OPTIONAL_SGLANG_TOKEN"
# base_url = "http://localhost:30000/v1"
# model = "deepseek-ai/DeepSeek-V4-Pro"     # or deepseek-ai/DeepSeek-V4-Flash

# Self-hosted vLLM OpenAI-compatible server
[providers.vllm]
# api_key = "OPTIONAL_VLLM_TOKEN"
# base_url = "http://localhost:8000/v1"
# model = "deepseek-ai/DeepSeek-V4-Pro"     # or deepseek-ai/DeepSeek-V4-Flash

# Self-hosted Ollama OpenAI-compatible server
[providers.ollama]
# api_key = "OPTIONAL_OLLAMA_TOKEN"
# base_url = "http://localhost:11434/v1"
# model = "deepseek-coder:1.3b"             # or any local Ollama tag

# ─────────────────────────────────────────────────────────────────────────────────
# Web Search Provider
# ─────────────────────────────────────────────────────────────────────────────────
# Choose which backend `web_search` uses. Default is DuckDuckGo HTML scraping
# with Bing fallback — no API key needed. Bing remains selectable for users who
# explicitly prefer it. Switch to Tavily, Bocha, Metaso, or Baidu for
# API-backed search.
#
# [search]
# provider = "duckduckgo"  # duckduckgo | bing | tavily | bocha | metaso | baidu | volcengine
#                            # duckduckgo: HTML scrape with Bing fallback
#                            # bing:       HTML scrape, no API key
#                            # tavily:     https://tavily.com — AI search, needs api_key
#                            # bocha:      https://bochaai.com — 博查AI搜索，国内友好，需api_key
#                            # metaso:     https://metaso.cn — 秘塔AI搜索，每天 100 次免费
#                            #             设置 METASO_API_KEY 或 [search] api_key 可提升额度
#                            # baidu:      百度 AI Search via qianfan.baidubce.com，需 api_key
#                            # volcengine: 火山引擎 Ark web_search (免费 2 万次/月), 需 api_key
#                            #             也回退到 VOLCENGINE_API_KEY / VOLCENGINE_ARK_API_KEY / ARK_API_KEY 环境变量
# api_key = "YOUR_SEARCH_KEY" # required for tavily, bocha, and baidu; optional for metaso
#                             # WARNING: treat config.toml like a secret file when
#                             # storing API keys. Prefer env vars for local smoke tests.
#
# Env-var overrides:
#   DEEPSEEK_SEARCH_PROVIDER → search.provider
#   DEEPSEEK_SEARCH_API_KEY  → search.api_key
#   METASO_API_KEY           → metaso key fallback
#   BAIDU_SEARCH_API_KEY     → baidu key fallback

# ─────────────────────────────────────────────────────────────────────────────────
# Network Policy (#135)
# ─────────────────────────────────────────────────────────────────────────────────
# Per-domain allow/deny rules for outbound network calls made by the TUI's
# tools (`fetch_url`, `web_search`) and the MCP HTTP transport. Stdio MCP
# servers and direct LLM API calls are unaffected.
#
# Precedence: deny wins. A host listed in both `allow` and `deny` is denied.
#
# Host-matching rules:
#   - Exact match: `api.deepseek.com` matches only `api.deepseek.com`.
#   - Subdomain wildcard: an entry starting with `.` (e.g. `.example.com`)
#     matches `api.example.com` and `a.b.example.com` but not the apex
#     `example.com`. To cover both, list both. `*.example.com` is also accepted.
#
# Defaults are intentionally conservative: when this section is absent, no
# policy is enforced (mirrors pre-v0.7.0 behavior). To opt in:
#
# [network]
# default = "prompt"     # allow | deny | prompt
# allow = ["api.deepseek.com", "github.com", ".githubusercontent.com"]
# deny = []
# audit = true            # one line per call to ~/.codewhale/audit.log

# ─────────────────────────────────────────────────────────────────────────────────
# Skills (#140)
# ─────────────────────────────────────────────────────────────────────────────────
# Settings for the `/skill install <spec>` community-skill installer.
#   * registry_url           — curated index.json that resolves bare names to
#                              `github:owner/repo` specs. Override to point at
#                              a private fork or internal mirror.
#   * max_install_size_bytes — per-skill uncompressed size cap. Tarballs that
#                              exceed this limit are rejected during validation.
#                              Default: 5 MiB.
#
# `/skill install` is gated by `[network]`. Make sure `github.com` and
# `raw.githubusercontent.com` are reachable (default `prompt` is fine — you'll
# be asked once and can persist) before running it.
#
# [skills]
# registry_url = "https://raw.githubusercontent.com/Hmbown/deepseek-skills/main/index.json"
# max_install_size_bytes = 5_242_880

# ─────────────────────────────────────────────────────────────────────────────────
# TUI
# ─────────────────────────────────────────────────────────────────────────────────
[tui]
alternate_screen = "auto"   # auto/always use the TUI screen; never uses terminal scrollback
mouse_capture = true        # true copies only transcript user/assistant text; false uses raw terminal selection/copy
terminal_probe_timeout_ms = 500 # optional startup terminal-mode timeout (100-5000ms)
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
# notification_condition = "always" # always | never — overrides [notifications].threshold_secs.
#                                    "always" = notify on every successful turn (no threshold);
#                                    "never"  = suppress all turn-completion notifications;
#                                    unset    = use [notifications] defaults (recommended).
# locale = "auto"           # UI chrome language: auto | en | ja | zh-Hans | pt-BR
#                           # "auto" reads LC_ALL → LC_MESSAGES → LANG; falls back to English.
#                           # Override: `locale = "zh-Hans"` for Simplified Chinese regardless of OS locale.
#                           # Also settable at runtime: /config locale zh-Hans
#                           # Note: this only affects TUI labels/chrome — it does NOT change model output language.

# ─────────────────────────────────────────────────────────────────────────────────
# Feature Flags
# ─────────────────────────────────────────────────────────────────────────────────
[features]
shell_tool = true
subagents = true
web_search = true # enables canonical web.run plus the compatibility web_search alias
apply_patch = true
mcp = true
exec_policy = true
# vision_model = false  # enable vision model for image_analyze tool

# ─────────────────────────────────────────────────────────────────────────────────
# Vision Model Configuration (optional)
# ─────────────────────────────────────────────────────────────────────────────────
# Uses an OpenAI-compatible vision model API for the `image_analyze` tool.
# api_key inherits from the main config if not specified.
#
# [vision_model]
# model = "gemini-3.1-flash-lite-preview"  # Required: vision-capable model ID
# api_key = "YOUR_API_KEY"                 # Optional: defaults to main api_key
# base_url = "https://generativelanguage.googleapis.com/v1beta/openai/"  # Optional
#
# Xiaomi MiMo image understanding can be configured through the same tool:
# model = "mimo-v2.5"
# api_key = "YOUR_XIAOMI_KEY"
# base_url = "https://api.xiaomimimo.com/v1"

# ─────────────────────────────────────────────────────────────────────────────────
# Retry Configuration
# ─────────────────────────────────────────────────────────────────────────────────
[retry]
enabled = true
max_retries = 3
initial_delay = 1.0
max_delay = 60.0
exponential_base = 2.0

# ─────────────────────────────────────────────────────────────────────────────────
# Context Compaction
# ─────────────────────────────────────────────────────────────────────────────────
# Auto-compaction is a saved UI setting edited with `/config` (`auto_compact`).
# There is no config-file `[compaction]` table yet; detailed thresholds are
# chosen by the TUI from the active model/context budget.

# Append-only Flash seams are experimental and opt-in while the v0.7.5
# context/cache audit validates prefix-cache behavior.
[context]
enabled = false
verbatim_window_turns = 16
# Thresholds are based on the active request input estimate, not lifetime
# summed API usage.
l1_threshold = 192000
l2_threshold = 384000
l3_threshold = 576000
# Hard cycle reserves the normal 262144-token internal turn budget plus 1024
# safety tokens, separate from V4's official 384000 max-output metadata.
cycle_threshold = 768000
seam_model = "deepseek-v4-flash"

# ─────────────────────────────────────────────────────────────────────────────────
# Workshop / Large-Output Routing (#548)
# ─────────────────────────────────────────────────────────────────────────────────
# Tool outputs exceeding `large_output_threshold_tokens` are routed through a
# V4-Flash synthesis sub-agent.  Only the synthesis reaches the parent context;
# the raw text is stored in the workshop variable `last_tool_result` so the
# parent can call `promote_to_context` later if it needs the full content.
#
# Per-tool overrides let high-volume tools (e.g. exec_shell) use tighter
# thresholds without changing the global default.
#
# Add `raw = true` to any tool call to bypass routing for that invocation.
#
# [workshop]
# large_output_threshold_tokens = 4096
# [workshop.per_tool_thresholds]
# exec_shell  = 2048   # shell output synthesised aggressively
# grep_files  = 2048
# web_search  = 8192   # web results can be large; give them more room

# ─────────────────────────────────────────────────────────────────────────────────
# Capacity Controller (runtime pressure guardrails)
# ─────────────────────────────────────────────────────────────────────────────────
[capacity]
enabled = false
low_risk_max = 0.50
medium_risk_max = 0.62
severe_min_slack = -0.25
severe_violation_ratio = 0.40
refresh_cooldown_turns = 6
replan_cooldown_turns = 5
max_replay_per_turn = 1
min_turns_before_guardrail = 4
profile_window = 8
deepseek_v3_2_chat_prior = 3.9
deepseek_v3_2_reasoner_prior = 4.1
deepseek_v4_pro_prior = 3.5
deepseek_v4_flash_prior = 4.2
fallback_default_prior = 3.8

# ─────────────────────────────────────────────────────────────────────────────────
# Profile Example (for multiple environments)
# ─────────────────────────────────────────────────────────────────────────────────
# Select a profile with `deepseek --profile <name>` or `DEEPSEEK_PROFILE=<name>`.
[profiles.work]
api_key = "WORK_DEEPSEEK_API_KEY"
base_url = "https://api.deepseek.com/beta"

[profiles.dev]
api_key = "DEV_DEEPSEEK_API_KEY"
allow_shell = true

[profiles.nvidia-nim]
provider = "nvidia-nim"
api_key = "YOUR_NVIDIA_API_KEY"
base_url = "https://integrate.api.nvidia.com/v1"
default_text_model = "deepseek-ai/deepseek-v4-pro"

# ─────────────────────────────────────────────────────────────────────────────────
# Desktop Notifications (OSC 9 / BEL on long agent-turn completion)
# ─────────────────────────────────────────────────────────────────────────────────
# Emits an escape sequence to the terminal when a turn **completes successfully**
# and took longer than `threshold_secs`. Failed or cancelled turns are
# intentionally silent. Useful when you tab away from the TUI and want an alert
# for "your task is ready".
#
# method        = "auto"   # auto | osc9 | bel | off
#                 auto: OSC 9 for iTerm.app / Ghostty / WezTerm.
#                       On macOS / Linux, falls back to BEL.
#                       On Windows, falls back to "off" — BEL maps to the
#                       system error chime (SystemAsterisk / MB_OK), which
#                       sounds like an error popup. Set method = "bel"
#                       explicitly to opt back in (#583).
#                 osc9: \x1b]9;<msg>\x07 (iTerm2-style; shows macOS notification)
#                 bel:  plain \x07 beep
#                 off:  disable entirely
# threshold_secs = 30      # only notify when the turn took >= this many seconds
# include_summary = false  # include elapsed time + cost in the notification body
[notifications]
# method = "auto"
# threshold_secs = 30
# include_summary = false

# ─────────────────────────────────────────────────────────────────────────────────
# Workspace Snapshots (#137)
# ─────────────────────────────────────────────────────────────────────────────────
# Each turn the TUI takes a `pre-turn:<seq>` and `post-turn:<seq>` snapshot of
# your workspace into a side-git repo at:
#
#     ~/.codewhale/snapshots/<project_hash>/<worktree_hash>/.git
#
# Your own `.git` is never touched — `--git-dir` and `--work-tree` are always
# set together when shelling out to git. Use `/restore N` (slash command) or
# the `revert_turn` tool to roll the working tree back. Conversation history
# is unaffected.
#
# Disk footprint: ~1-2 GB worst case for a 100 MB workspace × 12 turns/day,
# typically far less thanks to git's content-addressed storage. The session
# boot prunes anything older than `max_age_days` (default 7).
#
# [snapshots]
# enabled = true        # Snapshot workspace pre/post each turn for /restore
# max_age_days = 7      # Older snapshots pruned at session start
# max_workspace_gb = 2  # Snapshots self-disable on first init when the
#                         # non-excluded workspace exceeds this size in GB
#                         # (v0.8.32). Default 2 GB protects against running
#                         # codewhale in directories with hundreds of GB
#                         # of datasets / model weights / docker dumps where
#                         # `git add -A` would hang the TUI for hours. Set
#                         # to 0 to disable the cap (v0.8.31 behaviour);
#                         # raise to a higher number for legitimate large
#                         # monorepos.

# ─────────────────────────────────────────────────────────────────────────────────
# LSP Diagnostics (post-edit) (#136)
# ─────────────────────────────────────────────────────────────────────────────────
# After every successful file edit (`edit_file`, `apply_patch`, `write_file`),
# the engine asks an LSP server for diagnostics on the file and injects them
# as a synthetic system message before the next API call. This lets the agent
# see compile breaks immediately without round-tripping through the user.
#
# Enabled by default. Failure modes are non-blocking: a missing LSP binary,
# a crashed server, or a timeout simply skips the post-edit hook for that
# turn — the agent's work is never blocked.
#
# Built-in language → server defaults:
#   rust       → rust-analyzer
#   go         → gopls serve
#   python     → pyright-langserver --stdio
#   typescript → typescript-language-server --stdio
#   java       → jdtls
#   vue        → vue-language-server --stdio
#   c, cpp     → clangd
#
# Java support uses Eclipse JDT LS via the `jdtls` command. IntelliJ IDEA is
# not required, and installing IntelliJ IDEA alone does not install `jdtls`.
#
# Override the defaults via the `servers` table below.
[lsp]
# enabled = true
# poll_after_edit_ms = 5000
# max_diagnostics_per_file = 20
# include_warnings = false
# [lsp.servers]
# rust = ["rust-analyzer"]
# go = ["gopls", "serve"]
# java = ["jdtls"]
# vue = ["vue-language-server", "--stdio"]

# ─────────────────────────────────────────────────────────────────────────────────
# Hooks (optional)
# ─────────────────────────────────────────────────────────────────────────────────
# Hooks run shell commands on lifecycle events (session start/end, tool calls, etc.).
# Configure as `[[hooks.hooks]]` under a `[hooks]` table.
#
# Available events: session_start, session_end, message_submit,
# tool_call_before, tool_call_after, mode_change, on_error, shell_env.
#
# `shell_env` (#456) is special: the hook runs immediately before each
# `exec_shell` invocation and its stdout is parsed as `KEY=VALUE\n` lines.
# Those vars are merged into the spawned process environment (later hooks
# override earlier ones). Use this for ephemeral credentials, per-skill
# PATH adjustments, or short-lived tokens. The resolved KEY names (NEVER
# values) are written to `~/.codewhale/audit.log` so each session can be
# reconciled later. Hook failure / timeout simply contributes no vars —
# it does not abort the shell call.
#
# [hooks]
# enabled = true
# default_timeout_secs = 30
#
# [[hooks.hooks]]
# event = "session_start"
# command = "echo 'CodeWhale session started'"
#
# # Inject ephemeral creds into every shell call. Output one
# # KEY=VALUE per line on stdout (export prefix optional).
# [[hooks.hooks]]
# name = "aws-creds"
# event = "shell_env"
# command = "aws-vault export my-profile --format=env"
# # Optionally limit to specific tool names / categories:
# # condition = { type = "tool_category", category = "shell" }

# ─────────────────────────────────────────────────────────────────────────────────
# Runtime API (`deepseek serve --http`) (#561)
# ─────────────────────────────────────────────────────────────────────────────────
# Tuning knobs for the local HTTP/SSE daemon. The server binds to 127.0.0.1
# by default and is intended for local UIs (whalescale-desktop, dashboards,
# automation scripts). Today this section only controls the CORS allow-list;
# host/port/workers stay on `--host`, `--port`, and `--workers` flags.
#
# Built-in defaults always include:
#   http://localhost:3000   http://127.0.0.1:3000
#   http://localhost:1420   http://127.0.0.1:1420
#   tauri://localhost
#
# Use `cors_origins` to add extra dev origins (e.g. Vite's default `:5173`).
# User entries STACK on top of the defaults — they do not replace them. The
# CLI flag `--cors-origin URL` (repeatable) and env var
# `DEEPSEEK_CORS_ORIGINS=url1,url2` resolve to the same merged list.
#
# [runtime_api]
# cors_origins = ["http://localhost:5173", "http://127.0.0.1:5173"]

# ─────────────────────────────────────────────────────────────────────────────────
# Tool Overrides & Plugins ([tools])
# ─────────────────────────────────────────────────────────────────────────────────
# The `[tools]` table lets you replace any built-in tool with a custom
# implementation (script or command) or disable it entirely — without
# forking or recompiling the binary.
#
# Plugin scripts dropped in the plugin directory are auto-discovered and
# registered as model-visible tools alongside the built-in ones.
#
# Scripts receive the tool's JSON input on **stdin** and must return a
# JSON `ToolResult` (`{"content": "...", "success": true}`) on **stdout**.
#
# [tools]
# # Custom plugin directory (defaults to `~/.codewhale/tools/`)
# plugin_dir = "~/.codewhale/tools"
#
# [tools.overrides]
# # Disable a tool entirely — removes it from the model-visible catalog.
# "code_execution" = { type = "disabled" }
#
# # Replace a tool with a script. Relative paths resolve against plugin_dir.
# "exec_shell" = { type = "script", path = "audit-exec-shell.sh" }
#
# # Replace a tool with a command (binary on PATH or absolute path).
# "read_file" = { type = "command", command = "bat", args = ["--paging=never"] }
#
# # Scripts can also accept static arguments before the JSON input:
# "fetch_url" = { type = "script", path = "cached-fetch.sh", args = ["--ttl", "300"] }

# ──────────── Enterprise example: audit-logging exec_shell wrapper ──────────────
# Drop `audit-exec-shell.sh` in `~/.codewhale/tools/` and enable with:
#
# [tools.overrides]
# "exec_shell" = { type = "script", path = "audit-exec-shell.sh" }
#
# The wrapper logs every request to `~/.codewhale/audit/exec_shell.log`, then
# delegates to your own approved shell executor. Do not pipe the raw JSON
# request into `sh -s`; parse the command field and enforce your policy first.
#
# ```sh
# #!/usr/bin/env sh
# # name: exec_shell
# # description: Audit-logging wrapper for exec_shell
# # approval: required
# LOGDIR="${HOME}/.codewhale/audit"
# mkdir -p "$LOGDIR"
# LOGFILE="$LOGDIR/exec_shell.log"
# input=$(cat)
# echo "[$(date -Iseconds)] $input" >> "$LOGFILE"
# printf '%s\n' '{"content":"audit wrapper placeholder: configure an executor","success":false}'
# ```

# ─────────────────────────────────────────────────────────────────────────────────
# Requirements (admin constraints) example file
# ─────────────────────────────────────────────────────────────────────────────────
# allowed_approval_policies = ["on-request", "untrusted", "never"]
# allowed_sandbox_modes = ["read-only", "workspace-write"]