sample_config.toml

default_model = "bodega-current"
default_thinking = true

# ── Bodega models ─────────────────────────────────────────────────────────────
#
# These route to the local Bodega inference engine (http://localhost:44468).
#
# ⚠️  IMPORTANT — TWO SEPARATE CONFIG SYSTEMS:
#
# 1. THIS FILE (axe-cli config.toml) — what axe-cli reads.
#    The [models.*] blocks here only accept these fields:
#
#      provider          = "bodega"          (required)
#      model             = "<model-id>"       (required — the ID you gave it when loading)
#      alias             = "<nickname>"       (optional)
#      max_context_size  = <int>              (required)
#      capabilities      = ["thinking", ...]  (optional)
#
#    That's it. Fields like tool_call_parser, reasoning_parser, continuous_batching,
#    max_concurrency, cb_*, etc. are NOT read by axe-cli and will be silently ignored
#    or cause a validation error if added here.
#
# 2. BODEGA SERVER — how to load models into the running Bodega app.
#    Bodega runs as an app; you load models into it via the API endpoint.
#    This is where all the rich model-loading options live:
#
#      model_path, model_id, model_type          — what to load and how to name it
#      tool_call_parser, reasoning_parser         — structured output parsers
#      enable_auto_tool_choice                    — automatic tool selection
#      max_concurrency, queue_timeout             — concurrency and backpressure
#      context_length                             — per-model context window
#      continuous_batching, cb_*                  — high-throughput batching engine
#      draft_model_path, num_draft_tokens         — speculative decoding
#      prompt_cache_size                          — KV-cache reuse for recurring prefixes
#
#    Load a model at runtime via POST /v1/admin/load-model:
#
#      curl -X POST http://localhost:44468/v1/admin/load-model \
#        -H "Content-Type: application/json" \
#        -d '{
#          "model_path": "srswti/bodega-raptor-8b-mxfp4",
#          "model_id":   "srswti/bodega-raptor-8b-mxfp4",
#          "model_type": "lm",
#          "context_length": 32768,
#          "max_concurrency": 1,
#          "enable_auto_tool_choice": true,
#          "tool_call_parser": "qwen3",
#          "reasoning_parser": "qwen3"
#        }'
#
#    📖 Full reference: docs/bodega-inference-engine.md
#       Covers all load parameters, CB config, parsers, speculative decoding, hardware tuning.
#
# ──────────────────────────────────────────────────────────────────────────────

# "current" — proxy to whatever Bodega model is currently active.
# Useful as a catch-all when running a single model via the CLI launcher.
[models.bodega-current]
provider = "bodega"
model = "current"
max_context_size = 32768
capabilities = ["thinking"]

# Raptor 8B — best general-purpose laptop model.
# Load on Bodega side with: tool_call_parser="qwen3", reasoning_parser="qwen3"
[models.bodega-raptor]
provider = "bodega"
model = "srswti/bodega-raptor-8b-mxfp4"
max_context_size = 32768
capabilities = ["thinking"]

# Raptor 90M — ultra-fast edge model.
# Load on Bodega side with: continuous_batching=true, cb_completion_batch_size=32
[models.bodega-raptor-small]
provider = "bodega"
model = "srswti/bodega-raptor-90m"
max_context_size = 32768
capabilities = ["thinking"]

# Orion 0.6B — sub-1B, ~1350 tok/s system throughput on M4 Max with continuous batching, batch =64.
# Load on Bodega side with: continuous_batching=true, tool_call_parser="qwen3"
[models.bodega-orion]
provider = "bodega"
model = "srswti/bodega-orion-0.6b"
max_context_size = 32768
capabilities = ["thinking"]

# Centenario 21B — high-capacity production model.
# Load on Bodega side with: max_concurrency=3, tool_call_parser="harmony"
[models.bodega-centenario]
provider = "bodega"
model = "srswti/bodega-centenario-21b-mxfp4"
max_context_size = 65536
capabilities = ["thinking"]

# axe-stealth-37b — high-capacity stealth model, qwen3_coder tool parser.
# Load on Bodega side with: tool_call_parser="qwen3_coder", reasoning_parser="qwen3_5"
[models.axe-stealth-37b]
provider = "bodega"
model = "srswti/axe-stealth-37b"
max_context_size = 32768
capabilities = ["thinking"]

# Blackbird 21B — uncensored 21B, 128K context.
# Load on Bodega side with: tool_call_parser="harmony", reasoning_parser="harmony"
[models.bodega-blackbird]
provider = "bodega"
model = "srswti/blackbird-she-doesnt-refuse-21b"
max_context_size = 128000
capabilities = ["thinking"]

[models.kimi-k2-thinking]
provider = "openrouter"
model = "moonshotai/kimi-k2-thinking"
max_context_size = 262144
capabilities = ["thinking"]

[models.nemotron-nano]
provider = "openrouter"
model = "nvidia/nemotron-3-nano-30b-a3b"
max_context_size = 32000

[models.glm-4-flash]
provider = "openrouter"
model = "z-ai/glm-4.7-flash:nitro"
max_context_size = 128000

[models.deepseek-r1]
provider = "openrouter"
model = "deepseek/deepseek-r1"
max_context_size = 64000
capabilities = ["thinking", "always_thinking"]

[models.deepseek-r1-distill-qwen]
provider = "openrouter"
model = "deepseek/deepseek-r1-distill-qwen-32b"
max_context_size = 64000
capabilities = ["thinking"]

[models.qwen-coder-32b]
provider = "openrouter"
model = "qwen/qwen-2.5-coder-32b-instruct"
max_context_size = 32000

[models.codestral]
provider = "openrouter"
model = "mistralai/codestral-latest"
max_context_size = 32000

[models.claude-sonnet]
provider = "anthropic"
model = "claude-3-5-sonnet-latest"
max_context_size = 200000
capabilities = ["thinking", "image_in"]

[models.claude-4-5-sonnet]
provider = "anthropic"
model = "claude-sonnet-4-5-20250929"
max_context_size = 200000
capabilities = ["thinking", "image_in"]


[models.gpt-nano]
provider = "openai"
model = "gpt-5-nano-2025-08-07"
max_context_size = 200000
capabilities = ["thinking", "always_thinking"]

[models.gpt-5-2]
provider = "openai"
model = "gpt-5.2-2025-12-11"
max_context_size = 200000
capabilities = ["thinking", "always_thinking"]

[providers.openrouter]
type = "openai_legacy"
base_url = "https://openrouter.ai/api/v1"
api_key = "sk-or-v1-"
reasoning_key = "reasoning"

[providers.openrouter.custom_headers]
HTTP-Referer = "https://github.com/SRSWTI/axe"
X-Title = "axe"

[providers.openai]
type = "openai_responses"
base_url = "https://api.openai.com/v1"
api_key = "sk-proj-...."
reasoning_key = "reasoning"

[providers.anthropic]
type = "anthropic"
base_url = "https://api.anthropic.com/v1"
api_key = "sk-ant-api...."
reasoning_key = "reasoning"

# Bodega provider — local inference engine running on Apple Silicon.
#
# The provider block only carries connection settings (base_url, api_key).
# All the rich model-loading options (tool_call_parser, reasoning_parser,
# continuous_batching, max_concurrency, context_length, etc.) are set when
# loading models into the Bodega app via:
#
#   POST /v1/admin/load-model
#
# See docs/bodega-inference-engine.md for the full parameter reference
# and example curl commands for every model type and configuration.
#
# Environment variable overrides (applied at runtime):
#   BODEGA_BASE_URL      Override base_url
#   BODEGA_MODEL         Override the model identifier for a specific entry
#   BODEGA_TEMPERATURE   Set generation temperature
#   BODEGA_MAX_TOKENS    Cap output tokens
[providers.bodega]
type = "bodega"
base_url = "http://localhost:44468"
api_key = ""

[loop_control]
max_steps_per_turn = 100
max_retries_per_step = 3
max_ralph_iterations = 0
reserved_context_size = 50000

[services]

[mcp.client]
tool_call_timeout_ms = 300000  # 5 minutes