.codecontext.toml

# ---------------------------------------------------------------------------
# CodeContext example configuration
# ---------------------------------------------------------------------------
# Copy this file to the root of your project as `.codecontext.toml` and edit
# the sections that apply to you. Every field below is optional — CodeContext
# ships with sensible defaults and will work with an empty file.
#
# Secrets (API keys) are NEVER stored here. Instead, set `api_key_env` to the
# name of the environment variable that holds your key, and CodeContext will
# read it at call time. This keeps keys out of git history and crash dumps.
# ---------------------------------------------------------------------------


# ===========================================================================
# Project scanning
# ===========================================================================
# Which files CodeContext indexes. Globs are evaluated in order. Sensible
# defaults already exclude .git, node_modules, build artifacts, binaries,
# and the .codecontext/ data directory itself.

# include = ["**/*"]
# exclude = ["vendor/**", "fixtures/**", "*.generated.py"]

# Hard cap on individual file size (bytes). Files larger than this are
# skipped to keep the index fast.
# max_file_bytes = 512000


# ===========================================================================
# Retrieval & context packing
# ===========================================================================
# Default token budget for the context pack the gateway sends to the LLM.
# Smaller budgets = cheaper requests but less context. 4000 is a reasonable
# starting point for code Q&A; raise for large refactors.
# default_context_budget_tokens = 4000

# Whether to use semantic embeddings on top of BM25. Requires the
# `codecontext[embeddings]` extra (sentence-transformers). If the extra is
# not installed CodeContext falls back to BM25-only retrieval automatically.
# enable_embeddings = true

# Hard cap on escalation cost per single request, in USD. Requests whose
# estimated external-LLM cost exceeds this are routed locally or blocked.
# This is the primary safety rail against runaway spend.
# max_escalation_cost_per_request = 0.08


# ===========================================================================
# LLM client — pick ONE of the three [llm_client] blocks below
# ===========================================================================
# The gateway speaks every provider's HTTP API directly using urllib, so you
# do NOT need to install the openai or anthropic SDKs. `enabled = false` (the
# default) means the gateway returns the outbound payload without calling an
# external model — useful for dry runs and CI.


# ---- Option A: OpenAI --------------------------------------------------
# Cheapest starting point. `gpt-4o-mini` is the default and is priced at
# $0.15 input / $0.60 output per 1M tokens.
#
# Before running:
#   export OPENAI_API_KEY="sk-..."

[llm_client]
enabled       = true
provider      = "openai"
model         = "gpt-4o-mini"
api_key_env   = "OPENAI_API_KEY"
max_tokens    = 1024
temperature   = 0.2
timeout_seconds = 60.0


# ---- Option B: Anthropic ----------------------------------------------
# Swap in by commenting out [llm_client] above and uncommenting this block.
# Claude Haiku is the cheapest tier; Sonnet is the balanced default.
#
# Before running:
#   export ANTHROPIC_API_KEY="sk-ant-..."

# [llm_client]
# enabled       = true
# provider      = "anthropic"
# model         = "claude-3-5-haiku-20241022"
# api_key_env   = "ANTHROPIC_API_KEY"
# max_tokens    = 1024
# temperature   = 0.2
# timeout_seconds = 60.0
# anthropic_version = "2023-06-01"


# ---- Option C: Ollama (fully local, free) -----------------------------
# Point CodeContext at a local Ollama server. No API key needed. Perfect for
# privacy-sensitive work or offline development. Pricing stays $0.
#
# Before running:
#   ollama pull llama3.1:8b
#   ollama serve

# [llm_client]
# enabled       = true
# provider      = "ollama"
# model         = "llama3.1:8b"
# base_url      = "http://127.0.0.1:11434"
# max_tokens    = 1024
# temperature   = 0.2
# timeout_seconds = 120.0   # local models are usually slower than APIs


# ---- Option D: OpenAI-compatible endpoint (vLLM, LM Studio, Groq,
#      Together, Fireworks, DeepSeek, a RunPod-hosted Qwen, etc.) -------
# Any server that speaks the OpenAI /v1/chat/completions shape works here.
# Set base_url to the server's /v1 root. api_key_env is optional: if the
# server doesn't require auth, leave it unset or point at an empty variable.

# [llm_client]
# enabled       = true
# provider      = "openai_compatible"
# model         = "Qwen/Qwen3-32B-FP8"
# base_url      = "https://your-runpod-endpoint.proxy.runpod.net/v1"
# api_key_env   = "RUNPOD_API_KEY"   # optional
# max_tokens    = 1024
# temperature   = 0.2


# ===========================================================================
# Pricing overrides (advanced, all providers)
# ===========================================================================
# If you've negotiated custom rates, self-host, or use a model CodeContext
# doesn't recognise, override the per-1M-token pricing here. Format:
# "model-prefix" = [input_usd_per_1m, output_usd_per_1m]. Prefix match is
# longest-wins.

# [llm_client.pricing]
# "gpt-4o-mini"           = [0.15, 0.60]
# "claude-3-5-sonnet"     = [3.00, 15.00]
# "Qwen/Qwen3-32B-FP8"    = [0.00, 0.00]   # self-hosted, compute tracked elsewhere


# ===========================================================================
# Extra headers (advanced)
# ===========================================================================
# Sent with every LLM request. Useful for routing/tagging through a proxy
# like Helicone or LiteLLM.

# [llm_client.extra_headers]
# "Helicone-Auth"         = "Bearer ${HELICONE_API_KEY}"
# "Helicone-Property-App" = "codecontext"