Providers

demur supports any OpenAI-compatible endpoint with a configurable base URL (covering hosted APIs, gateways, and self-hosted runtimes such as vLLM or Ollama) and the Anthropic API natively. Every pipeline role names a model, and every role may use a different provider.

Provider configuration

A named provider lives under [providers.<name>]:

[providers.openai]
family = "openai"                      # openai or anthropic
base_url = "https://api.openai.com/v1"
key_env = "OPENAI_API_KEY"             # environment variable holding the key

[providers.anthropic]
family = "anthropic"
base_url = "https://api.anthropic.com"
key_env = "ANTHROPIC_API_KEY"

Keys

The key comes from the environment variable named by key_env. On a local run you may instead name a file with key_file; the first line of the file is used. The key is never logged, never published in a review or check run, and never appears in an error message: error text that echoes request material is redacted before it can reach any output.

When using key_file, give an absolute path, because ~ is not expanded. Create the file with the key on the first line and restrict its permissions:

mkdir -p ~/.secrets
printf '%s\n' 'sk-your-key' > ~/.secrets/ajam-key
chmod 600 ~/.secrets/ajam-key

The environment variable takes precedence: when the variable named by key_env is set, the file is not read.

Model roles and prices

Every role is required, and every model requires a price, because budgets are denominated in money:

[models.triage]
provider = "openai"
name = "gpt-4o-mini"
input_price = 0.15    # USD per million input tokens
output_price = 0.60   # USD per million output tokens

There is no bundled price table. A stale table would silently understate spend, which is the failure a cost-control feature must not have, so a model without a price fails the run instead of being guessed at.

Model-specific parameters

Reasoning effort and thinking budgets are translated into each provider's dialect:

# OpenAI-compatible: reasoning effort
[models.deep]
provider = "openai"
name = "gpt-4o"
input_price = 2.50
output_price = 10.00
reasoning_effort = "high"   # minimal, low, medium, high

# Anthropic: thinking budget
[models.deep]
provider = "anthropic"
name = "claude-sonnet-4-5"
input_price = 3.00
output_price = 15.00
thinking_budget = 8000      # tokens

Passthrough parameters

Anything the bot does not model is forwarded untouched, so provider-specific controls are reachable without waiting for support:

[providers.openai.extra_body]
top_p = 0.9

[providers.openai.extra_headers]
"X-Custom-Header" = "value"

[models.deep.extra_body]
custom_sampling_control = 42

Prompt caching

The pipeline sends the same pull request context across several passes, which is the shape prompt caching is built for. Prompts put the stable context first and mark it cacheable where the provider supports it (Automatic Prompt Caching on OpenAI-compatible endpoints, cache control markers on Anthropic). Reported cached tokens are priced at cached_input_price, which defaults to half the input price and can be set explicitly per model.