Configuration

Settings live in ~/.ex5/config.toml. Override the whole data root with EX5_HOME. Project-level overrides go in <repo>/.ex5/local.toml. Precedence: CLI flag / env > config file > built-in default.

You normally do not write this by hand — `ex5 login --key ex5-live-…` generates it from the model catalogue.

toml

# ~/.ex5/config.toml
default_model = "ex5/ex5-security-27b"

[providers.ex5]
type = "openai"                    # OpenAI Chat Completions
base_url = "https://ex5.ai/v1"
api_key = "ex5-live-…"

[models."ex5/ex5-security-27b"]
provider = "ex5"
model = "ex5-security-27b"
max_context_size = 131072
max_output_size = 8192             # ÇIKIŞ tavanı — bağlamdan ayrı, aşağıya bak
capabilities = ["tool_use"]
display_name = "EX5-Security 27B"

[thinking]
enabled = false

Output limit

max_output_size is separate from max_context_size on purpose. Our models are served by vLLM, where the prompt and the reply share one window (max_model_len). If a client sets max_tokens to the full context size, the request is rejected and every agent step fails. Leave it as login wrote it unless you know the server accepts more.

Thinking

Reasoning models produce a chain of thought before answering. When the model declares supports_reasoning, ex5 renders that stream in a separate collapsed block instead of printing it as the answer. Toggle it with [thinking].enabled; leaving it off keeps turns cheaper and the tool loop faster.

Server side matters too: if the endpoint is not started with a reasoning parser, the thinking is not separated into reasoning_content and lands in the answer text. That is a server configuration issue, not a model defect.

Other providers

ex5 ships pointed at ex5.ai but is not locked to it. Add any provider as a sibling table and any model under [models.<alias>]. Supported type values: openai, openai_responses, anthropic, google-genai, vertexai, kimi. Anything that speaks OpenAI Chat Completions works — OpenRouter, Together, a local vLLM, Ollama — by pointing base_url at it.

toml

[providers.openai]
type = "openai"
base_url = "https://api.openai.com/v1"
api_key = "sk-…"

[models."openai/gpt-x"]
provider = "openai"
model = "gpt-x"
max_context_size = 128000
max_output_size = 16384
capabilities = ["tool_use"]

Tool calling is required. ex5 is an agent: without tool_calls it can talk but cannot read or edit your repo. A self-hosted vLLM needs --enable-auto-tool-choice and --tool-call-parser; without those flags it accepts tools, returns 200, and never emits a call.