claude_cli_server#

Inference server backed by the Claude Code CLI in headless mode.

Satisfies the same call contract as InferenceServer so it can be dropped into config.inference via hydra _target_ with no other changes:

inference:

_target_: kernelfoundry.algorithm.claude_cli_server.ClaudeCLIServer model_name: opus num_completions: 1

Why this exists: the stock InferenceServer reads os.environ["OPENAI_API_KEY"] (or the Anthropic equivalent) at construction, so every run is billed against a metered API key. The claude CLI authenticates against a Claude subscription instead, and -p/--print makes it non-interactive.

Classes

ClaudeCLIServer([model_name, ...])

Generate completions by invoking claude -p as a subprocess.

Exceptions

ClaudeCLIError

Raised when the CLI cannot be run or returns something unusable.
exception kernelfoundry.algorithm.claude_cli_server.ClaudeCLIError[source]#

Raised when the CLI cannot be run or returns something unusable.

class kernelfoundry.algorithm.claude_cli_server.ClaudeCLIServer(model_name: str | None = None, num_completions: int = 1, max_tokens: int | None = None, temperature: float | None = None, timeout: int = 900, max_retry: int = 3, claude_exe: str = 'claude', disallow_tools: bool = True, max_turns: int = 8, effort: str | None = None, verbose: bool = False, **kwargs)[source]#

Generate completions by invoking claude -p as a subprocess.

Parameters:

  • model_name – Value for --model. An alias such as opus or sonnet tracks the latest model; a dated id pins one. None leaves the CLI default.

  • num_completions – Completions to return per call. The CLI returns one response per invocation, so N completions cost N sequential invocations.

  • max_tokens – Recorded for reporting only. The CLI exposes no output-token ceiling, so unlike the API path this cannot truncate a response.

  • temperature – Accepted and ignored, because the CLI exposes no sampling controls. Diversity across branches comes from separate sessions and from the differing prompts the evolutionary loop constructs, not from temperature.

  • timeout – Per-invocation timeout in seconds.

  • max_retry – Retries per completion before giving up.

  • claude_exe – Path or name of the CLI binary.

  • disallow_tools – Run as a pure text completion with no tool use. Leave this on for the evolutionary loop, which wants a kernel back rather than an agent that edits files.

  • verbose – Log each invocation.

__init__(model_name: str | None = None, num_completions: int = 1, max_tokens: int | None = None, temperature: float | None = None, timeout: int = 900, max_retry: int = 3, claude_exe: str = 'claude', disallow_tools: bool = True, max_turns: int = 8, effort: str | None = None, verbose: bool = False, **kwargs)[source]#
__call__(messages: list[dict], single_question: bool = False, **kwargs)[source]#

Return (outputs, metadata) matching InferenceServer.

Parameters:

  • messages – OpenAI-style {"role", "content"} dicts.

  • single_question – One completion, ignoring num_completions.

Returns:

(list[str], dict) where the dict carries model, token counts and the indices of any truncated completions.