hackagent.orchestrator.campaign.overrides

Endpoint overrides and local-server readiness for running a campaign file.

A campaign file pins each model's endpoint, but the servers a run actually reaches are often decided at launch (a SLURM job, a notebook, a one-off CLI run). These helpers rewrite a campaign's model connections to point at the servers given at run time, and wait for local servers to come up and serve the models the campaign needs. They operate on the plain dict a :class:CampaignSpec dumps to, so a caller can edit connections in place and re-validate.

The CLI campaign command and scripts/run_campaign.py both build on these; neither logic lives in the other.

EndpointOverrides Objects

@dataclass(frozen=True)
class EndpointOverrides()

Run-time endpoints to point campaign models at.

ollama repoints every (Ollama) model. The OpenAI-compatible overrides split by role: openai moves the target and the judges sharing its server, judge moves the other judges, attacker moves attack role models. ollama cannot be combined with the OpenAI-compatible ones.

ServerReadiness Objects

@dataclass(frozen=True)
class ServerReadiness()

How long to wait for local servers, and the pids to watch.

campaign_models

def campaign_models(
        values: dict[str, Any]) -> Iterator[tuple[str, dict[str, Any]]]

Yield (kind, model) for every model a campaign names.

kind is target, role or judge. Models are the mutable dicts of values, so callers can rewrite their connections in place.

installed_models

def installed_models(endpoint: str,
                     timeout: float,
                     server_pid: int | None = None,
                     *,
                     provider: str = "ollama",
                     api_key: str | None = None) -> set[str]

Wait for readiness and return installed or served model identifiers.

api_key is sent as a bearer token to an OpenAI-compatible server, which answers 401 to an unauthenticated model listing when started with a key.

apply_endpoint_overrides

def apply_endpoint_overrides(values: dict[str, Any],
                             overrides: EndpointOverrides) -> None

Point a campaign's models at overrides' servers, in place.

Raises:

• ValueError - if the overrides do not fit the campaign's model connections (wrong wire type, no role model to point at, or Ollama mixed with OpenAI-compatible overrides).

wait_for_servers

def wait_for_servers(values: dict[str, Any],
                     readiness: ServerReadiness) -> None

Wait for every local server and check it serves the models it must.

Raises:

• ValueError - if a vLLM model has no endpoint. An Ollama model without one is checked at OLLAMA_HOST, where it will run.
• RuntimeError - if a server never becomes ready or lacks a model.