Skip to main content

hackagent.orchestrator.campaign.resolve

Turn a campaign spec into the objects a run needs.

Resolution loads the goals, creates every model client (target, attack roles, judges), and constructs each attack from its parameters plus a traced completion for each of its role models. No request is sent to any model. The evaluation panel is built here too, and the runner hands it to iterative attacks as they search.

EmbedderBuilder​

Builds the texts-in, vectors-out callable an embedder role receives.

ResolvedAttack Objects​

@dataclass(frozen=True)
class ResolvedAttack()

A ready-to-run attack and the configuration it was built from.

configuration​

Parameters and role model names, as recorded with the run.

ResolvedCampaign Objects​

@dataclass(frozen=True)
class ResolvedCampaign()

probes​

One reachability check per distinct model, for preflight.

resolve_campaign​

def resolve_campaign(
        spec: CampaignSpec,
        *,
        build: ModelBuilder = build_model,
        load: GoalLoader = load_goals,
        build_embed: EmbedderBuilder = build_embedder,
        settings: Optional[Settings] = None) -> ResolvedCampaign

Instantiate everything spec describes.

Gateway credentials come from settings, resolved once if omitted. An explicit model connection api_key takes precedence.

resolve_attack​

def resolve_attack(
        spec: AttackSpec,
        build: ModelBuilder,
        *,
        retries: int = 0,
        build_embed: EmbedderBuilder = build_embedder) -> ResolvedAttack

Construct an attack, binding each role model to the shape it declares.

A completion role receives a traced chat callable; an embedder role receives a texts-in, vectors-out callable from the embeddings endpoint.

build_panel​

def build_panel(spec: EvaluationSpec,
                build: ModelBuilder,
                *,
                retries: int = 0) -> Optional[Panel]

Bind every judge to its model. None when no judge is configured.

resolution_summary​

def resolution_summary(resolved: ResolvedCampaign) -> dict[str, str]

What resolving the campaign found, as label-value rows.

The facts a run is about to act on: how many goals were loaded, which attacks were built, and which optional parts are switched on. The CLI prints them as a table and the builder logs them before it starts, so they are written once here and cannot drift apart.