Skip to content

Domain Workflow

Domain-Expert Workflow Generation — reusable harness synthesis (spec: custom_agent_harness.md).

Cluster: Uncategorised | Type: component | MCP Tools: None

Overview

Public API

HarnessArtifacts

Which §6.1 harness-package artifacts are actually present (not assumed).

Field Type Default
runs_end_to_end_on_unseen bool False
produces_declared_format bool False
has_workflow_graph bool False
has_rule_ledger bool False
skill_file_count int 0
skills_have_routing_heuristics bool False
has_rubric bool False
has_eval_suite bool False
has_known_limitations bool False
has_versioned_manifest bool False
has_tool_policy bool False
has_injection_handling bool False
has_escalation_policy bool False
has_failure_mode_docs bool False
has_run_report bool False
has_user_docs bool False
has_maintainer_docs bool False
has_rollback_metadata bool False
composable bool False

EvalSummary

Results of running the harness's OWN generated eval suite (§8.2 manifest).

Field Type Default
pass_rate float \| None None
has_golden_tests bool False
has_adversarial_tests bool False
has_regression_tests bool False
has_hard_rule_tests bool False
hard_rule_violation_rate float 0.0
prompt_injection_success_rate float 0.0
unsafe_external_action_rate float 0.0
required_escalation_miss_rate float 0.0
outperforms_baseline bool \| None None
unresolved_hard_rule_conflicts int 0

Methods:

blocking_failures(minimum_pass_rate: float = 0.9) -> list[str]

The §8.2 blocking requirements that are violated (empty == none).

HarnessCard

The §14.1 registry card — honest summary incl. the real eval status.

Field Type Default
harness_id str required
version str required
domain str required
task str required
achieved_tier str required
reliability_label str required
latest_pass_rate float \| None required
blocking_failures int required
when_to_use list[str] field(default_factory=list)
when_not_to_use list[str] field(default_factory=list)
missing_for_next_tier list[str] field(default_factory=list)

Methods:

to_dict() -> dict

HarnessAssessment

Field Type Default
achieved_tier str required
reliability_label str required
constraint_complete bool required
missing_constraints list[str] field(default_factory=list)
blocking_reasons list[str] field(default_factory=list)
missing_for_next_tier list[str] field(default_factory=list)
prompt_generator_trap bool False
over_claimed bool False

Methods:

to_dict() -> dict

Functions

check_constraint_completeness(artifacts: HarnessArtifacts) -> tuple[bool, list[str]]

Return (complete, missing[]) per §12.2. complete iff nothing is missing.

is_prompt_generator_trap(artifacts: HarnessArtifacts) -> bool

§20 — output that lacks the structural harness core (graph + a skill + an eval suite)

assess_harness(artifacts: HarnessArtifacts, evals: EvalSummary, claimed_tier: str | None = None, minimum_pass_rate: float = 0.9) -> HarnessAssessment

Deterministically assess the harness's REAL tier + honest reliability label (§18).

build_harness_card(harness_id: str, version: str, domain: str, task: str, assessment: HarnessAssessment, evals: EvalSummary, when_to_use: list[str] | None = None, when_not_to_use: list[str] | None = None) -> HarnessCard

Assemble the §14.1 registry card with the HONEST eval status + achieved tier.