Cost

Ledger and unit pricing for completer calls and successful tool runs. Module: mechaharness.core.access. Stability: evolving.

Surface

Item	Value
Module	`mechaharness.core.access`
Event	`core:cost`
Stability	evolving

Ability units

Ability ranks simple → basic → intermediate → proficient → advanced cost 1, 2, 4, 8, 16 units. A CapabilityProfile sums those units (empty profile counts as simple / 1). Tool cost is that tool’s ability.units.

InMemoryCostAccountant.price_inference() / price_tool() append a CostEntry, update the ledger, and emit core:cost when agent_id and run_id are set. When a completer returns OpenAI-style usage, those token counts are stored on the entry and on the run CostReport (local Spark/$0 pricing can still use Ability units while tracking tokens).

Harness

AbstractHarness injects CostAccountant (default InMemoryCostAccountant on the same EventLog). Each inference turn is priced after complete() and passes through CompletionResponse.usage. Each successful tool run is priced after invoke. Denied or unknown tools are not priced. HarnessResult.cost is the ledger for that run. Thinking-model reasoning_content is kept on ChatMessage and recorded on core:inference.

from mechaharness.core.access import InMemoryCostAccountant
from mechaharness.core.events import InMemoryEventLog
from mechaharness.harness.base import HarnessConfig
from mechaharness.harness.pass_through import PassThroughHarness
from mechaharness.inference.mock import MockInferenceStrategy

log = InMemoryEventLog()
harness = PassThroughHarness(
    inference=MockInferenceStrategy(),
    config=HarnessConfig(model="mock"),
    event_log=log,
    cost=InMemoryCostAccountant(event_log=log),
)

Tool grants are documented in Access control.