The token total is measured. The cost per task is a house rule.

An allocation rule cannot fail, and that is the objection to using one. Divide a window's token total by commit count and the parts always sum to the whole, which looks like corroboration and is only division; the agreement comes from the rule, not from anything anyone observed ^[2]. The author's list of how the mapping actually breaks is the load-bearing part: one session can cover several tasks, one task can span several sessions, a branch can sit untouched while the operator investigates something else or reviews another person's work, and elapsed time says nothing about the computational weight of the prompts and tool calls inside it ^[3].

So the tool is built to be structurally incapable of the guess. `agent-cost measure` takes session IDs and nothing else, and the caller that already knows the task selects them ^[4]. The one adapter described, `spec-lane`, runs the binary as a subprocess and validates the JSON, the `measure/v1` protocol version, the schema, and the absence of forbidden personal dimensions, while the measurement side never learns what the task was ^[5].

The design the post sketches runs local observations, then auditable normalized facts, then explicit pricing status, then caller-selected sessions, then task-attribution policy, then an optional dashboard ^[15]. That puts four trust gates between a log line and a chart, with the chart last of six ^[16]. Most cost dashboards ship the first element and the last one and quietly interpolate the middle, which is where a session number becomes a task number without anybody signing for it.

The uncertainty handling is where the boundary earns its keep. An unknown model comes out `unpriced`. A Claude cache write with no TTL breakdown is priced at the cheaper five-minute rate and stamped `lower_bound`. Codex logs do not expose cache-write tokens, so no zero-valued cache-write row is invented ^[9]. Malformed events, unreadable files and cumulative counters that go backwards stay visible in `data_quality` ^[10]. "Fail closed" here means unsupported pricing or attribution does not become a confirmed value downstream ^[11], and unattributed usage is left unattributed rather than split into something that looks precise ^[12].

The stated requirements were logs staying on the machine, a small runtime dependency surface, auditable custom metrics, and machine-readable output another tool can consume ^[1]. Runtime makes no network calls and declares no Python dependencies, and the price catalog carries a version and SHA-256 digest into the output ^[7]. The author then limits his own claim: installing from PyPI still means trusting a registry, an installer, a build backend, a runtime and an operating system, and the tool needs read access to the logs ^[8]. Scoping a zero-network claim to runtime behaviour is rarer than making it.

This is one developer writing about his own package, `coding-agent-cost 0.1.0`, rerun in temporary `uvx` directories where the doctor command located the local sources ^[13]. He is explicit that broader CLIs, local visual dashboards and OpenTelemetry stacks are answering different questions rather than doing this badly ^[14]. Which leaves the reader's own test: if your per-feature agent cost cannot be traced to a named set of session IDs, it is a convention.