Measuring without the harness

The measurement report reads one file: the session ledger at ~/.local/state/agent-harness/usage.jsonl, written by the usage-log hook. Nothing in bin/harness usage needs a projected primitive, a merged settings file or a synced runtime, so the report is useful to someone who wants the numbers and not the way of working.

When the whole harness is the answer

• You want the rules the report scores. --rules counts detector hits against claude/hooks/rule-detectors.py, and a detector only means something when the rule it is named for is actually loaded into the session. Hit counts without the prose they measure are a share of nothing. See usage.md.
• You want stance groupings. --by stance reads the dimension: variant map the CLI resolves, so a checkout with no config has one variant per dimension and one group per report.
• You are already running harness sync. Then the hook is installed and the ledger fills without another step.

When the report alone is the answer

• You want spend per repository, model, role or day, and you run Claude Code or Codex as they come. Install the usage-log hook, leave everything else alone, and read usage.md for what each grouping means.
• You are evaluating the measurement before adopting the primitives. The ledger is local, it holds no message text, and removing the hook stops it.
• You are feeding an observability backend. harness usage export replays the ledger; see telemetry.md.

The packaged path

Installing the hook by hand from this repository is the only route today. A standalone package that carries the hook, the detectors and the report without the rest of the harness is planned in issue #453. Until it lands, a standalone reader copies two files out of this checkout and pins the version they came from.