feat(web): every inference on Luke's key is counted on the account's meter, a coding agent's and every title included (one runtime 6/n)

Merged
DA
6d ago
+172-76

One-agent-runtime migration, PR 6 (docs/proposals/one-agent-runtime.md §5, Decision 6: "Meter every inference on Luke's key, coders included").

What

Only the planning model and its workers counted the account's daily meter (spendHostedMeter). A coding agent or its worker running on Luke's key wasn't counted, and neither was any generated title.

  • server/hosted/agent-runtime/meter.ts (new): MeterSeams, meteredModel (moved unchanged from brain-host/model.ts), and accountMeter(spend, userId).
  • Luke's key carries the meter. ProviderKey.meter is set only on Luke's own key (coder-host/account-keys.ts). modelOnKey puts the meter in front of a model on Luke's key, and in front of nothing on the account's own key. Every step on Luke's key is counted, whoever asked:
    • the planning model and its workers (brainModelSelection hands its meter in with the key instead of wrapping afterwards);
    • a coding agent and its workers, which were never counted before.
  • Titles. titleModel(keys, meter) counts a title on Luke's Anthropic key. Both hosts' titleModel seam now takes the account (titleModel(userId)), so a coding agent's title, a delegated worker's title and a planning worker's title are each counted for the account they name.
  • BYOK steps are never counted: the account pays for them.

The count still enforces nothing. It never refuses, and a failed count is reported by its kind and the inference runs anyway (meter.ts header).

Tests

  • tests/coder-host-account-keys.test.ts, against providers faked at their fetch over a real dialect:
    • New: a step on Luke's key counts the account's meter once, even one the provider refuses, and a step on the account's own key counts nothing.
    • New: a title generated on Luke's key counts the account's meter once.
  • tests/hosted-brain-host-model.test.ts: the meter's own tests (one count per inference, a failed count reported by kind) now import it from agent-runtime/meter.
  • The whole web suite passes (153 files, 1,319 tests). The offline brain and coder evals pass against a local Postgres. Lint and repository checks pass.

This PR touches neither the Mac app nor the renderer.

🤖 Generated with Claude Code

Open workspace in Conductor

Devin Review