[kemalelmizan]

The gap

#74 made the planning pass's model calls visible in the run summary. The System One decision layer spends the same way and was still invisible:

• With --system1-backend openai, every step makes a real, billed completion through /chat/completions. Nothing in the run reported its tokens.
• Tool telemetry cannot see it: a decision is not a tool execution, and the decision layer is deliberately not on the ChatModel seam (src/decision/systemone.ts).
• The usage the remote backend already parses (SystemOneResponse.usage) was dropped in SystemOneAdvisor.interpret.

So a run that leans on the remote decision tier understated its own spend, the same way an --plan-interval run did before #74.

The fix

Report every call. SystemOneAdvisor gains onCall, which fires for each backend call with the step, latency, the backend's tokens (when it reports any) and whether it answered:

this.onCall?.({ step: input.step, durationMs: Date.now() - started, ok: false });   // the backend threw
this.onCall?.({ step: input.step, durationMs: Date.now() - started, usage, ok: true });

It reports a call whose verdict was discarded (every answer below the confidence floor, so the loop never applies the decision) too — the tokens were spent either way. That mirrors PlanningPassOptions.onCall from #74.

Omit tokens rather than print zero. Both transports default the counts to 0 when a response carries no usage, so 0 means "not reported"; decision_tokens=0 would claim the decision was free. Same rule as plan_tokens.

Print it where a user looks. AgentRuntime accumulates the reports and exposes decisionCallSummary(), which the CLI appends to both run-summary sites:

[agent] stop=completed steps=4 plans=2 plan_tokens=1234 plan_ms=4100 decisions=2 decision_tokens=52 decision_ms=180

"" when the decision layer never ran, so a run without --system1 prints exactly what it printed before.

And in the REPL. /decision now ends with the same running total — the REPL has no run summary, and the decision layer is what that command is about:

confidence 0.80, model alpha, decoding deterministic, backend ollama:tev1:0.8b
  model=alpha (p=0.80); decoding=deterministic
  this run: decisions=2 decision_tokens=52 decision_ms=180

Notes

• Tokens, not money. The remote tier's price lives with its provider, so the agent reports the count and does not invent a cost.
• A shared concept is not built yet. There are now two accumulators: PlanUsage (a CLI-side host policy, feat(steps): report what the planning pass costs #74) and decisionCallSummary() (the runtime, because the decision layer is the runtime's seam). If a third instance of "the run's own model call" appears, that is the moment to look for one shape instead of three — the same reasoning that deferred ADR-004's coordinator.

Verified

• SystemOneAdvisor: reports tokens on success (step included), reports a failed call with no tokens, reports a discarded verdict, and omits tokens a backend never reported.
• AgentRuntime: decisionCallSummary() is "" before any call, decisions=1 decision_ms=… with a silent backend, and decisions=2 decision_tokens=52 decision_ms=… across a silent and a reporting call; lastDecisionSummary() shows the running total.
• pnpm build, pnpm typecheck, pnpm lint, pnpm test:coverage — green: 335 passed / 5 skipped, coverage 90.4 / 84.0 / 89.4 / 90.4.

Pre-existing prettier drift in src/decision/advisor.ts, src/decision/advisor.test.ts and src/runtime.ts is left untouched (#49); the lines added here match prettier's output.