ROADMAP-06 — P0 — Turn token usage is written to every tool-call step row (double counting) #98

Series: ROADMAP — filed from the 2026-10-27 whole-project assessment · Backlog task: M14-T06 · Priority: P0 · Effort: extra small

Problem

When a turn requests multiple tool calls, the entire turn's usage is written onto every generated step row: sum(steps.input_tokens) > runs.total_input_tokens. Any dashboard or cost report built by summing steps lies; runs totals are correct by accident.

There is also no model/cost attribution anywhere (tracked separately in ROADMAP-12) — this issue fixes the attribution math only.

Evidence

• agentkit/core/agent.py:394–407 — per-step trace write copies response.usage.input_tokens/output_tokens onto each tool-call step
• agentkit/db/models.py — steps.input_tokens, runs.total_input_tokens

Suggested fix / acceptance criteria

• A turn's usage is recorded once (turn-level step), per-call step rows carry per-call latency/outcome only (0 for token columns or a shared-usage marker chosen by design)
• Invariant test: for any run, sum(steps.*) <= runs.total_* holds with equality semantics documented
• Existing trace consumers/tests updated

Suggested branch / commit

fix/step-token-attribution · fix(core): record turn token usage once, not per tool-call step