This page holds the decisions of the engine facts that follow Engine Facts, from D62. Each decision rests on a fact of the Claude Code engine that the code alone does not show. Read it before you change the slice repostate or the slice turngate. The index is Decisions.

Problem: the slice repostate needs to give facts to the model at the start, to find the model that served a response, and to bound a registry request.

Choice: classic.SessionStart returns additionalContext, which the engine adds to the context of the session. The served model is usage.model of turn.complete, and the requested model is $.session.model(). The status-line command gets no served model, so the mark goes through $.ui.status. $.http.fetch has no timeout option, so the hook races the request with $.clock.sleep(ms, { signal: next.signal }), and the signal ends the wait when the dispatch ends. A request that loses the race, fails, or answers with a status other than 2xx, 404, or 410 is no answer, so the hook asks nothing and writes a notice with $.ui.log.

Evidence: the types of the mod API in .claude-plugin/types/claude-code/index.d.ts: TurnUsage.model, SleepOptions.signal, and the HttpInit type with no timeout field. The lab test repostate.test.ts runs each case with stubs.

Problem: the turn gate makes one model call in a hook, and its cost must be measured.

Choice: Each judge call writes the usage of its $.model.complete result to the debug log with $.ui.log. The cost figure of the judge comes from that usage, and never from the total_cost_usd of the session.

Evidence: in all 4 runs of task 1.10, the result message did not count the model call of a hook in total_cost_usd (V7, Failure-Gaps-0.1.0). ModelCompleteResult.usage in the types of the mod API gives the tokens of one call. The lab test turngate.test.ts checks that the usage goes to the debug log.

• Engine Facts
• Repo State
• Turn Gate
• Decisions