each-run-should-record-the-model

Summary

Delivers itd-166: every model call records the side abcd asked for and the side that answered, on the receipts that already exist, with every field a route cannot read named with its reason. Decisions 1–10 on the intent are the design record.

Shape. Two seams, one per path a call takes, sharing one vocabulary.

• Shared vocabulary in internal/core/oracle (the runner package already imports oracle; oracle imports nothing from runner): Usage (prompt, completion, reasoning, cache-read, cache-write tokens), Cost (a pointer, so zero is not confused with unreported, with its unit fixed per route), and Unpinned (field name and reason: not_reported, self_reported, harness_computed, not_configured) checked against the fixed asked and served lists of decision 10.
• Provider adapter path (internal/adapter/openaiapi, internal/core/oracle): the request adds stream_options.include_usage unless the provider's connection turns it off (stream_usage: false, default on), and a refusal's error names that setting; it is recorded as adapter-added, not as a setting sent. Result, Usage and the stream chunk gain the generation id, a top-level provider, prompt_tokens_details.cached_tokens, completion_tokens_details.reasoning_tokens and cost; the parser keeps reading past the finish reason to [DONE] for the usage chunk. Every provider-supplied string is cleaned and bounded in admit, as the reported model is. complete() hashes the rendered messages. CallRecord gains prompt sha256, provider, generation id, finish reason, usage, cost and the unpinned list; the harness hand-off marks ReceiptRoute's served model self-reported. No request is added.
• Runner path (internal/core/runner): Answer gains usage, cost, served models and harness version, and its comment stops saying "or the one it was asked for". The claude runner reads the result event's usage and total_cost_usd (harness-computed), the served models from modelUsage's keys (cross-checked against the init event's model), and claude_code_version; reasoning tokens are listed not reported. The opencode runner stops stamping the asked model (resolves iss-2610161709506207): served model empty and not reported, tokens and cost summed across step_finish events (cost harness-computed; sub-session events are already filtered out, so the totals are a floor and say so). RouteRecord gains the asked model, usage, cost, harness version and unpinned list; Route.Model is the canonical served model and ReceiptRecord.Model stays as the legacy copy.
• Persistence: no new file. Run state moves to schema 14 (a version-13 file reads as not recorded and is written back at 14; a file at 13 or lower carrying the new fields is refused). Built as 13; renumbered when release v0.15.1 took 13 for a drain lane's refused. maskRoute names every new string and slice. The consistency answer's route.json carries the oracle receipt, cleaned at the adapter.

How each acceptance criterion is met. (1) adapter both sides: step 1; (2) claude: step 2; (3) opencode: step 2; (4) reasons against fixed lists: steps 1 and 2; (5) host hand-off self-reported: step 1; (6) usage opt-out: step 1; (7) receipts, cleaning, no new file: steps 1 and 3; (8) no added request: step 1.

Footprint

• packages: internal/adapter/openaiapi, internal/core/oracle, internal/core/runner, internal/core/implement/loop, internal/core/intent, internal/surface/cli
• tests: OpenRouter-shaped streamed and plain responses fill provider, generation id, finish reason, usage details and cost; a usage chunk after the finish reason is read; absent fields are listed with their reason; the request carries stream_options.include_usage unless the connection turns it off, and a refusing server's error names the setting; the request count is the same with and without the record; the prompt hash is stable and sensitive to one byte; control characters in provider, generation id and model are cleaned; a claude result event's usage, cost, modelUsage keys and version are read from a fixture taken from a real stream; an opencode stream's step-finish tokens and cost are summed and its served model is empty and not reported; the oracle's host hand-off marks the served model self-reported; state.json at 14 round-trips, at 13 reads as not recorded, and an older file carrying the new fields is refused; the mask covers each new field; drive passes the asked model and usage through

Steps

1. The provider adapter records both sides
   • packages: internal/adapter/openaiapi, internal/core/oracle
   • tests: OpenRouter-shaped streamed and plain responses fill the served side; the usage chunk after the finish reason is read; absent fields listed with reasons; the usage request and its opt-out, with the refusal naming it; no added request; prompt sha256; provider strings cleaned and bounded; the host hand-off's served model self-reported
2. The runners record both sides
   • packages: internal/core/runner
   • tests: claude result usage, harness-computed cost, modelUsage served models and version read from a real-stream fixture, reasoning listed not reported; opencode step-finish tokens and cost summed, served model empty and not reported, never the asked model; the asked model kept beside the served one
3. The sides travel with the existing receipts
   • packages: internal/core/implement/loop, internal/core/intent, internal/surface/cli
   • tests: run state schema 14 with the version-13 rule; the lane receipt and the consistency route file round-trip the new fields with no new file; maskRoute covers each new string and slice; drive passes the asked model and usage through; the route line render names the served model and cost when present