af18009d59
Wave 1 item 1.2 of plans/token-waste-waves.md. Three provider shapes collapsed into two stored outcomes, so "the provider said zero" was indistinguishable from "the provider said nothing": details present, cached_tokens: 0 -> stored 0 details present, no cached count -> stored NULL no details block at all -> stored NULL That distinction is the entire question the item exists to settle. If the field is omitted on a full cache miss, every cache rate computed from the rows that carry it is conditioned on a hit having occurred and is biased upward -- which would make the 0.924 figure on pruned turns meaningless. `cached_tokens_source TEXT` records which shape produced the row: 'reported' | 'details_no_count' | 'no_details'. Additive, idempotent ALTER alongside the existing one, NULL on existing rows -- which is honest, since a row written before the column cannot say, and inferring one would fabricate exactly the evidence being gathered. Invariant: cached_prompt_tokens IS NOT NULL exactly when source = 'reported'. A JSON null under an existing cached_tokens key groups with 'details_no_count', not with a zero; no value is invented anywhere. Two bugs fell out of writing it: - `prompt_tokens_details: null` called .get on None and raised AttributeError inside the streamed finally block. - cached tokens were read ONLY from the explicit `usage` argument, so every buffered call site (non-streaming dispatch, eval_proficiency, seed_energy) dropped the count even when the provider reported it. It now resolves its source the same way cost already did. Measured read-only on the live DB before writing this: the 12.8% NeuralWatt coverage in the plan is not sparse reporting. Of 425 post-restart NULL rows, 420 are seed_reference sweeps; real dispatch traffic carries a count on 282 of 287 rows, and 24 of those are an explicit 0. So the provider does report zeros, and the upward-bias worry is much smaller than the plan assumed -- but the column is what makes that checkable rather than argued. 7 new tests: the three shapes at parse level, JSON null, a non-dict details block, the buffered payload path, and the three shapes asserted distinct as STORED rows. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U
25 KiB
25 KiB