Name the number before judging it.
Newest provider-reported prompt sample.
Sample plus signed surface movement.
Current Session under fixed-density estimation.
Approximate system, tools, and message composition.
Route a large cumulative increase call by call.
Chart provider input buckets per call; cumulative usage never falls after compaction.
Measure the model-facing result, not only the source file or generated artifact.
Compare exact canonical arguments, failures, cache buckets, and changed variables.
Bounds accepted plain-text results, with documented bypasses including read and mixed content.
Rewrites oversized results only after compaction pressure or provider overflow.
Reduces future request history but adds an auxiliary summary call when pruning is insufficient.
Detects exact consecutive repeats at thresholds 3, 5, and 8; it is advisory.
Repair both consumers.
Preserve Session, compaction range, runtime artifact, and exact schema path.
Guard messageTokens and the sibling pressure surfaceTokens.
Discard cached rows and refold authoritative history from sequence zero.
Updating only messageTokens leaves the sibling projection exposed. Updating code without state migration leaves existing cached Sessions ambiguous.
Keep billing and occupancy separate.
Provider usage anchors a completed request. Heuristics price later surface movement. Composition rows remain approximate and need not sum to the provider-anchored headline.
Primary evidence.
- Cumulative token-cost analysis #4758
- Wide-compaction report #4674
- Independent report #4703
- rc.2 shipped base composition
- rc.2 token-meter contract
Measure the call, preserve the Session.
The complete guide covers provider buckets, spill and pruning timing, compaction cost, retry evidence, signed underflow recovery, and regression gates.
Read the complete guide