the context meter could not see a compaction
上下文仪表看不见压缩
The composer's [context meter](../feature/2026-08-05-composer-context-meter-breakdown.md) took its ring, percentage, and `~used / capacity` header from `contextPressure.pressureTokens`, the newest provider-reported prompt size. That number moves only when a request reports usage, and compaction reports none: `compaction-basic` summarizes through a direct `ctx.llm.stream()` call and appends `compaction/start`, `compac
English
Problem
The composer's context meter took its ring, percentage, and ~used / capacity header from contextPressure.pressureTokens, the newest provider-reported prompt size. That number moves only when a request reports usage, and compaction reports none: compaction-basic summarizes through a direct ctx.llm.stream() call and appends compaction/start, compaction/summary, the replacement user/message, and compaction/end — no assistant/message, no usage chunk.
So the meter was frozen across the one action taken to change it. Driving a real compactNow through the agent loop:
BEFORE compact: ring=4% header=~4227/100000 rows=[system 18, tools 0, messages 4365]
AFTER compact: ring=4% header=~4227/100000 rows=[system 18, tools 0, messages 286]
The composition rows, which fold the surface, dropped by 93%. The ring — the primary affordance, and the reason a user opens the panel right after compacting — did not move at all, and would not until an entire further turn completed. The panel then showed a header and rows disagreeing by more than an order of magnitude, at exactly the moment a reader was most likely to add the rows up.
Decision
contextPressure publishes a second numerator, projectedTokens: the provider sample plus the heuristic repricing of everything the surface gained or lost since that sample was taken, clamped at zero. The fold carries the priced surface through the shared surface-fold.ts and stamps sampledSurfaceTokens when a usage sample lands — before the same event joins the surface, so an assistant/message anchors against the surface its own request actually carried. stateVersion moves to 3.
Only the delta is estimated. The anchor stays provider-exact, which keeps the estimator's systematic CJK and JSON-schema underpricing out of the occupancy figure while still letting the number react the moment content lands or a span is shadowed. contextOccupancy reads projectedTokens and falls back to the bare sample, so a projection restored from a pre-field checkpoint degrades to the old behavior instead of vanishing.
This reverses the "the ring, header, and bar length stay provider-exact" half of the context meter decision. What that decision was protecting — not fabricating precision by scaling heuristic rows to a provider total — is preserved: the rows are still unscaled, and the header still does not equal their sum. What changed is the recognition that "provider-exact but describing a request two compactions ago" is not the more truthful figure.
Alternatives considered
Project measure().totalTokens instead. The measurement service already composes exactly this (baseline anchor plus signed surfaceDeltaTokens), and it reacts correctly — measured at 4383 → 304 across the same compaction. But it is a service over private replay state, not a pure fold, and a projection cannot call it. Reproducing its anchor inside a ProjectionDefinition needs _estimateProviderAssistant's random access to the chunk events cited by seq (session.events[seq]), which apply(state, event) does not have. Anchoring on the sampled surface total is the same idea reachable from a pure per-event fold.
Emit a synthetic usage record at the end of compaction. Would move pressureTokens itself, but the only usage compaction holds is the summarization request's own — a different prompt entirely. Recording it as the conversation's prompt size would be a lie in the durable log rather than in one display.
Let the UI subtract, exposing sampledSurfaceTokens and reading contextBreakdown.messageTokens. Splits one figure's arithmetic across two projections and the client. The host owns the vocabulary; it should publish the whole value.
Consequences
Occupancy now advances with every surface event rather than once per turn, so the ring creeps up as a turn produces tool results instead of jumping at its end — and drops the instant a compaction lands. That is more projection frames on the wire: one per surface event for contextPressure, the rate contextBreakdown already ran at.
The panel's composition rows still do not sum to the header, and now for one clearly-stated reason instead of two: the rows carry the estimator's error, the header's anchor does not. The remaining lever is estimator accuracy (CJK-aware weighting in estimate.ts), which changes no seam.
sampledSurfaceTokens assumes nothing joins the surface between a step's request and its usage report. The loop admits steering and context before buildRequest and drains tool results after assistant/message, so that holds; if it ever stops holding, the error is bounded by one message and self-corrects at the next sample.
Testing
packages/llm/token-meter/tests/token-usage-projection.spec.ts covers the carry-forward across surface growth and a compaction (the sample holding still while the projection shrinks) and the zero clamp when heuristic error would drive the figure negative. packages/client/ui-conversation/tests/context-meter.client.spec.tsx pins the ring reading the projected figure, and chat-stats.spec.tsx pins contextOccupancy's preference and its fallback. The end-to-end numbers above came from driving BasicCompactionEngine.compactNow through a real AgentLoop with the projection registry mounted.
中文
问题
composer 的上下文仪表的圆环、百分比与 ~已用 / 容量 标题都取自 contextPressure.pressureTokens，即提供方报告的最新提示词规模。这个数字只在某个请求报告用量时才会移动，而压缩（compaction）不报告用量：compaction-basic 通过直连的 ctx.llm.stream() 调用生成摘要，只追加 compaction/start、compaction/summary、用作替换的 user/message 和 compaction/end——没有 assistant/message，也没有用量分片。
于是在唯一一个专门用来改变它的操作面前，这块仪表纹丝不动。通过真实的 agent loop（智能体循环）驱动一次 compactNow：
BEFORE compact: ring=4% header=~4227/100000 rows=[system 18, tools 0, messages 4365]
AFTER compact: ring=4% header=~4227/100000 rows=[system 18, tools 0, messages 286]
折叠表层得出的组成明细行下降了 93%。而圆环——那个主要的可操作元素，也正是用户压缩完立刻会去点开面板的理由——完全没动，而且要等到又跑完一整个轮次才会动。此时面板上的标题与明细行相差一个数量级以上，恰恰发生在读者最可能去把明细行加总的时刻。
决策
contextPressure 发布第二个分子 projectedTokens：在提供方样本之上，加上自取样以来表层增减部分的启发式重新计价，下界钳制为零。该折叠通过共享的 surface-fold.ts 携带已计价的表层，并在用量样本落地时记下 sampledSurfaceTokens——记录时机在同一条事件加入表层之前，因此 assistant/message 锚定的正是它自己那次请求实际携带的表层。stateVersion 提升到 3。
只有增量部分是估算的。锚点保持提供方精确值，从而把估算器对 CJK 文本与 JSON Schema 的系统性低估挡在占用率数字之外，同时又让这个数字能在内容落地或某段区间被遮蔽的瞬间做出反应。contextOccupancy 读取 projectedTokens，并回退到裸样本，因此从不含该字段的检查点恢复出来的投影会退化为旧行为，而不是直接消失。
这推翻了上下文仪表决策中「圆环、标题与进度条总长保持提供方精确值」的那一半。那条决策真正想守住的东西——不要把启发式明细行按比例缩放到提供方总量、从而伪造精度——依然守住了：明细行仍未被缩放，标题仍不等于它们之和。改变的是这样一个认识：「提供方精确、但描述的是两次压缩之前那个请求」并不是更真实的数字。
备选方案
改为投影 measure().totalTokens。 测量服务本来就合成了正是这个量（baseline 锚点加有符号的 surfaceDeltaTokens），而且反应正确——同一次压缩前后实测为 4383 → 304。但它是一个建立在私有重放状态上的服务，不是纯折叠，投影无法调用它。要在 ProjectionDefinition 内部复现它的锚点，需要 _estimateProviderAssistant 对按 seq 引用的分片事件进行随机访问（session.events[seq]），而 apply(state, event) 拿不到。以取样时的表层总量作为锚点，是同一个思路在纯逐事件折叠中可达的版本。
在压缩结束时补写一条合成的用量记录。 这确实能推动 pressureTokens 本身，但压缩手上唯一的用量是摘要请求自己的用量——那是完全另一个提示词。把它记成本对话的提示词规模，等于把谎言写进持久日志，而不只是写进某一处展示。
让 UI 自己做减法：暴露 sampledSurfaceTokens，再读 contextBreakdown.messageTokens。 这会把一个数字的算术拆散到两个投影和客户端三处。词汇的所有者是宿主，就应当由它发布完整值。
影响
占用率现在随每个表层事件推进，而不再是每个轮次跳一次，因此一个轮次中产生工具结果时圆环会持续爬升，而不是等到轮次结束才跳变——压缩落地的瞬间它也会掉下来。代价是线路上多了投影帧：contextPressure 每个表层事件推一帧，也就是 contextBreakdown 本来就在跑的频率。
面板的组成明细行仍然加不出标题数字，但现在只剩一个能讲清楚的原因，而不是两个：明细行带着估算器的误差，标题的锚点不带。剩下的抓手是估算精度（在 estimate.ts 里做 CJK 感知加权），它不改动任何 seam。
sampledSurfaceTokens 依赖一个前提：在某个步骤的请求与它的用量报告之间，不会有新内容加入表层。agent loop 在 buildRequest 之前接纳 steering（中途引导）与上下文，在 assistant/message 之后才排空工具结果，因此该前提成立；即便将来不再成立，误差也被限制在一条消息以内，并在下一个样本处自行纠正。
测试
packages/llm/token-meter/tests/token-usage-projection.spec.ts 覆盖了投影值在表层增长与一次压缩期间的延续更新（样本保持不动而投影值缩小），以及启发式误差会把数字压到负数时的零钳制。packages/client/ui-conversation/tests/context-meter.client.spec.tsx 钉住圆环读取投影值这一点，chat-stats.spec.tsx 钉住 contextOccupancy 的优先级与回退。上面那组端到端数字来自在挂载了投影注册表的真实 AgentLoop 上驱动 BasicCompactionEngine.compactNow。