Merge master into fix/conversation-column-one-axis-scroll
This commit is contained in:
creatixchu
@@ -2,5 +2,5 @@
|
||||
# side as of the last confirmed-consistent state. Both languages carry equal authority;
|
||||
# after editing either side, bring the other along and re-record with:
|
||||
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-21-semantic-session-checkpoints.md
|
||||
2026-07-21-semantic-session-checkpoints.md: 927a4c5d6d2aad5dea460ea29f97c1686e9d5398
|
||||
2026-07-21-semantic-session-checkpoints.zh.md: c896ce55b23f08832cb5c80bf7aae264e366fee9
|
||||
2026-07-21-semantic-session-checkpoints.md: 697d878dccbfab1ded9f73e134d594ba6f0665e1
|
||||
2026-07-21-semantic-session-checkpoints.zh.md: c6184b0b00db16f9e318f4cbd148b1f0525f03a5
|
||||
|
||||
@@ -10,9 +10,9 @@ Persistence buffered every synchronous `session/event` until the loop's final tu
|
||||
|
||||
## Decision
|
||||
|
||||
`dsh-session-checkpoint-policy` owns semantic durability barriers as a zero-config plugin beside a persistence backend. At `agent/step`, it flushes pending prompt input or the preceding response/result batch before the next request is derived. It wraps `llm/stream` lazily and flushes the live session after `request/header` is logged but before the adapter stream is constructed. It wraps top-level `tools/execute` after ordered pre-execute policy and flushes the recorded `tool/call` before the tool body; nested dispatches reuse the outer model-visible call. The loop's final `turn/end` checkpoint remains the closing boundary and settles before another queued turn or idle observation.
|
||||
`dsh-session-checkpoint-policy` owns semantic durability barriers as a zero-config plugin beside a persistence backend. At `agent/pre-step`, it flushes pending prompt input or the preceding response/result batch before the next request is derived. It wraps `llm/stream` lazily and flushes the live session after `request/header` is logged but before the adapter stream is constructed. It wraps top-level `tools/execute` after ordered pre-execute policy and flushes the recorded `tool/call` before the tool body; nested dispatches reuse the outer model-visible call. The loop's final `turn/end` checkpoint remains the closing boundary and settles before another queued turn or idle observation.
|
||||
|
||||
Persistence and checkpoint scheduling remain separate Cordis plugins. A backend makes requested `session/flush` boundaries durable but does not choose them; loading it without this policy is valid and retains the loop's coarser checkpoints. First-party persisted apps and runtimes explicitly mount both, while a specialized deployment may intentionally omit or replace the policy. Registration order governs whether events appended by other `agent/step` listeners precede this checkpoint; prompt input and the preceding loop-owned assistant message and ordered results are already in the log.
|
||||
Persistence and checkpoint scheduling remain separate Cordis plugins. A backend makes requested `session/flush` boundaries durable but does not choose them; loading it without this policy is valid and retains the loop's coarser checkpoints. First-party persisted apps and runtimes explicitly mount both, while a specialized deployment may intentionally omit or replace the policy. Registration order governs whether events appended by other `agent/pre-step` listeners precede this checkpoint; prompt input and the preceding loop-owned assistant message and ordered results are already in the log.
|
||||
|
||||
Checkpoint failure and cancellation are fail-closed at effect boundaries. A rejected request checkpoint prevents adapter dispatch; a rejected tool checkpoint becomes an error result without invoking the tool body. If cancellation lands while the tool checkpoint is pending, the policy rechecks the signal and returns the canonical `ABORTED_BEFORE_DISPATCH` result. A rejected between-step checkpoint closes the turn before another model request. A rejected final turn checkpoint is reported live and does not prevent later queued work. Persistence serialization continues to belong to the coordinator, so concurrent tool checkpoints cannot duplicate event sequences.
|
||||
|
||||
|
||||
@@ -10,9 +10,9 @@ Status: implemented
|
||||
|
||||
## 决策
|
||||
|
||||
`dsh-session-checkpoint-policy` 以零配置插件的形式与持久化后端共同加载，并负责语义持久性屏障。在 `agent/step` 时，该插件会在推导下一个请求前刷新待持久化的提示词输入或前一批响应/结果。该插件惰性包装 `llm/stream`，在记录 `request/header` 之后、构造适配器流之前，刷新当前会话。该插件还在有序的执行前策略之后包装顶层 `tools/execute`，在进入工具主体前刷新已记录的 `tool/call`；嵌套分发则复用外层模型可见调用。循环的最终 `turn/end` 检查点仍是轮次的收尾边界，并会在处理另一个已排队轮次或观察到空闲状态之前完成。
|
||||
`dsh-session-checkpoint-policy` 以零配置插件的形式与持久化后端共同加载，并负责语义持久性屏障。在 `agent/pre-step` 时，该插件会在推导下一个请求前刷新待持久化的提示词输入或前一批响应/结果。该插件惰性包装 `llm/stream`，在记录 `request/header` 之后、构造适配器流之前，刷新活动会话。该插件还在有序的执行前策略之后包装顶层 `tools/execute`，在进入工具主体前刷新已记录的 `tool/call`；嵌套分发则复用外层模型可见调用。循环的最终 `turn/end` 检查点仍是轮次的收尾边界，并会在处理另一个已排队轮次或观察到空闲状态之前完成。
|
||||
|
||||
持久化与检查点调度仍是相互独立的 Cordis 插件。后端使请求的 `session/flush` 边界持久化，但不选择边界；只加载后端而不加载本策略仍是有效组合，并保留循环提供的较粗检查点。第一方持久化应用与运行时会显式加载两者，专用部署则可以有意省略或替换本策略。注册顺序决定其他 `agent/step` 监听器追加的事件是否先于本检查点；提示词输入以及前一批由循环自身记录的助手消息与有序结果都已在日志中。
|
||||
持久化与检查点调度仍是相互独立的 Cordis 插件。后端使请求的 `session/flush` 边界持久化，但不选择边界；只加载后端而不加载本策略仍是有效组合，并保留循环提供的较粗检查点。第一方持久化应用与运行时会显式加载两者，专用部署则可以有意省略或替换本策略。注册顺序决定其他 `agent/pre-step` 监听器追加的事件是否先于本检查点；提示词输入以及前一批由循环自身记录的助手消息与有序结果都已在日志中。
|
||||
|
||||
检查点失败与取消在副作用边界上采取失败关闭策略。请求检查点被拒绝时，系统不会分发给适配器；工具检查点被拒绝时，系统会返回错误结果，不调用工具主体。如果在工具检查点等待期间收到取消，策略会重新检查信号，并返回标准的 `ABORTED_BEFORE_DISPATCH` 结果。步骤间检查点被拒绝时，系统会在发起下一个模型请求前结束该轮次。轮次的最终检查点被拒绝时，系统会实时报告该失败，但不会阻止后续排队工作。持久化写入的串行化仍由协调器负责，因此并发的工具检查点不会产生重复的事件序列。
|
||||
|
||||
|
||||
@@ -2,5 +2,5 @@
|
||||
# side as of the last confirmed-consistent state. Both languages carry equal authority;
|
||||
# after editing either side, bring the other along and re-record with:
|
||||
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.md
|
||||
2026-07-29-human-transcript-append-origin.md: 4d8d66ea625d02f098da906ae58af1a2e5e41819
|
||||
2026-07-29-human-transcript-append-origin.zh.md: d4c1c4a83d868af5e27d7bda2a22827cf9d6c5b1
|
||||
2026-07-29-human-transcript-append-origin.md: 9804d3d8d67a4ad00c3d395c2081fd47e03c254a
|
||||
2026-07-29-human-transcript-append-origin.zh.md: 909f0756dd29c9deddb840a188eb19a773f92d29
|
||||
|
||||
@@ -6,7 +6,7 @@ English | [中文](2026-07-29-human-transcript-append-origin.zh.md)
|
||||
|
||||
## Problem
|
||||
|
||||
The terminal and the host history gateway both treated the model-visible surface as the human transcript. A successful compaction replaces a surface range with one checkpoint node, so the moment that replacement landed the terminal dropped every message it shadowed — conversation the user had already read — and re-ran that destructive rebuild on any later replacement. The same confusion reached pagination: `maxMessages` counted every `user/message`, `assistant/message`, and `steering/message` in the window, so a model-only replacement copy consumed a page slot the human never filled, and the cut could land between a compaction's log-only provenance and the replacement that cites it.
|
||||
The terminal and the host history gateway both treated the model-visible surface as the human transcript. A successful compaction replaces a surface range with one checkpoint node, so the moment that replacement landed the terminal dropped every message it shadowed — conversation the user had already read — and re-ran that destructive rebuild on any later replacement. The same confusion reached pagination: `maxMessages` counted every `user/message` and `assistant/message` in the window, so a model-only replacement copy consumed a page slot the human never filled, and the cut could land between a compaction's log-only provenance and the replacement that cites it.
|
||||
|
||||
Nothing was lost from the log. `Session.events` still held every original message and full tool result; the surface only decides what the model is sent next. The defect was entirely in the projection.
|
||||
|
||||
|
||||
@@ -6,7 +6,7 @@ Status: implemented
|
||||
|
||||
## Problem
|
||||
|
||||
终端与宿主历史网关都把模型可见的 surface 当作人类可读记录（transcript）。一次成功的压缩（compaction）会用一个检查点节点替换一段 surface 范围，因此该替换一落地，终端就丢弃了它所遮蔽的每条消息——那些是用户已经读过的对话——并在此后任何替换到来时重新执行这次破坏性重建。同样的混淆也波及分页：`maxMessages` 统计窗口内的每个 `user/message`、`assistant/message` 和 `steering/message`，于是仅供模型使用的替换副本占用了一个人类从未填充的页面额度，而切分点还可能落在压缩的仅日志溯源信息与引用它的替换之间。
|
||||
终端与宿主历史网关都把模型可见的 surface 当作人类可读记录（transcript）。一次成功的压缩（compaction）会用一个检查点节点替换一段 surface 范围，因此该替换一落地，终端就丢弃了它所遮蔽的每条消息——那些是用户已经读过的对话——并在此后任何替换到来时重新执行这次破坏性重建。同样的混淆也波及分页：`maxMessages` 统计窗口内的每个 `user/message` 和 `assistant/message`，于是仅供模型使用的替换副本占用了一个人类从未填充的页面额度，而切分点还可能落在压缩的仅日志溯源信息与引用它的替换之间。
|
||||
|
||||
日志本身没有丢失任何内容。`Session.events` 仍保存着每条原始消息和完整的工具结果；surface 只决定接下来发送给模型的内容。缺陷完全在投影层。
|
||||
|
||||
|
||||
@@ -0,0 +1,6 @@
|
||||
# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
|
||||
# side as of the last confirmed-consistent state. Both languages carry equal authority;
|
||||
# after editing either side, bring the other along and re-record with:
|
||||
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-08-04-load-pre-react-loop-sessions.md
|
||||
2026-08-04-load-pre-react-loop-sessions.md: b7dcad1ff7fe0aa8239ac03f52b50dfa55417515
|
||||
2026-08-04-load-pre-react-loop-sessions.zh.md: 14c6dcc7b4da016ebef424a66fae0b96fb469cd4
|
||||
@@ -0,0 +1,40 @@
|
||||
# Agent Note: Load sessions from the pre-react-loop format
|
||||
|
||||
Status: implemented
|
||||
|
||||
English | [中文](2026-08-04-load-pre-react-loop-sessions.zh.md)
|
||||
|
||||
## Problem
|
||||
|
||||
The react-loop simplification changed durable events while retaining `SESSION_FORMAT_VERSION` 0. Stored sessions from the change's base contain `steering/message` and `turn/start.trigger`; their terminal reasons also use coarse `aborted`, separate `disposed`, and two older error payloads. Current surface and turn invariants cannot replay those records directly.
|
||||
|
||||
The new durable inbox is not part of this compatibility problem. The base emitted process-local inbox notifications but no `agent/inbox/*` session events, so replaying old history as pending work would resurrect already claimed or discarded prompts.
|
||||
|
||||
## Decision
|
||||
|
||||
`PersistenceCoordinator` recognizes the exact pre-react-loop shapes after backend decoding and projects them into the current read view. It removes the obsolete `turn/start.trigger`, converts `steering/message` to the same identified `user/message`, maps old failure facts into the current structured error, folds `disposed` into an aborted turn with the `disposed` cause, and represents coarse aborted records with the persistence-only `{ kind: 'legacy' }` cause because their caller is unavailable.
|
||||
|
||||
The coordinator applies the projection to `load`, `inspect`, adoption, HMR prefix comparison, and `readFrom`. A seek-capable `readFrom` normally reads only its suffix; when that suffix contains a legacy event needing an earlier replacement identity, the coordinator loads and normalizes the complete prefix before returning the requested seq range.
|
||||
|
||||
The importer does not synthesize inbox splices. A resumed pre-react-loop agent begins with empty pending lists, matching the base runtime's inability to persist pending inbox work. The stored artifact remains append-only and later events use the current format.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
**Treat the same-version records as unsupported.** This follows the pre-release default but strands sessions produced by the PR base even though the removed steering content and terminal facts have complete mappings.
|
||||
|
||||
**Replay old inbox notifications into durable splices.** Those notifications were not session events and do not provide a trustworthy pending-state snapshot. Inferring insertions without every claim and discard would re-run consumed work.
|
||||
|
||||
**Assign coarse aborted records to an existing caller.** Mapping them to `user`, `parent`, or `hook` would manufacture provenance. A dedicated `legacy` cause keeps the stop classification without making a false audit claim.
|
||||
|
||||
**Rewrite stored JSONL and SQLite records.** A rewrite would violate the append-only contract and require backend-specific atomic migration machinery for a read compatibility boundary.
|
||||
|
||||
## Consequences
|
||||
|
||||
Sessions written in the refactor's base format resume through the current AgentLoop with their steering content, turn boundaries, error facts, and stop classification intact. The shared coordinator contract covers in-memory, JSONL, and SQLite `load`/`inspect`/`readFrom`, including the SQLite suffix fallback; an assembled JSONL Agent resume verifies that the historical transcript is visible while both new inbox lists start empty.
|
||||
|
||||
This exception supports the base format, not intermediate formats produced during development of the refactor. In particular, it defines no migration for earlier experimental `agent/inbox/spliced` payloads. Exact-shape recognition keeps malformed current-looking records on their rejection path instead of guessing them into validity.
|
||||
|
||||
## Related
|
||||
|
||||
- [Load sessions persisted before message identity](2026-07-28-load-pre-identity-session-messages.md) — owns deterministic identities and the general read-only import boundary for another same-version format change.
|
||||
- [Session persistence as an abstract service](../architecture/2026-06-14-session-persistence.md) — owns append-only backend storage and resume.
|
||||
@@ -0,0 +1,40 @@
|
||||
# Agent Note: 加载 react-loop 重构前格式的会话
|
||||
|
||||
Status: implemented
|
||||
|
||||
[English](2026-08-04-load-pre-react-loop-sessions.md) | 中文
|
||||
|
||||
## 问题
|
||||
|
||||
react-loop 简化在保持 `SESSION_FORMAT_VERSION` 为 0 的同时更改了持久事件。该变更基线所存储的会话包含 steering（中途引导）事件 `steering/message` 和 `turn/start.trigger`；其终止原因还使用粗粒度 `aborted`、独立的 `disposed` 和两种旧版错误载荷。当前 surface 和轮次不变量无法直接回放这些记录。
|
||||
|
||||
新的持久 inbox 不属于此兼容性问题。该基线会发出进程本地 inbox 通知，但不会产生 `agent/inbox/*` 会话事件，因此将旧历史回放为待处理工作会让已经领取或丢弃的提示词再次执行。
|
||||
|
||||
## 决策
|
||||
|
||||
`PersistenceCoordinator` 会在后端解码后识别 react-loop 重构前的确切形状，并将其投影为当前读取视图。它移除已废弃的 `turn/start.trigger`，把 `steering/message` 转换为同一条带标识的 `user/message`，将旧版失败事实映射为当前结构化错误，把 `disposed` 折叠为带 `disposed` 原因的已中止轮次，并用仅供持久化导入使用的 `{ kind: 'legacy' }` 原因表示粗粒度中止记录，因为无法获得其调用方。
|
||||
|
||||
协调器会把该投影应用于 `load`、`inspect`、接管、HMR 前缀比较和 `readFrom`。可寻址的 `readFrom` 通常只读取后缀；如果后缀包含需要更早替换标识的旧版事件，协调器会先加载并规范化完整前缀，再返回所请求的 seq 范围。
|
||||
|
||||
导入器不会合成 inbox splice。恢复后的 react-loop 重构前 agent 从空的待处理列表开始，这与基线运行时无法持久化待处理 inbox 工作的行为一致。已存储产物仍然仅追加，后续事件使用当前格式。
|
||||
|
||||
## 考虑过的替代方案
|
||||
|
||||
**将同版本记录视为不受支持。** 这符合预发布阶段的默认立场，但会使 PR 基线产生的会话无法恢复，尽管已移除的 steering 内容和终止事实都有完整映射。
|
||||
|
||||
**将旧 inbox 通知回放为持久 splice。** 这些通知不是会话事件，也无法提供可信的待处理状态快照。如果无法获知每一次领取和丢弃，就推断插入操作，会让已消费的工作再次执行。
|
||||
|
||||
**将粗粒度中止记录归因于现有调用方。** 将其映射到 `user`、`parent` 或 `hook` 会虚构来源。专用的 `legacy` 原因既能保留停止分类，也不会产生虚假的审计事实。
|
||||
|
||||
**重写已存储的 JSONL 和 SQLite 记录。** 重写会违反仅追加契约，并要求为读取兼容边界建立后端专用的原子迁移机制。
|
||||
|
||||
## 后果
|
||||
|
||||
以重构基线格式写入的会话可以通过当前 AgentLoop 恢复，并完整保留 steering 内容、轮次边界、错误事实和停止分类。共享协调器契约覆盖内存、JSONL 和 SQLite 的 `load`／`inspect`／`readFrom`，包括 SQLite 后缀回退；组装后的 JSONL agent 恢复用例会验证历史 transcript（文本记录）可见，同时两个新 inbox 列表都从空状态开始。
|
||||
|
||||
此例外支持基线格式，不支持重构开发期间产生的中间格式。具体而言，它没有为更早的实验性 `agent/inbox/spliced` 载荷定义迁移。通过确切形状识别，当前格式外观相似但结构错误的记录仍会走拒绝路径，不会被猜测性地转换为有效记录。
|
||||
|
||||
## 相关资料
|
||||
|
||||
- [加载消息标识机制引入前持久化的会话](2026-07-28-load-pre-identity-session-messages.md)：负责另一项同版本格式变更的确定性标识和通用只读导入边界。
|
||||
- [以抽象服务实现会话持久化](../architecture/2026-06-14-session-persistence.md)：负责仅追加后端存储和恢复。
|
||||
@@ -0,0 +1,6 @@
|
||||
# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
|
||||
# side as of the last confirmed-consistent state. Both languages carry equal authority;
|
||||
# after editing either side, bring the other along and re-record with:
|
||||
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-08-05-context-meter-blind-to-compaction.md
|
||||
2026-08-05-context-meter-blind-to-compaction.md: ab39ae4e109f238960fd60de5e5b61075344f525
|
||||
2026-08-05-context-meter-blind-to-compaction.zh.md: c93aa509530e48acf906b18ba85bab4c7d355d69
|
||||
@@ -0,0 +1,46 @@
|
||||
# Agent Note: the context meter could not see a compaction
|
||||
|
||||
Status: implemented
|
||||
|
||||
English | [中文](2026-08-05-context-meter-blind-to-compaction.zh.md)
|
||||
|
||||
## Problem
|
||||
|
||||
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: `compact-basic` summarizes through a direct `ctx.llm.stream()` call and appends `compact/start`, `compact/summary`, the replacement `user/message`, and `compact/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](../feature/2026-08-05-composer-context-meter-breakdown.md). 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 provenance (`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.spec.tsx` pins the ring reading the projected figure, and `chat-stats-bash-sample.spec.tsx` pins `contextOccupancy`'s preference and its fallback. The end-to-end numbers above came from driving `BasicCompactService.compactNow` through a real `AgentLoop` with the projection registry mounted.
|
||||
@@ -0,0 +1,46 @@
|
||||
# Agent Note：上下文仪表看不见压缩
|
||||
|
||||
Status: implemented
|
||||
|
||||
[English](2026-08-05-context-meter-blind-to-compaction.md) | 中文
|
||||
|
||||
## 问题
|
||||
|
||||
composer 的[上下文仪表](../feature/2026-08-05-composer-context-meter-breakdown.md)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，而压缩不报告用量：`compact-basic` 通过直连的 `ctx.llm.stream()` 调用生成摘要，只追加 `compact/start`、`compact/summary`、用作替换的 `user/message` 和 `compact/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 agent loop 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轮才会动。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](../feature/2026-08-05-composer-context-meter-breakdown.md)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` 对分片来源的随机访问（`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，而不再是每轮跳一次，因此一轮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。循环在 `buildRequest` 之前接纳 steering 与 context，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` 覆盖了样本在表层增长与一次压缩上的推进（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.spec.tsx` 钉住圆环读取投影值这一点，`chat-stats-bash-sample.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 `BasicCompactService.compactNow`。
|
||||
Reference in New Issue
Block a user