fix(compact): harden pruning integration (round 2)
This commit is contained in:
Hypatia May
@@ -107,11 +107,11 @@ Each step renders one prompt assembly. Plugins contribute ordered sections, tool
|
||||
|
||||
Post-tool context follows all results, preserving call/result adjacency. Steering drains before `agent/post-step`, which observes durable output, results, context, and steering while the step signal remains open. Leftover steering becomes next-turn input. `agent/turn-stop` is terminal through close and flush: later steering is discarded, while ordinary queued prompts survive.
|
||||
|
||||
When loaded, `dsh-compact-basic` consumes that post-step checkpoint for `ctx.tokenMeter` pressure under the actual routed header. Once pressure or canonical context overflow qualifies, it runs optional `ctx.toolResultPrune` rewriting before summary selection and remeasures the replayed surface. Overflow recovery authorizes retry after either pruning or tool-balanced summary compaction advances `surface.replaceGeneration`. The same turn signal owns both paths.
|
||||
When loaded, `dsh-compact-basic` consumes that post-step checkpoint for `ctx.tokenMeter` pressure under the actual routed header. Once pressure or canonical context overflow qualifies, it runs optional `ctx.toolResultPrune` rewriting before summary selection and remeasures the replayed surface. Overflow recovery authorizes retry after either pruning or tool-balanced summary compaction advances `surface.replaceGeneration`, including when later summary work fails after a prune. The same turn signal owns both paths, and cancellation still wins.
|
||||
|
||||
### Failure Boundaries
|
||||
|
||||
The turn is the containment boundary. `LlmService` preserves and privately tags errors from final adapter selection, dispatch, and iteration. Those errors and terminal in-band error/aborted finishes close the failed step before `agent/request-error`; retry reconstructs the next numbered step from the log, while decline or failed recovery preserves the provider error. Attempts count consecutive failures and reset after success.
|
||||
The turn is the containment boundary. `LlmService` preserves and privately tags errors from final adapter selection, dispatch, and iteration. Those errors and terminal in-band error/aborted finishes close the failed step before `agent/request-error`; retry reconstructs the next numbered step from the log, while decline or recovery failure before any replacement preserves the provider error. Attempts count consecutive failures and reset after success.
|
||||
|
||||
Prompt, middleware, result, tool, post-step, and continuation failures remain ordinary `agent/error` failures. Cancellation and disposal beat recovery. Durable undispatched tool calls receive synthetic `ABORTED` results, preventing dangling replay. `cancel()` clears queues and aborts active work; disposal awaits quiescence before unregistering.
|
||||
|
||||
|
||||
@@ -58,6 +58,6 @@ export type CompactionTrigger = 'pressure' | 'context-overflow'
|
||||
|
||||
`CompactService` exposes `compactIfNeeded(agent, trigger, signal)` for automatic `pressure` or `context-overflow` policy, returning `null` when no safe work exists, and `compactRegion(...)` for an explicit inclusive surface range. Implementations must forward the supplied signal to summarization. The seam owns no pricing API: the singleton [`ctx.tokenMeter`](token-meter.md) directly owns estimation and replay, while `dsh-compact-basic` owns retention, event sequencing, routed summarization calls, and their configuration.
|
||||
|
||||
Pressure compaction runs at serial `agent/post-step`, after successful assistant output, tool results, buffered context, and steering are durable but before `step/end`. Once pressure or canonical overflow qualifies, compact-basic invokes optional [`ctx.toolResultPrune`](../../packages/compact/tool-result-prune/README.md) before range selection, remeasures through `ctx.tokenMeter`, and can advance the surface without a summary. Failed-request recovery runs through `agent/request-error` after the failed step closes and authorizes a fresh numbered-step retry only when the surface replacement generation advances. Region boundaries preserve tool-call/result pairing but not whole turns, allowing early closed steps of one oversized turn to compact. `dsh-compact-basic` owns thresholds, retained-tail policy, overflow caps, and failure handling.
|
||||
Pressure compaction runs at serial `agent/post-step`, after successful assistant output, tool results, buffered context, and steering are durable but before `step/end`. Once pressure or canonical overflow qualifies, compact-basic invokes optional [`ctx.toolResultPrune`](../../packages/compact/tool-result-prune/README.md) before range selection, remeasures through `ctx.tokenMeter`, and can advance the surface without a summary. Failed-request recovery runs through `agent/request-error` after the failed step closes and authorizes a fresh numbered-step retry only when the surface replacement generation advances, even if later summary work throws after pruning; cancellation still wins. Region boundaries preserve tool-call/result pairing but not whole turns, allowing early closed steps of one oversized turn to compact. `dsh-compact-basic` owns thresholds, retained-tail policy, overflow caps, and failure handling.
|
||||
|
||||
The seam exports `toolPairingBalancedBefore(session, node)` and `toolPairingBalancedAfter(session, node)` for those edge checks. Both validate current surface membership, reject stale or missing seqs and orphan results, and ignore a caller-retained `node.next`; the [package contract](../../packages/compact/compact/README.md#tool-pairing-boundaries) owns their cache semantics.
|
||||
|
||||
@@ -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
|
||||
2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md: 99dc7625b8e185464d7a8a1ea8eda5baf0674df7
|
||||
2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md: 4f84d4435341005c058e32582e6d26b2a9f29bc1
|
||||
2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md: deedb81f8cf75ab80b70e2ef3148ba76d1278886
|
||||
2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md: 51d28fa243f522eedcac29002670575889d48369
|
||||
|
||||
@@ -34,9 +34,9 @@ If cancellation lands after assistant tool calls are durable but before all call
|
||||
|
||||
For `pressure`, compact-basic applies the service-wide threshold and retained-tail policy to one unified `ctx.tokenMeter.measure()` result. Below pressure it returns without pruning. Once pressure qualifies, optional `ctx.toolResultPrune` rewrites oversized current results and compact-basic remeasures through the same meter; safe pressure skips the model call, while remaining pressure selects and summarizes from the pruned surface. The same singleton meter owns range pricing, provenance, shadowed token counts, and non-shrinking-summary rejection. The common defaults remain threshold ratio `0.8`, retained history `floor(contextWindow × 0.16)`, summarization model `''`, `maxTokens: 8192`, `compactionRetries: 1`, and `auto: true`.
|
||||
|
||||
For canonical overflow, compact-basic bypasses scalar pressure and the normal retained-token budget. It prunes first, then chooses the maximal tool-balanced head range while leaving the newest indivisible unit and attempts one shrinking summary compaction under the same signal when a range exists. The automatic listener snapshots `session.surface.replaceGeneration` and returns `{ action: 'retry' }` whenever pruning or summarization increases it. A backend returning a result without replacement cannot authorize retry, while pruning-only progress can authorize a retry without a `CompactionResult`.
|
||||
For canonical overflow, compact-basic bypasses scalar pressure and the normal retained-token budget. It prunes first, then chooses the maximal tool-balanced head range while leaving the newest indivisible unit and attempts one shrinking summary compaction under the same signal when a range exists. The automatic listener snapshots `session.surface.replaceGeneration` and returns `{ action: 'retry' }` whenever pruning or summarization increases it. This remains true when pruning lands before later summary work throws; cancellation still wins. A backend returning a result without replacement cannot authorize retry, while pruning-only progress can authorize a retry without a `CompactionResult`.
|
||||
|
||||
`maxOverflowRetries` is optional and defaults to `1`; `0` disables overflow recovery without disabling pressure. `auto: false` registers neither automatic listener. Noncanonical errors, exhausted attempts, an already-aborted signal, a missing routed model, no safe range, no generation change, and recovery throws all delegate to the next listener. With no later recovery, the loop reports the original provider error object and code. Cancellation or disposal remains authoritative even if recovery work completes concurrently.
|
||||
`maxOverflowRetries` is optional and defaults to `1`; `0` disables overflow recovery without disabling pressure. `auto: false` registers neither automatic listener. Noncanonical errors, exhausted attempts, an already-aborted signal, a missing routed model, no safe range, no generation change, and recovery throws before any replacement all delegate to the next listener. With no later recovery, the loop reports the original provider error object and code. A recovery throw after generation advances authorizes retry from durable progress; cancellation or disposal remains authoritative even if recovery work completes concurrently.
|
||||
|
||||
The default summarizer still resolves explicit configuration, then the latest logged route, then agent options. Because direct `llm/stream` middleware may reroute that auxiliary call, `compact/summary.model` records the final mutable `GenerateOptions.model` observed after dispatch rather than the pre-waterfall candidate.
|
||||
|
||||
@@ -44,7 +44,7 @@ The default summarizer still resolves explicit configuration, then the latest lo
|
||||
|
||||
Lifecycle tests pin post-step ordering after durable tool/context/steering work, content-less and max-token successes, final-adapter dispatch/iterator/in-band boundaries, retry numbering, attempt reset, cancellation, disposal, synthetic tool results, and original error identity.
|
||||
|
||||
Compact tests pin low-friction service-wide defaults, actual routed-model selection, unlisted-model measurement, unified pressure-and-retention decisions, pressure-gated pruning, pruning-only relief, summarization from pruned input, optional-plugin fallback, pruning-only and summarized overflow recovery, newest tool-pair retention, non-shrinking rejection, generation proof, caps, disabled listeners, single downstream delegation, and auxiliary summary routing provenance. Real-loop composition covers both thrown and in-band overflow: the failed step closes, compaction lands between attempts, and the next numbered request is reconstructed from the replacement surface.
|
||||
Compact tests pin low-friction service-wide defaults, actual routed-model selection, unlisted-model measurement, unified pressure-and-retention decisions, pressure-gated pruning, pruning-only relief, summarization from pruned input, optional-plugin fallback, pruning-only and summarized overflow recovery, prune-then-summary-failure progress, newest tool-pair retention, non-shrinking rejection, generation proof, caps, disabled listeners, single downstream delegation, and auxiliary summary routing provenance. Real-loop composition covers both thrown and in-band overflow: the failed step closes, compaction lands between attempts, and the next numbered request is reconstructed from the replacement surface.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
|
||||
@@ -34,9 +34,9 @@ Status: implemented
|
||||
|
||||
对于 `pressure`，compact-basic 把服务级阈值与保留尾部策略应用到一次统一的 `ctx.tokenMeter.measure()` 结果。低于压力时直接返回，不执行剪枝。压力达到条件后，可选的 `ctx.toolResultPrune` 会改写当前表层中过大的工具结果，compact-basic 再通过同一个 meter 重新计量；若压力恢复安全则跳过模型调用，否则从已剪枝表层选择范围并生成摘要。范围定价、来源、被遮蔽 token 数与非缩小摘要拒绝也由同一个单例 meter 完成。通用默认值保持为阈值比例 `0.8`、保留历史 `floor(contextWindow × 0.16)`、摘要模型 `''`、`maxTokens: 8192`、`compactionRetries: 1` 与 `auto: true`。
|
||||
|
||||
对于规范化溢出，compact-basic 绕过标量压力与普通保留 token 预算。它先执行剪枝，再在保留最新不可分割单元的同时选择最大的工具配对平衡头部范围；存在范围时，才在同一 signal 下尝试一次缩小摘要压缩。自动监听器先记录 `session.surface.replaceGeneration`，剪枝或摘要让 generation 增加时就返回 `{ action: 'retry' }`。后端若只返回结果但没有替换表层，不能授权重试；只有剪枝取得进展时，即使没有 `CompactionResult` 也可以授权重试。
|
||||
对于规范化溢出，compact-basic 绕过标量压力与普通保留 token 预算。它先执行剪枝，再在保留最新不可分割单元的同时选择最大的工具配对平衡头部范围；存在范围时，才在同一 signal 下尝试一次缩小摘要压缩。自动监听器先记录 `session.surface.replaceGeneration`，剪枝或摘要让 generation 增加时就返回 `{ action: 'retry' }`。即使剪枝先落盘而后续摘要工作抛错，这条规则仍然成立；取消依然优先。后端若只返回结果但没有替换表层，不能授权重试；只有剪枝取得进展时，即使没有 `CompactionResult` 也可以授权重试。
|
||||
|
||||
`maxOverflowRetries` 可选且默认为 `1`；`0` 只禁用溢出恢复，不会禁用压力检查。`auto: false` 不注册任何自动监听器。非规范化错误、尝试耗尽、已经中止的 signal、缺失路由模型、没有安全范围、generation 未变化，以及恢复抛错都会委托给下一个监听器。若没有后续恢复，循环报告原始提供方错误对象与代码。即使恢复工作并发完成，取消或销毁仍具有最终优先级。
|
||||
`maxOverflowRetries` 可选且默认为 `1`；`0` 只禁用溢出恢复，不会禁用压力检查。`auto: false` 不注册任何自动监听器。非规范化错误、尝试耗尽、已经中止的 signal、缺失路由模型、没有安全范围、generation 未变化，以及在任何替换之前恢复抛错，都会委托给下一个监听器。若没有后续恢复，循环报告原始提供方错误对象与代码。generation 增加后的恢复抛错会基于持久进展授权重试；即使恢复工作并发完成，取消或销毁仍具有最终优先级。
|
||||
|
||||
默认摘要器仍依次解析显式配置、最近记录的路由与 agent options。因为直接 `llm/stream` 中间件可以重新路由该辅助调用，`compact/summary.model` 记录分发后最终可变的 `GenerateOptions.model`，而不是 waterfall 之前的候选值。
|
||||
|
||||
@@ -44,7 +44,7 @@ Status: implemented
|
||||
|
||||
生命周期测试固定 post-step 位于持久工具、上下文与 steering 工作之后，覆盖无内容与达到 token 上限的成功、最终适配器分发/迭代器/带内边界、重试编号、尝试重置、取消、销毁、合成工具结果与原始错误身份。
|
||||
|
||||
压缩测试固定低摩擦服务级默认值、实际路由模型选择、未列出模型计量、统一压力与保留决策、压力门控剪枝、剪枝独立解除压力、从已剪枝输入生成摘要、可选插件回退、仅剪枝与剪枝后摘要两类溢出恢复、最新工具配对保留、非缩小拒绝、generation 证明、上限、禁用监听器、单次下游委托与辅助摘要路由来源。真实循环组合同时覆盖抛出式和带内溢出：失败 step 关闭，压缩落在两次尝试之间，下一个编号请求从替换表层重建。
|
||||
压缩测试固定低摩擦服务级默认值、实际路由模型选择、未列出模型计量、统一压力与保留决策、压力门控剪枝、剪枝独立解除压力、从已剪枝输入生成摘要、可选插件回退、仅剪枝与剪枝后摘要两类溢出恢复、剪枝后摘要失败的持久进展、最新工具配对保留、非缩小拒绝、generation 证明、上限、禁用监听器、单次下游委托与辅助摘要路由来源。真实循环组合同时覆盖抛出式和带内溢出：失败 step 关闭，压缩落在两次尝试之间，下一个编号请求从替换表层重建。
|
||||
|
||||
## 考虑过的替代方案
|
||||
|
||||
|
||||
Reference in New Issue
Block a user