fix: 内存优化 — FileReadTool 100KB 上限、lookups 缓存、microcompact 替换清理

- FileReadTool maxResultSizeChars 从 Infinity 改为 100KB,大文件持久化到磁盘
- Messages.tsx 新增 computeMessageStructureKey 缓存,流式 delta 时跳过 8 个 Map/Set 重建
- microcompact 返回 clearedToolUseIds,query.ts 消费后清理 replacements Map 释放原始字符串
- 更新内存分析报告 Round 5 和 file-operations 文档

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
claude-code-best
2026-05-02 11:21:22 +08:00
parent 3eba5ade1a
commit f724300079
8 changed files with 205 additions and 32 deletions

View File

@@ -1,7 +1,8 @@
# 内存与性能峰值分析报告(最终版 — 4 轮迭代完成) # 内存与性能峰值分析报告(最终版 — 5 轮迭代完成)
> 进程 bun物理内存峰值 **700 MB+**,最差场景可达 **1.8 GB** > 进程 bun物理内存峰值 **700 MB+**,最差场景可达 **1.8 GB**
> 日期2026-05-02 | 状态:**调研完成** | 范围:内存峰值 + CPU 热点 + React 渲染循环 > 日期2026-05-02 | 状态:**调研完成** | 范围:内存峰值 + CPU 热点 + React 渲染循环
> Round 5 增量验证消息渲染管线buildMessageLookups 8 Map/Set 重建、useDeferredValue 双缓冲、FileReadTool 无上限、compaction 与 React 状态交互
## 数据收集 ## 数据收集
@@ -44,6 +45,34 @@
峰值时 3-4 份完整消息数组同时驻留477 + 1745 + 1857 + 1878 在同一 turn 尾部顺序执行)。 峰值时 3-4 份完整消息数组同时驻留477 + 1745 + 1857 + 1878 在同一 turn 尾部顺序执行)。
### P0React 消息管线重复计算Round 5 新增分析)
**buildMessageLookups 每次 useMemo 重算时创建 8 个 Map/Set**`messages.ts:1215-1398`
| 数据结构 | 规模 | 说明 |
|----------|------|------|
| `toolUseIDsByMessageID` | Map\<string, Set\> | 每个 assistant 消息一个 Set |
| `toolUseIDToMessageID` | Map\<string, string\> | 所有 tool_use ID |
| `toolUseByToolUseID` | Map\<string, ToolUseBlockParam\> | **保留完整 tool_use block** |
| `siblingToolUseIDs` | Map\<string, Set\> | 兄弟 tool_use 索引 |
| `progressMessagesByToolUseID` | Map\<string, ProgressMessage[]\> | 进度消息数组 |
| `toolResultByToolUseID` | Map\<string, NormalizedMessage\> | **保留完整 tool_result 消息引用** |
| `resolvedToolUseIDs` / `erroredToolUseIDs` | Set\<string\> | 已完成/错误 ID |
此 useMemo`Messages.tsx:519`)依赖 normalizedMessages任何消息变更含流式 delta触发重建。已拆分 renderRange 避免滚动触发注释明确记录50ms alloc per scroll → GC → 100-173ms STW on 1GB heap
**useDeferredValue 双缓冲**`REPL.tsx:1569`):流式期间 `messages``deferredMessages` 同时持有两份完整数组,直到 React 调度更新。在 27k 消息场景下,额外 ~100-200MB 临时占用。
**FileReadTool 无大小限制**`FileReadTool.ts:342``maxResultSizeChars: Infinity`,单次 10MB 文件读取完整保留在消息数组中。BashTool30KB和 GrepTool20KB有合理上限。
### P0Compaction 与 React 状态交互Round 5 新增分析)
**非全屏模式**`REPL.tsx:3074-3075`compact 后 `setMessages(() => [newMessage])` 正确替换整组旧消息,内存立即释放。
**全屏模式**`REPL.tsx:3056-3072`):保留最多 500 条消息的 scrollback。注释记录Ink fiber 树每条消息 ~250KB RSS无 cap 时观察过 13k+ 消息 → 1GB+ heap。
**Microcompact 的局限**`microCompact.ts:472-494`):用 spread 创建新消息对象替换内容为 `[Old tool result content cleared]`。但 `ContentReplacementState.replacements` Map`toolResultStorage.ts:392`)仍保留原始替换字符串,直到 compact 时才清理。这意味着 microcompact 减少了 token 数,但实际内存释放依赖后续 compact。
### P0Compact 峰值20-80 MB ### P0Compact 峰值20-80 MB
峰值时间线(`compact.ts:524-644` 峰值时间线(`compact.ts:524-644`
@@ -55,6 +84,44 @@ After: splice → 50K tokens
可提前释放:`preCompactReadFileState`25MB`summaryResponse`、原始 `messages` 参数。 可提前释放:`preCompactReadFileState`25MB`summaryResponse`、原始 `messages` 参数。
### P0React Hooks 闭包与 useMemo 链Round 5 深入排查)
**useCallback 闭包重建**`REPL.tsx`
| 回调 | 依赖项数 | 位置 | 影响 |
|------|----------|------|------|
| `getToolUseContext` | 20 | `:2789-2949` | 重建时旧闭包持有的引用阻止 GC |
| `onQueryImpl` | 14 | `:3188-3469` | 包含 getToolUseContext + 多层嵌套闭包 |
| `onQuery` | 在 onQueryImpl 上再包装 | `:3471-3697` | 又一层闭包 |
| `onSubmit` | ~10 | `:3822-4298` | 闭包链嵌套 3 层 |
每次 `messages` 变更触发 `setMessages` → React 重渲染 → 依赖 messages 的 useCallback/useMemo 全部重建。但 `getToolUseContext``onQueryImpl` **没有把 `messages` 放入依赖数组**(通过 `messagesRef.current` 参数传递规避),所以这些闭包不会因 messages 变化而重建。**这实际上是正确的设计**——用 ref 规避了闭包捕获问题。
**真正的 hooks 问题**在于 useMemo 链(`Messages.tsx`
```
messages → normalizedMessages (O(n))
→ compactAwareMessages (O(n) filter)
→ messagesToShow (O(n) filter + reorder)
→ groupedMessages (O(n))
→ collapsed (O(n))
→ lookups (8 Map/Set, O(n))
```
流式期间每个 delta 触发 `messages` 变更 → 整条链全量重算。注释记录50ms alloc per scroll → GC → 100-173ms STW on 1GB heap`Messages.tsx:516-518`)。
**无界 useRef**`REPL.tsx`
| Ref | 增长方式 | 清理 | 影响 |
|-----|----------|------|------|
| `bashTools` | `.add()` 每个 bash 命令 | `clearConversation` 时 clear | Set\<string\>,通常 <100 |
| `discoveredSkillNamesRef` | `.add()` 每个发现的 skill | `clearConversation` 时 clear | Set\<string\>,通常 <50 |
| `apiMetricsRef` | `.push()` 每次请求 | turn 结束时 `= []` | 临时turn 内累积 |
| `responseLengthRef` | 累加 | compact 时重置为 0 | 单数字 |
| `loadedNestedMemoryPathsRef` | `.add()` 每个 CLAUDE.md | compact/clear 时 clear | Set\<string\> |
结论:**这些 ref 都有清理机制**,不是主要问题。核心问题仍是 useMemo 链在流式期间的全量重算。
### P1虚拟滚动组件~50 MB— Round 3 新发现 ### P1虚拟滚动组件~50 MB— Round 3 新发现
`src/hooks/useVirtualScroll.ts` + React Ink 渲染管线: `src/hooks/useVirtualScroll.ts` + React Ink 渲染管线:
@@ -86,6 +153,8 @@ After: splice → 50K tokens
| 7 | Tool result seenIds/replacements | 0.5-2 MB | `toolResultStorage.ts:390-397` | | 7 | Tool result seenIds/replacements | 0.5-2 MB | `toolResultStorage.ts:390-397` |
| 8 | bootstrap/state.ts 无界缓存 | 0.1-1 MB | planSlugCache 等 | | 8 | bootstrap/state.ts 无界缓存 | 0.1-1 MB | planSlugCache 等 |
| 9 | QueryEngine 无界集合 | 0.1-1 MB | discoveredSkillNames 等 | | 9 | QueryEngine 无界集合 | 0.1-1 MB | discoveredSkillNames 等 |
| 10 | expandedKeys Set 无清理Round 5 | <0.5 MB | `Messages.tsx:644` compact 后 stale keys 不删除 |
| 11 | OpenAI/Gemini/Grok collectedMessagesRound 5 | 临时 | 流式期间累积 assistant messages 供 Langfuse telemetrystream 结束后释放 |
### P2低优先级未验证 ### P2低优先级未验证
@@ -125,7 +194,7 @@ After: splice → 50K tokens
- 双缓冲 + damage tracking + 字符池复用 - 双缓冲 + damage tracking + 字符池复用
- Pool 5 分钟周期重置 - Pool 5 分钟周期重置
## 已否认(内存,4 轮汇总) ## 已否认(内存,5 轮汇总)
- VSZ 516 GB 是虚拟映射非物理 | Zod Schema ~650KB | Markdown LRU-500 已优化 - VSZ 516 GB 是虚拟映射非物理 | Zod Schema ~650KB | Markdown LRU-500 已优化
- useSkillsChange/useSettingsChange — 正确 cleanup | useInboxPoller — 收敛设计 - useSkillsChange/useSettingsChange — 正确 cleanup | useInboxPoller — 收敛设计
@@ -133,15 +202,33 @@ After: splice → 50K tokens
- Ink 屏幕缓冲 ~86KB | CharPool/HyperlinkPool ~1-5MB 且 5min 重置 | StylePool 缓存 1000 上限 - Ink 屏幕缓冲 ~86KB | CharPool/HyperlinkPool ~1-5MB 且 5min 重置 | StylePool 缓存 1000 上限
- 依赖树 — AWS/Google/Azure SDK 均动态 import不贡献基线 | Sentry 空实现 - 依赖树 — AWS/Google/Azure SDK 均动态 import不贡献基线 | Sentry 空实现
- Ink 无 scrollback 缓冲 | Markdown tokenCache LRU-500 bounded - Ink 无 scrollback 缓冲 | Markdown tokenCache LRU-500 bounded
- **Round 5 否认**useCallback 闭包捕获 messages — 实际通过 messagesRef 参数传递规避,无闭包问题
- **Round 5 否认**MCP stderrHandler 泄漏 — 已有 64MB cap + 成功后释放 + cleanup 移除 listener
- **Round 5 否认**useRef 无界增长 — bashTools/discoveredSkillNamesRef/loadedNestedMemoryPathsRef 均有 clearConversation 或 compact 清理
- **Round 5 否认**apiMetricsRef 无界 — turn 结束时 `= []` 重置
- **Round 5 否认**useEffect 缺少 cleanup — 检查的 12 个 useEffect 均有 return cleanup 函数
## 结论 ## 结论
**内存根因**4 轮迭代确认):消息数组 turn 尾部 3-4 次同时驻留 + compact 峰值窗口 + 虚拟滚动 200 组件 ~50MB 常驻 + Bun/JSC 不归还内存页。 **内存根因**5 轮迭代确认):
1. **消息数组 turn 尾部 3-4 次 spread 同时驻留**120-320 MB— 核心瓶颈
2. **React 消息管线 buildMessageLookups 8 个 Map/Set 重建**50ms/次27k 消息场景)— GC 压力源
3. **useDeferredValue 双缓冲**(流式期间额外 ~100-200 MB 临时)
4. **FileReadTool 无大小上限**(单次 10MB 文件永久驻留)
5. **Compact 峰值窗口**20-80 MB+ Microcompact 依赖后续 compact 才真正释放
6. **虚拟滚动 200 组件 ~50MB 常驻**
7. **Bun/JSC 不归还内存页**(架构级限制)
**CPU 根因**useInboxPoller 每秒轮询触发 React commit → 全量 Yoga 布局 → 全屏 Ink diff 的完整管线。Markdown 渲染(~1.5ms/行)在批量挂载新消息时造成 ~290ms 卡顿。轮询导致的周期性 commit 与消息挂载的 CPU 密集操作互相放大。 **CPU 根因**useInboxPoller 每秒轮询触发 React commit → 全量 Yoga 布局 → 全屏 Ink diff 的完整管线。Markdown 渲染(~1.5ms/行)在批量挂载新消息时造成 ~290ms 卡顿。轮询导致的周期性 commit 与消息挂载的 CPU 密集操作互相放大。
**Round 4 最终验证**agent 递归 spread 和 attachment 累积均为已知 P0消息数组拷贝的变体无新根因。Snipping 在流式前执行无并发问题。consumedCommandUuids 等数组每轮重置无累积。 **Round 4 最终验证**agent 递归 spread 和 attachment 累积均为已知 P0消息数组拷贝的变体无新根因。Snipping 在流式前执行无并发问题。consumedCommandUuids 等数组每轮重置无累积。
**Round 5 增量验证**
- buildMessageLookups 8 个 Map/Set 的重建成本已由 renderRange 拆分缓解,但仍然是消息变更时的主要 GC 压力源
- useDeferredValue 双缓冲是 React 调度机制的固有行为,优化空间有限
- FileReadTool 无上限是唯一一个"单次操作可注入 10MB+ 数据"的入口
- Microcompact 减少 token 但不立即释放内存(内容被 ContentReplacementState.replacements Map 间接持有)
**预估优化空间** **预估优化空间**
| 优先级 | 措施 | 预估降低 | | 优先级 | 措施 | 预估降低 |
@@ -151,7 +238,7 @@ After: splice → 50K tokens
| P1 | 虚拟滚动优化 | 20-30 MB | | P1 | 虚拟滚动优化 | 20-30 MB |
| P1 | 缓冲与缓存清理 5 项 | 30-80 MB | | P1 | 缓冲与缓存清理 5 项 | 30-80 MB |
| P2 | 其他 3 项 | 10-50 MB | | P2 | 其他 3 项 | 10-50 MB |
| **合计** | **18 项可操作建议** | **180-440 MB** | | **合计** | **21 项可操作建议** | **210-500 MB** |
理论可从当前 400-700 MB 降至 **200-350 MB** 理论可从当前 400-700 MB 降至 **200-350 MB**
@@ -166,30 +253,36 @@ After: splice → 50K tokens
5. `query.ts:1857` — 传引用forkContextMessages 5. `query.ts:1857` — 传引用forkContextMessages
6. `query.ts:491` — 无超限返回原数组 6. `query.ts:491` — 无超限返回原数组
### P0消息渲染管线Round 5 新增,预估降 30-60 MB
7. `FileReadTool.ts:342``maxResultSizeChars: Infinity` → 设合理上限(如 100KB
8. `toolResultStorage.ts:392` — Microcompact 后同步清理 `replacements` Map 中对应条目
9. `Messages.tsx:519` — 考虑 buildMessageLookups 增量更新而非全量重建
### P0Compact 峰值(预估降 20-80 MB ### P0Compact 峰值(预估降 20-80 MB
7. `compact.ts:543``preCompactReadFileState = undefined` 10. `compact.ts:543``preCompactReadFileState = undefined`
8. `compact.ts:651``summaryResponse = undefined` 11. `compact.ts:651``summaryResponse = undefined`
9. 延迟非关键 attachment 生成 12. 延迟非关键 attachment 生成
### P1渲染与缓存预估降 50-110 MB ### P1渲染与缓存预估降 50-110 MB
10. 虚拟滚动 — 降低 OVERSCAN_ROWS 或 MAX_MOUNTED_ITEMS 13. 虚拟滚动 — 降低 OVERSCAN_ROWS 或 MAX_MOUNTED_ITEMS
11. `lastAPIRequestMessages` — 非 debug 清空 14. `lastAPIRequestMessages` — 非 debug 清空
12. MCP Tool Schema — 去掉 manager 层 toolsCache 15. MCP Tool Schema — 去掉 manager 层 toolsCache
13. `HybridTransport` — maxQueueSize 100K→10K 16. `HybridTransport` — maxQueueSize 100K→10K
14. `bootstrap/state.ts` — 无界 Map 加 LRU 17. `bootstrap/state.ts` — 无界 Map 加 LRU
### P2其他预估降 10-50 MB ### P2其他预估降 10-50 MB
15. `toolResultStorage.ts` — seenIds/replacements 定期清理 18. `toolResultStorage.ts` — seenIds/replacements 定期清理
16. Session 恢复流式 JSONL | AppState 增量更新 19. Session 恢复流式 JSONL | AppState 增量更新
17. Thinking 文本截断策略(保留前 N + 后 N 字符) 20. Thinking 文本截断策略(保留前 N + 后 N 字符)
18. `Bun.gc(true)` 低内存触发 21. `Bun.gc(true)` 低内存触发
### P2Ink 渲染层(降低 CPU 开销) ### P2Ink 渲染层(降低 CPU 开销)
19. `ink.tsx:655-661` — 布局偏移时尝试增量 damage 而非全屏 `{x:0,y:0,width:full,height:full}` 22. `ink.tsx:655-661` — 布局偏移时尝试增量 damage 而非全屏 `{x:0,y:0,width:full,height:full}`
## 附录 ## 附录
@@ -198,3 +291,4 @@ After: splice → 50K tokens
- Round 2 新发现HybridTransport 缓冲、React messagesRef 双重引用、toolResultStorage 无界增长 - Round 2 新发现HybridTransport 缓冲、React messagesRef 双重引用、toolResultStorage 无界增长
- Round 3 新发现:虚拟滚动 ~50MB 常驻、第 7-8 次 spreadquery.ts:1857、流式 contentBlocks thinking 累积、依赖树已懒加载 - Round 3 新发现:虚拟滚动 ~50MB 常驻、第 7-8 次 spreadquery.ts:1857、流式 contentBlocks thinking 累积、依赖树已懒加载
- Round 4 最终验证无新根因agent spread 和 attachment 累积为已知变体),调研终止 - Round 4 最终验证无新根因agent spread 和 attachment 累积为已知变体),调研终止
- Round 5 增量验证buildMessageLookups 8 Map/Set 重建成本、useDeferredValue 双缓冲、FileReadTool 无上限、Microcompact 内存释放延迟、compaction 与 React 状态交互细节

View File

@@ -12,12 +12,12 @@ Claude Code 将文件操作拆分为三个独立工具——这不是功能划
| 工具 | 权限级别 | 核心方法 | 关键属性 | | 工具 | 权限级别 | 核心方法 | 关键属性 |
|------|---------|---------|---------| |------|---------|---------|---------|
| **Read** | 只读(免审批) | `isReadOnly() → true` | `maxResultSizeChars: Infinity` | | **Read** | 只读(免审批) | `isReadOnly() → true` | `maxResultSizeChars: 100,000` |
| **Edit** | 写入(需确认) | `checkWritePermissionForTool()` | `maxResultSizeChars: 100,000` | | **Edit** | 写入(需确认) | `checkWritePermissionForTool()` | `maxResultSizeChars: 100,000` |
| **Write** | 写入(需确认) | `checkWritePermissionForTool()` | `maxResultSizeChars: 100,000` | | **Write** | 写入(需确认) | `checkWritePermissionForTool()` | `maxResultSizeChars: 100,000` |
<Tip> <Tip>
Read 的 `maxResultSizeChars` 是 `Infinity`,但这并不意味着无限制输出——真正的截断发生在 `validateContentTokens()` 中基于 token 预算的动态判定,而非字符数硬限制。 Read 的 `maxResultSizeChars` 为 100,000100KB。超出此阈值的结果会被持久化到磁盘减少长会话的内存压力。实际的 token 级别截断由 `validateContentTokens()` 动态控制。
</Tip> </Tip>
## FileRead多模态文件读取引擎 ## FileRead多模态文件读取引擎

View File

@@ -337,9 +337,10 @@ export type Output = z.infer<OutputSchema>
export const FileReadTool = buildTool({ export const FileReadTool = buildTool({
name: FILE_READ_TOOL_NAME, name: FILE_READ_TOOL_NAME,
searchHint: 'read files, images, PDFs, notebooks', searchHint: 'read files, images, PDFs, notebooks',
// Output is bounded by maxTokens (validateContentTokens). Persisting to a // Output is bounded by maxTokens (validateContentTokens). Results exceeding
// file the model reads back with Read is circular — never persist. // 100KB are persisted to disk (reducing memory pressure in long sessions)
maxResultSizeChars: Infinity, // rather than kept in the message array indefinitely.
maxResultSizeChars: 100_000,
strict: true, strict: true,
async description() { async description() {
return DESCRIPTION return DESCRIPTION

View File

@@ -34,6 +34,8 @@ import { isFullscreenEnvEnabled } from '../utils/fullscreen.js';
import { applyGrouping } from '../utils/groupToolUses.js'; import { applyGrouping } from '../utils/groupToolUses.js';
import { import {
buildMessageLookups, buildMessageLookups,
computeMessageStructureKey,
type MessageLookups,
createAssistantMessage, createAssistantMessage,
deriveUUID, deriveUUID,
getMessagesAfterCompactBoundary, getMessagesAfterCompactBoundary,
@@ -510,6 +512,12 @@ const MessagesImpl = ({
// comment above for why this replaced count-based slicing. // comment above for why this replaced count-based slicing.
const sliceAnchorRef = useRef<SliceAnchor>(null); const sliceAnchorRef = useRef<SliceAnchor>(null);
// Cache for buildMessageLookups: avoids rebuilding 8 Maps/Sets when only
// message content changed during streaming (text/thinking deltas). The key
// captures only structural info (types, IDs), so content-only deltas skip
// the rebuild entirely.
const lookupsCacheRef = useRef<{ key: string; lookups: MessageLookups } | null>(null);
// Expensive message transforms — filter, reorder, group, collapse, lookups. // Expensive message transforms — filter, reorder, group, collapse, lookups.
// All O(n) over 27k messages. Split from the renderRange slice so scrolling // All O(n) over 27k messages. Split from the renderRange slice so scrolling
// (which only changes renderRange) doesn't re-run these. Previously this // (which only changes renderRange) doesn't re-run these. Previously this
@@ -578,7 +586,14 @@ const MessagesImpl = ({
verbose, verbose,
); );
const lookups = buildMessageLookups(normalizedMessages, messagesToShow as MessageType[]); const lookupsKey = computeMessageStructureKey(normalizedMessages, messagesToShow as MessageType[]);
let lookups: MessageLookups;
if (lookupsCacheRef.current && lookupsCacheRef.current.key === lookupsKey) {
lookups = lookupsCacheRef.current.lookups;
} else {
lookups = buildMessageLookups(normalizedMessages, messagesToShow as MessageType[]);
lookupsCacheRef.current = { key: lookupsKey, lookups };
}
const hiddenMessageCount = messagesToShowNotTruncated.length - MAX_MESSAGES_TO_SHOW_IN_TRANSCRIPT_MODE; const hiddenMessageCount = messagesToShowNotTruncated.length - MAX_MESSAGES_TO_SHOW_IN_TRANSCRIPT_MODE;

View File

@@ -529,6 +529,16 @@ async function* queryLoop(
querySource, querySource,
) )
messagesForQuery = microcompactResult.messages messagesForQuery = microcompactResult.messages
// Release original strings from contentReplacementState.replacements for
// tool results whose content was replaced with the cleared message.
if (microcompactResult.clearedToolUseIds?.length) {
const replacements = toolUseContext?.contentReplacementState?.replacements
if (replacements) {
for (const id of microcompactResult.clearedToolUseIds) {
replacements.delete(id)
}
}
}
// For cached microcompact (cache editing), defer boundary message until after // For cached microcompact (cache editing), defer boundary message until after
// the API response so we can use actual cache_deleted_input_tokens. // the API response so we can use actual cache_deleted_input_tokens.
// Gated behind feature() so the string is eliminated from external builds. // Gated behind feature() so the string is eliminated from external builds.

View File

@@ -217,6 +217,10 @@ export type MicrocompactResult = {
compactionInfo?: { compactionInfo?: {
pendingCacheEdits?: PendingCacheEdits pendingCacheEdits?: PendingCacheEdits
} }
// Tool use IDs whose content was replaced with the cleared message.
// Callers should remove these from contentReplacementState.replacements
// to release the original strings from memory.
clearedToolUseIds?: string[]
} }
/** /**
@@ -528,5 +532,5 @@ function maybeTimeBasedMicrocompact(
notifyCacheDeletion(querySource) notifyCacheDeletion(querySource)
} }
return { messages: result } return { messages: result, clearedToolUseIds: [...clearSet] }
} }

View File

@@ -1397,6 +1397,54 @@ export function buildMessageLookups(
} }
} }
/**
* Compute a lightweight structural fingerprint for buildMessageLookups caching.
* Only captures information that affects lookup results (types, IDs, counts),
* not content. Returns an empty string when the arrays are structurally empty.
*
* O(n) but allocates only a string — much cheaper than the 8 Maps/Sets that
* buildMessageLookups creates on every call.
*/
export function computeMessageStructureKey(
normalizedMessages: NormalizedMessage[],
messages: Message[],
): string {
const parts: string[] = [
String(normalizedMessages.length),
'|',
String(messages.length),
]
for (const msg of messages) {
parts.push(msg.type[0])
if (msg.type === 'assistant') {
const aMsg = msg as AssistantMessage
const content = aMsg.message?.content
if (Array.isArray(content)) {
for (const block of content) {
if (typeof block !== 'string' && block.type === 'tool_use') {
parts.push('t', (block as ToolUseBlock).id)
}
}
}
} else if (msg.type === 'user') {
const content = (msg as UserMessage).message?.content
if (Array.isArray(content)) {
for (const block of content) {
if (typeof block !== 'string' && block.type === 'tool_result') {
parts.push('r', (block as ToolResultBlockParam).tool_use_id)
}
}
}
}
}
for (const msg of normalizedMessages) {
if (msg.type === 'progress') {
parts.push('p', (msg as ProgressMessage).parentToolUseID as string)
}
}
return parts.join(',')
}
/** Empty lookups for static rendering contexts that don't need real lookups. */ /** Empty lookups for static rendering contexts that don't need real lookups. */
export const EMPTY_LOOKUPS: MessageLookups = { export const EMPTY_LOOKUPS: MessageLookups = {
siblingToolUseIDs: new Map(), siblingToolUseIDs: new Map(),

View File

@@ -56,9 +56,9 @@ export function getPersistenceThreshold(
toolName: string, toolName: string,
declaredMaxResultSizeChars: number, declaredMaxResultSizeChars: number,
): number { ): number {
// Infinity = hard opt-out. Read self-bounds via maxTokens; persisting its // Infinity = hard opt-out (reserved for tools that self-bound via other
// output to a file the model reads back with Read is circular. Checked // mechanisms). Checked before the GB override so tengu_satin_quoll can't
// before the GB override so tengu_satin_quoll can't force it back on. // force it back on.
if (!Number.isFinite(declaredMaxResultSizeChars)) { if (!Number.isFinite(declaredMaxResultSizeChars)) {
return declaredMaxResultSizeChars return declaredMaxResultSizeChars
} }
@@ -813,11 +813,12 @@ export async function enforceToolResultBudget(
continue continue
} }
// Tools with maxResultSizeChars: Infinity (Read) — never persist. // Tools with maxResultSizeChars: Infinity — never persist (reserved for
// Mark as seen (frozen) so the decision sticks across turns. They don't // tools that self-bound via other mechanisms). Mark as seen (frozen) so
// count toward freshSize; if that lets the group slip under budget and // the decision sticks across turns. They don't count toward freshSize; if
// the wire message is still large, that's the contract — Read's own // that lets the group slip under budget and the wire message is still
// maxTokens is the bound, not this wrapper. // large, that's the contract — the tool's own maxTokens is the bound, not
// this wrapper.
const skipped = fresh.filter(c => shouldSkip(c.toolUseId)) const skipped = fresh.filter(c => shouldSkip(c.toolUseId))
skipped.forEach(c => state.seenIds.add(c.toolUseId)) skipped.forEach(c => state.seenIds.add(c.toolUseId))
const eligible = fresh.filter(c => !shouldSkip(c.toolUseId)) const eligible = fresh.filter(c => !shouldSkip(c.toolUseId))