From 23a3c4518ce73608fd8b03b13e7035647f9e8b3d Mon Sep 17 00:00:00 2001 From: uchouT Date: Wed, 29 Apr 2026 21:01:19 +0800 Subject: [PATCH 1/2] fix(session): sanitize incomplete tool_call groups in prompt history When tool execution is aborted mid-turn, the assistant message with tool_calls is already persisted but the matching tool result messages never are. The next prompt sent to the LLM then violates the OpenAI/Anthropic protocol's tool_call group invariant and the API returns 400. - Replace local tail-only trimIncompleteToolCallGroup with a general removeIncompleteToolCallGroups exported from context-utils. It scans the full history and drops any assistant(toolCalls) whose tool_call_id responses are missing, plus any orphan tool messages. - Apply at history->prompt boundaries: assembleSessionContext, assembleMainSessionContext, and after replay merge in session send/stream (replay path keeps load-time history intact so the envelope's synthesized tool results can complete the group). - Update fork() to use the general version (also handles middle orphans in inherited parent history). Tests: tail orphan recovery, middle orphan recovery, complete-group preservation. --- packages/session/src/__tests__/abort.test.ts | 124 +++++++++++++++++++ packages/session/src/context-utils.ts | 49 +++++++- packages/session/src/create-session.ts | 54 ++------ 3 files changed, 182 insertions(+), 45 deletions(-) diff --git a/packages/session/src/__tests__/abort.test.ts b/packages/session/src/__tests__/abort.test.ts index 9b056d5..7877f27 100644 --- a/packages/session/src/__tests__/abort.test.ts +++ b/packages/session/src/__tests__/abort.test.ts @@ -1,5 +1,7 @@ import { describe, it, expect } from 'vitest' import { makeSession } from './helpers.js' +import { loadSession } from '../create-session.js' +import { InMemoryStorageAdapter } from '../mocks/in-memory-storage.js' import type { LLMAdapter, LLMChunk, LLMCompleteOptions, LLMResult, Message } from '../types/llm.js' /** 让 fetch-style adapter 监听 signal 的最小 LLMAdapter */ @@ -157,3 +159,125 @@ describe('Session.stream() AbortSignal', () => { expect(llm.calls[0]!.signal).toBe(controller.signal) }) }) + +/** + * 当 tool 执行被 abort 中断后，storage 中会残留 assistant(toolCalls) 但缺对应 tool 结果。 + * 下一次 send/stream 加载历史送给 LLM 时，必须把这种孤儿组过滤掉， + * 否则 OpenAI-compat adapter 会因协议不一致返回 400（assistant 有 tool_calls 缺响应）。 + */ +describe('orphaned tool_calls sanitization (abort recovery)', () => { + /** 直接往 storage 注入一个携带 orphan 历史的 session */ + async function seedSession( + storage: InMemoryStorageAdapter, + sessionId: string, + records: Message[], + ): Promise { + const now = new Date().toISOString() + await storage.putSession({ + id: sessionId, + label: 'Test', + role: 'standard', + status: 'active', + tags: [], + metadata: {}, + createdAt: now, + updatedAt: now, + }) + for (const rec of records) { + await storage.appendRecord(sessionId, rec) + } + } + + /** 抓取 LLM 收到的 prompt messages 的 mock adapter */ + function createCapturingLLM(reply: LLMResult = { content: 'ok' }): LLMAdapter & { calls: Message[][] } { + const calls: Message[][] = [] + const adapter: LLMAdapter = { + maxContextTokens: 1_000_000, + async complete(messages) { + calls.push(messages) + return reply + }, + } + return Object.assign(adapter, { calls }) + } + + it('tail orphan：abort 留下 assistant(toolCalls) 后下一轮 send 不应把孤儿带进 prompt', async () => { + const storage = new InMemoryStorageAdapter() + const sessionId = 'tail-orphan' + await seedSession(storage, sessionId, [ + { role: 'user', content: 'do X' }, + { + role: 'assistant', + content: '', + toolCalls: [{ id: 'tc-tail', name: 'foo', input: {} }], + }, + ]) + + const llm = createCapturingLLM() + const session = await loadSession(sessionId, { storage, llm }) + expect(session).not.toBeNull() + await session!.send('follow-up') + + const hasOrphan = llm.calls[0]!.some( + (m) => m.role === 'assistant' && m.toolCalls?.some((tc) => tc.id === 'tc-tail'), + ) + expect(hasOrphan).toBe(false) + // 跟进的 user 消息应当是最末一条 + expect(llm.calls[0]![llm.calls[0]!.length - 1]).toMatchObject({ role: 'user', content: 'follow-up' }) + }) + + it('middle orphan：被夹在干净消息中间的 orphan 也必须过滤（仅裁尾不够）', async () => { + const storage = new InMemoryStorageAdapter() + const sessionId = 'middle-orphan' + await seedSession(storage, sessionId, [ + { role: 'user', content: 'first' }, + { + role: 'assistant', + content: '', + toolCalls: [{ id: 'tc-orphan', name: 'foo', input: {} }], + }, + { role: 'user', content: 'second' }, + { role: 'assistant', content: 'clean response' }, + ]) + + const llm = createCapturingLLM() + const session = await loadSession(sessionId, { storage, llm }) + expect(session).not.toBeNull() + await session!.send('third') + + const hasOrphan = llm.calls[0]!.some( + (m) => m.role === 'assistant' && m.toolCalls?.some((tc) => tc.id === 'tc-orphan'), + ) + expect(hasOrphan).toBe(false) + // 干净的 assistant 应保留 + const hasClean = llm.calls[0]!.some((m) => m.role === 'assistant' && m.content === 'clean response') + expect(hasClean).toBe(true) + }) + + it('完整 tool call 组应原样保留（不能误伤）', async () => { + const storage = new InMemoryStorageAdapter() + const sessionId = 'complete-group' + await seedSession(storage, sessionId, [ + { role: 'user', content: 'do Y' }, + { + role: 'assistant', + content: '', + toolCalls: [{ id: 'tc-ok', name: 'bar', input: {} }], + }, + { role: 'tool', content: 'result', toolCallId: 'tc-ok' }, + { role: 'assistant', content: 'final' }, + ]) + + const llm = createCapturingLLM() + const session = await loadSession(sessionId, { storage, llm }) + expect(session).not.toBeNull() + await session!.send('next') + + const hasAssistantWithCall = llm.calls[0]!.some( + (m) => m.role === 'assistant' && m.toolCalls?.some((tc) => tc.id === 'tc-ok'), + ) + const hasToolResult = llm.calls[0]!.some((m) => m.role === 'tool' && m.toolCallId === 'tc-ok') + expect(hasAssistantWithCall).toBe(true) + expect(hasToolResult).toBe(true) + }) +}) diff --git a/packages/session/src/context-utils.ts b/packages/session/src/context-utils.ts index ce86290..3ef8d52 100644 --- a/packages/session/src/context-utils.ts +++ b/packages/session/src/context-utils.ts @@ -2,6 +2,49 @@ import type { Message, LLMAdapter } from './types/llm.js' import type { SessionStorage } from './types/storage.js' import type { CompressFn } from './types/functions.js' +/** + * 从历史中移除所有不完整的 tool call 组： + * 任何 `assistant(toolCalls=[A,B,...])` 后续若缺少与之 `toolCallId` 对应的 `tool` 消息， + * 则整组（assistant + 已写入的 tool 消息）一并丢弃；同时丢弃没有匹配 assistant 的孤立 tool 消息。 + * + * 触发场景： + * - tool 执行被 AbortSignal 中断 → assistant 已写入但 tool result 永远不再回灌（最常见） + * - 进程崩溃 / 手动改库 / 旧版 bug 残留 + * + * 这是历史→prompt 的不变量保证：送给 LLM 的 messages 必须满足 OpenAI/Anthropic 协议 + * 对 tool call group 完整性的要求，否则 OpenAI-compat adapter 会返回 400。 + */ +export function removeIncompleteToolCallGroups(records: Message[]): Message[] { + const result: Message[] = [] + let i = 0 + while (i < records.length) { + const rec = records[i]! + if (rec.role === 'assistant' && rec.toolCalls && rec.toolCalls.length > 0) { + const expectedIds = new Set(rec.toolCalls.map((tc) => tc.id)) + let j = i + 1 + while (j < records.length && records[j]!.role === 'tool') { + const t = records[j]! + if (t.toolCallId) expectedIds.delete(t.toolCallId) + j++ + } + if (expectedIds.size === 0) { + for (let k = i; k < j; k++) result.push(records[k]!) + } + // 不完整 → 整组丢弃 + i = j + continue + } + if (rec.role === 'tool') { + // 没有前导 assistant(toolCalls) 的孤立 tool 消息 → 丢弃 + i++ + continue + } + result.push(rec) + i++ + } + return result +} + /** 内置默认压缩提示词 */ const BUILTIN_COMPRESS_PROMPT = `你是对话压缩助手。请将以下对话历史压缩为一段简洁的摘要，保留关键上下文信息。 要求： @@ -155,7 +198,8 @@ export async function assembleSessionContext( const userTimestamp = new Date().toISOString() const userMessage: Message = { role: 'user', content: userContent, timestamp: userTimestamp } - const history = await storage.listRecords(sessionId) + // 净化历史：移除中断/崩溃残留的不完整 tool call 组，保证送给 LLM 的 prompt 协议合法 + const history = removeIncompleteToolCallGroups(await storage.listRecords(sessionId)) // 估算全量 token 数 const fullMessages = [...prefixMessages, ...history, userMessage] @@ -266,7 +310,8 @@ export async function assembleMainSessionContext( const userTimestamp = new Date().toISOString() const userMessage: Message = { role: 'user', content: userContent, timestamp: userTimestamp } - const history = await storage.listRecords(sessionId) + // 净化历史：与 assembleSessionContext 对称，避免不完整 tool call 组流入 prompt + const history = removeIncompleteToolCallGroups(await storage.listRecords(sessionId)) // 估算全量 token 数 const fullMessages = [...prefixMessages, ...history, userMessage] diff --git a/packages/session/src/create-session.ts b/packages/session/src/create-session.ts index 4611b10..e6ae7e9 100644 --- a/packages/session/src/create-session.ts +++ b/packages/session/src/create-session.ts @@ -4,47 +4,7 @@ import { SessionArchivedError } from './types/session-api.js' import type { SessionMeta, SessionMetaUpdate, ForkOptions } from './types/session.js' import type { Message } from './types/llm.js' import type { CreateSessionOptions, LoadSessionOptions, SendResult, StreamResult } from './types/functions.js' -import { assembleSessionContext, buildSessionIdentityMessages, createBuiltinCompressFn, type CompressionCache } from './context-utils.js' - -/** 裁掉尾部不完整的 tool call 组（assistant 有 toolCalls 但缺少对应 tool 结果） */ -function trimIncompleteToolCallGroup(records: Message[]): Message[] { - if (records.length === 0) return records - let end = records.length - while (end > 0) { - const last = records[end - 1]! - if (last.role === 'assistant' && last.toolCalls && last.toolCalls.length > 0) { - // assistant 有 toolCalls 但后面没有 tool 消息 → 裁掉 - end-- - continue - } - if (last.role === 'tool') { - // tool 消息，向前找到对应的 assistant - let assistantIdx = end - 2 - while (assistantIdx >= 0 && records[assistantIdx]!.role === 'tool') { - assistantIdx-- - } - if (assistantIdx >= 0) { - const assistant = records[assistantIdx]! - if (assistant.role === 'assistant' && assistant.toolCalls && assistant.toolCalls.length > 0) { - const expectedIds = new Set(assistant.toolCalls.map(tc => tc.id)) - for (let j = assistantIdx + 1; j < end; j++) { - const rec = records[j]! - if (rec.role === 'tool' && rec.toolCallId) { - expectedIds.delete(rec.toolCallId) - } - } - if (expectedIds.size > 0) { - // 不完整 → 裁掉整个组 - end = assistantIdx - continue - } - } - } - } - break - } - return end === records.length ? records : records.slice(0, end) -} +import { assembleSessionContext, buildSessionIdentityMessages, createBuiltinCompressFn, removeIncompleteToolCallGroups, type CompressionCache } from './context-utils.js' interface ToolResultEnvelope { toolResults: Array<{ @@ -114,6 +74,9 @@ async function assembleSessionReplayContext( messages.push({ role: 'system', content: memory }) } + // 注意：此处刻意不调用 removeIncompleteToolCallGroups。 + // replay 路径会把"assistant(toolCalls) + 由 envelope 合成的 tool 消息"拼接成完整组， + // 在加载阶段过早裁剪反而会把回灌目标删掉。完整组校验放在拼接后由调用方做。 const history = await storage.listRecords(sessionId) messages.push(...history) return { messages, insightConsumed } @@ -227,6 +190,9 @@ function buildSession( if (replayContext.insightConsumed) { await storage.clearInsight(currentMeta.id) } + // 替换为 replay 上下文后，原 assembled.messages 里的 sanitize 不再生效； + // 在拼好"assistant + tool 结果"完整组之后，再做一次孤儿组清理（防御中段 orphan）。 + promptMessages = removeIncompleteToolCallGroups(promptMessages) } // 调 LLM — adapter 抛 AbortError 时直接向上传播，下方 L3 写入分支整体跳过 @@ -299,6 +265,8 @@ function buildSession( if (replayContext.insightConsumed) { await storage.clearInsight(currentMeta.id) } + // 拼好完整组之后再清孤儿，防御中段 orphan（与 send() 对称） + promptMessages = removeIncompleteToolCallGroups(promptMessages) } if (!options.llm) { @@ -440,8 +408,8 @@ function buildSession( if (ctx !== 'none') { const parentRecords = await storage.listRecords(currentMeta.id) const selected = ctx === 'inherit' ? parentRecords : await ctx(parentRecords) - // 裁掉尾部不完整的 tool call 组（fork 发生在 tool 执行中，tool result 还没写入） - const records = trimIncompleteToolCallGroup(selected) + // 净化掉不完整的 tool call 组（fork 在 tool 执行中、或父历史里夹有中段 orphan 时都需要） + const records = removeIncompleteToolCallGroups(selected) for (const record of records) { await storage.appendRecord(childId, record) } From 7a25bb01251a8e184265b8fa408914f6ad520e97 Mon Sep 17 00:00:00 2001 From: uchouT Date: Thu, 30 Apr 2026 14:31:18 +0800 Subject: [PATCH 2/2] =?UTF-8?q?chore(release):=20bump=20session=200.7.0=20?= =?UTF-8?q?=E2=86=92=200.7.1?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- packages/session/package.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/packages/session/package.json b/packages/session/package.json index 86cf75b..b6588e9 100644 --- a/packages/session/package.json +++ b/packages/session/package.json @@ -1,6 +1,6 @@ { "name": "@stello-ai/session", - "version": "0.7.0", + "version": "0.7.1", "description": "Session layer for Stello — conversation topology engine", "license": "Apache-2.0", "author": "Stello Contributors",