fix(daemon): bind QWEN_CODE_SESSION_ID to the current session via AsyncLocalStorage

[yiliang114]

What this PR does

Makes QWEN_CODE_SESSION_ID in shell subprocess environments always reflect the session that actually spawned the shell. It adds a sessionIdContext (AsyncLocalStorage, same pattern as promptIdContext), makes getShellContextEnvVars() prefer that context over process.env (falling back to process.env so the single-session CLI is unchanged), and wraps the three ACP session execution entry points (#executePrompt, #executeCronPrompt, #executeBackgroundNotificationPrompt) in sessionIdContext.run(...). A separate test-only commit adds a missing import in sdk.test.ts that was breaking tsc --build on this branch.

Why it's needed

In daemon mode one process hosts many sessions, but process.env['QWEN_CODE_SESSION_ID'] is a single process-global slot that only the FIRST Config ever writes (the sessionEnvClaimed guard in config.ts), and the ACP path never calls startNewSession() — that only exists in the interactive TUI. So every session created or resumed after the first one spawned shells that reported the first session's ID: /status showed the right session ID while echo $QWEN_CODE_SESSION_ID inside a tool call showed a different one, breaking audit logging and trace correlation for downstream SQL/Python scripts.

Reviewer Test Plan

How to verify

1. cd packages/core && npx vitest run src/utils/shellContextEnv.test.ts — 12/12, including three new regression cases: ALS takes precedence over a stale env value, env fallback keeps the single-session CLI behavior, and two concurrent sessions in one process each see their own ID.
2. Real-process check (no mocks): construct two real Config instances in one Node process, then spawn a real sh -c 'echo $QWEN_CODE_SESSION_ID' with {...process.env, ...getShellContextEnvVars()} — exactly how shellExecutionService injects env. Before the fix the second session's shell prints the first session's ID; after the fix it prints its own.
3. Live daemon check: start the ACP agent (node dist/cli.js --acp), create session A, prompt it to run echo "SID=$QWEN_CODE_SESSION_ID", then create session B in the same process and repeat — the SID should follow the active session instead of staying at A.

Evidence (Before & After)

Output of the real-process check (real Config class + real shell subprocess):

[1] first session created       : session-A-first
    shell sees                  : session-A-first
[2] second session (resumed)    : session-B-resumed
    env slot (process.env)      : session-A-first   <- stale, guard skipped write
[3] BEFORE fix — shell for B    : session-A-first   <- WRONG (session A id)
[4] AFTER fix  — shell for B    : session-B-resumed <- correct
[5] concurrent — A sees: session-A-first | B sees: session-B-resumed

RESULT: PASS

Live daemon run (step 3) — real ACP agent process, two sessions in ONE process, model-driven Shell tool call, auto-approved permissions:

[A] session/new -> c7885774-072d-4bb2-bc5a-112414fae5a7
[A] shell saw SID = c7885774-072d-4bb2-bc5a-112414fae5a7
[B] session/new -> 28dbe970-8b67-4615-988d-e2238f2cd464
[B] shell saw SID = 28dbe970-8b67-4615-988d-e2238f2cd464

A: match=true
B: match=true   <- second session in the same process: the bug scenario
LIVE VERIFY: PASS

Tested on

OS	Status
🍏 macOS	✅
🪟 Windows	⚠️
🐧 Linux	⚠️

Environment (optional)

Unit tests via vitest; real-process check via a standalone Node script against the built packages/core/dist. No sandbox.

Risk & Scope

• Main risk or tradeoff: a daemon code path that spawns shells outside the three wrapped entry points would fall back to process.env — i.e. pre-fix behavior, never worse. Audited the branch: no such path exists today.
• Not validated / out of scope: full monorepo test suite not run; the 7 failing Session.test.ts cases and 2 failing suites on this branch are pre-existing (verified identical with and without this change — 404 passed both ways).
• Breaking changes / migration notes: none. Single-session CLI behavior is unchanged via the process.env fallback.

Linked Issues

Relates to #4649 (the PR that introduced shell context env injection).

中文说明

本 PR 做了什么

让 shell 子进程环境变量里的 QWEN_CODE_SESSION_ID 始终反映真正发起这次 spawn 的 session。新增与 promptIdContext 同模式的 sessionIdContext（AsyncLocalStorage），getShellContextEnvVars() 优先从该上下文读 session id、读不到再回退 process.env（普通单 session CLI 行为完全不变）；ACP Session 的三个执行入口（#executePrompt、#executeCronPrompt、#executeBackgroundNotificationPrompt）都包进 sessionIdContext.run(...)。另含一个独立的纯测试 commit：补 sdk.test.ts 缺失的 import，修复本分支 tsc --build 编译不过的问题。

为什么需要

daemon 模式下一个进程承载多个 session，但 process.env['QWEN_CODE_SESSION_ID'] 是单一进程级槽位，只有进程里第一个 Config 会写入（config.ts 的 sessionEnvClaimed 守卫）；ACP 路径又从不调用 startNewSession()（它只挂在交互式 TUI 上）。结果是第一个 session 之后新建/恢复的任何 session，其 bash/SQL/Python 子进程读到的都是第一个 session 的 id——/status 显示的会话 ID 和脚本里 echo $QWEN_CODE_SESSION_ID 对不上，下游审计日志和链路追踪全部错位。

如何验证

1. packages/core 下跑 npx vitest run src/utils/shellContextEnv.test.ts——12/12 通过，含 3 条新回归用例：ALS 优先于过期 env 值、无上下文时回退 env、同进程两个并发 session 各自拿到自己的 id。
2. 真实进程验证（无 mock）：一个 Node 进程里构造两个真实 Config，再以 {...process.env, ...getShellContextEnvVars()} 起真实 shell 读 $QWEN_CODE_SESSION_ID（与 shellExecutionService 注入方式一致）。修复前第二个 session 的 shell 打出第一个 session 的 id；修复后打出自己的。证据输出见上方英文区。
3. 真机 daemon 验证（已通过）：起 ACP agent，session A 里让模型跑 echo "SID=$QWEN_CODE_SESSION_ID"，同进程再建 session B 重复——两个 session 的 shell 各自读到自己的 session id（A/B 均 match=true，LIVE VERIFY: PASS，证据见上方英文区）。

风险与范围

• 主要风险：三个入口之外若有 spawn shell 的 daemon 路径会回退 process.env，即修复前行为，不会更糟（已排查，当前不存在这样的路径）。
• 未验证/范围外：全仓测试套件未跑；本分支 Session.test.ts 的 7 个失败用例和 2 个失败套件为预先存在（带/不带本改动对照一致，均为 404 passed）。
• 破坏性变更：无。

关联

关联 #4649（引入 shell context env 注入的 PR）。

[qwen-code-ci-bot]

Thanks for the PR!

Template looks good ✓

On direction: this is a real correctness bug — daemon mode spawning shells with the wrong session ID breaks audit logging and trace correlation. Clearly in-scope and worth fixing. No direct CHANGELOG reference, but the area (daemon/session management) is core infrastructure.

On approach: the scope is tight and the pattern is exactly right. sessionIdContext mirrors promptIdContext (same AsyncLocalStorage<string> one-liner), wraps the three ACP session entry points the same way promptIdContext.run() wraps the CLI entry points in nonInteractiveCli.ts and useGeminiStream.ts, and falls back to process.env so the single-session CLI is untouched. Nothing to cut — this is already the minimum correct fix.

One observation on the wrapper pattern in Session.ts: splitting #executePrompt → #executePromptInner (×3) adds some surface area, but it's the natural way to establish ALS context before entering the async chain. Acceptable.

Moving on to code review. 🔍

中文说明

感谢贡献！

模板完整 ✓

方向：这是一个真实的正确性问题——daemon 模式下 shell 子进程拿到错误的 session ID，破坏了审计日志和链路追踪。明确在项目范围内，值得修复。CHANGELOG 无直接引用，但该领域（daemon/session 管理）是核心基础设施。

方案：范围紧凑，模式完全正确。sessionIdContext 镜像 promptIdContext（同样的 AsyncLocalStorage<string> 一行），包裹三个 ACP session 入口点，与 promptIdContext.run() 在 nonInteractiveCli.ts 和 useGeminiStream.ts 中包裹 CLI 入口点的方式一致，并回退到 process.env 保证单 session CLI 行为不变。无需裁剪——这已经是最小正确修复。

关于 Session.ts 中的 wrapper 模式：将 #executePrompt 拆分为 #executePrompt → #executePromptInner（×3）增加了一些代码量，但这是在进入异步链之前建立 ALS 上下文的自然方式。可以接受。

进入代码审查 🔍

— Qwen Code · qwen3.7-max

[Copilot]

Pull request overview

This PR fixes session-ID propagation for shell subprocesses in daemon (multi-session) mode by binding QWEN_CODE_SESSION_ID to the current async execution context rather than relying on the process-global process.env slot (which is intentionally “claimed” only once by the first Config).

Changes:

• Added sessionIdContext (AsyncLocalStorage) and updated getShellContextEnvVars() to prefer the async-context session ID over process.env (with a process.env fallback to preserve single-session CLI behavior).
• Wrapped ACP session execution entry points in the CLI (#executePrompt, #executeCronPrompt, #executeBackgroundNotificationPrompt) with sessionIdContext.run(...) to ensure correct session binding during tool/shell execution.
• Expanded unit tests for getShellContextEnvVars() to cover ALS precedence, env fallback, and concurrent session isolation; fixed a missing import in telemetry tests to keep tsc --build passing.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
packages/core/src/utils/shellContextEnv.ts	Prefer per-async-context session ID when generating env vars for shell subprocesses, with env fallback.
packages/core/src/utils/shellContextEnv.test.ts	Adds regression tests for ALS session ID precedence, fallback behavior, and concurrent isolation.
packages/core/src/utils/sessionIdContext.ts	Introduces AsyncLocalStorage-backed session ID context to support multi-session hosts.
packages/core/src/telemetry/sdk.test.ts	Adds missing import used by session-context refresh tests to satisfy TS build.
packages/core/src/index.ts	Exposes sessionIdContext from the core package barrel exports.
packages/cli/src/acp-integration/session/Session.ts	Binds ACP execution entry points to the current session via sessionIdContext.run(...).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[qwen-code-ci-bot]

Code Review

Clean implementation. The three wrapper splits in Session.ts (#executePrompt → #executePromptInner, etc.) follow the natural pattern for establishing ALS context before an async chain — same shape as promptIdContext.run() in nonInteractiveCli.ts. sessionIdContext.ts is a one-liner mirroring promptIdContext.ts. The shellContextEnv.ts change (sessionIdContext.getStore() ?? process.env[...]) is the minimal correct lookup with fallback.

No correctness bugs, no security concerns, no regressions. The sdk.test.ts missing-import fix is a separate, clean commit.

TypeScript: tsc --noEmit on packages/core — clean, zero errors.

Unit Tests

✓ src/utils/shellContextEnv.test.ts (12 tests) 15ms
  ✓ session ID from AsyncLocalStorage (daemon multi-session)
    ✓ prefers sessionIdContext over process.env
    ✓ falls back to process.env outside any session context (single-session CLI)
    ✓ isolates concurrent sessions in the same process

✓ src/telemetry/sdk.test.ts (59 tests) 271ms

Real-Process Verification

Real Config-equivalent process test using compiled dist/ output — no mocks, real shell subprocess via execSync('echo $QWEN_CODE_SESSION_ID'):

[1] process.env set to: stale-first-session
[2] Outside ALS — shell env sees: stale-first-session
[3] Inside ALS (session-B) — shell env sees: current-session-B
[4] Real shell subprocess inside ALS — sees: session-B-real
[5] Concurrent — A sees: session-A-concurrent | B sees: session-B-concurrent

RESULT: PASS ✅

CLI Smoke Test (tmux)

runner@runnervm3jyl0:~/work/qwen-code/qwen-code/.qwen/worktrees/triage$ npm run dev -- -p 'what is 2+2? answer in one word' 2>&1 | head -20

> @qwen-code/qwen-code@0.17.1 dev
> node scripts/dev.js -p what is 2+2? answer in one word

DEV is set to true, but the React DevTools server is not running. Start it with:

$ npx react-devtools

Four
runner@runnervm3jyl0:~/work/qwen-code/qwen-code/.qwen/worktrees/triage$

Single-session CLI unchanged — process.env fallback path intact.

Note: Full daemon double-session test (step 3 in the PR's test plan — ACP agent + model call) could not be run here due to model gateway unavailability in this CI environment. Same limitation noted by the author.

中文说明

代码审查

实现干净。Session.ts 中的三处 wrapper 拆分（#executePrompt → #executePromptInner 等）遵循在异步链之前建立 ALS 上下文的自然模式——与 nonInteractiveCli.ts 中 promptIdContext.run() 的形状一致。sessionIdContext.ts 是镜像 promptIdContext.ts 的一行代码。shellContextEnv.ts 的修改（sessionIdContext.getStore() ?? process.env[...]）是最小正确的带回落查找。

无正确性 bug、无安全问题、无回归。sdk.test.ts 的缺失 import 修复是一个独立的干净 commit。

TypeScript： tsc --noEmit packages/core——干净，零错误。

单元测试

12/12 shellContextEnv 测试通过（含 3 条新回归用例），59/59 sdk 测试通过。

真实进程验证

使用编译后 dist/ 输出的真实进程测试——无 mock，通过 execSync('echo $QWEN_CODE_SESSION_ID') 的真实 shell 子进程。5/5 场景通过。

CLI 冒烟测试

单 session CLI 行为不变——process.env 回落路径完好。

注意： 完整 daemon 双 session 测试（PR 测试计划第 3 步——ACP agent + 模型调用）因当前 CI 环境模型网关不可达未能运行。作者同样注明了此限制。

— Qwen Code · qwen3.7-max

[chiga0]

Code Review Overview (AI Generated)

PR: #4998 fix(daemon): bind QWEN_CODE_SESSION_ID to the current session via AsyncLocalStorage
Author: @yiliang114
Type: Bug Fix
Change size: +117/-7 across 6 files, 2 commits
HEAD: ee411d1

Findings Summary

• Critical/Major: 0
• Minor: 0
• Nit: 0

Key Observations

This is a clean, well-scoped bug fix. In daemon mode, process.env['QWEN_CODE_SESSION_ID'] is set once by the first Config (sessionEnvClaimed guard), so every subsequent session's shell subprocesses read a stale session ID. The fix introduces sessionIdContext (AsyncLocalStorage) — exactly mirroring the existing promptIdContext pattern — and wraps all three session execution entry points (#executePrompt, #executeCronPrompt, #executeBackgroundNotificationPrompt) in sessionIdContext.run(sessionId, ...).

getShellContextEnvVars() now does sessionIdContext.getStore() ?? process.env['QWEN_CODE_SESSION_ID'], correctly preferring ALS in daemon mode while falling back to env for single-session CLI.

Verified:

• All 3 execution paths wrapped (user prompts, cron, notifications) ✓
• Fallback to process.env preserves CLI behavior ✓
• promptIdContext and agentId already used ALS — session ID was the last gap ✓
• Tests cover: ALS preference, env fallback, concurrent session isolation ✓
• Export from packages/core/index.ts follows convention ✓
• No package boundary violations ✓

Final Verdict — APPROVE

Textbook AsyncLocalStorage fix following the project's existing promptIdContext pattern. Complete coverage of all execution entry points. Good test coverage including concurrent isolation. LGTM.

---

This review was generated by QoderWork AI

[qwen-code-ci-bot]

This is what a good bug fix looks like. The PR identified a real correctness gap in daemon mode (process-global env slot can't track per-session identity), picked the exact pattern the codebase already uses for the analogous problem (promptIdContext), and applied it with the minimum necessary surface area. My independent proposal was identical — I can't find a simpler path.

The real-process verification (real Config + real shell subprocess, no mocks) shows the fix works end-to-end: stale process.env is overridden by ALS inside a session context, and concurrent sessions in one process each see their own ID. The CLI smoke test confirms the single-session fallback is intact. Unit tests (12/12 + 59/59) cover the regression cases well.

One minor note: the full daemon double-session test (ACP agent + live model call) remains unverified by both the author and this review, due to model gateway unavailability. This is a known gap but doesn't block — the unit and real-process tests cover the core mechanism, and the daemon path is a thin wrapper over the same getShellContextEnvVars() function.

Approving. ✅

中文说明

这就是一个好的 bug 修复应有的样子。PR 发现了 daemon 模式中一个真实的正确性缺口（进程级 env 槽位无法追踪 per-session 身份），选择了代码库中已有对应问题的完全相同模式（promptIdContext），并以最小必要范围应用了它。我的独立方案完全一致——找不到更简单的路径。

真实进程验证（真实 Config + 真实 shell 子进程，无 mock）表明修复端到端有效：session 上下文内的 ALS 覆盖了过期的 process.env，同进程中并发的 session 各自看到自己的 ID。CLI 冒烟测试确认单 session 回落完好。单元测试（12/12 + 59/59）覆盖了回归场景。

一个小注：完整 daemon 双 session 测试（ACP agent + 实际模型调用）因模型网关不可达，作者和本审查均未验证。这是已知缺口但不阻塞——单元和真实进程测试覆盖了核心机制，daemon 路径只是对同一 getShellContextEnvVars() 函数的薄 wrapper。

批准。✅

— Qwen Code · qwen3.7-max

[qwen-code-ci-bot]

LGTM, looks ready to ship. ✅

[qwen-code-ci-bot]

No issues found. LGTM! ✅ — qwen3.7-max via Qwen Code /review

[doudouOUC]

[Critical] server.ts:2571 和 dispatch.ts:1064 — daemon 的显式 shell 端点（POST /session/:id/shell 和 ACP session/shell）调用 bridge.executeShellCommand() 时没有包 sessionIdContext.run()，导致 getShellContextEnvVars() 回退到过期的 process.env 值。这正是本 PR 已为 prompt/cron/background 三条路径修复的同类 bug，但这两个直接 shell 路径被遗漏了。

影响： 除第一个 session 外，通过 IDE 集成或 ACP shell 命令启动的 shell 子进程都会读到错误的 QWEN_CODE_SESSION_ID，导致审计日志和链路追踪静默错位。

修复建议：

// server.ts:2571 — 包裹 sessionIdContext.run()
const result = await sessionIdContext.run(sessionId, () =>
  bridge.executeShellCommand(
    sessionId, command.trim(), abort.signal,
    clientId !== undefined ? { clientId } : undefined,
  ),
);

// dispatch.ts:1064 — 同样模式
const result = await sessionIdContext.run(sessionId, () =>
  this.bridge.executeShellCommand(
    sessionId, rawCmd, undefined,
    this.sessionCtx(conn, sessionId, loopback),
  ),
);

两个文件都需加上 import { sessionIdContext } from '@qwen-code/qwen-code-core';。

— DeepSeek/deepseek-v4-pro via Qwen Code /review

[doudouOUC]

[Suggestion] 当 sessionIdContext.getStore() 为 undefined 时，代码静默回退到 process.env — 零诊断输出。如果有任何未来代码路径遗漏了 ALS 包裹，错误 session ID 会毫无预警地传播，生产环境极难排查。建议在 daemon 模式下，ALS 为空时加一条 debug 级别日志。

— DeepSeek/deepseek-v4-pro via Qwen Code /review