fix(cli): emit deferred stream-json startup warnings

[barry166]

Summary

• write startup warnings generated during deferred stream-json initialization to stderr
• preserve the existing early warning output by only emitting warnings added during initialization
• cover the warning delta and deduplication behavior with a session regression test

Fixes #7158.

Testing

• npm run test --workspace=packages/cli -- src/nonInteractive/session.test.ts
• npm run typecheck --workspace=packages/cli
• npm run lint --workspace=packages/cli -- src/nonInteractive/session.ts src/nonInteractive/session.test.ts
• npm exec prettier -- --check packages/cli/src/nonInteractive/session.ts packages/cli/src/nonInteractive/session.test.ts
• git diff --check

[qwen-code-ci-bot]

Thanks for the PR!

Template: the body uses ## Summary / ## Testing instead of the repository's PR template headings (## What this PR does, ## Why it's needed, ## Reviewer Test Plan, etc.). The key information is present and the linked issue is well-documented, so not blocking on this — but please use the template headings on future PRs to help reviewers navigate quickly.

Problem: observed bug with a clear reproduction in #7158. Stream-json mode silently drops startup warnings generated during deferred config.initialize(). Plain headless and interactive modes surface them correctly. The issue provides a concrete repro sketch and identifies the root cause.

Direction: aligned — startup warnings should reach users regardless of input format. SDK and automation clients consuming stream-json are exactly the audience that can't discover misconfiguration any other way. This matches the existing pattern in gemini.tsx (the non-stream-json branch already does snapshot + delta emission).

Size: 34 additions, 0 deletions across 2 files (production: ~10 lines, test: ~24 lines). Not applicable for core module checks.

Approach: minimal and focused. Snapshot pre-init warnings into a Set, emit only the delta after initialize(). Mirrors the established pattern in gemini.tsx line-for-line. No scope creep.

Moving on to code review. 🔍

中文说明

感谢贡献！

模板： PR 正文使用了 ## Summary / ## Testing，而非仓库模板要求的 ## What this PR does、## Why it's needed、## Reviewer Test Plan 等标题。关键信息已提供，关联 issue 文档清晰，不因此阻塞——但未来请使用模板标题以便 reviewer 快速定位。

问题： 已在 #7158 中观测到并提供复现。Stream-json 模式下，config.initialize() 期间产生的启动警告被静默丢弃。普通 headless 和交互模式正常输出。Issue 提供了具体复现步骤和根因分析。

方向： 对齐——无论输入格式如何，启动警告都应触达用户。Stream-json 的消费者（SDK 和自动化客户端）恰恰是最需要发现配置异常的一方。与 gemini.tsx 中的已有模式一致。

规模： 34 行新增，0 行删除，涉及 2 个文件（生产代码约 10 行，测试约 24 行）。不触及核心模块保护规则。

方案： 最小且聚焦。将初始化前的警告快照到 Set 中，initialize() 后仅输出增量。与 gemini.tsx 中的既有模式逐行一致。无范围蔓延。

进入代码审查 🔍

— Qwen Code · qwen3.7-max

_{Reviewed at 947c91f794969a6c6035ec07904675a21f6a0755 · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

Code Review

Independent proposal (before reading the diff): snapshot config.getWarnings() into a Set before config.initialize(), then iterate post-init warnings and emit only the delta to stderr — mirroring the existing gemini.tsx pattern.

The PR's implementation matches this proposal exactly. In ensureConfigInitialized:

1. Snapshot pre-init warnings into a Set before config.initialize()
2. After init, iterate config.getWarnings(), skip any already in the Set, emit the rest to stderr, and add them to the Set for dedup

This is line-for-line consistent with the non-stream-json branch in gemini.tsx (line ~1061). Clean, minimal, no unnecessary abstraction.

One minor note: gemini.tsx uses writeStderrLine(warning) while this PR uses process.stderr.write(\${warning}\n`). Functionally identical, but the helper exists. Not blocking — the inline write is clearer in context and the \n` matches the helper's behavior.

No correctness bugs, security holes, or regressions found. Test covers both legs of the dedup: pre-existing warning NOT re-emitted, new warning IS emitted.

Real-Scenario Testing

Ran stream-json mode end-to-end in tmux. No startup warnings were triggered in this CI environment (no legacy plan files or team-memory config present), so the before/after comparison shows clean stream-json output on both sides. The fix targets a code path that the unit test directly exercises.

Before (installed build, v0.19.11)

$ echo '{"type":"user","session_id":"test","message":{"role":"user","content":"say hello"}}' | qwen --input-format stream-json --output-format stream-json 2>before-stderr.log
{"type":"system","subtype":"init","uuid":"3b29f8d5-...","session_id":"3b29f8d5-...","model":"qwen3.7-max",...}
{"type":"assistant","uuid":"2677c7d3-...","message":{"content":[{"type":"text","text":"Hello! How can I help you today?"}],...}}
{"type":"result","subtype":"success","result":"Hello! How can I help you today?","duration_ms":8238,...}
---STDERR---
(empty — no warnings triggered)

After (PR code via npm run dev)

$ echo '{"type":"user","session_id":"test","message":{"role":"user","content":"say hi"}}' | npm run dev -- --input-format stream-json --output-format stream-json 2>after-stderr.log
{"type":"system","subtype":"init","uuid":"51753308-...","session_id":"51753308-...","model":"qwen3.7-max",...}
---STDERR---
DEV is set to true, but the React DevTools server is not running. Start it with:
$ npx react-devtools

Both runs complete successfully. Stderr is clean (the DevTools message is a dev-environment notice, not related to this fix). The unit test (36/36 passing) directly verifies the warning-delta behavior.

中文说明

代码审查

独立方案（阅读 diff 前）：在 config.initialize() 之前将 config.getWarnings() 快照到 Set 中，初始化后遍历警告集合，仅将增量输出到 stderr——与 gemini.tsx 中的既有模式一致。

PR 的实现与此方案完全一致。在 ensureConfigInitialized 中：

1. 在 config.initialize() 前将已有警告快照到 Set
2. 初始化后遍历 config.getWarnings()，跳过已在 Set 中的，输出新增的，并加入 Set 去重

与 gemini.tsx（约第 1061 行）的非 stream-json 分支逐行一致。干净、最小化、无不必要的抽象。

一个细节： gemini.tsx 使用 writeStderrLine(warning) 而本 PR 使用 process.stderr.write(\${warning}\n`)`。功能等价，内联写入在此上下文中更清晰。不阻塞。

无正确性缺陷、安全漏洞或回归。测试覆盖了去重的两个分支：已有警告不被重复输出，新增警告被正确输出。

真实场景测试

在 tmux 中端到端运行了 stream-json 模式。该 CI 环境未触发启动警告（无遗留 plan 文件或团队记忆配置），因此 before/after 对比显示两边均为干净的 stream-json 输出。单元测试直接覆盖了修复目标代码路径。

两次运行均成功完成，stderr 无异常输出。单元测试 36/36 通过，直接验证了警告增量行为。

— Qwen Code · qwen3.7-max

_{Reviewed at 947c91f794969a6c6035ec07904675a21f6a0755 · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

Confidence: 5/5 — Clean across every stage; a textbook bugfix.

This is what a good external contribution looks like: a real bug with a filed issue (#7158, P2), a minimal fix that mirrors the established pattern in gemini.tsx, and a test that actually verifies the behavior rather than just exercising the code path.

The fix is 10 lines of production code that snapshot pre-init warnings and emit only the delta after config.initialize() — exactly the right approach, and exactly what the non-stream-json branch already does. No scope creep, no unnecessary abstraction, no drive-by refactors.

The test is well-designed: it sets up pre-existing warnings, mocks initialize() to add a new one, and asserts both legs (old warning NOT re-emitted, new warning IS emitted). All 36 session tests pass.

Stream-json mode runs cleanly end-to-end in both the installed build and the PR code.

Approving. ✅

中文说明

信心度：5/5 — 各阶段均通过；教科书级别的 bugfix。

这是一个优秀外部贡献的典范：真实 bug 已有 P2 issue（#7158），最小化修复与 gemini.tsx 中的既有模式一致，测试实际验证了行为而非仅仅跑通代码路径。

修复仅 10 行生产代码：在 config.initialize() 前快照已有警告，初始化后仅输出增量——正是正确的做法，也是非 stream-json 分支已有的做法。无范围蔓延、无不必要的抽象、无顺手重构。

测试设计良好：预设已有警告，mock initialize() 新增一条，断言两个分支（旧警告不重复输出，新警告正确输出）。36 个 session 测试全部通过。

Stream-json 模式在安装版本和 PR 代码中均端到端正常运行。

已批准 ✅

— Qwen Code · qwen3.7-max

_{Reviewed at 947c91f794969a6c6035ec07904675a21f6a0755 · re-run with @qwen-code /triage}

[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]

No issues found. LGTM! ✅

— qwen3.7-max via Qwen Code /review