fix(core): remove greeting-responder example from agent tool prompt#4923
Merged
Conversation
The greeting-responder example in the agent tool description could mislead some models into launching a subagent for simple greetings instead of responding with plain text. Remove the example and its associated agent description entry to eliminate this risk.
Collaborator
|
Thanks for the PR! Template looks good ✓ On direction: This is a clean, targeted fix. The On approach: Scope is minimal — 9 lines deleted, no code logic touched. The remaining Moving on to code review. 🔍 中文说明感谢贡献！ 模板完整 ✓ 方向：这是一个干净、精准的修复。 方案：范围最小——删除 9 行，不涉及代码逻辑。剩余的 进入代码审查 🔍 — Qwen Code · qwen3.7-max |
Collaborator
Code ReviewThe diff is clean and matches what I'd do independently: remove the Real-Scenario TestingRan Before (main — greeting-responder example present)After (PR branch — greeting-responder example removed)Summary
Both runs produce identical user-visible behavior — plain text response, no agent tool invocation. The PR reduces prompt size by ~56 input tokens. The change is preventive: it removes a misleading few-shot example that could trigger agent misuse on other models, even though qwen3.7-max is not affected. 中文说明代码审查Diff 干净，与我的独立方案完全一致：删除 真实场景测试在两个分支上运行了 两次运行产生了相同的用户可见行为——纯文本响应，无 agent 工具调用。PR 减少了约 56 个输入 token。该更改是预防性的：它移除了一个可能在其他模型上触发 agent 误用的误导性 few-shot 示例，即使 qwen3.7-max 不受影响。 — Qwen Code · qwen3.7-max |
Collaborator
|
This is a clean, well-justified PR. The author identified a real risk — the greeting-responder few-shot example teaches models to delegate trivial input to a subagent — and applied the simplest possible fix: delete it. The remaining test-runner example already covers the agent tool pattern, so nothing is lost. Testing confirms no behavioral regression on qwen3.7-max (the model responds with plain text in both cases), and the prompt shrinks by ~56 tokens. The author's own 40-run cross-model test couldn't reproduce the issue either, but the preventive removal of a misleading example is sound engineering — it eliminates a footgun for models we haven't tested and future model versions. No concerns. Ships it. ✅ 中文说明这是一个干净、充分论证的 PR。作者识别出了一个真实风险——greeting-responder 的 few-shot 示例教会模型将简单输入委托给子代理——并应用了最简单的修复：删除它。剩余的 test-runner 示例已经覆盖了 agent 工具的使用模式，没有任何损失。 测试确认在 qwen3.7-max 上没有行为回归（两种情况下模型都用纯文本响应），提示词缩小了约 56 个 token。作者自己的 40 次跨模型测试也无法复现问题，但预防性移除误导性示例是合理的工程实践——消除了未测试模型和未来模型版本的隐患。 无顾虑，通过。✅ — Qwen Code · qwen3.7-max |
qwen-code-ci-bot
approved these changes
Jun 10, 2026
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, looks ready to ship. ✅
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does
Removes the
greeting-responderexample from the agent tool's model-facing prompt. This includes the fake agent description entry ("greeting-responder": use this agent to respond to user greetings with a friendly joke) and the accompanying few-shot example that shows the model launching a subagent in response to "Hello".Why it's needed
We encountered reports of some models misusing the agent tool for simple greeting messages — launching a subagent instead of responding with plain text. The
greeting-responderexample in the agent tool prompt is the likely culprit: it teaches the model that greetings should be delegated to an agent.We ran a cross-model headless test (8 models × 5 repeats = 40 runs) sending "hi" and checking for agent tool invocations via both JSON output and raw API request logs. None of the tested models (claude-opus-4-6, gpt-5.5, deepseek-v4-pro, deepseek-v4-flash, qwen3.7-max, qwen3.7-plus, glm-5.1, kimi-k2.6) triggered the agent tool in our tests. However, the example still presents an unnecessary risk — it's the only few-shot that teaches a "delegate to agent" pattern for trivial user input, and models we haven't tested (or future model versions) may be more susceptible.
Reviewer Test Plan
How to verify
npm run build && npm run bundlenode dist/cli.js "hi" --approval-mode yolo --output-format json 2>/dev/null— confirm the response is plain text with no agent tool call.--openai-logging --openai-logging-dir /tmp/api-logsand inspect the API log to confirmresponse.choices[0].message.tool_callsis empty or absent.Evidence (Before & After)
N/A — prompt-level change, no TUI impact. The removed text was part of the model-facing tool description, not user-visible UI.
Tested on
Risk & Scope
中文说明
从 agent 工具的模型提示词中移除了
greeting-responder示例。该示例包括一个虚构的 agent 描述条目和一个展示模型在收到 "Hello" 时启动子代理的 few-shot 示例。部分模型会误用 agent 工具来处理简单的问候消息（启动子代理而非直接回复文本）。虽然我们在 8 个模型 × 5 次重复 = 40 次测试中未能复现此问题，但该示例仍然存在不必要的风险——它是唯一一个教导模型将简单用户输入委托给 agent 的 few-shot 示例。