fix(agent): 强化 prompt 可靠性——结果验证、budget 语义、安全边界与认识论纪律#1354
Merged
CodFrm merged 4 commits intorelease/v1.4-agentfrom Apr 23, 2026
Merged
fix(agent): 强化 prompt 可靠性——结果验证、budget 语义、安全边界与认识论纪律#1354CodFrm merged 4 commits intorelease/v1.4-agentfrom
CodFrm merged 4 commits intorelease/v1.4-agentfrom
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
该 PR 通过更新主 agent / sub-agent / compact 总结器的提示词文本与测试断言，强化 agent 编排可靠性（子代理结果验收、budget 语义、不可逆操作边界、以及输出置信度与验证纪律），以减少“静默失败传播”和“未验证即宣称成功”等系统性失效模式。
Changes:
- 在主 agent prompt 中补充子代理结果验收规范、依赖任务的 fallback 指引，并扩展不可逆操作确认清单（加入 userscript 场景）。
- 在 sub-agent prompt 中修正 tool-call budget 语义，并为 researcher/page_operator/general 三类角色追加输出纪律（置信度分层、动作/结果分离、取舍透明与失败诚实）。
- 更新 system prompt 相关测试，新增断言覆盖本次大部分文本变更。
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| src/app/service/agent/core/system_prompt.ts | 主/子代理系统提示词更新：safety 增强、子代理结果接收规范、依赖任务 fallback、sub-agent budget 语义修正。 |
| src/app/service/agent/core/system_prompt.test.ts | 为 system prompt 与 sub-agent system prompt 新增断言，覆盖新增段落与关键措辞。 |
| src/app/service/agent/core/sub_agent_types.ts | 为 researcher/page_operator/general 追加更严格的输出与验证规范（置信度、结果验证、失败诚实等）。 |
| src/app/service/agent/core/compact_prompt.ts | 强化 compact 总结器对“中途修正指令”记录的优先级与逐字记录要求。 |
| const SECTION_SAFETY = `## Safety | ||
|
|
||
| - **Confirm before irreversible actions**: submitting forms, making purchases, deleting data, posting content. | ||
| - **Confirm before irreversible actions**: submitting forms, making purchases, deleting data, posting content, installing or modifying userscripts. For userscripts specifically, show the script's \`@match\` patterns and a summary of what it does before installing — a userscript runs on every matching page after installation and cannot be easily recalled. |
There was a problem hiding this comment.
主 Agent 的 Safety 段新增了 userscript 的不可逆操作确认要求，但子代理 prompt 里的 SUB_AGENT_SECTION_SAFETY 仍只覆盖表单/购买/删除/发帖等场景。因为实际执行点击/安装流程更可能发生在 page_operator 子代理里，这里缺口会导致子代理在收到“安装脚本”任务时缺少同等级的安全约束。建议同步把 “installing or modifying userscripts” 纳入子代理 Safety，并明确：子代理无法 ask_user 时应停止并把 @match + 功能摘要回报给父代理等待确认。
Suggested change
| - **Confirm before irreversible actions**: submitting forms, making purchases, deleting data, posting content, installing or modifying userscripts. For userscripts specifically, show the script's \`@match\` patterns and a summary of what it does before installing — a userscript runs on every matching page after installation and cannot be easily recalled. | |
| - **Confirm before irreversible actions**: submitting forms, making purchases, deleting data, posting content, installing or modifying userscripts. For userscripts specifically, show the script's \`@match\` patterns and a summary of what it does before installing — a userscript runs on every matching page after installation and cannot be easily recalled. If the executor cannot ask the user directly, it must stop and report the \`@match\` patterns plus the summary back to the parent agent for confirmation instead of proceeding. |
Comment on lines
17
to
+20
| 3. **User Messages** | ||
| - List ALL user messages that are not tool results | ||
| - These are critical for understanding the user's feedback and changing intent | ||
| - Include any mid-conversation corrections or preference changes | ||
| - List ALL user messages that are not tool results, in order | ||
| - **Mid-task corrections are highest priority** — if the user interrupted an ongoing operation with a correction (e.g. "stop", "do it differently", "that's wrong"), record these verbatim. These messages are the most commonly lost in long conversations and the most damaging to skip: a resumed agent will repeat the exact mistake that was already corrected. | ||
| - Include preference changes, clarifications, and any instruction that overrides an earlier one |
There was a problem hiding this comment.
buildCompactUserPrompt 的 Section 3 文本发生了语义增强（强调 mid-task corrections 最高优先级、要求逐字记录），但现有 compact_prompt.test.ts 只断言包含 8 个段落标题，并未覆盖这条关键约束。建议在对应测试文件中补充断言（例如包含 “Mid-task corrections are highest priority” / “record these verbatim” 等关键短语），避免后续 prompt 回归。
Member
Code reviewNo issues found. Checked for bugs and CLAUDE.md compliance. 🤖 Generated with Claude Code - If this code review was useful, please react with 👍. Otherwise, react with 👎. |
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
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.
背景
在对 agent 行为的实际观察中，发现四类系统性失效模式：
此外，compact 摘要器在长对话中频繁丢失用户中途修正指令，导致恢复后 agent 重复已纠正的错误；并行 sub-agent 有依赖关系时，下游 agent 在上游未成功的情况下也会静默继续执行。
变更内容
system_prompt.ts新增：sub-agent 结果接收规范
在
SECTION_SUB_AGENT的### Anti-Patterns之后插入### Receiving Sub-Agent Results段，明确要求：Issues字段，有问题则显式决策（重试 / 换 agent / 上报用户），不得静默并入新增：并行任务 fallback 指引
在
### Writing Sub-Agent Prompts末尾追加：若 sub-agent 依赖上游输出（如 OPFS 文件），必须在委托 prompt 中写明输入缺失时的 fallback 行为，不得假设上游已成功。扩展：不可逆操作确认清单
SECTION_SAFETY第一条在posting content后追加installing or modifying userscripts，并说明原因：userscript 安装后在所有匹配页面持续运行，安装前须展示@match模式与功能摘要供用户确认。sub_agent_types.tsSUB_AGENT_SECTION_TOOL_USAGE：budget 语义修正将旧措辞：
替换为：
明确 sub-agent budget 仅针对当前子任务，与主 agent 独立，消除"省 budget"导致的过早放弃行为。
researcher.systemPromptAddition：置信度分层输出末尾追加规范：输出中须区分三类信息——
不得将三者混入单一叙述，主 agent 需要可区分的置信度信号才能正确决策。
page_operator.systemPromptAddition：动作与结果分离末尾追加规范：「点击了提交按钮」与「表单已成功提交」是两个不同事实。每次操作后须通过
get_tab_content或execute_script验证结果，无法确认时如实说明，不得推断为成功。general.systemPromptAddition：选择透明度与失败诚实末尾追加规范：存在多种可行方案时须简述取舍理由；方案失败时报告为失败，不得包装成"部分成功"。
compact_prompt.tsbuildCompactUserPrompt：中途修正指令优先级Section 3
User Messages改为：影响范围
system_prompt.tssub_agent_types.tscompact_prompt.tssystem_prompt.test.ts所有改动均为 prompt 文本，不涉及架构调整、新 agent 类型或 TypeScript 运行时逻辑。
测试
新增断言已覆盖本次所有文本变更，包括：
### Receiving Sub-Agent Results段及关键措辞covers this subtask only）；旧措辞（Use them wisely）已断言不再出现installing or modifying userscripts及@match说明Mid-task corrections are highest priority及逐字记录要求