fix(core): 用户中断 reasoning 时丢弃未执行的工具调用

[hanydd]

fix: #2470

AgentScope-Java Version

2.0.1-SNAPSHOT

Description

修改概述

修复用户中断 ReActAgent 的 reasoning 阶段时，尚未执行的工具调用被写入 memory，导致后续请求携带孤立 tool_use、会话上下文无法恢复的问题。

问题原因

reasoning 产生的消息会在工具执行前写入 memory。用户中断恰好发生在工具调用流式生成期间，或 reasoning 已完成但 acting 尚未开始时，当前实现仍会保留其中的 ToolUseBlock。由于对应工具没有执行，memory 中不会产生配对的工具结果；重建 Agent 或发起下一轮模型请求时，严格校验消息历史的模型会拒绝这段上下文。

修复方案

• 用户中断 reasoning 时，仅持久化已经生成的文本和 thinking 内容，丢弃所有尚未执行的 ToolUseBlock。
• 将正常 reasoning 消息的完整持久化延后到 acting 前的中断检查之后，覆盖“工具调用已生成、但 acting 尚未开始”的窗口。
• 保持系统中断以及 PartialReasoningPolicy 的既有语义不变。

测试覆盖

• reasoning 流式生成部分工具参数时发生用户中断。
• reasoning 已生成完整工具调用、acting 开始前发生用户中断。
• 用户中断纯文本 reasoning 时保留已生成文本。
• 用户中断仅包含工具调用的 reasoning 时不持久化空消息。
• 系统中断分别使用 PartialReasoningPolicy.SAVE 和 DISCARD。
• gotoReasoning 时分别处理非空和空 reasoning 消息。
• 基于中断后的 memory 重建 Agent，并在严格历史校验下继续对话。
• 本地 JaCoCo 报告中，本 PR 修改的所有可执行行和条件分支均已覆盖。
• 完整执行 agentscope-core 测试：2228 tests，0 failures，0 errors，8 skipped。

Checklist

Please check the following items before code is ready to be reviewed.

• Code has been formatted with mvn spotless:apply
• All tests are passing (mvn -pl agentscope-core test)
• Javadoc comments are complete and follow project conventions
• Related documentation has been updated (not applicable; this fix does not change the public API)
• Code is ready for review

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[oss-maintainer]

Review: PR #2471

fix(core): 用户中断 reasoning 时丢弃未执行的工具调用

Summary

Fixes a correctness issue where user-interrupted reasoning could persist orphaned ToolUseBlock entries into the conversation history, making subsequent model calls fail due to unresolvable tool-use references.

Analysis

Root cause: The reasoning message (including any in-flight ToolUseBlock) was persisted to memory before the acting phase. If a user interrupt arrived during the window between tool-call generation and acting start, the tool call was written to state without a corresponding ToolResultBlock, leaving the conversation in an unrecoverable state.

Fix approach — correct and minimal:

1. persistInterruptedReasoningMessage() — new method that strips ToolUseBlock entries for USER interrupts while preserving text/thinking content. SYSTEM interrupts retain existing PartialReasoningPolicy semantics (unchanged).
2. Deferred persistence — moves the contextMutable().add(eventMsg) call to after checkInterrupted() in the pre-acting path, closing the race window.
3. withoutToolUseBlocks() — clean static helper that filters out tool-use blocks; returns null if nothing remains (avoiding empty message persistence).

Code quality:

• Clean separation between user and system interrupt handling
• No changes to public API or PartialReasoningPolicy contract
• Proper null handling throughout
• toList() (immutable) used correctly

Test coverage (420 lines, 7 scenarios):

• Partial streaming tool call during user interrupt
• Completed tool call before acting during user interrupt
• Text-only reasoning preserved on user interrupt
• Tool-call-only reasoning not persisted as empty message
• System interrupt with SAVE and DISCARD policies
• gotoReasoning with non-null and null reasoning messages
• Agent rebuild from interrupted state with strict history validation

All 2228 core tests pass. CI is fully green.

Verdict: Approved ✅ — well-scoped fix with thorough test coverage.