fix(core): normalize tool call result transcripts

[Zbhbb]

AgentScope-Java Version

Current main branch.

Description

Fixes #2191.

Background

After an interruption, a tool-use block can remain in history while its tool result is missing or appears later in the transcript. Providers that require adjacent tool calls and results reject that payload.

Changes

• Normalize outbound model transcripts without mutating stored agent state.
• Move known tool results next to their owning assistant tool calls.
• Insert a terminal error result only when no result exists, so the next model call can recover safely.
• Apply the normalization to both normal ReAct reasoning and summary generation.

Validation

• mvn -q -pl agentscope-core -am -Dtest=MessageUtilsToolTranscriptTest,HookStopAgentTest,ReActAgentNewLoopE2ETest test
• mvn -q -T1 -pl agentscope-core -am verify

The full repository verify ran through its Java test phase, but the unrelated agentscope-builder step could not download npm 10.9.2 because the npm registry TLS handshake was terminated.

Checklist

• Code has been formatted with mvn spotless:apply rules
• Relevant tests are passing
• Javadoc follows existing conventions
• Documentation update is not needed
• Code is ready for review

[Copilot]

Pull request overview

This PR addresses provider rejections (e.g., OpenAI/Anthropic 400) caused by mismatched tool-call/tool-result ordering after interruptions by normalizing the outbound transcript (without mutating stored agent state) before sending messages to the model.

Changes:

• Add MessageUtils.normalizeToolCallResults(...) to relocate tool results adjacent to their owning assistant tool calls and synthesize a terminal error tool result when missing.
• Apply transcript normalization to both the main ReAct reasoning call path and the summarization call path in ReActAgent.
• Add/extend tests to validate transcript normalization behavior and ensure agent state isn’t mutated by normalization.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
agentscope-core/src/main/java/io/agentscope/core/util/MessageUtils.java	Introduces transcript normalization logic for tool results/tool calls.
agentscope-core/src/main/java/io/agentscope/core/ReActAgent.java	Applies normalization before model calls in reasoning + summary.
agentscope-core/src/test/java/io/agentscope/core/util/MessageUtilsToolTranscriptTest.java	Adds unit tests covering tool transcript normalization scenarios.
agentscope-core/src/test/java/io/agentscope/core/agent/ReActAgentNewLoopE2ETest.java	Adds an E2E assertion that normalization occurs before model invocation and does not mutate stored history.

[Zbhbb]

Fixed in 5510e8c. Replaced the fully-qualified IntStream expression with a simple loop to keep the test readable.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[oss-maintainer]

LGTM. Looks good.

---

Automated review by "github-manager-bot"

[AgentScopeJavaBot]

🤖 AI Review

This PR adds a model-agnostic normalization layer for tool call/result message sequences before they're sent to LLM providers. It addresses a real problem: various LLM APIs reject or misbehave when tool call/result transcripts have structural issues. The normalization is inserted at both model call sites in ReActAgent (reasoning and summarizing), and is backed by 98 comprehensive tests.

[AgentScopeJavaBot]

🤖 AI Review

This PR adds a model-agnostic normalization layer for tool call/result message sequences before they're sent to LLM providers. It addresses a real problem: various LLM APIs reject or misbehave when tool call/result transcripts have structural issues. The normalization is inserted at both model call sites in ReActAgent (reasoning and summarizing), and is backed by 98 comprehensive tests.

[AgentScopeJavaBot]

🤖 AI Review

This PR adds a model-agnostic normalization layer for tool call/result message sequences before they're sent to LLM providers. It addresses a real problem: various LLM APIs reject or misbehave when tool call/result transcripts have structural issues. The normalization is inserted at both model call sites in ReActAgent (reasoning and summarizing), and is backed by 98 comprehensive tests.