fix(autocontextmemory):large message compression format adaptation

[shiyiyue1102]

This pull request enhances the handling of large tool messages during the auto-context memory compression process to prevent model errors caused by tool message structures. It introduces logic to convert tool messages into plain text format before sending them for compression, and then reconstructs the appropriate tool message structure after summarization. Additionally, comprehensive tests are added to verify this behavior.

Improvements to tool message handling and compression:

• In AutoContextMemory.java, large tool messages are now converted to plain text format before being sent to the model for compression, avoiding errors related to tool call/result detection during model inference. After summarization, the original tool message structure is preserved in the compressed message. [1] [2] [3] [4] [5]
• A new helper method convertToolMessageToText is added to perform the conversion of tool messages to a text-based format suitable for model input.
• The signature of generateLargeMessageSummary is updated to accept both the converted message and the original message, ensuring the original structure can be restored after compression.

Testing and verification:

• A new test, testLargeToolMessageConvertedToTextDuringCompression, is added to verify that large tool messages are correctly converted to text during compression and that the model receives them in the expected format.
• A CapturingTestModel test double is introduced to capture and inspect messages sent to the model during tests, ensuring the conversion logic works as intended.

Minor changes:

• Import for ContentBlock added to support the new logic.
• Minor whitespace cleanup in pom.xml.## AgentScope-Java Version

[The version of AgentScope-Java you are working on, e.g. 1.0.11, check your pom.xml dependency version or run mvn dependency:tree | grep agentscope-parent:pom(only mac/linux)]

Description

[Please describe the background, purpose, changes made, and how to test this PR]

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 test)
• Javadoc comments are complete and follow project conventions
• Related documentation has been updated (e.g. links, examples, etc.)
• Code is ready for review

[Copilot]

Pull request overview

This PR adjusts AutoContextMemory’s large-message compression flow to avoid model errors triggered by tool-call/tool-result message structures by converting tool messages to plain text for summarization and then reconstructing tool message structures afterward.

Changes:

• Detect large messages via character count across all content-block types (including tool blocks) and summarize them with the model.
• Convert tool messages into plain-text messages for the compression prompt context, then rebuild tool-structured messages post-summarization.
• Add tests with a capturing model double to inspect messages sent to the model during compression.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 7 comments.

File	Description
agentscope-extensions/agentscope-extensions-autocontext-memory/src/main/java/io/agentscope/core/memory/autocontext/AutoContextMemory.java	Adds tool-message-to-text conversion for compression context and reconstructs tool message structures after summarization.
agentscope-extensions/agentscope-extensions-autocontext-memory/src/test/java/io/agentscope/core/memory/autocontext/AutoContextMemoryTest.java	Adds a capturing Model test double and a new test covering tool-message conversion during large-message compression.
agentscope-examples/quickstart/pom.xml	Minor whitespace cleanup.

[Copilot]

In the ToolUseBlock branch, the generated LLM summary (finalContent, including the offload tag) is never incorporated into the returned message. As written, a “summarized” large ToolUse message will keep the original ToolUseBlock(s) unchanged (potentially still huge) and provides no text for the model to see or for context_reload discovery. Consider embedding the summary/offload hint into the returned content (e.g., add a TextBlock alongside a minimized ToolUseBlock, or return a pure text Msg for tool-use summaries) so size is actually reduced and reload remains possible.

Suggested change

	// Preserve ToolUseBlock structure with compressed summary
	List<ContentBlock> compressedBlocks = new ArrayList<>();
	for (ContentBlock originalBlock : originalMessage.getContent()) {
	if (originalBlock instanceof ToolUseBlock) {
	ToolUseBlock originalToolUse = (ToolUseBlock) originalBlock;
	// Keep original tool use block (tool calls are usually not too large)
	// Preserve ToolUseBlock structure and include the summary/offload hint as text
	List<ContentBlock> compressedBlocks = new ArrayList<>();
	if (!finalContent.isEmpty()) {
	compressedBlocks.add(TextBlock.builder().text(finalContent).build());
	}
	for (ContentBlock originalBlock : originalMessage.getContent()) {
	if (originalBlock instanceof ToolUseBlock) {
	ToolUseBlock originalToolUse = (ToolUseBlock) originalBlock;
	// Keep original tool use block while surfacing the compressed summary to the model

[Copilot]

When rebuilding a summarized ToolResult message, this loop only retains ToolResultBlock instances and drops any other ContentBlocks that may be present in the original message (e.g., TextBlocks before/after the tool result, which exist elsewhere in the codebase). This risks losing information during compression. Consider preserving non-ToolResult blocks (or at least appending them) while replacing only the ToolResultBlock.output with the compressed summary.

[Copilot]

The rebuilt ToolResultBlock does not preserve originalToolResult.getMetadata() (contrast with the offload path later in this class that explicitly keeps ToolResult metadata). This can drop semantic flags like agentscope_suspended and other provider metadata. Preserve the original ToolResultBlock metadata when constructing the compressed ToolResultBlock.

Suggested change

	// Replace output with compressed summary
	ToolResultBlock compressedToolResult =
	ToolResultBlock.builder()
	.id(originalToolResult.getId())
	.name(originalToolResult.getName())
	.output(List.of(TextBlock.builder().text(finalContent).build()))
	// Replace output with compressed summary while preserving tool result metadata
	ToolResultBlock compressedToolResult =
	ToolResultBlock.builder()
	.id(originalToolResult.getId())
	.name(originalToolResult.getName())
	.output(List.of(TextBlock.builder().text(finalContent).build()))
	.metadata(originalToolResult.getMetadata())

[Copilot]

For ToolUseBlock conversion, this uses toolUse.getContent() (raw streaming content) to populate “Arguments”, which is often null/empty for normal tool calls. This means the converted text sent to the model can omit the actual tool inputs (toolUse.getInput()). Consider serializing getInput() (and optionally also getContent() when present) so compression has the arguments needed for an accurate summary.

Suggested change

	if (toolUse.getContent() != null && !toolUse.getContent().isEmpty()) {
	textContent.append("Arguments: ").append(toolUse.getContent()).append("\n");
	}
	Object toolInput = toolUse.getInput();
	String rawContent = toolUse.getContent();
	if (toolInput != null) {
	textContent.append("Arguments: ").append(String.valueOf(toolInput)).append("\n");
	}
	if (rawContent != null && !rawContent.isEmpty()
	&& (toolInput == null || !rawContent.equals(String.valueOf(toolInput)))) {
	textContent.append("Raw Content: ").append(rawContent).append("\n");
	}

[Copilot]

ToolResult conversion only extracts output from TextBlock instances. Tool outputs can contain other ContentBlock types (e.g., images), which would be silently omitted from the compression context and can degrade summarization. Consider adding a fallback representation for non-TextBlock outputs (type name, URL/metadata, or toString()), or at least a placeholder indicating omitted non-text content.

[Copilot]

This test computes foundConvertedToolMessage but never asserts it, so the test can pass even if no tool message was converted (it only asserts callCount > 0, twice). Add an assertion that foundConvertedToolMessage is true (and remove the redundant second callCount assertion) to ensure the test actually validates the conversion behavior.

[Copilot]

The comment // disable LLM compression is misleading here: minCompressionTokenThreshold only gates Strategy 1 (tool invocation compression). Strategy 5 (large message summarization) still calls the model regardless. Update the comment (or the config) so the test setup accurately reflects what is being disabled/enabled.

Suggested change

	.minCompressionTokenThreshold(Integer.MAX_VALUE) // disable LLM compression
	.minCompressionTokenThreshold(
	Integer.MAX_VALUE) // disable Strategy 1 token-threshold gating only

[codecov]

Codecov Report

❌ Patch coverage is 57.14286% with 36 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...ope/core/memory/autocontext/AutoContextMemory.java	57.14%	24 Missing and 12 partials ⚠️

📢 Thoughts on this report? Let us know!

[LearningGp]

Msg has already been processed by convertToolMessageToText in the outer method. No need for redundant processing.

[LearningGp]

This drops the content and only retains the tool use.