Python: Bound tool result compaction summaries

[scarab-systems]

Motivation & Context

ToolResultCompactionStrategy replaces older tool-call groups with a synthetic [Tool results: ...] message, but that digest can include the full text of large tool results. In read-heavy workloads, the strategy can report that compaction ran while preserving most of the payload in the newly inserted message.

The digest should also respect existing exclusion state: if an earlier compaction pass already excluded a message from context, a later tool-result summary should not restore that payload.

This change keeps the existing digest behavior for small included results while preventing one generated summary message from carrying an unbounded payload or reintroducing already-excluded content.

Description & Review Guide

• What are the major changes? Build tool-result digest content from the included messages in the selected group, keep full-group provenance links, bound the synthesized summary text with a truncation marker, and add regression coverage for both large included results and already-excluded results.
• What is the impact of these changes? Large included tool-result digests no longer include the full payload verbatim, already-excluded tool results stay out of the inserted summary, and small-result summaries remain unchanged. There is no public API change.
• What do you want reviewers to focus on? Whether the boundary between digest content and provenance is right: digest text uses included messages only, while provenance still records the full group that was compacted.

Related Issue

Fixes #7393

Related: #7391/#7387 covers excluded-message restoration in the same digest path; this PR includes that guard because it affects the summary-building logic touched here.

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

🟡 Not ready to approve

The current digest-building logic can still reintroduce tool results from already-excluded messages into the new summary message, undermining exclusion semantics.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Pull request overview

This PR updates the Python ToolResultCompactionStrategy to prevent the synthesized [Tool results: ...] replacement message from embedding unbounded tool-result payload text, keeping small-result summaries unchanged while bounding large digests and adding a regression test.

Changes:

• Add a bounded summary-text helper (_summary_text) with a truncation marker and a 4096-character cap for tool-result digests.
• Update tool-result compaction to use the bounded helper when generating [Tool results: ...] messages.
• Add a regression test to ensure large tool-result payloads are not embedded verbatim and are truncated to the cap.

File summaries

File	Description
python/packages/core/agent_framework/_compaction.py	Adds bounded tool-result digest construction and routes ToolResultCompactionStrategy summaries through it.
python/packages/core/tests/core/test_compaction.py	Adds a regression test asserting large tool-result summaries are truncated and capped.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 2
• Review effort level: Low

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.

[Copilot]

🟡 Not ready to approve

There’s a correctness risk in exclusion filtering (is not True identity check) and a brittleness issue in the new truncation test that should be addressed before approval.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Review details

Comments suppressed due to low confidence (2)

python/packages/core/agent_framework/_compaction.py:903

• included_group_msgs checks the exclusion flag using is not True, which can incorrectly treat non-bool truthy values (e.g., numpy.bool_(True) or other truthy sentinels) as included. Elsewhere in this module (included_messages, _included_group_ids, set_excluded) exclusion is treated as a boolean via get(EXCLUDED_KEY, False). Align this filter to the same boolean semantics so excluded messages can’t leak into the digest.

            group_msgs = grouped.get(group_id, [])
            included_group_msgs = [msg for msg in group_msgs if msg.additional_properties.get(EXCLUDED_KEY) is not True]
            # Build a call_id → function_name map from function_call contents.

python/packages/core/tests/core/test_compaction.py:778

• The "large summary payload" test hard-codes the tool-result size via 1_000 lines. That makes the test brittle if _SUMMARY_MAX_CHARS is ever increased enough that this payload no longer triggers truncation, causing assertions like "[truncated]" in summary_text and "file-end" not in summary_text to fail even though the implementation is correct. Consider constructing large_result based on ToolResultCompactionStrategy._SUMMARY_MAX_CHARS so the test always exercises the truncation branch.

    """Summary text should not embed an oversized tool result verbatim."""
    large_result = "file-start\n" + ("line contents\n" * 1_000) + "file-end"
    messages = [

• Files reviewed: 2/2 changed files
• Comments generated: 0 new
• Review effort level: Low

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.

[Copilot]

🟢 Ready to approve

The changes directly address the reported compaction issues (unbounded summary payload and excluded-content restoration) with targeted implementation updates and clear regression tests.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0 new
• Review effort level: Low

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.