[Kip (uragoodman)]

Motivation & Context

ToolResultCompactionStrategy collapses old tool-call groups into a short
summary, but its result extraction only handles function_result,
mcp_server_tool_result, and search_tool_result. Groups containing
shell_tool_result, code_interpreter_tool_result,
image_generation_tool_result, or computer_tool_result are still excluded,
yet their summary says [Tool results: no results]. The tool output is silently
lost from the compacted conversation.

This is reachable through ContextWindowCompactionStrategy, which uses
ToolResultCompactionStrategy internally, as well as by callers using the
strategy directly.

Based on upstream main at commit ca65085 (agent-framework-core 1.24.0), merged into the branch after the first CI run.

Fixes #9886

The issue was filed with a complete reproduction. This PR is submitted
concurrently for early direction feedback; happy to adjust the approach or wait
if maintainers prefer to discuss the direction first.

Description & Review Guide

What are the major changes?

• Add result extraction for the four hosted result types in
  ToolResultCompactionStrategy.
• Add _hosted_tool_result_summary to render:
  • shell stdout/stderr/exit code/timed-out state;
  • code interpreter text, error, and status output;
  • image generation and computer results without embedding data URIs or
    base64 payloads;
  • an explicit no output returned fallback instead of no results when
    there is genuinely no result text.
• Add regression tests covering all four types, status-only/timeout cases,
  mixed hosted results, multiple outputs/errors, original-message exclusion,
  and payload exclusion.

What is the impact of these changes?

Only older tool-call groups containing these hosted result types change. The
function, MCP, and search paths are unchanged. No public API changes. Hosted
image payloads are represented by metadata (media_type, file_id, or
screenshot) rather than copied into the summary.

This PR intentionally scopes itself to ToolResultCompactionStrategy.
SummarizationStrategy has a related rendering gap and is left for a separate
follow-up.

What do you want reviewers to focus on?

• Whether labeling hosted results by content type (rather than call/tool name)
  is acceptable, since these content types have no separate tool name.
• Whether the metadata-only rendering for image/computer results and the
  explicit no output returned fallback are the right level of detail.

AI Assistance

• This is an AI-assisted contribution. I reviewed, understood, and verified all submitted content and accept responsibility for it.

AI assistance details: an AI coding assistant helped with code-path analysis,
implementation, tests, and drafting this PR description. I reviewed and
verified the change and ran the focused tests below.

Contribution Checklist

• The changed package imports cleanly and the focused compaction suite passes:
  pytest tests/core/test_compaction.py -q — 168 passed on Python 3.12.16.
• New regression tests were added and pass.
• ruff check agent_framework/_compaction.py tests/core/test_compaction.py
  passes with the repository-pinned Ruff 0.17.1.
• Full workspace build and the full unit-test suite were not run locally; CI will run them.
• This is not a breaking change.
• No ready-for-implementation label was present when this PR was opened; it is submitted for early direction feedback.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

[Kip (uragoodman)]

@microsoft-github-policy-service agree

[Kip (uragoodman)]

Updated to address both failing CI jobs, and merged main (ca65085e1) so the branch is current:

• Package Checks (pyright): _hosted_tool_result_summary built output_items with a comprehension over content.outputs. That attribute is typed list[Content] | Any | None, so isinstance(outputs, Sequence) narrows it to Sequence[Unknown] and the item was reported as Content | Unknown (reportUnknownVariableType). It now accumulates Content items explicitly, matching the existing cast("Sequence[Any]", ...) pattern used elsewhere in the module.
• Test Typing Checks (mypy/zuban): excluded was first assigned as [] in the parametrized regression test, so it needed an explicit annotation.

Locally verified with the CI commands: poe check-packages (core pyright: 0 errors) and ci-test-typing (core green under mypy, zuban, pyright, ty, and pyrefly; the only failures are POSIX-only symbols in the tools package that don't exist on Windows). The focused compaction suite passes, and the workflow/checkpoint tests pass when run serially — the parallel failures are pre-existing temp-file contention on Windows, with a different set failing each run.

Workflow runs are back to action_required after this push — could a maintainer re-approve them when convenient?

[Eduard van Valkenburg (eavanvalkenburg)]

Thanks for contributing this work. We are closing this PR while we plan #9886, which remains open. The issue form asks contributors to wait for explicit maintainer agreement before implementing a non-trivial change, and this PR was opened before that agreement. A /plan request is now pending to settle the hosted-result summary contract and scope. Please wait for that agreement before further implementation; the existing work can inform the plan, and a maintainer can coordinate the next step once the direction is settled. This closure is a process/planning decision, not a code-review assessment.