Python: fix(core): include tool trajectory details in summarizer input

[JHf0912]

Problem

SummarizationStrategy feeds the summarizer LLM a transcript rendered by _format_summary_message (python/packages/core/agent_framework/_compaction.py), which only uses Message.text. Since Message.text (python/packages/core/agent_framework/_types.py) concatenates TextContent only, tool rounds collapse to placeholders like 2. [assistant] function_call and 3. [tool] function_result. The summarizer therefore never sees the tool name, arguments, results, exceptions, or call_id, and produces summaries that silently discard the tool trajectory.

Fixes #8086

Solution

Add a per-content renderer _format_summary_content that serializes tool trajectory contents into the summary input transcript, and route _format_summary_message through it:

• function_call → function_call <name>(<arguments>) [call_id=<id>]
• function_result → function_result: <result> [call_id=<id>] (with error(<exception>) prefix when the call failed)
• mcp_server_tool_call / mcp_server_tool_result → same shape with the MCP tool name and output
• function_approval_request / function_approval_response → nested call name, approval id, and decision

Design decisions:

• Zero public API change. Only private formatting helpers in _compaction.py change; SummarizationStrategy's trigger conditions, summary message shape, and trace links are untouched.
• Group-atomic input selection unchanged. _select_summary_input_groups still selects whole groups by token budget; enriched groups simply carry their true payload, so the budget now reflects what the summarizer actually receives.
• Legacy rendering preserved for text-only messages (byte-identical, locked in by a regression test).

Before → after (real output from the end-to-end reproduction):

# before
2. [assistant] function_call
3. [tool] function_result

# after
2. [assistant] function_call get_weather({"city":"Seattle"}) [call_id=call_1]
3. [tool] function_result: sunny, 22C [call_id=call_1]

Changes

• python/packages/core/agent_framework/_compaction.py
  • New _format_summary_content dispatch over tool-call / result / MCP / approval content types (reuses existing _tool_result_text).
  • _format_summary_message now combines structured renderings with Message.text, falling back to the legacy content-type list only when nothing else is available.
• python/packages/core/tests/core/test_compaction.py
  • 8 new tests: function call details, results with exceptions, results without call_id, mixed text/tool messages, text-only golden rendering, MCP tool details, approval request, approval response.

Testing

cd python/packages/core
uv run pytest tests/core/test_compaction.py -q   # 87 passed
uv run pytest tests/core -q                       # full core suite: passed
uv run ruff check agent_framework/_compaction.py tests/core/test_compaction.py
uv run poe pyright                               # no new errors in touched files

• End-to-end: a SummarizationStrategy run over a 6-message conversation with two tool rounds shows the summarizer input containing names, arguments, results, and call_ids, and the summary message replacing the excluded originals with trace links intact.
• Regression anchors: existing budget/selection tests (test_summarization_strategy_bounds_summary_input_to_complete_groups, test_summary_input_selection_does_not_retokenize_selected_transcript) pass unchanged; group-atomic selection semantics are preserved.

Notes for Reviewer

• Token budget: enriched groups are larger, so a budgeted summarization round may select fewer groups than before. This is intentional — the payload was always part of the context; it was just invisible to the summarizer. Flagged for review in case maintainers prefer an include_tool_details opt-in/opt-out.
• Trust boundary: tool arguments/results now reach the summarizer LLM. The class docstring already warns that the summarizer must be trusted as much as the primary model; this PR widens the data surface sent to it. A redaction hook could be a follow-up.
• Out of scope (deliberately): text_reasoning rendering, structured JSON summary output, and no-call_id adjacency pairing in group_messages are left unchanged; the last touches pairing semantics covered by spec docs/specs/004-python-function-calling-loop.md and is proposed separately.
• Orthogonal to open PR Python: fix: preserve synthetic compaction summaries #7944 (synthetic summary reconciliation in before_run).

[Copilot]

🟡 Changes recommended

A critical crash path and two moderate rendering defects remain unresolved.

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

Pull request overview

Adds structured function, MCP, and approval trajectory details to Python summarizer transcripts.

Changes:

• Renders tool calls, results, exceptions, IDs, and approval decisions.
• Preserves text-only formatting.
• Adds focused regression tests.

Required fixes:

• Critical (1 vote): Safely stringify non-JSON-serializable MCP result mappings to prevent compaction crashes.
• Moderate (2 votes): Preserve chronological ordering of text and structured content.
• Moderate (1 vote): Use tool_name when rendering approvals wrapping MCP calls.

File summaries

File	Description
python/packages/core/tests/core/test_compaction.py	Adds formatter regression coverage.
python/packages/core/agent_framework/_compaction.py	Adds structured summary rendering; contains the unresolved findings above.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 3
• Review effort level: Balanced

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[JHf0912]

Addressed the three findings in 6b5360d:

1. Crash path: non-JSON-serializable MCP result mappings now use the same \json.dumps(..., default=str)\ with \str()\ fallback as \Content.from_function_result, so _select_summary_input_groups\ can no longer raise.
2. Time order: _format_summary_message\ now emits consecutive text blocks in place while iterating \message.contents, preserving temporal order for mixed messages; pure-text rendering stays byte-identical.
3. MCP approval identity: nested calls in approval contents fall back to \ ool_name\ when
   ame\ is absent.

Added three regression tests (non-JSON MCP output, mixed-content ordering, MCP approval tool name); all 90 tests in \ ests/core/test_compaction.py\ pass, ruff clean.

copilot-pull-request-reviewer please re-review.

[Ricky-7-Yan]

Verified the current 6b5360da head independently against the exact two-file diff:

• the complete tests/core/test_compaction.py suite passes;
• Ruff check and format check pass for both changed files;
• the three earlier findings are resolved: mapping outputs now use the established default=str fallback, mixed text/tool contents retain their original order, and MCP approvals use tool_name when name is absent.

I also checked compatibility with the existing Message.text behavior for text-only, consecutive-text, text-plus-unrendered-content, and non-text fallback cases. No new blocking issue found.

[Evan Mattson (moonbox3)]

JHf0912 please respond to each open PR review comments as to whether they're addressed or not.

[JHf0912]

Thanks Evan Mattson (@moonbox3) — done. I've replied to all three open review threads. All three are addressed in 6b5360d, each with a dedicated regression test:

1. Crash path (non-JSON-serializable MCP result mappings) — _tool_result_text\ now uses \json.dumps(..., default=str)\ with a \str()\ fallback, matching \Content.from_function_result.
2. Time order for mixed content — _format_summary_message\ emits consecutive text blocks in place while iterating \message.contents; pure-text rendering stays byte-identical.
3. MCP approval identity — nested calls now fall back to \ ool_name\ when
   ame\ is absent.

All 90 tests in \ ests/core/test_compaction.py\ pass; ruff clean.