[hylin (linhongyu510)]

Title

fix(python): preserve non-adjacent tool pairs during history truncation

Motivation and Context

• Problem: ChatHistoryTruncationReducer could retain a FunctionResultContent while dropping its matching FunctionCallContent when the two were separated by non-function messages.
• Existing contract: adjacent tool blocks were already protected, and matched ID pairs are used elsewhere by the reducer utilities. A retained tool result should not become an orphan after truncation.
• Why this is minimal: the fix chooses the existing user-threshold candidate boundary first, then applies a final matched-ID pair closure.
• Scope: offline chat-history reduction only; no provider, network, E2E, public API, dependency, lockfile, or documentation-only refactor.
• Issue: self-discovered during audit; independent dedup found no equivalent open issue/PR.

Description

locate_safe_reduction_index now builds an O(n) suffix minimum-call map from get_call_result_pairs(history). After the existing adjacency and threshold-window candidate selection, it moves the cut back only when a retained result requires an earlier matched call. Chained non-adjacent pairs are closed transitively. If the required call is before offset_count, reduction returns None rather than crossing offset/summary territory.

Offset, threshold, system/developer preservation, unknown/unpaired content, and existing user-boundary search semantics are otherwise unchanged.

Local verification

• Base: 483ffb487a33a1644886992d56599feae3b69b76.
• Pre-fix formal regression: 4 failed, 4 passed.
• Post-fix formal regression: 8 passed.
• Production-only revert cycle: red exit 1, restore green exit 0.
• Related chat-history suite: 105 passed.
• Ruff: 0.9.10; check and format check pass.
• Mypy: targeted changed production file passes.
• Multi-Python matrix and wheel build were not run in this preparation step.
• AI assistance was used; an independent local review was rerun before publication. No external CLI review is claimed.

Contribution Checklist

• The code builds clean without any errors or warnings in the focused local checks.
• The PR follows the Semantic Kernel Contribution Guide and local Ruff/Mypy checks.
• All unit tests pass, and I added focused tests for the non-adjacent call/result regression.
• I did not break anyone.
• AI assistance is disclosed.

[hylin (linhongyu510)]

Rebased onto upstream main snapshot b5b526d9b8bd39da41db4a5c0f1aff20d26f7ec8 (2026-10-27).

Local verification on this rebased HEAD (4c489f2d44a9d41164d40c2b27aaf59333421b70), frozen lock env (uv 0.13.1, uv sync --frozen --all-extras --dev):

• Formal regression suite: 8 passed.
• Production-only revert (base bytes restored via try/finally): exactly 4 failed / 4 passed, exit 1; restored HEAD green 8 passed.
• Focused reducer suite: 43 passed.
• Related chat-history suite: 212 passed.
• Official type gate mypy -p semantic_kernel --config-file mypy.ini: no issues found in 556 source files (HEAD and base).
• Ruff check: clean.
• Official pre-commit (--all-files): 13 hooks passed (exit 0).
• Warnings: pytest-timeout config warning resolved by installing the locked plugin; Pydantic deprecation warnings are pre-existing on base (40 in focused, 52 in related suite).
• AI assistance was used and reviewed; this is an independent local verification, not a maintainer review. Multi-Python matrix / wheel build not run in this step.

[Copilot]

🟡 Changes recommended

History truncation can still orphan tool results because only the first matched result per message is collected and backward cut adjustments can split adjacent tool blocks.

2 open issues

• Collect all batched function results before calculating the cut · New
• Reapply tool-block adjacency after backward cut adjustments · New

What changed in this PR

This PR updates Python chat-history truncation to retain matching function calls when their results appear later, separated by other messages.

Changes:

• Adds matched-ID boundary adjustment after existing adjacency and user-threshold selection.
• Handles chained pair dependencies and rejects cuts that require crossing the offset boundary.
• Adds regression tests for pair preservation, thresholds, offsets, and system-message retention.

File	Description
python/​tests/​unit/​contents/​test_chat_history_truncation_pair_invariant.py	Adds focused truncation and pair-preservation regression tests.
python/​semantic_kernel/​contents/​history_reducer/​chat_history_reducer_utils.py	Adjusts truncation boundaries using suffix minima of matched call indices.

_{Review effort: Balanced}

[hylin (linhongyu510)]

Fixed in c814f99. Removed the early break so every FunctionResultContent item in a message is collected; del call_ids_seen[call_id] already prevents double-matching.

Two new test files: test_chat_history_multi_result_pairing.py (internal helper, 2 cases) and test_chat_history_truncation_reversed_results.py — a public ChatHistoryTruncationReducer regression where calls sit at different message indices (c2 at msg1, c3 at msg3) with results reversed in one tool message, orphaning the earlier call on base.

Local validation on the source-equivalent blob: focused 11 passed (3 new + 8 existing); base ablation 2 fail / 1 pass; restore 11 passed; full contents 428 passed, 66 warnings. Ruff passed on the three changed files; an independent typecheck of the changed production file passed. The earlier producer package-wide check covered 556 files (not independently rerun). Warning set unchanged from base. AI assistance was used.

[hylin (linhongyu510)]

Investigated the matched-pair invariant: for each matched FunctionResultContent retained in the tail, its paired call is also retained. The local check used one history shape (an adjacent call/result block plus a non-adjacent pair) at target_count 2/3/4/5, with zero matched-pair orphans.

These checks cover ID-matched FunctionResultContent pairs, not the text-only TOOL result in the example you provided. They therefore do not establish that adjacent-block preservation remains intact after the closure moves the cut backward. I have not yet validated that text-only TOOL case or made an adjacency fix for this point, and am leaving this conversation unresolved rather than treating the limited checks as evidence that the concern is invalid. AI assistance was used.