Python: exclude encrypted reasoning payloads from compaction token counts

[Lubaoshuai]

Motivation & Context

The compaction token counter serialises each message to JSON and tokenises the string. That serialisation included provider reasoning payloads: opaque encrypted blobs (OpenAI Responses encrypted_content, Anthropic thinking signature) and the JSON-serialised reasoning_details stored by the Chat Completions client. Encrypted blobs are replayed opaquely and never tokenised by the model, so they inflated local token estimates and could trigger premature compaction (#8348).

Description & Review Guide

• What are the major changes?
  • _serialize_content excludes provider-encrypted reasoning payloads from the token-estimation serialisation:
    • Opaque protected_data / additional_properties["encrypted_content"] blobs are dropped.
    • JSON reasoning_details (Chat Completions) keep their clear-text members (summary, reasoning_text, nested reasoning.text) — the client replays those to the provider as input, so only their opaque encrypted_content members are excluded (review follow-up; previously the whole structure was dropped).
  • Group token-count annotations are stamped with a serialization basis version (token_count_basis). Counts cached under an older basis — e.g. sessions annotated before this change, since annotations survive Message.to_dict()/from_dict() while annotate_token_counts skips messages that already have a count — are recomputed instead of reused (review follow-up).
• What is the impact of these changes?
  • Token estimates match the text the provider actually processes; transcripts annotated by older versions are lazily re-tokenised on their next compaction pass instead of carrying inflated counts that could truncate valid history or fire summarisation early.
• What do you want reviewers to focus on?
  • The recursive opaque-member filter (_strip_opaque_reasoning_payload) and the basis-version handling in _write_group_annotation / _token_count / _write_token_count.

Related Issue

Fixes #8348

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]

🔵 Needs a closer look

Previously cached token counts can remain inflated and must be invalidated or recomputed.

Pull request overview

This PR prevents opaque encrypted reasoning payloads from inflating Python compaction token estimates.

Changes:

• Excludes protected_data and additional_properties["encrypted_content"].
• Adds regression tests for both payload locations.
• Cached token-count annotations still require invalidation or recomputation.

File summaries

File	Summary
python/packages/core/tests/core/test_compaction.py	Adds regression tests for both encrypted payload locations.
python/packages/core/agent_framework/_compaction.py	Filters encrypted reasoning payloads during token serialization.

Review details

Suppressed comments (1)

python/packages/core/agent_framework/_compaction.py:707

• This only changes newly computed counts. Token counts are cached in the _group annotation, and annotate_token_counts skips messages whose count is already present; those annotations are also preserved by Message.to_dict()/from_dict(). A transcript annotated before this fix (or one whose reasoning payload is added after annotation) will therefore continue to use the inflated count and can still trigger premature compaction. Please invalidate/version the cached count when the serialization basis changes, or otherwise force a retokenization for such stale annotations.

    payload.pop("protected_data", None)

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite (auto)

Note

Copilot is running an experiment and ran this review at Lite.

---

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

[Lubaoshuai]

@microsoft-github-policy-service agree

[Eduard van Valkenburg (eavanvalkenburg)]

Thanks for the update. Before this is ready, could you please:

• follow up on and resolve the remaining review discussions: python/packages/core/agent_framework/_compaction.py:707, python/packages/core/agent_framework/_compaction.py:707
• get the non-green checks passing: team_check (CANCELLED), review (CANCELLED)
• replace the PR body with the current repository template, retain its four required sections, provide substantive Motivation and Description answers, and complete the applicable checklist items

Once those are addressed, please re-request review. Thanks!

[Lubaoshuai]

Thanks Eduard van Valkenburg (@eavanvalkenburg) Evan Mattson (@moonbox3) — all three points are addressed in efb44d8:

1. Remaining review discussions: both resolved above (clear-text reasoning preserved via recursive encrypted_content stripping in _serialize_content; stale group token counts invalidated via the new token_count_basis version stamp, so pre-change sessions are lazily re-tokenized).
2. Checks: re-triggered by the new push; watching for green.
3. PR body: rewritten per the repository template — all four sections present, checklist completed.

Tests: the two original regression tests remain, plus two new ones (clear-text reasoning counted while opaque members are excluded; stale-basis counts recomputed and re-stamped). pytest packages/core/tests/core/test_compaction.py: 101 passed; ruff check / ruff format clean.

Re-requesting review.

[Evan Mattson (moonbox3)]

Please fix the failing CI/CD checks.

[Eduard van Valkenburg (eavanvalkenburg)]

Thanks for the update. The current head is not green: Package Checks and Test Typing Checks fail on pyright/mypy errors in _compaction.py and test_compaction.py (unknown dictionary key/value types and an unsupported comparison involving an optional integer), and Merge Gatekeeper fails consequently. Could you please fix the typing failures and re-request review? Thanks!

[Lubaoshuai]

Thanks Eduard van Valkenburg (@eavanvalkenburg) Evan Mattson (@moonbox3) — the pyright/mypy failures are fixed in 48bde32:

• _strip_opaque_reasoning_payload iterates explicitly cast dict[Any, Any] / list[Any] sources — the isinstance narrowing on Any was leaking Unknown key/item types under pyright strict ("unknown dictionary key/value types").
• additional_properties is cast to dict[str, Any] before the encrypted_content filter (same root cause).
• The clear-text regression test now asserts non-None counts before the ordering comparison ("unsupported comparison involving an optional integer").

Local verification: pyright agent_framework/_compaction.py → 0 errors; pytest packages/core/tests/core/test_compaction.py → 111 passed; ruff check / ruff format clean. The branch was also rebased on the current main merge.

One note: tests/core/test_agent_hooks.py has a pre-existing Import "agent_hooks" could not be resolved in the static sweep — the module is created by the test fixture at runtime, the file is untouched by this PR and the error appears on main identically.

CI is re-running on the new head; re-requesting review.

[Eduard van Valkenburg (eavanvalkenburg)]

Lubaoshuai Reviewed the complete diff and found no actionable issues. All reported checks for the current head are green.