fix: exclude reasoning_details from preflight token estimate (#73298)

[JonthanaHanh]

Summary

The reasoning_details field (OpenRouter/Anthropic thinking blocks + opaque cryptographic signature blobs) inflates the rough preflight token estimate by ~4x, causing premature compression on thinking models.

Providers do not bill these envelope bytes as prompt tokens — in a measured Kimi K3 session, reasoning_details held 2,124K chars vs 281K chars of actual thinking text. The estimator reported ~533K tokens when real prompt_tokens was ~140K, triggering compression at ~27% of the configured threshold.

Fix

Skip reasoning_details in both _estimate_message_chars and _estimate_message_tokens_without_images, alongside the existing _anthropic_content_blocks exclusion. This is the same treatment already applied to Anthropic content blocks (which also contain opaque metadata not billed as tokens).

Changes

• agent/model_metadata.py: Add reasoning_details to the skip set in both _estimate_message_chars() (L2788) and _estimate_message_tokens_without_images() (L2817)

Test Plan

• All 132 tests in tests/agent/test_model_metadata.py pass
• Syntax check passes

Fixes #73298

[alt-glitch]

This was generated by AI during triage.

Related: #72087 already changes the same estimator hunk with a broader, regression-tested Anthropic replay-field deduplication. This PR is the narrower reasoning_details-only repair for #73298; a maintainer should choose or consolidate the overlapping approaches.

[PRATHAMESH75]

Reviewed against #73298 — this correctly fixes the root cause. Excluding reasoning_details from the shadow in both _estimate_message_chars and _estimate_message_tokens_without_images (agent/model_metadata.py) drops exactly the ~2.1M-char signed envelope the issue identified as phantom weight, while reasoning_content (the ~281K of actual thinking text) is still counted — so the estimate lands near the ~188K reconstruction instead of the phantom ~533K. This is the issue's suggested direction #1, and updating _estimate_message_chars too is right since it feeds other budget paths.

Two notes for a reviewer:

• No regression test. Worth adding one: a message dict carrying a large reasoning_details blob should no longer inflate estimate_messages_tokens_rough / _estimate_message_tokens_without_images (assert the estimate with vs. without the blob is ~equal). That locks in the fix and guards the exact regression [Bug]: Preflight token estimate counts reasoning_details envelope at chars/4 — compression fires at ~27% of real usage on thinking models #73298 describes.
• Undercount direction is safe here. Fully excluding the envelope means the rough estimate no longer sees those bytes, but production evidence in the issue (real prompt_tokens ~140K while the envelope alone was 2.1M chars) shows the provider does not bill it at chars/4, and should_defer_preflight_to_real_usage prefers the real reading once one exists — so this won't cause the opposite (late-compression) failure.

LGTM once a test is added.

[teknium1]

Thanks for the focused preflight repair. The current-main premise is verified: agent/model_metadata.py:3005-3027 places reasoning_details into the estimator shadow, and agent/conversation_loop.py:1764 uses that estimate for request pressure.

Problems

• The PR changes both estimator helpers but adds no regression coverage. A future refactor could reintroduce the signed-envelope inflation without failing tests/agent/test_model_metadata.py.

Suggested changes

• Add a behavioral test comparing otherwise identical messages with and without a large reasoning_details payload. Assert that estimate_messages_tokens_rough is unchanged; directly test _estimate_message_chars too, since this diff changes it.
• The member comment correctly notes overlap with #72087, which edits the same estimator hunk with an Anthropic-specific deduplication strategy. Consolidate the intended behavior before landing either patch.

Automated hermes-sweeper review.

[teknium1]

Please add a regression test showing that a large reasoning_details envelope leaves estimate_messages_tokens_rough unchanged while ordinary counted reasoning content still affects the result. This line changes the production preflight path but currently has no direct coverage.

[teknium1]

Merged via salvage PR #75884 (#75884) — your preflight estimator fix was cherry-picked with authorship preserved (composed into main's newer _wire_message_shadow() helper during rebase), plus a companion commit covering the second site @israellot documented (_REPLAY_BUDGET_KEYS tail-budget walk), counting thinking text but excluding signature envelopes per the #51800 discussion. Fixes #73298. Thanks!