fix(agent): count codex_reasoning_items in tail-protection token estimate

[ms-alan]

Closes #55572

Summary

The tail-protection estimator _estimate_msg_budget_tokens() was counting message content and tool_calls but not codex_reasoning_items. For sessions running on the OpenAI Responses/Codex backend, those items carry the model's encrypted_content reasoning blobs — real wire payload replayed to the provider. The estimator was blind to them, so the protected tail under-measured Responses sessions by 100K+ tokens (27% of session size in observed case), causing compaction to trigger too late.

Root cause

codex_reasoning_items was simply absent from the token-counting loop in _estimate_msg_budget_tokens, despite being counted indirectly via full-dict stringify in the preflight estimator estimate_messages_tokens_rough.

Fix

Add iteration over codex_reasoning_items using the same len(str(it)) // _CHARS_PER_TOKEN formula already used for tool_calls:

Observed impact

A real 214-turn session carried 162 reasoning items totalling ~459 KB (~115K tokens) — none of which the old tail estimator counted. The protected tail (~20K-token budget) silently held ~100K+ extra wire tokens.

Note

Privacy is unaffected: _serialize_for_summary only reads content, tool_calls, and tool_call_id, so encrypted blobs are never sent to the auxiliary summarizer.

[alt-glitch]

This was generated by AI during triage.

Related to #45485 (the broader OPEN fix in the same _estimate_msg_budget_tokens tail-budget walk in agent/context_compressor.py). This PR adds only codex_reasoning_items to the count; #45485 already counts the fuller Codex/reasoning replay field set (reasoning_content/reasoning_details/codex_reasoning_items/codex_message_items) — a superset. A subset of a broader open PR is related, not a duplicate. Also related to the issue family #55572/#51800 and competing fix #51822. A human should pick the canonical fix (#45485 looks most complete).

[teknium1]

Thanks for the focused regression fix. This is already implemented on current main by a broader canonical fix.

• Automated hermes-sweeper review verified that agent/context_compressor.py:330-365 includes codex_reasoning_items in _REPLAY_BUDGET_KEYS and accounts for its serialized payload in _estimate_msg_budget_tokens().
• The same implementation also covers reasoning, reasoning_content, reasoning_details, and codex_message_items, matching the broader related work noted in the discussion.
• Commit 78ee0aa36703dad6b25a33303e5fd19c4b3bc0d7 was merged via PR fix(compression): count Codex replay fields in tail-protection budget (#55572) #59804; tests/agent/test_context_compressor.py:410-515 covers both combined and per-field replay-payload tail budgeting.
• The fix shipped in v2026.7.7.

Closing as implemented on main.