fix(agent): route reasoning model disconnects to timeout, not context_overflow

[AlexFucuson9]

Problem

When a reasoning model (o1/o3, DeepSeek R1, Nemotron, Grok reasoning, QwQ) hits a transport disconnect on a large session, the error classifier incorrectly classifies it as context_overflow with should_compress=True.

Reasoning models can think for minutes before producing tokens. A transport disconnect during thinking is almost always an upstream proxy idle-kill (NVIDIA NIM ~120s, OpenAI/Anthropic stream-idle), NOT a real context overflow. Classifying it as context_overflow triggers destructive compression on a phantom error, silently deleting conversation history.

Fix

Add reasoning model detection before the context_overflow classification at agent/error_classifier.py:718. For reasoning models, route to timeout instead so retries/fallback can handle the disconnect properly.

# Reasoning model prefixes that should skip context_overflow classification
_reasoning_prefixes = (
    "o1", "o3", "o4",
    "deepseek-r", "deepseek-reasoner",
    "nemotron", "qwq", "qwen3",
    "grok-4",
)

Files changed

File	Change
agent/error_classifier.py	Add reasoning model check before context_overflow classification (+21 lines)

Fixes #52271

[tonydwb]

Code Review Summary

Verdict: Approved

Clean fix for reasoning model error classification. Transport disconnects on reasoning models now correctly route to timeout instead of context_overflow, preventing destructive compression on phantom errors. Good explanation of the root cause.

Changes:

• agent/error_classifier.py: 21 additions — reasoning model detection before context_overflow classification

LGTM.

Reviewed by Hermes Agent

[kshitijk4poor]

Issue #52271 was fixed in #52795 (merged). This PR addresses Layer 1 of the same issue (classifier routing) but the full fix has landed. Thanks for the contribution!