fix: #38652 - detect OpenRouter/Nous "in the output" format in parse_available_output_tokens_from_error()

[lfgogogo]

Summary

When using OpenRouter-compatible providers (Nous Research inference, OpenRouter), the parse_available_output_tokens_from_error() function fails to detect "output cap too large" errors, causing an infinite auto-reset loop.

This fix adds regex-based detection for the OpenRouter/Nous error format:

maximum context length is 256000 … 5683 of text input, 13410 of tool input, 262000 in the output

It calculates available_output = context_length - text_input - tool_input, allowing Hermes to auto-reduce max_tokens and retry instead of falling into the input-compression recovery path (which fails on fresh sessions).

Before: "Session auto-reset" loops forever because /new doesn't fix the root cause (the max_tokens config value).

After: Hermes detects the output-cap error, reduces max_tokens, retries successfully, and warns the user.

Changes

• agent/model_metadata.py — parse_available_output_tokens_from_error():
  • Broadened is_output_cap_error guard to also match \d+\s+in\s+the\s+output
  • Added OpenRouter/Nous regex extraction to compute available = context_length - text_input - tool_input
  • Anthropic-format extraction preserved as fallback

Closes #38652

🤖 Generated with Claude Code

[alt-glitch]

Duplicate of #38659 (first competing open fix for #38652, same file/function parse_available_output_tokens_from_error() in agent/model_metadata.py). Several open alternates also exist: #38689, #39175, #39957. Consolidating on one fix would help the maintainer pick. Flagging for dedup; closure goes through review.

[teknium1]

Closing as already implemented on main.

Automated hermes-sweeper review found the same OpenRouter/Nous output-cap parsing is already shipped:

• agent/model_metadata.py:1055 now recognizes the OpenRouter/Nous "in the output" + "maximum context length" error shape.
• agent/model_metadata.py:1086 computes available output from the breakdown: context_length - text_input - tool_input.
• tests/test_output_cap_parsing.py:8 covers the OpenRouter/Nous breakdown format.
• The implementation landed in 1fb99b1f229a700f8ee1a3e90ff1d44f85e963d4 (fix(stream+output-cap): guard empty streams and parse OpenRouter output-cap errors (#40589)) and is contained in v2026.6.19.

Thanks for the fix; the prior duplicate discussion was right that this needed consolidation, and the current main branch now covers the requested behavior.