fix: parse OpenRouter/Nous "in the output" error format in parse_available_output_tokens_from_error()

[Xeron2000]

Summary

parse_available_output_tokens_from_error() in agent/model_metadata.py only recognized Anthropic's output-cap error format ("available_tokens: N"). OpenRouter-compatible providers (Nous Research, OpenRouter) use a different format:

5683 of text input, 13410 of tool input, 262000 in the output

The function returned None for these errors because:

1. The is_output_cap_error guard required literal "max_tokens" — OpenRouter format doesn't contain it
2. No extraction path existed for computing available = context - text_input - tool_input

Bug impact

On fresh sessions:

1. User configures max_tokens larger than context_length - input_tokens
2. API returns 400 → classified as context_overflow ✅
3. parse_available_output_tokens_from_error() returns None → falls through to input-too-large compression path ❌
4. Nothing to compress on fresh session → returns "Context length exceeded (4,882 tokens). Cannot compress further."
5. Gateway triggers Session auto-reset
6. Next message → same error → infinite loop

User sees repeated "Session auto-reset" and /new has no effect because the actual fix is reducing max_tokens, not clearing history.

Changes

• Guard: is_output_cap_error now also matches '(\d+) in the output' as an output-cap error indicator
• Extraction: New regex path computes available_output = context_length - text_input - tool_input from the OpenRouter/Nous error format
• Tests: Added 7 tests covering Anthropic format, OpenRouter/Nous format, and negative cases

Fixes #38652

[tonydwb]

Review of PR #38659

Title: fix: parse OpenRouter/Nous "in the output" error format in parse_available_output_tokens_from_error()

Looks Good

• Clean diff (85 additions / 1 deletions across 2 files).

---

Reviewed by Hermes Agent

[teknium1]

This appears to be implemented on current main by a later merged fix.

Automated hermes-sweeper review evidence:

• agent/model_metadata.py:1053 now recognizes the OpenRouter/Nous output-cap phrasing via "in the output" plus "maximum context length".
• agent/model_metadata.py:1086 parses maximum context length is N ... (A of text input, B of tool input, C in the output) and computes available = context - text - tool.
• tests/test_output_cap_parsing.py:5 covers the OpenRouter/Nous breakdown format and the existing Anthropic available_tokens format.
• The implementation landed in 1fb99b1f229a700f8ee1a3e90ff1d44f85e963d4 (fix(stream+output-cap): guard empty streams and parse OpenRouter output-cap errors) and is contained in v2026.6.19.

Closing as implemented on main. Thanks for the clear report and PR.