fix(model_metadata): parse OpenRouter/Nous 'in the output' format in output-cap errors #38652

[kyssta-exe]

Summary

only recognized Anthropic's keyword, causing it to return for OpenRouter/Nous errors that use the breakdown format. This led to the input-too-large recovery path (compression) being triggered instead of the correct output-cap reduction, resulting in infinite auto-reset loops.

Fix

Add a second detection strategy that:

1. Identifies errors mentioning alongside
2. Extracts the context limit from
3. Sums input tokens from patterns like and
4. Returns as the available output

Example

For the error:

Previously: returned → fell through to compression path → infinite loop
Now: returns (256000 - 5683 - 13410) → retries with

Tests

• All 27 existing tests in pass
• All 155 tests in pass
• Added manual verification for OpenRouter/Nous error format

Closes #38652

[teknium1]

This has already been fixed on main.

Automated hermes-sweeper review evidence:

• agent/model_metadata.py:1053 now recognizes the OpenRouter/Nous output-cap format by checking for "in the output" plus "maximum context length".
• agent/model_metadata.py:1086 parses the breakdown form (... of text input, ... of tool input, ... in the output) and returns context - text_input - tool_input as the available output budget.
• agent/conversation_loop.py:3040 consumes that parsed budget by setting _ephemeral_max_output_tokens and retrying without shrinking context_length, which is the recovery path this PR was aiming for.
• tests/test_output_cap_parsing.py:8 has regression coverage for the OpenRouter/Nous breakdown format.
• The implementation landed in 1fb99b1f229a700f8ee1a3e90ff1d44f85e963d4 via fix: guard empty streams + parse OpenRouter output-cap errors #40589 and is contained in tag v2026.6.19.

The linked issue #38652 was also already closed by a maintainer as fixed on main. Thanks for the original report and fix direction.