fix(agent): parse OpenRouter/Nous output-cap errors to prevent infinite reset loop

[liuhao1024]

What does this PR do?

Adds OpenRouter / Nous Research inference error format support to parse_available_output_tokens_from_error(). Previously, the function only recognized Anthropic's "available_tokens" keyword. When using OpenRouter-compatible providers, errors with "N in the output" format were misclassified as prompt-too-long errors, causing infinite session reset loops when max_tokens exceeded the available context.

Related Issue

Fixes #38652

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/model_metadata.py: Extended parse_available_output_tokens_from_error() to detect OpenRouter/Nous error format ("N of text input, M of tool input, K in the output") and calculate available output tokens as context_length - text_input - tool_input. Refactored detection into two phases: (1) identify error format, (2) extract available tokens.
• tests/test_ctx_halving_fix.py: Added 4 tests covering the OpenRouter/Nous format: canonical format with all components, format without tool input, format with only text input, and format without context length (returns None).

How to Test

1. Run pytest tests/test_ctx_halving_fix.py -v — all 31 tests should pass (4 new + 27 existing)
2. Run pytest tests/agent/test_model_metadata.py -v — all 93 tests should pass (no regressions)
3. Manual verification: configure a provider with max_tokens exceeding the context window and confirm the error recovery loop no longer occurs

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

Code Intelligence

• Analyzed: parse_available_output_tokens_from_error in agent/model_metadata.py (callers: conversation_loop.py:2989, anthropic_adapter.py:2119)
• Blast radius: LOW — pure parsing function, no side effects; callers already handle None return gracefully
• Related patterns: parse_context_limit_from_error() (sibling function, same file), _ephemeral_max_output_tokens (consumer in conversation loop)

[alt-glitch]

Duplicate of #38659 — both fix #38652 by extending parse_available_output_tokens_from_error() in agent/model_metadata.py to parse the OpenRouter/Nous "N in the output" error format. Same root cause and same code path; maintainers should pick one.

[liuhao1024]

Thanks for the flag @alt-glitch. Comparing the two PRs:

• fix: parse OpenRouter/Nous "in the output" error format in parse_available_output_tokens_from_error() #38659 adds OpenRouter/Nous format detection inline (85 additions, 2 files)
• This PR (fix(agent): parse OpenRouter/Nous output-cap errors to prevent infinite reset loop #38667) rewrites the function with phase-based structure and broader documentation (92 additions, 2 files)

The production code changes are functionally equivalent — both extend parse_available_output_tokens_from_error() to handle the OpenRouter/Nous "N in the output" format. #38659's approach is more minimal, which is generally preferred for upstream contributions. Closing in favor of #38659.