What does this PR do?

Passes config_context_length to all get_model_context_length() call sites so that custom OpenAI-compatible endpoints (e.g., self-hosted, Infini-AI) that don't expose context_length via their /models API show and use the correct context window from custom_providers[].models..context_length in config.yaml.

Problem

get_model_context_length() accepts a config_context_length parameter that lets callers pass an explicit context length from custom_providers[].models.<model>.context_length in config.yaml. However, several display and utility call sites never pass this value, causing the function to fall through to the 128K default when custom endpoints don't return context_length from their /models API.

The AIAgent.__init__ path in run_agent.py correctly resolves context_length from custom_providers, but these call sites miss it:

The remaining 6 call sites already pass config_context_length:

• gateway/run.py L4067: already passes _hyg_config_context_length
• run_agent.py L1513: already passes _config_context_length
• run_agent.py L1800: already passes getattr(self, "_config_context_length", None)
• run_agent.py L2014: already passes _aux_context_config
• agent/context_compressor.py L305: constructor parameter, already received
• hermes_cli/web_server.py L624: intentionally passes None (wants auto-detected value)

Duplicate Check

I searched existing upstream PRs for the exact code path and issue terms before opening:

• config_context_length
• get_model_context_length passthrough
• custom_providers context_length
• 128K default context

I did not find an open PR that fixes this specific passthrough gap.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Security fix
• Documentation update
• Tests (adding or improving test coverage)
• Refactor (no behavior change)

Changes Made

• gateway/run.py: Added config_context_length=getattr(self, "_config_context_length", None) at L3777 (@context expansion path).
• gateway/run.py: Added inline custom_providers resolution at L4718 (/modelinfo display) and L5577 (/model switch display) since these paths don't have self._config_context_length cached. The resolution logic mirrors AIAgent.__init__.
• cli.py: Added config_context_length=getattr(self, "_config_context_length", None) at L4750, L4979, and L7934.
• run_agent.py: Added config_context_length=getattr(self, "_config_context_length", None) at L6437 (fallback model switch path).

Impact

@context reference expansion (gateway + CLI)

@context expansion uses the returned context length to set injection limits (hard_limit = context_length * 0.50, soft_limit = context_length * 0.25). With the 128K default, a @context reference on a 200K model hits the hard limit at 64K tokens — 36K tokens short of the correct 100K limit. File content beyond 64K tokens is silently refused injection.

Fallback model switch (run_agent.py)

When the primary model fails and the agent falls back, get_model_context_length() at L6437 is called without config_context_length. If the fallback model is also a custom endpoint, the context compressor updates to an incorrect 128K threshold, causing premature compression.

Display-only (no functional impact)

/modelinfo and /model switch displays show 128K instead of the configured 200K — misleading but not functionally harmful.

How to Test

1. Configure a custom OpenAI-compatible endpoint that does not return context_length from its /models API:

   # config.yaml
   custom_providers:
     - base_url: "https://your-custom-endpoint/v1"
       models:
         your-model:
           context_length: 200000

2. Start Hermes with that model and run:

   /modelinfo
   

   Before: shows 128K context.
   After: shows 200K context (from config).

3. Test @context reference expansion with a large file:

   @context Read the file at /path/to/large/file.txt
   

   Before: hard limit = 64K tokens; large files are refused injection.
   After: hard limit = 100K tokens; correct injection.

4. Test /model switch display — switch to the custom model and verify context length shown is correct.

Testing

Verified with Infini-AI custom endpoint:

• Before: /modelinfo showed 128K, @context hard limit = 64K
• After: /modelinfo shows 200K (configured), @context hard limit = 100K

Full call-site audit: all 13 get_model_context_length() invocations across gateway/run.py, cli.py, run_agent.py, agent/context_compressor.py, and hermes_cli/web_server.py have been verified — 7 fixed, 6 already correct.

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 (no unrelated commits)
• I've verified all call sites of the affected function are covered

Documentation & Housekeeping

• I've updated relevant documentation — N/A (internal fix, no API change)
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact — no platform-specific behavior