Summary

Fixes #5089 (also partially addresses concerns raised in #2513, #12977 for the banner-display path).

The session-reset / info banner in _format_session_info (gateway/run.py) resolved the context window only from the top-level model.context_length key. When users configure context_length under either the newer providers.<name>.models.<model> dict schema or the legacy custom_providers[].models.<model> list schema, the banner fell through to get_model_context_length()'s probe chain. Remote OpenAI-compatible proxies frequently omit context_length from /models responses, so the probe failed silently and the banner displayed:

◆ Context: 128K tokens (default — set model.context_length in config to override)

— even though the runtime ContextCompressor was already budgeting with the correct value (resolved by AIAgent.__init__ at run_agent.py:1608-1643). A display-vs-truth inconsistency rather than a functional bug, but confusing and repeatedly reported (#5089, #2513).

Fix

After the existing top-level lookup, walk get_compatible_custom_providers(cfg), match the entry by base_url against the resolved runtime base_url, and read entry["models"][model]["context_length"]. Pass that through to get_model_context_length() as config_context_length; step 0 short-circuits the probe.

This mirrors the exact resolution AIAgent.__init__ already performs, so the banner's displayed value matches what the compressor actually uses.

Why this is an improvement over existing open PRs

Four other open PRs target #5089:

• fix(gateway): honor custom provider context length #5096, fix(gateway): read custom provider context length in status #8240, fix: honor custom-provider context overrides in session info #10690: each hand-rolls traversal of the legacy custom_providers: list only. They do not cover the newer providers: dict schema that users land on via hermes model (and which config.yaml templates now use).
• fix: pass config_context_length to /model display and gateway fallback paths #12380: broader in scope (touches /model display + gateway fallback paths) but still traverses lists manually.

This PR routes through get_compatible_custom_providers() — the existing compat shim at hermes_cli/config.py:2078 that already bridges both schemas for the rest of the runtime. One lookup, both schemas, stays consistent with AIAgent.__init__, and eliminates future drift between the banner and the compressor.

Reproduction

~/.hermes/config.yaml (new-schema form — the legacy list form reproduces identically):

model:
  default: kimi-k2.6
  provider: my-proxy
providers:
  my-proxy:
    name: my-proxy
    api: https://my-openai-proxy.example/v1
    api_key: ${MY_PROXY_KEY}
    models:
      kimi-k2.6:
        context_length: 262144

Restart gateway, trigger /reset:

• Before: ◆ Context: 128K tokens (default — set model.context_length in config to override)
• After: ◆ Context: 262K tokens (config)

Verification

Simulated the patched block against a real-world config (exact code shape of the fix, not a standalone re-implementation):

after top-level: None
after nested fallback: 262144
Banner would display: 262K tokens (config)

No functional change to the compressor or any runtime code path — this only affects what _format_session_info displays.

Test plan

• Existing tests in tests/gateway/test_session_info.py still pass.
• Manual: gateway restart with new-schema providers: config → banner shows nested context_length.
• Manual: gateway restart with legacy custom_providers: list config → banner shows nested context_length (unchanged behavior vs fix(gateway): honor custom provider context length #5096 et al).
• Manual: gateway restart without any nested context_length → probe still runs as before.