fix(config): resolve provider-level context_length on /model switch

[vanhoof]

Problem

/model switch always shows Context: 256,000 tokens for custom providers configured via providers: in config.yaml, regardless of the actual context_length value set on the provider entry.

/model vertex-opus46
  ✓ Model switched: claude-opus-4.6
    Provider: vertex-opus46
    Context: 256,000 tokens    ← wrong, should be 1,000,000

This affects all custom provider setups (vertex-proxy, LiteLLM, or any providers: entry) using the common config shape:

providers:
  vertex-opus46:
    base_url: http://127.0.0.1:8788/anthropic
    model: claude-opus-4.6
    context_length: 1000000
    api_mode: anthropic_messages

Root Cause

Three bugs in the resolution chain:

1. get_compatible_custom_providers() early-returns on dict-shaped custom_providers

When custom_providers: {} exists in config.yaml (an empty dict, not a list), the not isinstance(custom_providers, list) guard returns [], skipping the providers: dict loop below it. All user-configured providers become invisible to the context_length lookup.

2. get_custom_provider_context_length() only checks nested models.<model>.context_length

The common config shape has model and context_length as siblings at the provider level, with no nested models: dict. The lookup returns None and the chain falls through to DEFAULT_FALLBACK_CONTEXT = 256_000.

3. resolve_display_context_length() never loads custom_providers

The function accepts custom_providers as a parameter, but neither call site in cli.py passes it. When custom_providers is None, the resolver has nothing to check against and falls through to the default.

Fix

hermes_cli/config.py:

• get_compatible_custom_providers(): gracefully handle dict-shaped custom_providers by normalizing it the same way as the providers: dict, instead of early-returning.
• get_custom_provider_context_length(): after the existing models.<model>.context_length lookup, fall back to the provider-level context_length when the entry's default model matches the requested model.

hermes_cli/model_switch.py:

• resolve_display_context_length(): when custom_providers is None, load it from config via get_compatible_custom_providers(load_config()) so the resolver sees user-configured provider entries.

agent/agent_runtime_helpers.py:

• Persist the resolved new_context_length back to agent._config_context_length after switch_model completes, so the display path in cli.py reads the correct value on subsequent /model calls.

Testing

Manual: Tested with providers: config entries pointing at a local proxy (vertex-proxy on port 8788, Anthropic wire protocol). Gateway restarted between tests.

Before fix:

/model vertex-opus46   → Context: 256,000 tokens
/model vertex-haiku    → Context: 256,000 tokens

After fix:

/model vertex-opus46   → Context: 1,000,000 tokens
/model vertex-haiku    → Context: 200,000 tokens

Automated: All existing tests pass:

• test_config.py -- 87 passed
• test_config_validation.py -- 21 passed (dict-shaped custom_providers detection still works)
• test_model_switch_custom_providers.py -- 19 passed
• test_custom_provider_context_length -- 12 passed, 2 skipped
• Full custom_provider_context or model_switch_custom -- 31 passed, 2 skipped

Supersedes #36199 and #37712.

Signed-off-by: Chris van Hoof vanhoof@ouwish.com

[teknium1]

Thanks for tracing the provider-resolution chain. The remaining config behavior is still needed on current main: hermes_cli/config.py:4947-4955 returns early when custom_providers is a dict, and hermes_cli/config.py:5147-5161 ignores a provider-level model plus context_length.

Problems

• The PR changes three production files but adds no regression coverage for either config-schema case. Existing tests exercise nested models.<id>.context_length (tests/hermes_cli/test_custom_provider_context_length.py:15-143), not the provider-level fallback or custom_providers: {} interaction.

Suggested changes

• Add focused tests for both cases above, ideally exercising the normalized providers: path through get_compatible_custom_providers() and then context resolution.
• The resolver's implicit config reload should be reconsidered when salvaging: current CLI callers already pass agent._custom_providers at cli.py:7934 and cli.py:8243, via 9d848cc60.

Automated hermes-sweeper review.

[teknium1]

Please add regression coverage for this compatibility branch: custom_providers: {} plus a valid providers: entry must still return the normalized provider, rather than silently hiding it.

[teknium1]

Please add a paired regression test for the provider-level fallback, including a non-matching requested model so a provider's default context length cannot leak to another model at the same endpoint.

[vanhoof]

Closing as superseded upstream. The provider-level context_length gap on /model switch is now handled on main by #69712 (fix(cli): honor custom_providers in preflight shrink warning), which threads agent._custom_providers into the resolver at the call site — the same fix this PR proposed. Verified against current main during a runtime rebase.