fix(agent): preserve global model.context_length through switch_model and fallback

[TF0rd]

Problem

Both switch_model() and try_activate_fallback() unconditionally set agent._config_context_length = None before passing it to get_model_context_length(). This discards the global model.context_length from config.yaml, causing the context window to fall through to the 256K probe-down default even when the user has explicitly configured a larger value (e.g. 1048576).

The clearing was intended to prevent inheriting a stale per-model custom_providers override when switching models — but it also nukes the global config setting, which applies to ALL models.

Fix

Before clearing _config_context_length, re-read the global model.context_length from config.yaml and pass it to get_model_context_length(). The per-model override is still cleared (correct behavior), but the global setting is preserved.

Also consolidates the config load in switch_model() that was previously duplicated — once for the global context_length and once for custom_providers. Both now share a single load_config() call.

Files Changed

• agent/agent_runtime_helpers.py — switch_model(): preserve global context_length, deduplicate config load
• agent/chat_completion_helpers.py — try_activate_fallback(): same fix

Testing

• Existing tests pass: test_custom_provider_context_length, test_model_switch_context_display, test_session_info (28/28)
• Manually verified: no more "defaulting to 256,000 tokens" after model switch with model.context_length: 1048576 in config

[teknium1]

Thanks for identifying the lost global override. The premise is confirmed on current main: switch_model() clears _config_context_length at agent/agent_runtime_helpers.py:1848 and passes it to the resolver at :2056; fallback does the same at agent/chat_completion_helpers.py:1505 and :1626. That conflicts with the documented precedence where model.context_length is highest priority (website/docs/integrations/providers.md:1139-1149).

Problems

• The PR adds no regression tests. Existing tests/run_agent/test_switch_model_context.py:44-57 only verifies clearing a prior override and does not configure a global model.context_length, so it cannot protect the global-vs-per-model distinction.
• GitHub currently reports this branch as conflicting. The current implementations are at agent/agent_runtime_helpers.py:1772 and agent/chat_completion_helpers.py:1330, not the original hunk locations.

Suggested changes

• Salvage the two resolution changes into the current helper locations and add explicit switch and fallback tests for a global config override, while retaining the no-global case that clears a stale per-model override.

Automated hermes-sweeper review.