Summary

When a fallback provider is activated or /model switch is triggered, _config_context_length is unconditionally cleared (set to None). The subsequent get_model_context_length() call then resolves the context window from scratch. If the model isn't in any static table (DEFAULT_CONTEXT_LENGTHS) and live probing fails, the resolution falls through all 10 steps to DEFAULT_FALLBACK_CONTEXT (256K).

This causes the status bar to show 97.9K/256K even though the user configured context_length: 1000000 in config.yaml — most visible after a compaction interruption + resume cycle.

Root Cause

In try_activate_fallback() (chat_completion_helpers.py:1013) and switch_model() (agent_runtime_helpers.py:1367):

agent._config_context_length = None  # ← cleared unconditionally

Then get_model_context_length(config_context_length=None, ...) is called. Without the explicit config override, resolution falls through to 256K for custom endpoints that don't advertise context_length via /models.

Fix

Save the original _config_context_length before clearing it. When the fallback/switched model is the same as the primary (common for custom providers where the fallback is the same endpoint accessed via a different credential), restore the saved override so the resolution chain receives it at Step 0:

_effective_config_ctx = getattr(agent, "_config_context_length", None)
if _saved_config_ctx and fb_model == old_model:
    _effective_config_ctx = _saved_config_ctx

When the fallback model is different from the primary, the override is NOT restored — the new model should resolve its own context window.

Changes

Test plan

• 401 agent tests pass
• Same-model fallback preserves user's model.context_length
• Different-model fallback resolves fresh context_length (doesn't inherit override)
• restore_primary_runtime() continues to work correctly (restores from snapshot)

Closes #32423