fix: preserve configured context lengths across runtime paths

[tongguang2]

Summary

This PR fixes a set of context length inconsistencies that still affect custom-provider runtimes across different execution paths.

What it fixes

• Normalize persistent context-length cache keys so /v1 and /v1/ resolve to the same provider endpoint
• Preserve configured context-length overrides when switching models
• Preserve configured context-length overrides when activating fallback models
• Avoid incorrectly reusing a per-model custom_providers context override for a different target model
• Add regression tests for cache normalization, switch-model handling, fallback handling, and configured override persistence

Root cause

There were two separate issues:

1. Persistent context-length cache lookups treated .../v1 and .../v1/ as different keys, which caused cache misses for the same custom provider endpoint in different runtime paths.
2. Runtime context refresh paths (switch_model() / fallback activation) did not consistently resolve the configured context-length override for the active target model and endpoint.

As a result, custom providers that were correctly configured in config.yaml could still fall back to an incorrect detected context length in some paths.

Implementation details

• Added base URL normalization for persistent context-length cache writes
• Made cache reads backward-compatible with legacy slash variants
• Added runtime-target-aware context-length resolution in AIAgent
• Reused that runtime-target-aware resolution when refreshing model metadata during model switch and fallback activation

Tests

Verified with:

• tests/agent/test_model_metadata.py
• tests/run_agent/test_switch_model_context.py
• tests/run_agent/test_fallback_model.py
• tests/run_agent/test_provider_fallback.py
• tests/run_agent/test_invalid_context_length_warning.py
• tests/run_agent/test_compression_feasibility.py

Local result:

• 147 passed

Notes

This PR is intentionally scoped to the runtime-path inconsistency and cache-key normalization issue. It does not introduce broader changes to provider metadata discovery beyond what is needed to make configured context lengths behave consistently.

[alt-glitch]

Related to #11437 (context length display/runtime alignment), #8785 / #8786 (compression context overrides) — same family of context-length consistency bugs across different runtime paths.

[alt-glitch]

Related to #11437, #8785, #8786.

[alt-glitch]

Related to #11437, #8785, #8786.

[teknium1]

Thanks for digging into this, @tongguang2 — the context-length consistency area is genuinely fiddly and the investigation was solid. Closing without merge after verifying both halves against current main:

Runtime-path overrides (switch_model / fallback): already solved on main, differently. The runtime paths now re-read custom_providers live from config and pass them into get_model_context_length(), which resolves the per-model override internally via get_custom_provider_context_length() (trailing-slash tolerant). See agent/agent_runtime_helpers.py (switch_model) and agent/chat_completion_helpers.py (fallback). This landed as part of #15779. Snapshotting the config at __init__ (the _resolve_config_context_length / _custom_provider_context_configs approach) is actually the weaker design here — it can't see a config edit made mid-session, whereas the live re-read can.

Persistent cache-key /v1 vs /v1/ normalization: the premise doesn't hold against how base_url actually flows. Every cache read (get_cached_context_length) and write (save_context_length) in get_model_context_length() is keyed on the same base_url string passed into that one resolution call. A single endpoint config resolves to one canonical base_url per process, so the read and write always use the identical key — there's no runtime path where the same provider is written with /v1 and read back with /v1/. The added test only passes because it manually calls save with one form and get with the other; the real call sites never create that mismatch.

So both halves are either already covered or fix a scenario that doesn't occur at runtime. Closing as superseded by #15779. Appreciate the contribution and the careful write-up.