fix(llm): only attach prompt-cache points on routes LiteLLM serves

[devin-ai-integration]

Summary

cache_control_injection_points is LiteLLM's own request field, consumed by its Anthropic transform — but _prompt_cache_extra_args() attached it to any model name containing "claude", conflating model identity with provider routing. A bare claude-sonnet-4-5 with LLM_API_BASE set (i.e. Claude behind an OpenAI-compatible gateway) is served by the SDK's own OpenAI client, which forwards unknown request kwargs straight into AsyncCompletions.create(), so every turn dies before the first token:

TypeError: AsyncCompletions.create() got an unexpected keyword argument 'cache_control_injection_points'

The gate now asks who actually serves the request, mirroring StrixProvider._resolve_prefixed_model() (bare names and openai//any-llm/ → SDK clients; anything else prefixed → LiteLLM fallback):

-if not is_claude_model(model_name):
+if not is_claude_model(model_name) or not routes_through_litellm(model_name):
     return None

def routes_through_litellm(model_name: str | None) -> bool:
    name = (model_name or "").strip()
    if not name or codex.subscription_model(name):
        return False
    prefix, _, rest = name.partition("/")
    return bool(rest) and prefix.lower() not in {"openai", "any-llm"}

So anthropic/…, litellm/…, vertex_ai/… and the mapped Bedrock routes keep prompt caching unchanged; only the routes that would have crashed lose the field (they never cached anyway — the gateway decides caching there).

tests/test_models.py pins the helper against the provider itself: for each name it builds StrixProvider().get_model(...), unwraps the _NonStreamingModel/_TurnGuardModel wrappers, and asserts isinstance(inner, LitellmModel) is routes_through_litellm(name) — otherwise drift between the two makes every request on a route fail mid-turn, which is exactly the bug above.

Testing

• uv run pytest tests/test_models.py tests/test_inputs.py → 123 passed. Full suite passes except a pre-existing, unrelated tests/test_pricing.py::test_resolves_common_bare_model_names failure (resolve_litellm_model("MiniMax-M3") is None on main too); 1186 passed with it deselected.
• ruff check, ruff format --check, mypy, bandit on the changed files: clean.
• Runtime check against a packaged build with these two files in place, dispatching a real turn at an unreachable base URL: bare claude-sonnet-4-5 and openai/claude-sonnet-4-5 now reach the transport (APIConnectionError) instead of raising the TypeError, while anthropic/…, litellm/anthropic/… and bedrock/… still send the cache points.

Link to Devin session: https://app.devin.ai/sessions/54414ea3ecb34ade9189534e5c690969
Open in Devin Desktop: https://app.devin.ai/desktop/session/54414ea3ecb34ade9189534e5c690969?variant=devin

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

The PR prevents LiteLLM-specific prompt-cache arguments from being attached to model requests served by native SDK clients.

• Adds a shared helper that classifies whether a model name routes through LiteLLM.
• Uses that classification when constructing prompt-cache request arguments.
• Adds coverage for native SDK, LiteLLM, Anthropic, Bedrock, Ollama, and compatible gateway routes.

Confidence Score: 5/5

The PR appears safe to merge, with no actionable correctness or security issues identified.

The new gate follows the provider-routing boundary and preserves prompt-cache arguments on LiteLLM-backed Claude routes while removing an unsupported argument from native SDK requests.

Important Files Changed

Filename	Overview
strix/config/models.py	Adds model-route classification aligned with provider construction, including native SDK and subscription exclusions.
strix/core/inputs.py	Restricts LiteLLM prompt-cache injection arguments to requests actually served through LiteLLM.
tests/test_inputs.py	Verifies that native SDK Claude routes no longer receive LiteLLM-only cache arguments.
tests/test_models.py	Cross-checks the new classification helper against the concrete model type built by StrixProvider.

_{Reviews (1): Last reviewed commit: "fix(llm): only attach prompt-cache point..." | Re-trigger Greptile}