fix(report): use bare-model fallback in usage-ledger cost estimate

[JOhnsonKC201]

Problem

_estimate_litellm_entry_cost (strix/report/usage.py) builds a fallback list of model-name candidates so the LiteLLM cost lookup can fall back from the prefixed name to the bare model name:

candidates = [model]
if "/" in model:
    candidates.append(model.split("/", 1)[-1])

for candidate in candidates:
    cost = completion_cost(
        completion_response={"model": candidate, "usage": usage_payload},
        model=model,   # pins the lookup to the original name
    )

But the pricing lookup passed model=model, so completion_cost always used the original prefixed name and the bare-name candidate was never actually tried (LiteLLM prioritizes the model= kwarg). The loop's fallback was dead code.

With e.g. STRIX_LLM="openai/mistralai/Mistral-7B-Instruct-v0.3" (pointing LLM_API_BASE at a self-hosted OpenAI-compatible server serving HF-named models), LiteLLM keys the price under the bare Mistral-7B-Instruct-v0.3. Every candidate resolves to a prefixed name absent from the price map, cost stays None, and the entry is reported as $0.00.

The sibling estimator _estimate_response_cost in strix/report/state.py builds the same candidate list and correctly passes model=candidate (locked by test_cost_callback_estimates_cost_with_bare_model_fallback). This path was written with model=model, defeating its own fallback.

Fix

Pass model=candidate so the bare-name fallback is actually attempted, matching _estimate_response_cost. For single-candidate configs (candidate == model) behavior is unchanged.

Tests

Adds two tests to tests/test_cost_tracking.py (mirroring the existing bare-model-fallback test): one on _estimate_litellm_entry_cost, one end-to-end through _estimate_litellm_cost, both pricing only the bare name. uv run pytest tests/test_cost_tracking.py → 11 passed; reverting the fix makes both new tests fail (None == 0.0021).

[greptile-apps]

Greptile Summary

This PR enables bare-model fallback for usage-ledger cost estimates. The main changes are:

• Passes each model candidate to LiteLLM's explicit model lookup.
• Adds direct coverage for the entry-level fallback.
• Adds end-to-end coverage for prefixed OpenAI-compatible models.

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the changed code.
• Existing successful lookups remain the first attempt.
• The new tests cover the corrected fallback path.

Important Files Changed

Filename	Overview
strix/report/usage.py	Uses the current candidate for each cost lookup while preserving the original model as the first attempt.
tests/test_cost_tracking.py	Covers the bare-model fallback through both the helper and aggregate usage paths.

_{Reviews (1): Last reviewed commit: "fix(report): use bare-model fallback in ..." | Re-trigger Greptile}