fix(pricing): resolve MiniMax-M3 (provider prefix + pre-registration)

[shreyassks]

Description

Fixes the cost dashboard reporting $0.00 for every call when the upstream model is MiniMax-M3 (Anthropic-compatible endpoint served from the MiniMax provider).

Two root causes in headroom/pricing/litellm_pricing.py:

1. resolve_litellm_model() had no minimax/ provider prefix. LiteLLM's community pricing database stores MiniMax-M3 only under minimax/MiniMax-M3. The resolver never tried that prefix, so callers in proxy/cost.py, proxy/savings_tracker.py, and perf/analyzer.py silently fell back to the unresolved name.
2. The prefix check was case-sensitive. MiniMax's model name uses mixed case (MiniMax-M3), but every existing prefix pattern (claude-, gpt-, o1-, …) was lowercase, so even after adding "minimax-" the bare MiniMax-M3 wouldn't match.

This PR fixes both, plus adds a _register_minimax_pricing() helper that pre-populates litellm.model_cost["MiniMax-M3"] from minimax/MiniMax-M3 at module load — a safety net so estimate_cost() (which doesn't know the minimax/ prefix internally) succeeds even on a cold resolver cache or if LiteLLM drops the prefixed entry in a future release.

Net change: +97 / −1 lines across 2 files (one production file + one test file).

Closes #

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Changes Made

• Production (headroom/pricing/litellm_pricing.py):
  • Add "minimax-": "minimax/" to the provider-prefix table in _resolve_litellm_model_uncached() so the resolver knows about the MiniMax provider.
  • Compute model_lower = model.lower() and match prefixes against it instead of model, so the mixed-case bare name MiniMax-M3 resolves correctly. The existing prefixes (claude-, gpt-, o1-, o3-, o4-, gemini-) are already lowercase patterns matched against canonical lowercase names (claude-sonnet-4-5-…, gpt-4o, gemini-2.0-flash) — no regression.
  • Add _register_minimax_pricing(): if minimax/MiniMax-M3 is in litellm.model_cost and MiniMax-M3 is not, copy the pricing dict under the bare key. No-op on older LiteLLM (entry missing) or when the user has already customised MiniMax-M3.
  • Invoke _register_minimax_pricing() once at module import.
• Tests (tests/test_pricing_litellm.py):
  • Add test_litellm_minimax_mixed_case_with_provider_prefix — verifies resolve_litellm_model("MiniMax-M3") returns "minimax/MiniMax-M3" via the case-insensitive prefix match.
  • Add test_litellm_minimax_preregistration_safety_net — verifies the pre-registration populates the bare MiniMax-M3 key, that estimate_cost() returns the correct dollar figure (0.84 for 1M in + 100k out), and that a user-customised bare entry is never clobbered.

Testing

• Unit tests pass (pytest)
• Linting passes (ruff check .)
• Type checking passes (mypy headroom)
• New tests added for new functionality
• Manual testing performed

Test Output

$ uv run pytest tests/test_pricing_litellm.py -v
============================= test session starts ==============================
platform darwin -- Python 3.11.15, pytest-9.0.3, pluggy-1.6.0
configfile: pyproject.toml
plugins: anyio-4.12.1, langsmith-0.8.0, asyncio-1.3.0, cov-7.0.0
collected 7 items

tests/test_pricing_litellm.py::test_litellm_helpers_when_dependency_is_unavailable PASSED [ 14%]
tests/test_pricing_litellm.py::test_litellm_model_pricing_exact_match_and_defaults PASSED [ 28%]
tests/test_pricing_litellm.py::test_litellm_model_pricing_uses_provider_prefixes PASSED [ 42%]
tests/test_pricing_litellm.py::test_litellm_model_pricing_uses_aliases_and_zero_cost_defaults PASSED [ 57%]
tests/test_pricing_litellm.py::test_litellm_model_pricing_returns_none_for_unknown_models PASSED [ 71%]
tests/test_pricing_litellm.py::test_litellm_minimax_mixed_case_with_provider_prefix PASSED [ 85%]
tests/test_pricing_litellm.py::test_litellm_minimax_preregistration_safety_net PASSED [100%]

============================== 7 passed in 1.07s ===============================

$ uv run ruff check headroom/pricing/litellm_pricing.py tests/test_pricing_litellm.py
All checks passed!

$ uv run mypy headroom/pricing/litellm_pricing.py
Success: no issues found in 1 source file

Manual reproducer (matches the PR writeup):

$ uv run python -c "
from headroom.pricing.litellm_pricing import resolve_litellm_model, estimate_cost
import litellm

print('resolve_litellm_model(MiniMax-M3):', resolve_litellm_model('MiniMax-M3'))
print('MiniMax-M3 in litellm.model_cost :', 'MiniMax-M3' in litellm.model_cost)
print('estimate_cost (1M in, 100k out): ', estimate_cost('MiniMax-M3', 1_000_000, 100_000))
"

resolve_litellm_model(MiniMax-M3): minimax/MiniMax-M3
MiniMax-M3 in litellm.model_cost : True
estimate_cost (1M in, 100k out):  0.84

Real Behavior Proof

• Environment: macOS Darwin 25.5.0, Python 3.11.15, headroom-ai installed editable via uv from this branch, litellm pulled from PyPI on first run.
• Exact command / steps: after git checkout fix/minimax-pricing && uv sync --all-extras --dev, run (1) uv run python -c "from headroom.pricing.litellm_pricing import resolve_litellm_model, estimate_cost; import litellm; print(resolve_litellm_model('MiniMax-M3'), 'MiniMax-M3' in litellm.model_cost, estimate_cost('MiniMax-M3', 1_000_000, 100_000))", then (2) uv run pytest tests/test_pricing_litellm.py -v, then (3) uv run ruff check headroom/pricing/litellm_pricing.py tests/test_pricing_litellm.py, then (4) uv run mypy headroom/pricing/litellm_pricing.py.
• Observed result: (1) resolve_litellm_model('MiniMax-M3') returns minimax/MiniMax-M3 (was MiniMax-M3, unresolved); 'MiniMax-M3' in litellm.model_cost is True (proves _register_minimax_pricing() ran); estimate_cost('MiniMax-M3', 1_000_000, 100_000) returns 0.84 (matches $0.60/M in × 1M + $2.40/M out × 0.1M). (2) All 7 tests in tests/test_pricing_litellm.py pass (5 pre-existing + 2 new MiniMax-specific). (3) ruff reports All checks passed!. (4) mypy reports Success: no issues found in 1 source file.
• Not tested: end-to-end through the running proxy against a live MiniMax-M3 endpoint — no Anthropic-compatible key configured in this environment. The reproducer exercises the exact code path the proxy's cost accumulator uses, but I did not point the proxy at a real upstream.

Review Readiness

• I have performed a self-review
• This PR is ready for human review

Checklist

• My code follows the project's style guidelines
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation — N/A: no user-facing docs reference litellm_pricing.py directly; the only public API affected (estimate_cost) now returns correct values for a previously-unsupported model.
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have updated the CHANGELOG.md if applicable — N/A: this repo doesn't appear to use CHANGELOG.md (not present at repo root).

Additional Notes

• Why both fixes are needed. estimate_cost() calls get_model_pricing() directly, and get_model_pricing() has its own hardcoded prefix list ["openai/", "anthropic/", "google/", "mistral/", "deepseek/"] that does not include minimax/. So the prefix resolver alone is not enough for estimate_cost("MiniMax-M3") to return a non-None number — the pre-registration step is what makes the bare name resolve. The prefix resolver change matters for the proxy's cost/savings/perf code paths that call resolve_litellm_model() and then look up the prefixed string themselves.
• Why case-insensitive matching is safe. All existing prefixes are lowercase patterns matched against already-lowercase canonical model names — lower-casing before startswith() is a no-op for them. Only the new "minimax-" entry uses a mixed-case input.
• Pricing drift note. _register_minimax_pricing() mirrors upstream LiteLLM (input $0.60/M, output $2.40/M, cache read $0.12/M as of 2026-06). Re-check after LiteLLM updates; the function already short-circuits when the user has customised the entry.
• Did not run the full test suite, only tests/test_pricing_litellm.py. Wider CI will catch anything I missed.

[github-actions]

PR governance

This PR follows the template and is marked ready for human review.

[codecov]

Codecov Report

❌ Patch coverage is 63.63636% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
headroom/pricing/litellm_pricing.py	63.63%	2 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!

[JerrettDavis]

The pricing fix itself is the right direction: the mixed-case MiniMax resolver path is covered, and the bare-name pre-registration gives estimate_cost("MiniMax-M3", ...) a direct lookup path.

The branch cannot merge while lint is red, though. The latest lint job has ruff check . passing, then fails at formatting:

Would reformat: tests/test_pricing_litellm.py

Please run ruff format tests/test_pricing_litellm.py and push the formatted file so CI can go green.

[JerrettDavis]

The formatting blocker is resolved and the MiniMax pricing change looks good. The resolver now matches MiniMax-M3 case-insensitively through the minimax/ prefix, and the bare-name pre-registration makes estimate_cost("MiniMax-M3", ...) work even when callers do not resolve a prefixed LiteLLM key first. The tests cover both the resolver and the pre-registration safety net without clobbering a customized bare entry.

Lint/build/e2e are green. The current red test shards I inspected on similar heads are the shared HuggingFace cache issue in memory tests, not this pricing path.