feat(providers): add native MiniMax support

[octo-patch]

Reason: Add native MiniMax routing for the current text models with regional OpenAI-compatible endpoints.

• Register MiniMax through the shared OpenAI-compatible client with global and mainland-China endpoint selection.
• Add native catalog bindings with current upstream IDs, pricing, and cache rates while preserving existing route preference.
• Document the MiniMax settings and cover regional URL selection, provider dispatch, catalog resolution, and model rewriting.

Checks:

• go test ./...
• go vet ./internal/providers/... ./internal/router/catalog ./internal/router/rl ./internal/translate ./cmd/router
• git diff --check

[greptile-apps]

PR author is not in the allowed authors list.

[octo-patch]

Updated this branch with the latest main and resolved the configuration conflict while preserving the MiniMax environment settings. I ran go test ./..., go vet ./internal/providers/... ./internal/router/catalog ./internal/router/rl ./internal/translate ./cmd/router, and git diff --check.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 380b2b5. Configure here.}

[cursor]

M3 uses long-context prices

High Severity

The native MiniMax M3 binding is priced at $0.60/$2.40 with a $0.12 cache-read rate, which is MiniMax's long-context tier for inputs over 512K. The model-level window stays at 512_000 and the MiniMax binding does not override ContextWindow, so routed turns never enter that tier. Official standard rates at or below 512K are $0.30/$1.20 with $0.06 cache reads. Catalog prices feed billing and planner EV, so MiniMax M3 usage is charged and scored at twice the actual cost.

Additional Locations (1)

• internal/router/catalog/catalog_test.go#L154-L155

 

^{Reviewed by Cursor Bugbot for commit 380b2b5. Configure here.}

[devin-ai-integration]

Thank you so much for this, @octo-patch — a native MiniMax provider is a genuinely useful addition, and you followed the internal/providers/AGENTS.md "adding a provider" recipe to the letter (sibling *BaseURL constant in openaicompat instead of a new adapter, the three-map edit in provider.go, registerDeploymentKeyedProvider in the composition root, the affinity/roster coverage maps, docs). That made this easy to carry over.

I've re-landed it as #1188 with your commits preserved and credited (Co-authored-by). It's not a fault with your work — the two things I had to touch are internal and not obvious from the outside:

• Rebased onto main. The repo moved to the weave-os org and the module path changed to weave-os/router, and docs/CONFIGURATION.md gained new rows in the same spot, so the branch no longer merged cleanly.
• M3 native binding priced at the ≤512k tier (0.300 / 1.200, cache read 0.06) instead of 0.600 / 2.400. Per MiniMax's pay-as-you-go table the 0.60/2.40 numbers are the >512k long-context tier (or the pre-discount list price); the catalog's ContextWindow: 512_000 never reaches it, and internal/router/catalog/AGENTS.md notes this table feeds billing and planner EV directly, so it needs to be the effective rate — same as the adjacent Fireworks/OpenRouter rows. Cursor's bot flagged this too. Test updated to match.

Everything else is exactly your change. Thanks again for contributing — hope to see more from you!