MiniMax-M3 exposes a single binary thinking knob (adaptive | disabled) on its OpenAI-compatible endpoint at api.minimaxi.com. Today our provider unconditionally emits reasoning_effort, which M3 rejects — so users with a MiniMax provider can't actually configure the thinking level even though the EffortCapability plumbing implies they can.

This PR teaches the openai provider and the EffortCapability layer to recognize api.minimaxi.com as a separate wire shape, mirroring the existing DeepSeek handling.

Wire-format change

For a base_url of api.minimaxi.com (or any *.minimaxi.com subdomain), buildRequest now:

• emits thinking.type = "adaptive" when /effort auto (the empty-effort case, which is also the M3 model default)
• emits thinking.type = "adaptive" for the explicit adaptive level
• emits thinking.type = "disabled" for the disabled level
• omits reasoning_effort entirely — M3 has no level scale and rejects the field

This is in addition to the existing DeepSeek path, which still emits both thinking.type = "enabled" and reasoning_effort = high|max.

User-facing EffortCapability change

M3 entries now expose /effort auto|adaptive|disabled with default adaptive. Legacy level names from other vendors resolve to a valid M3 level so stale /effort values don't error:

Unknown values get a MiniMax-specific usage hint (/effort auto|adaptive|disabled) instead of the generic message.

Refactor bundled in

The host-matching logic was duplicated across internal/provider/openai/openai.go (private *BaseURL helpers) and internal/config/effort.go (private *Entry wrappers). Adding a new gateway would have required updating both files in lockstep, with no compiler-level guarantee they stay in sync.

I extracted the shared primitive into a new internal/provider/openai/host.go:

func IsDeepSeek(baseURL string) bool {
    return matchesVendorHost(baseURL, "deepseek.com", "api.deepseek.com")
}

func IsMiniMax(baseURL string) bool {
    return matchesVendorHost(baseURL, "minimaxi.com", "api.minimaxi.com")
}

The *Entry wrappers in effort.go now just gate on the openai kind and delegate. net/url is no longer imported in effort.go or openai.go.

The new host_test.go pins the matching rule (canonical hostname exact match, plus any subdomain of the apex) directly at the source.

Tests

• TestBuildRequestMiniMaxThinking — wire-shape coverage for all three levels (auto/adaptive/disabled), plus assertion that reasoning_effort is omitted
• TestNewMiniMaxEffortValidation — boot-time validation of accepted and rejected effort values
• TestNewMiniMaxSetsFlag — base-URL detection (with and without /v1 suffix)
• TestIsMiniMaxEntry — host detection edge cases (subdomain variants accepted, bare apex rejected, wrong-spelling rejected)
• TestEffortCapabilityMiniMax — /levels and default for M3 entries
• TestNormalizeEffortMiniMax — full table of legacy level remaps
• TestNormalizeEffortMiniMaxRejectsGarbage — MiniMax-specific error message
• TestEffectiveEffortMiniMax — EffectiveEffort resolution
• TestIsDeepSeek / TestIsMiniMax (new in host_test.go) — host-matching primitives

All MiniMax and DeepSeek tests pass; the existing TestRealDeepSeekCacheProbe and the 18+ other openai tests are unchanged and still green.

Out of scope / not changed

• The <think>…</think> block in content is handled by think.go (already in upstream from 19bf1872). This PR doesn't touch it.
• The HTTP client, retry logic, and SSE parsing are unchanged.
• Other OpenAI-compatible gateways (MiMo, vLLM, llama.cpp) fall into the default branch and are unaffected.

Test results

$ go vet ./...
(clean)

$ go test ./internal/provider/openai/ ./internal/config/ -run 'IsDeepSeek|IsMiniMax|MiniMax|DeepSeek|Effort'
ok  reasonix/internal/provider/openai  12.117s
ok  reasonix/internal/config           0.747s