internal/anthropic and internal/translate moved to the new
Plexum-anthropic-compat-client repo (extracted from MiniMax + Kimi
duplicates). cmd/plexum-minimax-provider-plugin imports them from
the shared module instead.
No behavioral change — same wire client, same translator.
Live re-verified after refactor: plexum say → "ready".
Field investigation against the live API:
- sk-cp- (Token Plan) + MiniMax-M2.7 → ✅ works
- sk-cp- (Token Plan) + MiniMax-M2.7-highspeed → ❌ "Token Plan Starter
(0/0 used)" — the base Token Plan EXCLUDES the highspeed variant
- sk-api- (pay-per-token) + either model → ✅ structurally
(needs balance > 0; otherwise "insufficient_balance (1008)")
Plugin code already handles both key types via the same
Authorization: Bearer header — no functional change needed. README now
makes the matrix explicit so operators don't try highspeed on a
plan-only key and wonder why "0/0 used" comes back.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Plexum ProviderPlugin that serves MiniMax models through MiniMax's
Anthropic-compatible HTTP endpoint
(https://api.minimax.io/anthropic, or CN api.minimaxi.com). Inspired
by openclaw's extensions/minimax provider-registration, but rewritten
in Go for Plexum's SDK.
internal/anthropic/ (~210 LOC + 6 tests):
- minimal HTTP+SSE Anthropic Messages client (POST /v1/messages,
stream:true, parses event:/data: SSE frames)
- handles non-2xx as HTTP error; stream errors land as Event{Type:"error"}
- 1 MiB SSE line cap; per-conn 5min timeout
internal/translate/ (~220 LOC):
- CanonicalToAnthropic: canonical.TurnRequest → MessagesRequest
- blockToAnthropic: TextBlock / ToolUseBlock / ToolResultBlock /
ThinkingBlock → loose ContentBlock map; preserves signatures + cache
control
- Translator: per-turn state machine; consumes anthropic.Event stream
and emits canonical.TurnEvent stream (handles thinking blocks +
tool_use input_json_delta accumulation + signature_delta capture)
cmd/plexum-minimax-provider-plugin/:
- Plugin manifest declares provider.models = [MiniMax-M2.7,
MiniMax-M2.7-highspeed]
- Backend fixed to "api" (per scope); region "global"|"cn" + base_url
override supported via config
- HostConfig from <profile>/plugins/plexum-minimax-provider/config.json
{api_key, region?, base_url?, max_tokens_default?}
scripts/install.sh: build + manifest emit; operator writes config.json
+ allows plugin + adds an agent + restarts.
End-to-end verified against the real key:
1. plexum say --agent-id mini ... → "Hi, I'm MiniMax!"
2. Multi-turn continuity: agent recalled the prior reply
3. Via gateway socket: {"outcome":"text","text":"\n\npong"}
4. Via Fabric channel (alice posts → plugin inbound → mini agent
→ MiniMax → outbound REST → reply visible in bt2-clean seq=11):
"Hi there! 👋 Fun fact: Octopuses have three hearts, blue blood,
and neurons distributed throughout their arms—so their tentacles
can 'think'"
The MiniMax-M2.7-highspeed variant works the same way but hit a
Code Plan rate-limit ceiling during testing (not a plugin issue).
Deferred:
- OAuth (Code Plan portal) — not in v1 scope per request
- MiniMax Portal provider (separate provider id minimax-portal)
- Image / TTS / video / music providers (separate plugins later)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
