@molecule/api-ai-minimax 1.0.0 → 1.1.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md ADDED
@@ -0,0 +1,188 @@
1
+ <!--
2
+ AUTO-GENERATED — DO NOT EDIT THIS FILE.
3
+ Generated by `mlcl sync-docs` from the package's src/index.ts JSDoc + mlcl/registry.json.
4
+ Edits here are overwritten on the next commit (molecule's pre-commit hook regenerates).
5
+ To change this document, edit the module-level JSDoc in src/index.ts.
6
+ Generated: 2026-08-04T01:47:34.069Z
7
+ -->
8
+
9
+ # @molecule/api-ai-minimax
10
+
11
+ > **Auto-generated, AI-first package reference** for the [molecule.dev](https://molecule.dev) ecosystem.
12
+ > It is written to be read by coding agents as much as by people, and is generated from this
13
+ > package's source — edit `src/index.ts` JSDoc, not this file.
14
+
15
+ MiniMax AI provider for molecule.dev.
16
+
17
+ ## Type
18
+
19
+ `provider`
20
+
21
+ ## Installation
22
+
23
+ ```bash
24
+ npm install @molecule/api-ai-minimax @molecule/api-ai @molecule/api-bond @molecule/api-i18n @molecule/api-secrets
25
+ ```
26
+
27
+ ## API
28
+
29
+ ### Interfaces
30
+
31
+ #### `MiniMaxConfig`
32
+
33
+ Configuration for MiniMax.
34
+
35
+ ```typescript
36
+ interface MiniMaxConfig {
37
+ /** Called on each rate-limited/overloaded upstream response, before any retry sleep. */
38
+ onRateLimit?: AiRateLimitCallback
39
+ /** API key. Defaults to MINIMAX_API_KEY env var. */
40
+ apiKey?: string
41
+ /** Default model. Defaults to 'minimax-m3'. */
42
+ defaultModel?: string
43
+ /** Maximum tokens for completions. */
44
+ maxTokens?: number
45
+ /** Base URL override (for proxies). Defaults to 'https://api.minimax.io' (MiniMax's INTERNATIONAL host; use 'https://api.minimaxi.com' for mainland China — keys are scoped per host). */
46
+ baseUrl?: string
47
+ /**
48
+ * Chat-completions path appended to {@link baseUrl}. Defaults to
49
+ * `/v1/chat/completions` (MiniMax-direct puts the version in the path). Set
50
+ * this when the same open-weight model is served by a US OpenAI-compatible
51
+ * host whose version prefix lives in the base URL instead — e.g. DeepInfra:
52
+ * `baseUrl='https://api.deepinfra.com/v1/openai'` + `completionsPath='/chat/completions'`.
53
+ */
54
+ completionsPath?: string
55
+ /**
56
+ * Optional catalog-id → upstream-model-id map, applied to the outbound request
57
+ * ONLY. Lets a US OpenAI-compatible host (DeepInfra) receive its namespaced id
58
+ * (`MiniMaxAI/MiniMax-M3`) while the rest of the platform — pricing, cost
59
+ * ceilings, display — keeps using the canonical catalog id (`minimax-m3`). An
60
+ * id not in the map passes through unchanged.
61
+ */
62
+ modelMap?: Record<string, string>
63
+ }
64
+ ```
65
+
66
+ #### `ProcessEnv`
67
+
68
+ Process Env interface.
69
+
70
+ ```typescript
71
+ interface ProcessEnv {
72
+ MINIMAX_API_KEY: string
73
+ /** Base URL override (for credential brokers / gateways / US OpenAI-compatible hosts). */
74
+ MINIMAX_BASE_URL?: string
75
+ /** Chat-completions path override (see {@link MiniMaxConfig.completionsPath}). */
76
+ MINIMAX_COMPLETIONS_PATH?: string
77
+ }
78
+ ```
79
+
80
+ ### Functions
81
+
82
+ #### `createProvider(config)`
83
+
84
+ Creates a MiniMax AI provider instance.
85
+
86
+ ```typescript
87
+ function createProvider(config?: MiniMaxConfig): AIProvider
88
+ ```
89
+
90
+ - `config` — MiniMax-specific configuration (API key, model, max tokens, base URL).
91
+
92
+ **Returns:** An `AIProvider` backed by the MiniMax Chat Completions API.
93
+
94
+ ### Constants
95
+
96
+ #### `aiMinimaxSecretDefinitions`
97
+
98
+ Secret definitions required by the MiniMax AI bond.
99
+
100
+ ```typescript
101
+ const aiMinimaxSecretDefinitions: SecretDefinition[]
102
+ ```
103
+
104
+ #### `provider`
105
+
106
+ The provider implementation.
107
+
108
+ ```typescript
109
+ const provider: AIProvider
110
+ ```
111
+
112
+ ## Core Interface
113
+
114
+ Implements `@molecule/api-ai` interface.
115
+
116
+ ## Bond Wiring
117
+
118
+ Setup function to register this provider with the bond system:
119
+
120
+ ```typescript
121
+ import { bond } from '@molecule/api-bond'
122
+ import { provider } from '@molecule/api-ai-minimax'
123
+
124
+ export function setupAiMinimax(): void {
125
+ bond('ai', 'minimax', provider)
126
+ }
127
+ ```
128
+
129
+ ## Injection Notes
130
+
131
+ ### Requirements
132
+
133
+ Peer dependencies:
134
+
135
+ - `@molecule/api-ai` ^1.0.1
136
+ - `@molecule/api-bond` ^1.0.1
137
+ - `@molecule/api-i18n` ^1.0.1
138
+ - `@molecule/api-secrets` ^1.0.1
139
+
140
+ ### Environment Variables
141
+
142
+ - `MINIMAX_API_KEY` _(required)_ — MiniMax API key
143
+ - Setup: Create an interface key in the MiniMax platform user center.
144
+ - Get it here: [https://platform.minimax.io/](https://platform.minimax.io/)
145
+
146
+ ### Runtime Dependencies
147
+
148
+ - `@molecule/api-ai`
149
+ - `@molecule/api-bond`
150
+ - `@molecule/api-i18n`
151
+ - `@molecule/api-secrets`
152
+
153
+ Config: `MINIMAX_API_KEY` (SERVER-side only) plus an optional default model id/base URL.
154
+
155
+ **Missing `MINIMAX_API_KEY` fails fast**: the provider throws naming the exact env var on
156
+ first use (the exported `provider` is a lazy proxy, so this fires on the first `chat()` call,
157
+ not at bond/module-load time) — it never silently sends an empty key.
158
+
159
+ **Error message disambiguation**: a plain 400 that ISN'T a context-length error (bad param,
160
+ malformed tool schema) gets its own non-retryable message distinct from the generic
161
+ "AI service error. Please try again." used for retryable failures.
162
+
163
+ ## E2E Tests
164
+
165
+ Integration checklist — drive the real UI (live preview, no mocks), adapt
166
+ each item to this app's actual chat/AI screens, and check every box off one
167
+ by one. A box you can't check is an integration bug to fix — not a skip. The
168
+ sandbox HAS an AI provider bonded, so the flow runs live end-to-end; AI
169
+ output is NON-DETERMINISTIC, so assert on STRUCTURE/behavior, not exact text:
170
+
171
+ - [ ] A message sent through the real chat UI comes back as a RELEVANT AI
172
+ reply — not an echo of the prompt, a hardcoded stub, or an empty bubble.
173
+ Ask something with a checkable answer (e.g. "What is 2 + 2?") and confirm
174
+ the response actually contains it ("4"), proving a live model answered.
175
+ - [ ] If the app streams, tokens render INCREMENTALLY — text grows word by
176
+ word in the UI, not one final blob dumped after a long frozen spinner. (A
177
+ streamed `chat()` yields `text` chunks then a final `done`; a single late
178
+ blob means the reply was awaited whole and streaming is broken.)
179
+ - [ ] Multi-turn CONTEXT is preserved: a follow-up that refers back to the
180
+ previous turn (e.g. after "2 + 2", ask "now double that" -> understood as 8) works — proving the full `messages` history is sent, not just the last
181
+ line.
182
+ - [ ] A provider failure (bad/missing key, rate limit, timeout) surfaces as
183
+ a graceful in-UI error message, NOT a crash, blank screen, a spinner that
184
+ never resolves, or an unhandled 500. Force one and watch the UI recover.
185
+ - [ ] The provider key + the provider call are SERVER-side only: the key
186
+ never reaches the browser (check the network tab, the JS bundle, and page
187
+ globals), and no route proxies arbitrary prompts to the model without auth
188
+ - a token cap — an open AI endpoint is an unbounded bill and abuse vector.
@@ -1 +1 @@
1
- {"version":3,"file":"provider.d.ts","sourceRoot":"","sources":["../src/provider.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EACV,UAAU,EAOX,MAAM,kBAAkB,CAAA;AASzB,OAAO,cAAc,CAAA;AAErB,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAA;AAymB/C;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,MAAM,CAAC,EAAE,aAAa,GAAG,UAAU,CAEjE"}
1
+ {"version":3,"file":"provider.d.ts","sourceRoot":"","sources":["../src/provider.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EACV,UAAU,EAOX,MAAM,kBAAkB,CAAA;AASzB,OAAO,cAAc,CAAA;AAErB,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAA;AA+oB/C;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,MAAM,CAAC,EAAE,aAAa,GAAG,UAAU,CAEjE"}
package/dist/provider.js CHANGED
@@ -72,7 +72,12 @@ class MiniMaxAIProvider {
72
72
  const canonicalModel = params.model ?? this.defaultModel;
73
73
  const model = this.modelMap[canonicalModel] ?? canonicalModel;
74
74
  const maxTokens = params.maxTokens ?? this.maxTokens;
75
- const messages = this.formatMessages(params.messages, params.system, params.cacheControl);
75
+ const messages = this.formatMessages(params.messages, params.system, params.cacheControl,
76
+ // M3 is natively multimodal (M2.x is text-only) — forward image blocks
77
+ // for it; both the native host and the DeepInfra re-host accept the
78
+ // OpenAI image_url data-URI form. Gate on the CANONICAL id: the mapped
79
+ // upstream id never matches the family anchor.
80
+ isM3Family(canonicalModel));
76
81
  const body = {
77
82
  // Abuse attribution (see ChatParams.endUserId): OpenAI-compatible APIs
78
83
  // scope enforcement by `user`, so an abusive tenant can be actioned
@@ -91,14 +96,33 @@ class MiniMaxAIProvider {
91
96
  const allTools = [...functions, ...serverTools];
92
97
  if (allTools.length > 0) {
93
98
  body.tools = allTools;
99
+ // Honor forced tool choice (discovery mode relies on it — an unenforced
100
+ // 'required' lets the model narrate the call as text instead of emitting
101
+ // tool_use). OpenAI-compatible forms; 'auto' is the API default → omit.
102
+ if (params.toolChoice === 'required') {
103
+ body.tool_choice = 'required';
104
+ }
105
+ else if (typeof params.toolChoice === 'object' && params.toolChoice.type === 'tool') {
106
+ body.tool_choice = { type: 'function', function: { name: params.toolChoice.name } };
107
+ }
94
108
  }
95
109
  // MiniMax's native API has no reasoning_effort / thinking budget. M3 takes
96
110
  // `thinking: {type: "adaptive" | "disabled"}` (the OpenAI-compatible
97
111
  // endpoint defaults to adaptive when omitted); M2.x thinking is always on
98
112
  // and takes no param. Synthase runs M3 as a non-thinking executor (like
99
113
  // DeepSeek), so explicitly disable unless thinking was requested.
100
- if (isM3Family(model)) {
114
+ // Family-gate on the CANONICAL id: `model` is the mapped upstream id on
115
+ // re-hosts (`MiniMaxAI/MiniMax-M3`), which the `minimax-m3` anchor never
116
+ // matches — that omission left thinking defaulted ON for every US turn.
117
+ if (isM3Family(canonicalModel)) {
101
118
  body.thinking = { type: params.thinking ? 'adaptive' : 'disabled' };
119
+ if (!params.thinking) {
120
+ // Open-weights re-hosts (DeepInfra) IGNORE the native `thinking` param
121
+ // and default reasoning ON — only `chat_template_kwargs.enable_thinking`
122
+ // reaches the template there. The native host accepts both together
123
+ // (verified live 2026-08-28), so send both and each host honors its own.
124
+ body.chat_template_kwargs = { enable_thinking: false };
125
+ }
102
126
  }
103
127
  // Keep thinking out of the visible message: with reasoning_split, thinking
104
128
  // arrives in `reasoning_content` (parsed below) instead of inline
@@ -212,15 +236,17 @@ class MiniMaxAIProvider {
212
236
  /**
213
237
  * Converts internal `ChatMessage` objects to the OpenAI-compatible message format.
214
238
  * System prompt is prepended as a system role message.
215
- * MiniMax does not support vision/images, so image blocks are skipped.
239
+ * Image blocks are forwarded only for vision-capable models (M3 family);
240
+ * text-only M2.x models get them skipped.
216
241
  *
217
242
  * @param messages - The chat messages to format.
218
243
  * @param system - Optional system prompt to prepend.
219
244
  * @param cacheControl - Optional cache hint for the system prompt block.
220
245
  * @param cacheControl.type - Ephemeral cache policy label expected by MiniMax.
246
+ * @param supportsImages - Whether the target model accepts image input.
221
247
  * @returns The formatted messages array for the MiniMax API.
222
248
  */
223
- formatMessages(messages, system, cacheControl) {
249
+ formatMessages(messages, system, cacheControl, supportsImages = false) {
224
250
  const formatted = [];
225
251
  if (system) {
226
252
  if (cacheControl) {
@@ -251,7 +277,15 @@ class MiniMaxAIProvider {
251
277
  parts.push({ type: 'text', text: block.text });
252
278
  break;
253
279
  case 'image':
254
- // MiniMax does not support vision/images — skip image blocks
280
+ // M3 family is natively multimodal; M2.x is text-only. Dropping an
281
+ // image SILENTLY (the pre-M3 behavior for every model) made the
282
+ // model answer blind to attachments with no error anywhere.
283
+ if (supportsImages) {
284
+ parts.push({
285
+ type: 'image_url',
286
+ image_url: { url: `data:${block.mediaType};base64,${block.data}` },
287
+ });
288
+ }
255
289
  break;
256
290
  case 'tool_use':
257
291
  toolCalls.push({
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@molecule/api-ai-minimax",
3
- "version": "1.0.0",
3
+ "version": "1.1.0",
4
4
  "description": "MiniMax AI provider for molecule.dev",
5
5
  "type": "module",
6
6
  "main": "dist/index.js",
@@ -17,7 +17,8 @@
17
17
  }
18
18
  },
19
19
  "files": [
20
- "dist"
20
+ "dist",
21
+ "README.md"
21
22
  ],
22
23
  "keywords": [
23
24
  "molecule",
@@ -25,18 +26,23 @@
25
26
  "minimax"
26
27
  ],
27
28
  "license": "Apache-2.0",
29
+ "repository": {
30
+ "type": "git",
31
+ "url": "https://github.com/molecule-dev/molecule.git",
32
+ "directory": "packages/api/bonds/ai/minimax"
33
+ },
28
34
  "devDependencies": {
29
- "@molecule/api-ai": "1.0.0",
30
- "@molecule/api-bond": "1.0.0",
35
+ "@molecule/api-ai": "1.2.0",
36
+ "@molecule/api-bond": "1.0.1",
31
37
  "@types/node": "26.1.2",
32
38
  "typescript": "6.0.3",
33
39
  "vitest": "4.1.10"
34
40
  },
35
41
  "peerDependencies": {
36
- "@molecule/api-ai": "^1.0.0",
37
- "@molecule/api-bond": "^1.0.0",
38
- "@molecule/api-i18n": "^1.0.0",
39
- "@molecule/api-secrets": "^1.0.0"
42
+ "@molecule/api-ai": "^1.0.1",
43
+ "@molecule/api-bond": "^1.0.1",
44
+ "@molecule/api-i18n": "^1.0.1",
45
+ "@molecule/api-secrets": "^1.0.1"
40
46
  },
41
47
  "peerDependenciesMeta": {
42
48
  "@molecule/api-ai": {