@hiper2d/ai-agents 0.28.0 → 0.29.0

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/dist/index.mjs CHANGED
@@ -1267,20 +1267,29 @@ var SupportedAiModels = {
1267
1267
  thinkingBudgetTokens: 1024,
1268
1268
  tags: ["slow", "cheap"]
1269
1269
  },
1270
- // MiniMax M3 (OpenAI-compatible endpoint, 1M context). Thinking-only entry: M3's `thinking`
1271
- // param defaults to adaptive (it decides per-request how much to think) and can be disabled,
1272
- // making it hybrid for cost purposes. The agent always sends `reasoning_split: true` so
1273
- // thinking arrives in `reasoning_content` instead of as `<think>` tags inside the answer.
1274
- // Note: unlike Qwen, M3 has NO thinking-budget parameter — adaptive is the only throttle.
1275
- // Speed from the 2026-08-05 live measurement (single sample): 25.3s → the >25s tier.
1276
- // Temperature: MiniMax range is [0,2], default 1.
1270
+ // MiniMax M3.1 Flash (OpenAI-compatible endpoint, 1M context). `minimax` repointed 2026-10-27
1271
+ // from MiniMax-M3 ($0.30/$1.20 after the permanent 50% off, 25s single-sample turn, 60s
1272
+ // timeouts in prod on 2026-10-25 and 10-27) to MiniMax-M3.1-Flash: half the price
1273
+ // ($0.15/$0.60), and the first MiniMax model with an effort dial — low|medium|high|xhigh|max,
1274
+ // provider default max, so it is pinned. Thinking is always on (`thinking: disabled` → 400),
1275
+ // thinking arrives in `reasoning_content` because the agent sends `reasoning_split: true`
1276
+ // (M3.1 rejects false). Temperature and top_p are documented as ignored ("not supported
1277
+ // yet", always 1 / 0.95), so the entry sets none and the agent sends none.
1278
+ // Live 2026-10-27 at medium: the all-providers sweep ran the schema ask in 3.1s, prose in
1279
+ // 1.7s and the 8-character scene (363 output tokens) in 2.4s — ~150 tok/s, the fastest
1280
+ // provider in the sweep that day (Kimi Highspeed's scene takes 8s); a 3.2k-token game
1281
+ // prompt with a two-sentence JSON reply took 1.9-2.1s at every effort level vs 3.5s for M3.
1282
+ // Tagged `fast` rather than `very-fast`: the sweep straddles the 3s line and production
1283
+ // turns carry 10k+ tokens of context. Launched 2026-09-27 as Token-Plan-only (the
1284
+ // 2026-10-02 probe got `unknown model` on the pay-as-you-go key); on the key since at
1285
+ // least 2026-10-27.
1277
1286
  [LLM_CONSTANTS.MINIMAX]: {
1278
- displayName: "MiniMax M3",
1279
- modelApiName: "MiniMax-M3",
1287
+ displayName: "MiniMax M3.1 Flash",
1288
+ modelApiName: "MiniMax-M3.1-Flash",
1280
1289
  apiKeyName: API_KEY_CONSTANTS.MINIMAX,
1281
1290
  hasThinking: true,
1282
- temperature: 1,
1283
- tags: ["very-slow", "cheap"]
1291
+ reasoningEffort: "medium",
1292
+ tags: ["fast", "cheap"]
1284
1293
  },
1285
1294
  // Meta Muse Spark 1.3 (added 2026-09-12) on Meta's own Model API — Standard tier, i.e.
1286
1295
  // the private model id (the `-contributor` id is a quarter of the price but Meta trains
@@ -1597,18 +1606,14 @@ var MODEL_PRICING = {
1597
1606
  outputPrice: 0.47,
1598
1607
  cacheHitPrice: 0.016
1599
1608
  },
1600
- // MiniMax M3. Rates from platform.minimax.io/docs/guides/pricing-paygo (2026-08-05, USD,
1601
- // "permanent 50% off" already applied): ≤512k and >512k input tiers. Caching is automatic
1602
- // (≥512 input tokens), hits reported in prompt_tokens_details.cached_tokens; no write fee
1603
- // for M3.
1609
+ // MiniMax M3.1 Flash. Rates from platform.minimax.io/docs/guides/pricing-paygo (2026-10-27,
1610
+ // USD): one tier across the 1M window (M3 had ≤512k / >512k tiers; the page lists none for
1611
+ // M3.1 Flash). Caching is automatic (≥512 input tokens), hits reported in
1612
+ // prompt_tokens_details.cached_tokens; no write fee listed.
1604
1613
  [SupportedAiModels[LLM_CONSTANTS.MINIMAX].modelApiName]: {
1605
- inputPrice: 0.3,
1606
- outputPrice: 1.2,
1607
- cacheHitPrice: 0.06,
1608
- extendedContextInputPrice: 0.6,
1609
- extendedContextOutputPrice: 2.4,
1610
- extendedContextCacheHitPrice: 0.12,
1611
- extendedContextThresholdTokens: 512e3
1614
+ inputPrice: 0.15,
1615
+ outputPrice: 0.6,
1616
+ cacheHitPrice: 0.015
1612
1617
  },
1613
1618
  // Meta Muse Spark 1.3, Standard tier. Rates from ai.developer.meta.com/docs/pricing-rate-limits
1614
1619
  // (2026-09-12): no long-context premium at any point of the 1M window; reasoning tokens bill
@@ -1638,7 +1643,7 @@ var HYBRID_THINKING_API_NAMES = /* @__PURE__ */ new Set([
1638
1643
  // hybrid by the same definition: we force reasoning on, so they carry the multiplier.
1639
1644
  SupportedAiModels[LLM_CONSTANTS.QWEN_MAX].modelApiName,
1640
1645
  SupportedAiModels[LLM_CONSTANTS.QWEN_FLASH].modelApiName,
1641
- SupportedAiModels[LLM_CONSTANTS.MINIMAX].modelApiName,
1646
+ // MiniMax left the set 2026-10-27: M3.1 Flash cannot turn thinking off (M3 could).
1642
1647
  // Mistral reasons only when asked (`reasoning_effort`), and we always ask — hybrid by the
1643
1648
  // same definition as Qwen. Added 2026-09-18 when the Mistral entries went thinking-on.
1644
1649
  SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName
@@ -1711,6 +1716,7 @@ var META_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh", "max"
1711
1716
  var MISTRAL_REASONING_EFFORTS = ["high"];
1712
1717
  var XAI_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
1713
1718
  var KIMI_REASONING_EFFORTS = ["low", "high", "max"];
1719
+ var MINIMAX_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
1714
1720
  function clampReasoningEffort(effort, allowed) {
1715
1721
  const rank = REASONING_EFFORT_SCALE.indexOf(effort);
1716
1722
  let best = allowed[0];
@@ -1734,6 +1740,7 @@ var toFuguEffort = (effort) => clampReasoningEffort(effort, FUGU_REASONING_EFFOR
1734
1740
  var toMistralEffort = (effort) => clampReasoningEffort(effort, MISTRAL_REASONING_EFFORTS);
1735
1741
  var toXaiEffort = (effort) => clampReasoningEffort(effort, XAI_REASONING_EFFORTS);
1736
1742
  var toKimiEffort = (effort) => clampReasoningEffort(effort, KIMI_REASONING_EFFORTS);
1743
+ var toMiniMaxEffort = (effort) => clampReasoningEffort(effort, MINIMAX_REASONING_EFFORTS);
1737
1744
 
1738
1745
  // src/pricing/token-usage-utils.ts
1739
1746
  function extractTokenUsage(response) {
@@ -5051,10 +5058,10 @@ var MiniMaxAgent = class extends AbstractAgent {
5051
5058
  get defaultParams() {
5052
5059
  return {
5053
5060
  model: this.model,
5054
- temperature: this.temperature,
5061
+ ...this.temperature !== void 0 ? { temperature: this.temperature } : {},
5055
5062
  stream: false,
5056
- // MiniMax deprecates max_tokens in favor of max_completion_tokens (M3 max is 512K,
5057
- // far above anything a turn needs).
5063
+ // MiniMax deprecates max_tokens in favor of max_completion_tokens (the M-series max is
5064
+ // 512K, far above anything a turn needs).
5058
5065
  max_completion_tokens: this.maxOutputTokens
5059
5066
  };
5060
5067
  }
@@ -5076,7 +5083,9 @@ var MiniMaxAgent = class extends AbstractAgent {
5076
5083
  thinkingParams() {
5077
5084
  return {
5078
5085
  thinking: { type: this.enableThinking ? "adaptive" : "disabled" },
5079
- reasoning_split: true
5086
+ reasoning_split: true,
5087
+ // Provider default is max; the catalog pins M3.1 Flash to medium. Older models ignore it.
5088
+ ...this.enableThinking && this.reasoningEffort ? { reasoning_effort: toMiniMaxEffort(this.reasoningEffort) } : {}
5080
5089
  };
5081
5090
  }
5082
5091
  convertToOpenAIMessages(messages) {
@@ -5489,7 +5498,7 @@ var AgentFactory = class {
5489
5498
  case LLM_CONSTANTS.QWEN_MAX:
5490
5499
  case LLM_CONSTANTS.QWEN_FLASH:
5491
5500
  return new QwenAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
5492
- // MiniMax M3 — adaptive thinking (the model decides per-request)
5501
+ // MiniMax M3.1 Flash — always-on adaptive thinking, effort from the catalog
5493
5502
  case LLM_CONSTANTS.MINIMAX:
5494
5503
  return new MiniMaxAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
5495
5504
  // Meta Muse Spark — Responses API with encrypted reasoning replay, effort from the catalog
@@ -5540,6 +5549,7 @@ export {
5540
5549
  LLM_CONSTANTS,
5541
5550
  MESSAGE_ROLE,
5542
5551
  META_REASONING_EFFORTS,
5552
+ MINIMAX_REASONING_EFFORTS,
5543
5553
  MISTRAL_REASONING_EFFORTS,
5544
5554
  MODEL_PRICING,
5545
5555
  MetaAgent,
@@ -5646,6 +5656,7 @@ export {
5646
5656
  toGlmEffort,
5647
5657
  toKimiEffort,
5648
5658
  toMetaEffort,
5659
+ toMiniMaxEffort,
5649
5660
  toMistralEffort,
5650
5661
  toOpenAIEffort,
5651
5662
  toXaiEffort,