@stackfactor/agent-utils 1.4.0 → 1.4.3

package/README.md

@@ -25,7 +25,7 @@ import {
 
 ## `langChain`
 
-Unified interface for running LLM prompts, managing LangChain agents, and generating images across OpenAI, Anthropic, Google, DeepSeek, Kimi (Moonshot), and GLM (Zhipu) providers.
+Unified interface for running LLM prompts, managing LangChain agents, and generating images across the OpenAI, Anthropic, and Google providers.
 
 ### `langChain.runPromptWithModel(modelName, config, prompt, onProgressReport?, minPercent?, maxPercent?, expectsJsonResponse?, schema?, agentName?, tools?, usageTracker?)`
 
@@ -39,8 +39,8 @@ When `expectsJsonResponse` is `true` (the default), JSON escape instructions are
 
 | Parameter             | Type                 | Default         | Description                                                                                              |
 | --------------------- | -------------------- | --------------- | -------------------------------------------------------------------------------------------------------- |
-| `modelName`           | `string`             | —               | Model identifier: `"gpt-4o"`, `"claude-3-5-sonnet"`, `"gemini-1.5-pro"`, `"deepseek-chat"`, `"kimi-k2-0905-preview"`, `"glm-4.6"`, etc. |
-| `config`              | `object`             | —               | API keys (`openAIAPIKey`, `anthropicAPIKey`, `googleAPIKey`, `deepSeekAPIKey`, `kimiAPIKey`, `glmAPIKey`), `temperature`, `agentic`, `recursionLimit`, `tavily`, `tavilyAPIKey` |
+| `modelName`           | `string`             | —               | Model identifier: `"gpt-4o"`, `"claude-3-5-sonnet"`, `"gemini-1.5-pro"`, etc. |
+| `config`              | `object`             | —               | API keys (`openAIAPIKey`, `anthropicAPIKey`, `googleAPIKey`), `temperature`, `agentic`, `recursionLimit`, `tavily`, `tavilyAPIKey` |
 | `prompt`              | `string \| object[]` | —               | Plain string or array of `{ role, content }` message objects                                             |
 | `onProgressReport`    | `function \| null`   | `null`          | Async callback receiving `{ message, progress }` updates                                                 |
 | `minPercent`          | `number`             | `0`             | Lower bound for progress percentage                                                                      |
@@ -338,7 +338,7 @@ A model that cannot tell it received half a regulation will reason over the half
 
 **Notes**
 
-- Works with every supported provider, including DeepSeek, Kimi and GLM, which have no native web search of their own.
+- Works with every supported provider.
 - Credits come from each response's own `usage` block rather than an estimate, and are accumulated on `usageTracker.tokens.tavilyCredits`, costed from the `tavily-credit-costs` constant (USD per credit; falls back to Tavily's list price of `0.008`).
 - URLs are collected on `usageTracker.webSearchSources`, deduplicated across the whole run — Tavily's terms require citing the original sources when its content is shown to end users.
 - Per-URL failures are surfaced in the tool result instead of being dropped, so a fetch failure is not mistaken for an absence of content.
@@ -355,11 +355,10 @@ A model that cannot tell it received half a regulation will reason over the half
 | OpenAI            | `gpt-`                 | `gpt-4o`, `gpt-4o-mini`              |
 | Anthropic         | `claude-`              | `claude-3-5-sonnet`, `claude-3-opus` |
 | Google            | `gemini-`              | `gemini-1.5-pro`, `gemini-2.0-flash` |
-| DeepSeek          | `deepseek-`            | `deepseek-chat`, `deepseek-reasoner` |
-| Kimi (Moonshot)   | `kimi-`, `moonshot-`   | `kimi-k2-0905-preview`, `moonshot-v1-8k` |
-| GLM (Zhipu)       | `glm-`                 | `glm-4.6`, `glm-4.5`                 |
 
-DeepSeek, Kimi, and GLM are routed through their OpenAI-compatible endpoints. Native structured output via `response_format` with JSON schema is only applied to `gpt-*` models; all other providers (including these three) use prompt-based JSON instructions.
+Native structured output is configured per provider when a Zod `schema` is passed: `response_format` with a JSON schema for `gpt-*`, `output_config.format` for `claude-*`, and JSON mode plus `responseSchema` for `gemini-*`.
+
+Output token limits are not configurable. Anthropic requires `max_tokens` on every request, so `claude-*` is pinned to `64000` with streaming always enabled; `gpt-*` and `gemini-*` are sent no limit and use each provider's own default, which is that model's maximum.
 
 ### Image Generation
 
@@ -379,11 +378,7 @@ The `config` object accepted by LangChain methods supports the following keys:
 | `openAIAPIKey`    | `string`  | OpenAI API key                              |
 | `anthropicAPIKey` | `string`  | Anthropic API key                           |
 | `googleAPIKey`    | `string`  | Google AI API key                           |
-| `deepSeekAPIKey`  | `string`  | DeepSeek API key                            |
-| `kimiAPIKey`      | `string`  | Kimi (Moonshot) API key                     |
-| `glmAPIKey`       | `string`  | GLM (Zhipu) API key                         |
 | `temperature`     | `number`  | Sampling temperature                        |
-| `maxTokens`       | `number`  | Maximum output tokens (default: `16384` for Claude, `200000` for other providers). For Claude, values above `21333` automatically enable LangChain streaming to bypass the Anthropic SDK's 10-minute non-streaming guard. |
 | `agentic`         | `boolean` | Enable agentic mode in `runPromptWithModel` |
 | `recursionLimit`  | `number`  | Max agent steps (default: `25`)             |
 | `tavilyAPIKey`    | `string`  | Tavily API key; required when `tavily` is set                                |

package/dist/cjs/langChain.d.ts

@@ -23,23 +23,22 @@ export type UsageTracker = {
 /**
  * Instantiates and returns the appropriate LangChain chat model based on the model
  * name prefix. `claude-` maps to `ChatAnthropic`, `gemini-` maps to
- * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. DeepSeek (`deepseek-`),
- * Kimi/Moonshot (`kimi-`, `moonshot-`), and GLM/Zhipu (`glm-`) models are routed
- * through `ChatOpenAI` against each provider's OpenAI-compatible endpoint. When a Zod
+ * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. When a Zod
  * `schema` is provided, native structured output is configured per provider: OpenAI
  * via `response_format` with `json_schema`, Anthropic via `output_config.format`, and
  * Gemini via JSON mode (`json: true`) plus `responseSchema`. In every case the model
  * emits JSON as the message text, so the caller's parse/validate pipeline is unchanged.
- * The schema is ignored for the OpenAI-compatible providers (DeepSeek/Kimi/GLM), which
- * have no native structured-output support here. Throws a `BAD_REQUEST` error for
- * unrecognised model names.
+ * Throws a `BAD_REQUEST` error for unrecognised model names.
+ *
+ * Output token limits are not caller-configurable. Anthropic requires
+ * `max_tokens` on every request and @langchain/anthropic defaults it to 4096,
+ * so the Claude branch pins it explicitly; OpenAI and Gemini are sent no limit
+ * and fall back to each provider's own default, which is that model's maximum.
  * @param modelName - The model identifier, e.g. `"gpt-4o"`, `"claude-3-5-sonnet"`,
- *   `"gemini-1.5-pro"`, `"deepseek-chat"`, `"kimi-k2-0905-preview"`, `"glm-4.6"`
+ *   `"gemini-1.5-pro"`
  * @param config - Configuration object containing API keys (`openAIAPIKey`,
- *   `anthropicAPIKey`, `googleAPIKey`, `deepSeekAPIKey`, `kimiAPIKey`, `glmAPIKey`),
- *   optional `maxTokens`, and optional `temperature`
+ *   `anthropicAPIKey`, `googleAPIKey`) and optional `temperature`
  * @param schema - Optional Zod schema used to configure native structured JSON output
- *   for GPT / Claude / Gemini models; ignored for OpenAI-compatible providers
  * @returns A configured LangChain chat model (or bound runnable) instance
  */
 export declare const getLLMModel: (modelName: string, config: any, schema?: any) => any;

package/dist/cjs/langChain.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA4GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AAkoBF;;;;;;;;;;;;;;;;;;;;;GAqBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAoHF,CAAC;;0CAhxB2C,GAAG,KAAG,IAAI;wBAmyB/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDAmTF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CA7zC8B,GAAG,KAAG,MAAM;+CA/wB9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCA+xBU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AAg2CT,wBAUE"}
+{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA6GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AA0lBF;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAwHF,CAAC;;0CA3uB2C,GAAG,KAAG,IAAI;wBA8vB/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDAmTF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CA7zC8B,GAAG,KAAG,MAAM;+CA1uB9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCA0vBU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AAg2CT,wBAUE"}

package/dist/cjs/langChain.js

@@ -30,9 +30,10 @@ Do NOT include raw newlines, tabs, or unescaped quotes inside JSON string values
  * Providers for which `getLLMModel` configures native, schema-constrained JSON
  * output (OpenAI `response_format`, Anthropic `output_config.format`, Gemini
  * `responseSchema` + JSON mode). For these the model emits JSON directly, so the
- * prompt-injected schema and the schema-validation retry are skipped. Other
- * providers (DeepSeek/Kimi/GLM via the OpenAI-compatible shim) have no native
- * support here and continue to rely on prompt instructions + validation.
+ * prompt-injected schema and the schema-validation retry are skipped. Every
+ * supported provider is listed here; the check remains so that adding one
+ * without native support falls back to prompt instructions + validation
+ * rather than silently skipping them.
  * @param modelName - The model identifier being routed
  * @returns `true` when native structured output is wired up for the provider
  */
@@ -543,36 +544,6 @@ const extractJSONFromResponse = (text) => {
     }
     return null;
 };
-/**
- * Maps a model name prefix to its OpenAI-compatible provider configuration. DeepSeek,
- * Kimi (Moonshot), and GLM (Zhipu) all expose OpenAI-compatible chat completion APIs
- * and can be routed through `ChatOpenAI` with a custom base URL and API key.
- * @param modelName - The model identifier to inspect
- * @param config - Configuration object containing provider-specific API keys
- * @returns An object with `apiKey` and `baseURL` when the model matches a known
- *   OpenAI-compatible provider, or `null` otherwise
- */
-const getOpenAICompatibleProvider = (modelName, config) => {
-    if (modelName.startsWith("deepseek-")) {
-        return {
-            apiKey: config.deepSeekAPIKey,
-            baseURL: "https://api.deepseek.com/v1",
-        };
-    }
-    if (modelName.startsWith("kimi-") || modelName.startsWith("moonshot-")) {
-        return {
-            apiKey: config.kimiAPIKey,
-            baseURL: "https://api.moonshot.ai/v1",
-        };
-    }
-    if (modelName.startsWith("glm-")) {
-        return {
-            apiKey: config.glmAPIKey,
-            baseURL: "https://open.bigmodel.cn/api/paas/v4/",
-        };
-    }
-    return null;
-};
 /**
  * Whether a model accepts the `temperature` sampling parameter at all. Some
  * models reject it with an HTTP 400, so it must be omitted rather than clamped:
@@ -582,7 +553,7 @@ const getOpenAICompatibleProvider = (modelName, config) => {
  *    accept it.
  *  - OpenAI reasoning models (the GPT-5 family and the o-series) reject it;
  *    gpt-4o still accepts it.
- * DeepSeek, Kimi/Moonshot, GLM, and Gemini all accept temperature.
+ * Gemini accepts temperature.
  * @param modelName - the model identifier being routed
  * @returns `true` when the model accepts a `temperature` parameter
  */
@@ -603,18 +574,14 @@ const modelSupportsTemperature = (modelName) => {
 };
 /**
  * Per-provider valid `temperature` range, used to clamp caller-supplied values.
- * Anthropic, Kimi (Moonshot), and GLM (Zhipu) cap at 1.0; OpenAI, Gemini, and
- * DeepSeek accept up to 2.0. Only consulted for models that support temperature
- * at all (see `modelSupportsTemperature`).
+ * Anthropic caps at 1.0; OpenAI and Gemini accept up to 2.0. Only consulted for
+ * models that support temperature at all (see `modelSupportsTemperature`).
  */
 const getTemperatureRange = (modelName) => {
-    if (modelName.startsWith("claude-") ||
-        modelName.startsWith("kimi-") ||
-        modelName.startsWith("moonshot-") ||
-        modelName.startsWith("glm-")) {
+    if (modelName.startsWith("claude-")) {
         return { min: 0, max: 1 };
     }
-    // gpt-, gemini-, deepseek-
+    // gpt-, gemini-
     return { min: 0, max: 2 };
 };
 /**
@@ -670,23 +637,22 @@ const extractTextContent = (content) => {
 /**
  * Instantiates and returns the appropriate LangChain chat model based on the model
  * name prefix. `claude-` maps to `ChatAnthropic`, `gemini-` maps to
- * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. DeepSeek (`deepseek-`),
- * Kimi/Moonshot (`kimi-`, `moonshot-`), and GLM/Zhipu (`glm-`) models are routed
- * through `ChatOpenAI` against each provider's OpenAI-compatible endpoint. When a Zod
+ * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. When a Zod
  * `schema` is provided, native structured output is configured per provider: OpenAI
  * via `response_format` with `json_schema`, Anthropic via `output_config.format`, and
  * Gemini via JSON mode (`json: true`) plus `responseSchema`. In every case the model
  * emits JSON as the message text, so the caller's parse/validate pipeline is unchanged.
- * The schema is ignored for the OpenAI-compatible providers (DeepSeek/Kimi/GLM), which
- * have no native structured-output support here. Throws a `BAD_REQUEST` error for
- * unrecognised model names.
+ * Throws a `BAD_REQUEST` error for unrecognised model names.
+ *
+ * Output token limits are not caller-configurable. Anthropic requires
+ * `max_tokens` on every request and @langchain/anthropic defaults it to 4096,
+ * so the Claude branch pins it explicitly; OpenAI and Gemini are sent no limit
+ * and fall back to each provider's own default, which is that model's maximum.
  * @param modelName - The model identifier, e.g. `"gpt-4o"`, `"claude-3-5-sonnet"`,
- *   `"gemini-1.5-pro"`, `"deepseek-chat"`, `"kimi-k2-0905-preview"`, `"glm-4.6"`
+ *   `"gemini-1.5-pro"`
  * @param config - Configuration object containing API keys (`openAIAPIKey`,
- *   `anthropicAPIKey`, `googleAPIKey`, `deepSeekAPIKey`, `kimiAPIKey`, `glmAPIKey`),
- *   optional `maxTokens`, and optional `temperature`
+ *   `anthropicAPIKey`, `googleAPIKey`) and optional `temperature`
  * @param schema - Optional Zod schema used to configure native structured JSON output
- *   for GPT / Claude / Gemini models; ignored for OpenAI-compatible providers
  * @returns A configured LangChain chat model (or bound runnable) instance
  */
 // Exported (but not re-exported from index.ts / the package's public entry
@@ -697,12 +663,17 @@ const getLLMModel = (modelName, config, schema = null) => {
     const modelSettings = resolveTemperatureSetting(modelName, config);
     // Claude models (Anthropic)
     if (modelName.startsWith("claude-")) {
+        // Anthropic requires `max_tokens` on every request and @langchain/anthropic
+        // defaults it to 4096 when omitted — low enough to truncate agentic turns
+        // mid-tool-call — so it is pinned here rather than left to the library.
+        // 64000 is Anthropic's recommended value for streamed requests and sits
+        // within the 128000-token output limit of every Claude model routed here.
+        //
         // Anthropic's SDK rejects non-streamed requests when max_tokens is large
-        // enough that the operation could exceed 10 minutes (threshold: 21333).
-        // Auto-enable LangChain streaming above that cutoff so `.invoke()` routes
-        // through the streaming endpoint internally — callers see no interface
-        // change, but the guard is bypassed.
-        const maxTokens = config.maxTokens || 16384;
+        // enough that the operation could exceed 10 minutes (threshold: 21333), so
+        // streaming is always on: `.invoke()` routes through the streaming endpoint
+        // internally and callers see no interface change.
+        const maxTokens = 64000;
         // Native structured output: Anthropic's `output_config.format` constrains
         // decoding to the schema while still emitting JSON as the message text, so
         // the downstream streaming + parse/validate pipeline is unchanged.
@@ -723,7 +694,7 @@ const getLLMModel = (modelName, config, schema = null) => {
             apiKey: config.anthropicAPIKey,
             maxTokens,
             modelName: modelName,
-            streaming: maxTokens > 16384,
+            streaming: true,
             invocationKwargs: { cache_control: { type: "ephemeral" } },
             ...outputConfig,
             ...modelSettings,
@@ -734,7 +705,9 @@ const getLLMModel = (modelName, config, schema = null) => {
     else if (modelName.startsWith("gemini-")) {
         const model = new google_genai_1.ChatGoogleGenerativeAI({
             apiKey: config.googleAPIKey,
-            maxOutputTokens: config.maxTokens || 200000,
+            // No `maxOutputTokens`: omitting it leaves Gemini's own default, which is
+            // the model's maximum output. Sending a value meant tracking a per-model
+            // ceiling that the API rejects outright rather than clamps.
             model: modelName,
             // `json: true` forces responseMimeType=application/json, guaranteeing the
             // model returns valid JSON text (never prose) for the parse pipeline.
@@ -756,17 +729,26 @@ const getLLMModel = (modelName, config, schema = null) => {
     else if (modelName.startsWith("gpt-")) {
         const openAISettings = {
             apiKey: config.openAIAPIKey,
-            max_tokens: config.maxTokens || 200000,
+            // No output limit is sent: OpenAI defaults to the model's maximum. The
+            // snake_case `max_tokens` that used to sit here was never a recognised
+            // ChatOpenAI field and was silently dropped by LangChain.
             modelName: modelName,
             ...modelSettings,
         };
-        // gpt-5.6's /v1/chat/completions endpoint 400s whenever function tools are
-        // bound, because it rejects the model's own server-side default reasoning
-        // effort in that combination ("use /v1/responses or set reasoning_effort
-        // to 'none'"); forcing it to "none" here avoids the collision. Scoped to
-        // exactly gpt-5.6 — sibling gpt-5.x models are not confirmed to share the
-        // bug.
-        const reasoningEffortOverride = modelName === "gpt-5.6" ? { reasoning_effort: "none" } : {};
+        // The gpt-5.6 family's /v1/chat/completions endpoint 400s whenever function
+        // tools are bound, because it rejects the model's own server-side default
+        // reasoning effort in that combination ("use /v1/responses or set
+        // reasoning_effort to 'none'"); forcing it to "none" here avoids the
+        // collision. The restriction belongs to the family, not to one id: the
+        // `gpt-5.6` alias resolves to `gpt-5.6-sol`, and `gpt-5.6-terra` /
+        // `gpt-5.6-luna` are separately addressable ids that hit the same 400, as
+        // do dated snapshots (`gpt-5.6-2026-08-22`). Matching the alias alone left
+        // every one of those unpatched, so match the whole family. Still excludes
+        // sibling releases (gpt-5.5, gpt-5.7): they are not confirmed to share the
+        // bug, and reasoning effort is worth keeping wherever it is accepted.
+        const reasoningEffortOverride = /^gpt-5\.6($|-)/.test(modelName)
+            ? { reasoning_effort: "none" }
+            : {};
         // Use native structured output with a JSON schema. `modelKwargs` is spread
         // verbatim into the Chat Completions request. Merge (not overwrite) with
         // the reasoning-effort override above, since both may apply at once.
@@ -789,17 +771,6 @@ const getLLMModel = (modelName, config, schema = null) => {
         }
         return new openai_1.ChatOpenAI(openAISettings);
     }
-    // OpenAI-compatible providers: DeepSeek, Kimi (Moonshot), GLM (Zhipu)
-    const openAICompatible = getOpenAICompatibleProvider(modelName, config);
-    if (openAICompatible) {
-        return new openai_1.ChatOpenAI({
-            apiKey: openAICompatible.apiKey,
-            maxTokens: config.maxTokens || 200000,
-            modelName: modelName,
-            configuration: { baseURL: openAICompatible.baseURL },
-            ...modelSettings,
-        });
-    }
     throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.BAD_REQUEST, const_js_1.default.ERROR.UNSUPPORTED_MODEL + ": " + modelName);
 };
 exports.getLLMModel = getLLMModel;

package/dist/esm/langChain.d.ts

@@ -23,23 +23,22 @@ export type UsageTracker = {
 /**
  * Instantiates and returns the appropriate LangChain chat model based on the model
  * name prefix. `claude-` maps to `ChatAnthropic`, `gemini-` maps to
- * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. DeepSeek (`deepseek-`),
- * Kimi/Moonshot (`kimi-`, `moonshot-`), and GLM/Zhipu (`glm-`) models are routed
- * through `ChatOpenAI` against each provider's OpenAI-compatible endpoint. When a Zod
+ * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. When a Zod
  * `schema` is provided, native structured output is configured per provider: OpenAI
  * via `response_format` with `json_schema`, Anthropic via `output_config.format`, and
  * Gemini via JSON mode (`json: true`) plus `responseSchema`. In every case the model
  * emits JSON as the message text, so the caller's parse/validate pipeline is unchanged.
- * The schema is ignored for the OpenAI-compatible providers (DeepSeek/Kimi/GLM), which
- * have no native structured-output support here. Throws a `BAD_REQUEST` error for
- * unrecognised model names.
+ * Throws a `BAD_REQUEST` error for unrecognised model names.
+ *
+ * Output token limits are not caller-configurable. Anthropic requires
+ * `max_tokens` on every request and @langchain/anthropic defaults it to 4096,
+ * so the Claude branch pins it explicitly; OpenAI and Gemini are sent no limit
+ * and fall back to each provider's own default, which is that model's maximum.
  * @param modelName - The model identifier, e.g. `"gpt-4o"`, `"claude-3-5-sonnet"`,
- *   `"gemini-1.5-pro"`, `"deepseek-chat"`, `"kimi-k2-0905-preview"`, `"glm-4.6"`
+ *   `"gemini-1.5-pro"`
  * @param config - Configuration object containing API keys (`openAIAPIKey`,
- *   `anthropicAPIKey`, `googleAPIKey`, `deepSeekAPIKey`, `kimiAPIKey`, `glmAPIKey`),
- *   optional `maxTokens`, and optional `temperature`
+ *   `anthropicAPIKey`, `googleAPIKey`) and optional `temperature`
  * @param schema - Optional Zod schema used to configure native structured JSON output
- *   for GPT / Claude / Gemini models; ignored for OpenAI-compatible providers
  * @returns A configured LangChain chat model (or bound runnable) instance
  */
 export declare const getLLMModel: (modelName: string, config: any, schema?: any) => any;

package/dist/esm/langChain.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA4GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AAkoBF;;;;;;;;;;;;;;;;;;;;;GAqBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAoHF,CAAC;;0CAhxB2C,GAAG,KAAG,IAAI;wBAmyB/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDAmTF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CA7zC8B,GAAG,KAAG,MAAM;+CA/wB9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCA+xBU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AAg2CT,wBAUE"}
+{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA6GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AA0lBF;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAwHF,CAAC;;0CA3uB2C,GAAG,KAAG,IAAI;wBA8vB/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDAmTF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CA7zC8B,GAAG,KAAG,MAAM;+CA1uB9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCA0vBU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AAg2CT,wBAUE"}

package/dist/esm/langChain.js

@@ -24,9 +24,10 @@ Do NOT include raw newlines, tabs, or unescaped quotes inside JSON string values
  * Providers for which `getLLMModel` configures native, schema-constrained JSON
  * output (OpenAI `response_format`, Anthropic `output_config.format`, Gemini
  * `responseSchema` + JSON mode). For these the model emits JSON directly, so the
- * prompt-injected schema and the schema-validation retry are skipped. Other
- * providers (DeepSeek/Kimi/GLM via the OpenAI-compatible shim) have no native
- * support here and continue to rely on prompt instructions + validation.
+ * prompt-injected schema and the schema-validation retry are skipped. Every
+ * supported provider is listed here; the check remains so that adding one
+ * without native support falls back to prompt instructions + validation
+ * rather than silently skipping them.
  * @param modelName - The model identifier being routed
  * @returns `true` when native structured output is wired up for the provider
  */
@@ -537,36 +538,6 @@ const extractJSONFromResponse = (text) => {
     }
     return null;
 };
-/**
- * Maps a model name prefix to its OpenAI-compatible provider configuration. DeepSeek,
- * Kimi (Moonshot), and GLM (Zhipu) all expose OpenAI-compatible chat completion APIs
- * and can be routed through `ChatOpenAI` with a custom base URL and API key.
- * @param modelName - The model identifier to inspect
- * @param config - Configuration object containing provider-specific API keys
- * @returns An object with `apiKey` and `baseURL` when the model matches a known
- *   OpenAI-compatible provider, or `null` otherwise
- */
-const getOpenAICompatibleProvider = (modelName, config) => {
-    if (modelName.startsWith("deepseek-")) {
-        return {
-            apiKey: config.deepSeekAPIKey,
-            baseURL: "https://api.deepseek.com/v1",
-        };
-    }
-    if (modelName.startsWith("kimi-") || modelName.startsWith("moonshot-")) {
-        return {
-            apiKey: config.kimiAPIKey,
-            baseURL: "https://api.moonshot.ai/v1",
-        };
-    }
-    if (modelName.startsWith("glm-")) {
-        return {
-            apiKey: config.glmAPIKey,
-            baseURL: "https://open.bigmodel.cn/api/paas/v4/",
-        };
-    }
-    return null;
-};
 /**
  * Whether a model accepts the `temperature` sampling parameter at all. Some
  * models reject it with an HTTP 400, so it must be omitted rather than clamped:
@@ -576,7 +547,7 @@ const getOpenAICompatibleProvider = (modelName, config) => {
  *    accept it.
  *  - OpenAI reasoning models (the GPT-5 family and the o-series) reject it;
  *    gpt-4o still accepts it.
- * DeepSeek, Kimi/Moonshot, GLM, and Gemini all accept temperature.
+ * Gemini accepts temperature.
  * @param modelName - the model identifier being routed
  * @returns `true` when the model accepts a `temperature` parameter
  */
@@ -597,18 +568,14 @@ const modelSupportsTemperature = (modelName) => {
 };
 /**
  * Per-provider valid `temperature` range, used to clamp caller-supplied values.
- * Anthropic, Kimi (Moonshot), and GLM (Zhipu) cap at 1.0; OpenAI, Gemini, and
- * DeepSeek accept up to 2.0. Only consulted for models that support temperature
- * at all (see `modelSupportsTemperature`).
+ * Anthropic caps at 1.0; OpenAI and Gemini accept up to 2.0. Only consulted for
+ * models that support temperature at all (see `modelSupportsTemperature`).
  */
 const getTemperatureRange = (modelName) => {
-    if (modelName.startsWith("claude-") ||
-        modelName.startsWith("kimi-") ||
-        modelName.startsWith("moonshot-") ||
-        modelName.startsWith("glm-")) {
+    if (modelName.startsWith("claude-")) {
         return { min: 0, max: 1 };
     }
-    // gpt-, gemini-, deepseek-
+    // gpt-, gemini-
     return { min: 0, max: 2 };
 };
 /**
@@ -664,23 +631,22 @@ const extractTextContent = (content) => {
 /**
  * Instantiates and returns the appropriate LangChain chat model based on the model
  * name prefix. `claude-` maps to `ChatAnthropic`, `gemini-` maps to
- * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. DeepSeek (`deepseek-`),
- * Kimi/Moonshot (`kimi-`, `moonshot-`), and GLM/Zhipu (`glm-`) models are routed
- * through `ChatOpenAI` against each provider's OpenAI-compatible endpoint. When a Zod
+ * `ChatGoogleGenerativeAI`, and `gpt-` maps to `ChatOpenAI`. When a Zod
  * `schema` is provided, native structured output is configured per provider: OpenAI
  * via `response_format` with `json_schema`, Anthropic via `output_config.format`, and
  * Gemini via JSON mode (`json: true`) plus `responseSchema`. In every case the model
  * emits JSON as the message text, so the caller's parse/validate pipeline is unchanged.
- * The schema is ignored for the OpenAI-compatible providers (DeepSeek/Kimi/GLM), which
- * have no native structured-output support here. Throws a `BAD_REQUEST` error for
- * unrecognised model names.
+ * Throws a `BAD_REQUEST` error for unrecognised model names.
+ *
+ * Output token limits are not caller-configurable. Anthropic requires
+ * `max_tokens` on every request and @langchain/anthropic defaults it to 4096,
+ * so the Claude branch pins it explicitly; OpenAI and Gemini are sent no limit
+ * and fall back to each provider's own default, which is that model's maximum.
  * @param modelName - The model identifier, e.g. `"gpt-4o"`, `"claude-3-5-sonnet"`,
- *   `"gemini-1.5-pro"`, `"deepseek-chat"`, `"kimi-k2-0905-preview"`, `"glm-4.6"`
+ *   `"gemini-1.5-pro"`
  * @param config - Configuration object containing API keys (`openAIAPIKey`,
- *   `anthropicAPIKey`, `googleAPIKey`, `deepSeekAPIKey`, `kimiAPIKey`, `glmAPIKey`),
- *   optional `maxTokens`, and optional `temperature`
+ *   `anthropicAPIKey`, `googleAPIKey`) and optional `temperature`
  * @param schema - Optional Zod schema used to configure native structured JSON output
- *   for GPT / Claude / Gemini models; ignored for OpenAI-compatible providers
  * @returns A configured LangChain chat model (or bound runnable) instance
  */
 // Exported (but not re-exported from index.ts / the package's public entry
@@ -691,12 +657,17 @@ export const getLLMModel = (modelName, config, schema = null) => {
     const modelSettings = resolveTemperatureSetting(modelName, config);
     // Claude models (Anthropic)
     if (modelName.startsWith("claude-")) {
+        // Anthropic requires `max_tokens` on every request and @langchain/anthropic
+        // defaults it to 4096 when omitted — low enough to truncate agentic turns
+        // mid-tool-call — so it is pinned here rather than left to the library.
+        // 64000 is Anthropic's recommended value for streamed requests and sits
+        // within the 128000-token output limit of every Claude model routed here.
+        //
         // Anthropic's SDK rejects non-streamed requests when max_tokens is large
-        // enough that the operation could exceed 10 minutes (threshold: 21333).
-        // Auto-enable LangChain streaming above that cutoff so `.invoke()` routes
-        // through the streaming endpoint internally — callers see no interface
-        // change, but the guard is bypassed.
-        const maxTokens = config.maxTokens || 16384;
+        // enough that the operation could exceed 10 minutes (threshold: 21333), so
+        // streaming is always on: `.invoke()` routes through the streaming endpoint
+        // internally and callers see no interface change.
+        const maxTokens = 64000;
         // Native structured output: Anthropic's `output_config.format` constrains
         // decoding to the schema while still emitting JSON as the message text, so
         // the downstream streaming + parse/validate pipeline is unchanged.
@@ -717,7 +688,7 @@ export const getLLMModel = (modelName, config, schema = null) => {
             apiKey: config.anthropicAPIKey,
             maxTokens,
             modelName: modelName,
-            streaming: maxTokens > 16384,
+            streaming: true,
             invocationKwargs: { cache_control: { type: "ephemeral" } },
             ...outputConfig,
             ...modelSettings,
@@ -728,7 +699,9 @@ export const getLLMModel = (modelName, config, schema = null) => {
     else if (modelName.startsWith("gemini-")) {
         const model = new ChatGoogleGenerativeAI({
             apiKey: config.googleAPIKey,
-            maxOutputTokens: config.maxTokens || 200000,
+            // No `maxOutputTokens`: omitting it leaves Gemini's own default, which is
+            // the model's maximum output. Sending a value meant tracking a per-model
+            // ceiling that the API rejects outright rather than clamps.
             model: modelName,
             // `json: true` forces responseMimeType=application/json, guaranteeing the
             // model returns valid JSON text (never prose) for the parse pipeline.
@@ -750,17 +723,26 @@ export const getLLMModel = (modelName, config, schema = null) => {
     else if (modelName.startsWith("gpt-")) {
         const openAISettings = {
             apiKey: config.openAIAPIKey,
-            max_tokens: config.maxTokens || 200000,
+            // No output limit is sent: OpenAI defaults to the model's maximum. The
+            // snake_case `max_tokens` that used to sit here was never a recognised
+            // ChatOpenAI field and was silently dropped by LangChain.
             modelName: modelName,
             ...modelSettings,
         };
-        // gpt-5.6's /v1/chat/completions endpoint 400s whenever function tools are
-        // bound, because it rejects the model's own server-side default reasoning
-        // effort in that combination ("use /v1/responses or set reasoning_effort
-        // to 'none'"); forcing it to "none" here avoids the collision. Scoped to
-        // exactly gpt-5.6 — sibling gpt-5.x models are not confirmed to share the
-        // bug.
-        const reasoningEffortOverride = modelName === "gpt-5.6" ? { reasoning_effort: "none" } : {};
+        // The gpt-5.6 family's /v1/chat/completions endpoint 400s whenever function
+        // tools are bound, because it rejects the model's own server-side default
+        // reasoning effort in that combination ("use /v1/responses or set
+        // reasoning_effort to 'none'"); forcing it to "none" here avoids the
+        // collision. The restriction belongs to the family, not to one id: the
+        // `gpt-5.6` alias resolves to `gpt-5.6-sol`, and `gpt-5.6-terra` /
+        // `gpt-5.6-luna` are separately addressable ids that hit the same 400, as
+        // do dated snapshots (`gpt-5.6-2026-08-22`). Matching the alias alone left
+        // every one of those unpatched, so match the whole family. Still excludes
+        // sibling releases (gpt-5.5, gpt-5.7): they are not confirmed to share the
+        // bug, and reasoning effort is worth keeping wherever it is accepted.
+        const reasoningEffortOverride = /^gpt-5\.6($|-)/.test(modelName)
+            ? { reasoning_effort: "none" }
+            : {};
         // Use native structured output with a JSON schema. `modelKwargs` is spread
         // verbatim into the Chat Completions request. Merge (not overwrite) with
         // the reasoning-effort override above, since both may apply at once.
@@ -783,17 +765,6 @@ export const getLLMModel = (modelName, config, schema = null) => {
         }
         return new ChatOpenAI(openAISettings);
     }
-    // OpenAI-compatible providers: DeepSeek, Kimi (Moonshot), GLM (Zhipu)
-    const openAICompatible = getOpenAICompatibleProvider(modelName, config);
-    if (openAICompatible) {
-        return new ChatOpenAI({
-            apiKey: openAICompatible.apiKey,
-            maxTokens: config.maxTokens || 200000,
-            modelName: modelName,
-            configuration: { baseURL: openAICompatible.baseURL },
-            ...modelSettings,
-        });
-    }
     throw errorHandlingHelper.create(constants.HTTP_CODES.BAD_REQUEST, constants.ERROR.UNSUPPORTED_MODEL + ": " + modelName);
 };
 /**

package/package.json

@@ -3,7 +3,7 @@
   "publishConfig": {
     "access": "public"
   },
-  "version": "1.4.0",
+  "version": "1.4.3",
   "description": "",
   "main": "dist/cjs/index.js",
   "module": "dist/esm/index.js",