@combycode/llm-sdk 2.2.0 → 2.2.2

package/CHANGELOG.md

@@ -4,6 +4,54 @@ All notable changes to `@combycode/llm-sdk` are documented here. The format foll
 [Keep a Changelog](https://keepachangelog.com/) and the project adheres to
 [Semantic Versioning](https://semver.org/).
 
+## [2.2.2] — 2026-08-17
+
+### Fixed
+
+- **Every Gemini model rejected any tool schema carrying `additionalProperties`.** Function
+  declarations were sent on Gemini's `parameters` field, which takes a narrow OpenAPI subset and
+  rejects anything outside it outright:
+
+  > `Invalid JSON payload received. Unknown name "additionalProperties" at 'tools[0].function_declarations[0].parameters'`
+
+  Since OpenAI strict mode *requires* `additionalProperties: false`, no single tool manifest could
+  satisfy both providers — a consuming app had to delete the property, degrading its OpenAI schema,
+  to keep Google working. Measured against the live API, the narrow field also rejects `$schema`,
+  `$ref`, `$defs`, `definitions`, `const`, `examples`, `exclusiveMinimum`/`Maximum`, `multipleOf`,
+  `uniqueItems`, `patternProperties`, `propertyNames`, `if`/`then`, `readOnly`, `deprecated`, and
+  `type` given as an array.
+
+  Schemas now go on `parametersJsonSchema`, which takes full JSON Schema unchanged. Sanitising into
+  the subset was the obvious alternative and is worse: it silently drops constraints and cannot
+  express a `$ref` at all. Verified end to end on every tool-capable model the catalog ships —
+  gemini 2.5 pro/flash/flash-lite, 3-flash, 3.1-pro (incl. customtools), 3.1-flash-lite,
+  3.5-flash/-lite, 3.6-flash, gemma-4-26b/-31b: 11/11 fail before the change and drive the tool
+  after it. The Interactions surface already accepted full JSON Schema and is unchanged.
+
+## [2.2.1] — 2026-08-17
+
+### Fixed
+
+- **Extended thinking returned a 400 on Claude 4.7 and later.** The Anthropic adapter sent
+  `thinking: {type:'enabled', budget_tokens: N}` to every model, on the reasoning that it was the
+  universally accepted shape — true when written, and since reversed. Anthropic removed
+  `budget_tokens` on 4.7+, so Sonnet 5, Opus 5/4.8/4.7 and Fable 5 rejected every thinking request
+  outright:
+
+  > `"thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive"`
+
+  There is no shape that works everywhere: Haiku 4.5, Sonnet 4.5 and the Opus 4.x line have no
+  adaptive mode at all and still require the budget (`adaptive thinking is not supported on this
+  model`), so a blanket switch would have broken the other half. The shape is now chosen per model
+  at 4.6 — the version that accepts both — with `effort` mapping to `output_config.effort` on the
+  adaptive side instead of a token budget. An unrecognised model id gets `adaptive`, since
+  `budget_tokens` is the shape being retired. Both halves verified against the live API.
+
+- **`complete()` silently dropped `thinking`.** The one-shot helper never declared the option, so a
+  reasoning request through the simplest entry point sent no thinking at all while `client.complete()`
+  and agents honoured it. Found while live-testing the fix above — the run came back green because
+  nothing was being sent.
+
 ## [2.2.0] — 2026-08-17
 
 ### Added

package/dist/helpers/one-shot.d.ts

@@ -15,7 +15,7 @@
  *  LLMClient.complete. Either way the helper destroys its created client
  *  before returning so callers don't leak. */
 import type { AgentTool } from '../agent/types';
-import type { CacheConfig } from '../llm/types/request';
+import type { CacheConfig, ThinkingConfig } from '../llm/types/request';
 import type { LLMClientConfig } from '../llm/client-config';
 import type { AudioOptions } from '../llm/types/audio';
 import type { ContentPart, Message } from '../llm/types/messages';
@@ -67,6 +67,10 @@ export interface CompleteOptions {
     /** Which output modalities to return. Default ['text']; add 'audio' for a spoken
      *  reply (surfaced as a media part on `response.media`). */
     outputModalities?: Array<'text' | 'audio'>;
+    /** Extended thinking. Missing from this helper until 2.2.1: `client.complete()` and
+     *  agents honoured `thinking` while a one-shot silently dropped it, so the simplest
+     *  entry point was the only one that could not reason. */
+    thinking?: ThinkingConfig;
     /** Service tier for this call. Also settable as a `model:tier` suffix (e.g.
      *  `anthropic/claude-opus-4.8:priority`); an explicit value here wins. */
     serviceTier?: ServiceTier;

package/dist/index.browser.js

@@ -26096,6 +26096,19 @@ var ANTHROPIC_THINKING_BUDGETS = {
   max: 16384
 };
 var DEFAULT_ANTHROPIC_THINKING_BUDGET = 2048;
+var ANTHROPIC_ADAPTIVE_THINKING_MIN = { major: 4, minor: 6 };
+function anthropicThinkingShape(model) {
+  const id = model.toLowerCase().replace(/^anthropic\//, "");
+  const modern = /^claude-[a-z]+-(\d+)(?:[-.](\d+))?/.exec(id);
+  if (modern) {
+    const major = Number(modern[1]);
+    const minor = modern[2] === void 0 ? 0 : Number(modern[2]);
+    const { major: minMajor, minor: minMinor } = ANTHROPIC_ADAPTIVE_THINKING_MIN;
+    return major > minMajor || major === minMajor && minor >= minMinor ? "adaptive" : "budgeted";
+  }
+  if (/^claude-\d/.test(id)) return "budgeted";
+  return "adaptive";
+}
 var ANTHROPIC_TOP_K_MODELS = /^claude-(opus-4-(1|5|6)|sonnet-4-(5|6)|haiku-4-5)(\b|-)/;
 function anthropicAcceptsTopK(model) {
   return ANTHROPIC_TOP_K_MODELS.test(model);
@@ -26477,6 +26490,16 @@ var AnthropicAdapter = class {
     }
     if (req.thinking) {
       if (req.thinking.mode === "off") {
+      } else if (anthropicThinkingShape(req.model) === "adaptive") {
+        const thinking = { type: "adaptive" };
+        if (req.thinking.visibility === "hidden") thinking.display = "omitted";
+        body.thinking = thinking;
+        if (req.thinking.effort) {
+          body.output_config = {
+            ...body.output_config ?? {},
+            effort: req.thinking.effort
+          };
+        }
       } else {
         const budget = req.thinking.effort ? ANTHROPIC_THINKING_BUDGETS[req.thinking.effort] ?? DEFAULT_ANTHROPIC_THINKING_BUDGET : DEFAULT_ANTHROPIC_THINKING_BUDGET;
         const thinking = { type: "enabled", budget_tokens: budget };
@@ -27080,7 +27103,30 @@ var GoogleAdapter = class {
           functionDeclarations: fnTools.map((t) => ({
             name: t.name,
             description: t.description,
-            parameters: t.parameters
+            // `parametersJsonSchema`, NOT `parameters`. The two are mutually exclusive
+            // (sending both is a 400) and accept different things:
+            //
+            //   parameters            a narrow OpenAPI subset. Anything outside it is
+            //                         rejected outright with `Unknown name "<keyword>"` —
+            //                         measured: additionalProperties (at any depth),
+            //                         $schema, $ref, $defs, definitions, const, examples,
+            //                         exclusiveMinimum/Maximum, multipleOf, uniqueItems,
+            //                         patternProperties, propertyNames, if/then,
+            //                         readOnly, deprecated, and `type` as an array.
+            //   parametersJsonSchema  full JSON Schema.
+            //
+            // We passed callers' schemas straight into `parameters`, so any tool defined
+            // with `additionalProperties: false` — which OpenAI's strict mode requires —
+            // failed on EVERY Gemini model. A consuming app had to delete it from its
+            // manifest, degrading its OpenAI schema to keep Google working.
+            //
+            // Sanitising into the subset was the obvious fix and is the wrong one: it
+            // silently drops constraints and cannot express a `$ref` at all. Verified on
+            // every tool-capable model we ship (2.5 pro/flash/flash-lite, 3-flash,
+            // 3.1-pro incl. customtools, 3.1-flash-lite, 3.5-flash/-lite, 3.6-flash,
+            // gemma-4-26b/-31b) that this field takes the schema unchanged and the model
+            // still calls the tool.
+            parametersJsonSchema: t.parameters
           }))
         });
       }
@@ -38859,7 +38905,8 @@ async function complete(opts) {
         serviceTier,
         cache: opts.cache,
         topK: opts.topK,
-        seed: opts.seed
+        seed: opts.seed,
+        thinking: opts.thinking
       });
     } else {
       res = await llm.complete(input, {
@@ -38873,7 +38920,8 @@ async function complete(opts) {
         serviceTier,
         cache: opts.cache,
         topK: opts.topK,
-        seed: opts.seed
+        seed: opts.seed,
+        thinking: opts.thinking
       });
     }
     const result = {

package/dist/index.js

@@ -26023,6 +26023,19 @@ var ANTHROPIC_THINKING_BUDGETS = {
   max: 16384
 };
 var DEFAULT_ANTHROPIC_THINKING_BUDGET = 2048;
+var ANTHROPIC_ADAPTIVE_THINKING_MIN = { major: 4, minor: 6 };
+function anthropicThinkingShape(model) {
+  const id = model.toLowerCase().replace(/^anthropic\//, "");
+  const modern = /^claude-[a-z]+-(\d+)(?:[-.](\d+))?/.exec(id);
+  if (modern) {
+    const major = Number(modern[1]);
+    const minor = modern[2] === void 0 ? 0 : Number(modern[2]);
+    const { major: minMajor, minor: minMinor } = ANTHROPIC_ADAPTIVE_THINKING_MIN;
+    return major > minMajor || major === minMajor && minor >= minMinor ? "adaptive" : "budgeted";
+  }
+  if (/^claude-\d/.test(id)) return "budgeted";
+  return "adaptive";
+}
 var ANTHROPIC_TOP_K_MODELS = /^claude-(opus-4-(1|5|6)|sonnet-4-(5|6)|haiku-4-5)(\b|-)/;
 function anthropicAcceptsTopK(model) {
   return ANTHROPIC_TOP_K_MODELS.test(model);
@@ -26404,6 +26417,16 @@ var AnthropicAdapter = class {
     }
     if (req.thinking) {
       if (req.thinking.mode === "off") {
+      } else if (anthropicThinkingShape(req.model) === "adaptive") {
+        const thinking = { type: "adaptive" };
+        if (req.thinking.visibility === "hidden") thinking.display = "omitted";
+        body.thinking = thinking;
+        if (req.thinking.effort) {
+          body.output_config = {
+            ...body.output_config ?? {},
+            effort: req.thinking.effort
+          };
+        }
       } else {
         const budget = req.thinking.effort ? ANTHROPIC_THINKING_BUDGETS[req.thinking.effort] ?? DEFAULT_ANTHROPIC_THINKING_BUDGET : DEFAULT_ANTHROPIC_THINKING_BUDGET;
         const thinking = { type: "enabled", budget_tokens: budget };
@@ -27007,7 +27030,30 @@ var GoogleAdapter = class {
           functionDeclarations: fnTools.map((t) => ({
             name: t.name,
             description: t.description,
-            parameters: t.parameters
+            // `parametersJsonSchema`, NOT `parameters`. The two are mutually exclusive
+            // (sending both is a 400) and accept different things:
+            //
+            //   parameters            a narrow OpenAPI subset. Anything outside it is
+            //                         rejected outright with `Unknown name "<keyword>"` —
+            //                         measured: additionalProperties (at any depth),
+            //                         $schema, $ref, $defs, definitions, const, examples,
+            //                         exclusiveMinimum/Maximum, multipleOf, uniqueItems,
+            //                         patternProperties, propertyNames, if/then,
+            //                         readOnly, deprecated, and `type` as an array.
+            //   parametersJsonSchema  full JSON Schema.
+            //
+            // We passed callers' schemas straight into `parameters`, so any tool defined
+            // with `additionalProperties: false` — which OpenAI's strict mode requires —
+            // failed on EVERY Gemini model. A consuming app had to delete it from its
+            // manifest, degrading its OpenAI schema to keep Google working.
+            //
+            // Sanitising into the subset was the obvious fix and is the wrong one: it
+            // silently drops constraints and cannot express a `$ref` at all. Verified on
+            // every tool-capable model we ship (2.5 pro/flash/flash-lite, 3-flash,
+            // 3.1-pro incl. customtools, 3.1-flash-lite, 3.5-flash/-lite, 3.6-flash,
+            // gemma-4-26b/-31b) that this field takes the schema unchanged and the model
+            // still calls the tool.
+            parametersJsonSchema: t.parameters
           }))
         });
       }
@@ -38786,7 +38832,8 @@ async function complete(opts) {
         serviceTier,
         cache: opts.cache,
         topK: opts.topK,
-        seed: opts.seed
+        seed: opts.seed,
+        thinking: opts.thinking
       });
     } else {
       res = await llm.complete(input, {
@@ -38800,7 +38847,8 @@ async function complete(opts) {
         serviceTier,
         cache: opts.cache,
         topK: opts.topK,
-        seed: opts.seed
+        seed: opts.seed,
+        thinking: opts.thinking
       });
     }
     const result = {

package/dist/llm/providers/anthropic/constants.d.ts

@@ -11,5 +11,35 @@ export declare const ANTHROPIC_THINKING_BUDGETS: Record<string, number>;
  * unrecognised.
  */
 export declare const DEFAULT_ANTHROPIC_THINKING_BUDGET = 2048;
+/**
+ * The version at which `thinking: {type:'adaptive'}` takes over from
+ * `{type:'enabled', budget_tokens}`.
+ *
+ * There is no shape that works everywhere, and the direction reversed under us. This
+ * adapter used to send the budgeted form to every model, on the reasoning that it was the
+ * universally accepted one — true when it was written. Anthropic then REMOVED
+ * `budget_tokens` on 4.7 and later: Sonnet 5, Opus 5/4.8/4.7 and Fable 5 reject it with a
+ * 400 ("thinking.type.enabled is not supported for this model"). Meanwhile the older half
+ * — Haiku 4.5, Sonnet 4.5, Opus 4.x — has no `adaptive` at all and still requires the
+ * budget. So the shape must be chosen per model.
+ *
+ * 4.6 is the boundary: it accepts both and prefers `adaptive`, everything above requires
+ * `adaptive`, everything below requires the budget.
+ */
+export declare const ANTHROPIC_ADAPTIVE_THINKING_MIN: {
+    readonly major: 4;
+    readonly minor: 6;
+};
+/**
+ * Pick the `thinking` shape for a model id.
+ *
+ * Parsed from the id rather than read from the catalog on purpose: the catalog is optional
+ * (an engine can run with none), `buildRequest` has no access to it, and its per-model
+ * `reasoning` block does not currently distinguish the two shapes anyway.
+ *
+ * An unrecognised id gets `adaptive`, because `budget_tokens` is the shape being retired —
+ * an id we do not recognise is far likelier to be newer than us than older.
+ */
+export declare function anthropicThinkingShape(model: string): 'adaptive' | 'budgeted';
 /** True when this Anthropic model still accepts `top_k` (see ANTHROPIC_TOP_K_MODELS). */
 export declare function anthropicAcceptsTopK(model: string): boolean;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@combycode/llm-sdk",
-  "version": "2.2.0",
+  "version": "2.2.2",
   "description": "Unified, pluggable AI SDK for accessing the LLMs of every major provider (Anthropic, OpenAI, Google, xAI, OpenRouter) through one API. Cross-environment: Node, Bun, and the browser.",
   "type": "module",
   "main": "./dist/index.js",