@jxburros/ai-nugget 0.4.1 → 0.7.0

package/nugget/dist/util.js

@@ -0,0 +1,114 @@
+export function asRecord(value) {
+    return value && typeof value === 'object' && !Array.isArray(value) ? value : undefined;
+}
+export function asString(value) {
+    return typeof value === 'string' ? value : undefined;
+}
+export function asNumber(value) {
+    return typeof value === 'number' && Number.isFinite(value) ? value : undefined;
+}
+export function promptChars(messages) {
+    return messages.reduce((sum, message) => {
+        if (typeof message.content === 'string')
+            return sum + message.content.length;
+        return sum + message.content.reduce((partSum, part) => partSum + (part.text?.length ?? 0) + (part.imageBase64?.length ?? 0), 0);
+    }, 0);
+}
+export function textFromMessages(messages) {
+    return messages.map((message) => {
+        if (typeof message.content === 'string')
+            return message.content;
+        return message.content.map((part) => part.text ?? '').join('');
+    }).join('\n');
+}
+export function globalEnv() {
+    const maybeProcess = globalThis;
+    return maybeProcess.process?.env ?? {};
+}
+/**
+ * Merges app-supplied {@link ChatRequest.providerOptions} into a provider
+ * request body: shallow at the top level, and one level deep for each key in
+ * `nestedKeys` where both sides are plain objects (Ollama `options`, Google
+ * `generationConfig`). Values in `providerOptions` win on collision.
+ */
+export function applyProviderOptions(body, providerOptions, nestedKeys = []) {
+    if (!providerOptions)
+        return body;
+    const merged = { ...body };
+    for (const [key, value] of Object.entries(providerOptions)) {
+        const base = asRecord(merged[key]);
+        if (nestedKeys.includes(key) && base && asRecord(value)) {
+            merged[key] = { ...base, ...value };
+        }
+        else {
+            merged[key] = value;
+        }
+    }
+    return merged;
+}
+/** Joins a base URL and path without producing a double slash, even if `base` ends with `/`. */
+export function joinUrl(base, path) {
+    return `${base.replace(/\/+$/, '')}/${path.replace(/^\/+/, '')}`;
+}
+/**
+ * Runs `fn` over `items` with at most `limit` in flight at once, preserving
+ * input order in the returned array. Used to probe many local models
+ * concurrently without opening an unbounded number of sockets.
+ */
+export async function mapWithConcurrency(items, limit, fn) {
+    const results = new Array(items.length);
+    let next = 0;
+    const workerCount = Math.max(1, Math.min(limit, items.length));
+    const worker = async () => {
+        for (;;) {
+            const index = next;
+            next += 1;
+            if (index >= items.length)
+                return;
+            results[index] = await fn(items[index], index);
+        }
+    };
+    await Promise.all(Array.from({ length: workerCount }, worker));
+    return results;
+}
+export function sleep(ms, signal) {
+    if (ms <= 0)
+        return Promise.resolve();
+    return new Promise((resolve, reject) => {
+        let settled = false;
+        const cleanup = () => signal?.removeEventListener('abort', onAbort);
+        const timer = setTimeout(() => {
+            settled = true;
+            cleanup();
+            resolve();
+        }, ms);
+        const onAbort = () => {
+            if (settled)
+                return;
+            settled = true;
+            clearTimeout(timer);
+            cleanup();
+            reject(new DOMException('Sleep aborted', 'AbortError'));
+        };
+        if (signal) {
+            if (signal.aborted)
+                onAbort();
+            else
+                signal.addEventListener('abort', onAbort, { once: true });
+        }
+    });
+}
+/**
+ * Token budgets behind the {@link ReasoningEffort} tiers for providers that
+ * express effort as a thinking-token budget (Anthropic `budget_tokens`, Google
+ * `thinkingBudget`). `'none'` is 0. Deliberately conservative: they are floors
+ * a model may stop under, not ceilings it fills.
+ */
+export const REASONING_BUDGET_TOKENS = {
+    none: 0,
+    minimal: 1024,
+    low: 2048,
+    medium: 8192,
+    high: 16384,
+};
+//# sourceMappingURL=util.js.map

package/nugget/dist/util.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"util.js","sourceRoot":"","sources":["../src/util.ts"],"names":[],"mappings":"AAEA,MAAM,UAAU,QAAQ,CAAC,KAAc;IACrC,OAAO,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAgC,CAAC,CAAC,CAAC,SAAS,CAAC;AACpH,CAAC;AAED,MAAM,UAAU,QAAQ,CAAC,KAAc;IACrC,OAAO,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AACvD,CAAC;AAED,MAAM,UAAU,QAAQ,CAAC,KAAc;IACrC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AACjF,CAAC;AAED,MAAM,UAAU,WAAW,CAAC,QAAuB;IACjD,OAAO,QAAQ,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,OAAO,EAAE,EAAE;QACtC,IAAI,OAAO,OAAO,CAAC,OAAO,KAAK,QAAQ;YAAE,OAAO,GAAG,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC;QAC7E,OAAO,GAAG,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,OAAO,EAAE,IAAI,EAAE,EAAE,CAAC,OAAO,GAAG,CAAC,IAAI,CAAC,IAAI,EAAE,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,WAAW,EAAE,MAAM,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IAClI,CAAC,EAAE,CAAC,CAAC,CAAC;AACR,CAAC;AAED,MAAM,UAAU,gBAAgB,CAAC,QAAuB;IACtD,OAAO,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE;QAC9B,IAAI,OAAO,OAAO,CAAC,OAAO,KAAK,QAAQ;YAAE,OAAO,OAAO,CAAC,OAAO,CAAC;QAChE,OAAO,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjE,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAChB,CAAC;AAED,MAAM,UAAU,SAAS;IACvB,MAAM,YAAY,GAAG,UAA4F,CAAC;IAClH,OAAO,YAAY,CAAC,OAAO,EAAE,GAAG,IAAI,EAAE,CAAC;AACzC,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,oBAAoB,CAClC,IAA6B,EAC7B,eAAoD,EACpD,aAAgC,EAAE;IAElC,IAAI,CAAC,eAAe;QAAE,OAAO,IAAI,CAAC;IAClC,MAAM,MAAM,GAA4B,EAAE,GAAG,IAAI,EAAE,CAAC;IACpD,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,eAAe,CAAC,EAAE,CAAC;QAC3D,MAAM,IAAI,GAAG,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;QACnC,IAAI,UAAU,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,IAAI,IAAI,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;YACxD,MAAM,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,IAAI,EAAE,GAAI,KAAiC,EAAE,CAAC;QACnE,CAAC;aAAM,CAAC;YACN,MAAM,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;QACtB,CAAC;IACH,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,gGAAgG;AAChG,MAAM,UAAU,OAAO,CAAC,IAAY,EAAE,IAAY;IAChD,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,EAAE,CAAC;AACnE,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAO,KAAmB,EAAE,KAAa,EAAE,EAA0C;IAC3H,MAAM,OAAO,GAAG,IAAI,KAAK,CAAI,KAAK,CAAC,MAAM,CAAC,CAAC;IAC3C,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC;IAC/D,MAAM,MAAM,GAAG,KAAK,IAAmB,EAAE;QACvC,SAAS,CAAC;YACR,MAAM,KAAK,GAAG,IAAI,CAAC;YACnB,IAAI,IAAI,CAAC,CAAC;YACV,IAAI,KAAK,IAAI,KAAK,CAAC,MAAM;gBAAE,OAAO;YAClC,OAAO,CAAC,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,KAAK,CAAC,KAAK,CAAE,EAAE,KAAK,CAAC,CAAC;QAClD,CAAC;IACH,CAAC,CAAC;IACF,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,MAAM,CAAC,CAAC,CAAC;IAC/D,OAAO,OAAO,CAAC;AACjB,CAAC;AAED,MAAM,UAAU,KAAK,CAAC,EAAU,EAAE,MAAoB;IACpD,IAAI,EAAE,IAAI,CAAC;QAAE,OAAO,OAAO,CAAC,OAAO,EAAE,CAAC;IACtC,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACrC,IAAI,OAAO,GAAG,KAAK,CAAC;QACpB,MAAM,OAAO,GAAG,GAAG,EAAE,CAAC,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;QACpE,MAAM,KAAK,GAAG,UAAU,CAAC,GAAG,EAAE;YAC5B,OAAO,GAAG,IAAI,CAAC;YACf,OAAO,EAAE,CAAC;YACV,OAAO,EAAE,CAAC;QACZ,CAAC,EAAE,EAAE,CAAC,CAAC;QACP,MAAM,OAAO,GAAG,GAAG,EAAE;YACnB,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,YAAY,CAAC,KAAK,CAAC,CAAC;YACpB,OAAO,EAAE,CAAC;YACV,MAAM,CAAC,IAAI,YAAY,CAAC,eAAe,EAAE,YAAY,CAAC,CAAC,CAAC;QAC1D,CAAC,CAAC;QACF,IAAI,MAAM,EAAE,CAAC;YACX,IAAI,MAAM,CAAC,OAAO;gBAAE,OAAO,EAAE,CAAC;;gBACzB,MAAM,CAAC,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;QACjE,CAAC;IACH,CAAC,CAAC,CAAC;AACL,CAAC;AAED;;;;;GAKG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAmE;IACrG,IAAI,EAAE,CAAC;IACP,OAAO,EAAE,IAAI;IACb,GAAG,EAAE,IAAI;IACT,MAAM,EAAE,IAAI;IACZ,IAAI,EAAE,KAAK;CACZ,CAAC"}

package/nugget/src/adapters/engines/anthropic.ts

@@ -1,9 +1,9 @@
 import { AIError } from '../../errors.js';
 import { estimatedUsage } from '../../tokens.js';
-import { postResponse, sseLines } from '../../transport.js';
-import type { ChatMessage, ChatRequest, ChatResult, ProviderAdapter, ResolvedConnection, StreamEvent, ToolCall } from '../../types.js';
-import { asNumber, asRecord, asString, textFromMessages } from '../../util.js';
-import { streamError, streamTimeout } from './base.js';
+import { fetchJson, postResponse, sseLines } from '../../transport.js';
+import type { ChatMessage, ChatRequest, ChatResult, ModelInfo, ProviderAdapter, ResolvedConnection, StreamEvent, ToolCall } from '../../types.js';
+import { applyProviderOptions, REASONING_BUDGET_TOKENS, asNumber, asRecord, asString, joinUrl, textFromMessages } from '../../util.js';
+import { DEFAULT_TIMEOUT_MS, parseArgs, randomId, safeParse, streamAnomaly, streamError, streamTimeout } from './base.js';
 
 const JSON_MODE_TOOL = 'json_output';
 
@@ -31,11 +31,22 @@ export class AnthropicAdapter implements ProviderAdapter {
     let inputTokens: number | undefined;
     let outputTokens: number | undefined;
     let stopReason: string | undefined;
+    let sawTerminal = false;
     const emittedTools: ToolCall[] = [];
     // Partial tool_use blocks keyed by content-block index (input_json_delta accumulation).
     const blocks = new Map<number, { id: string; name: string; raw: string }>();
     const timeout = streamTimeout(conn, req.signal);
     yield { type: 'start', callId: '', provider: conn.provider, model: req.model };
+    // Anthropic JSON mode is implemented as a forced `json_output` tool, which
+    // cannot coexist with the caller's own tools — same silent-downgrade shape
+    // as Gemini, so it gets the same signal rather than quietly dropping.
+    if (req.responseFormat?.type === 'json' && req.tools?.length) {
+      yield {
+        type: 'context',
+        kind: 'json_mode_downgraded',
+        data: { reason: 'Anthropic JSON mode uses a forced tool and cannot be combined with caller tools; the request was sent without JSON mode', provider: conn.provider },
+      };
+    }
     try {
       const res = await postResponse(`${conn.baseUrl}/v1/messages`, body(req, jsonMode), conn.headers, timeout.signal, conn.provider);
       const contentType = res.headers.get('content-type') ?? '';
@@ -45,6 +56,7 @@ export class AnthropicAdapter implements ProviderAdapter {
         inputTokens = parsed.inputTokens;
         outputTokens = parsed.outputTokens;
         stopReason = parsed.stopReason;
+        sawTerminal = true;
         if (jsonMode) {
           text = jsonTextFrom(parsed.toolCalls, parsed.text);
           if (text) yield { type: 'delta', text };
@@ -58,6 +70,7 @@ export class AnthropicAdapter implements ProviderAdapter {
         }
       } else {
         for await (const line of sseLines(res)) {
+          timeout.bump();
           const record = asRecord(safeParse(line));
           if (!record) continue;
           const type = asString(record.type);
@@ -81,6 +94,10 @@ export class AnthropicAdapter implements ProviderAdapter {
                 text += piece;
                 yield { type: 'delta', text: piece };
               }
+            } else if (delta?.type === 'thinking_delta') {
+              // Extended-thinking tokens ride the reasoning channel, kept out of `text`.
+              const piece = asString(delta.thinking) ?? '';
+              if (piece) yield { type: 'reasoning', text: piece };
             } else if (delta?.type === 'input_json_delta') {
               const partial = blocks.get(index);
               if (partial) partial.raw += asString(delta.partial_json) ?? '';
@@ -100,14 +117,18 @@ export class AnthropicAdapter implements ProviderAdapter {
                 yield { type: 'tool_call', call };
               }
             }
+          } else if (type === 'message_stop') {
+            sawTerminal = true;
           } else if (type === 'message_delta') {
             const delta = asRecord(record.delta);
+            if (asString(delta?.stop_reason)) sawTerminal = true;
             stopReason = asString(delta?.stop_reason) ?? stopReason;
             const usage = asRecord(record.usage);
             outputTokens = asNumber(usage?.output_tokens) ?? outputTokens;
           }
         }
       }
+      if (!sawTerminal) yield streamAnomaly('stream ended without a stop_reason or message_stop');
       const hasTools = emittedTools.length > 0;
       yield { type: 'done', result: {
         text,
@@ -124,6 +145,25 @@ export class AnthropicAdapter implements ProviderAdapter {
       timeout.done();
     }
   }
+
+  /**
+   * Lists models from Anthropic's `/v1/models` endpoint (auth + version headers
+   * already applied to `conn.headers`). The endpoint does not report a context
+   * window, so `contextWindow` is left undefined rather than guessed.
+   */
+  async listModels(conn: ResolvedConnection): Promise<ModelInfo[]> {
+    const { data } = await fetchJson(joinUrl(conn.baseUrl, '/v1/models'), {
+      method: 'GET',
+      headers: conn.headers,
+      timeoutMs: Math.min(conn.timeoutMs ?? DEFAULT_TIMEOUT_MS, 15_000),
+      provider: conn.provider,
+    });
+    const rows = Array.isArray(asRecord(data)?.data) ? asRecord(data)!.data as unknown[] : [];
+    return rows
+      .map((row) => asString(asRecord(row)?.id))
+      .filter((id): id is string => Boolean(id))
+      .map((id) => ({ id, source: { provider: conn.provider, connectionId: conn.id, baseUrl: conn.baseUrl } }));
+  }
 }
 
 function body(req: ChatRequest, jsonMode: boolean): Record<string, unknown> {
@@ -149,7 +189,25 @@ function body(req: ChatRequest, jsonMode: boolean): Record<string, unknown> {
     else if (req.toolChoice === 'auto') base.tool_choice = { type: 'auto' };
     else if (req.toolChoice === 'none') base.tool_choice = { type: 'none' };
   }
-  return base;
+  if (req.reasoningEffort !== undefined) {
+    // First-class reasoning effort → extended thinking. Anthropic requires
+    // `budget_tokens < max_tokens` and rejects temperature/top_p alongside
+    // thinking, so the budget is fitted under max_tokens (raising max_tokens
+    // when the caller's value leaves no room) and the samplers are dropped.
+    if (req.reasoningEffort === 'none') {
+      base.thinking = { type: 'disabled' };
+    } else {
+      const budget = REASONING_BUDGET_TOKENS[req.reasoningEffort];
+      const maxTokens = typeof base.max_tokens === 'number' ? base.max_tokens : 4096;
+      if (maxTokens <= budget) base.max_tokens = budget + 1024;
+      base.thinking = { type: 'enabled', budget_tokens: budget };
+      delete base.temperature;
+      delete base.top_p;
+    }
+  }
+  // providerOptions carries Anthropic-native fields (`thinking`, `metadata`,
+  // top-level `cache_control` extras, `service_tier`, …) without a release.
+  return applyProviderOptions(base, req.providerOptions);
 }
 
 function toAnthropicMessages(messages: ChatMessage[]): Record<string, unknown>[] {
@@ -173,28 +231,30 @@ function toAnthropicMessages(messages: ChatMessage[]): Record<string, unknown>[]
 }
 
 function toAnthropicMessage(m: ChatMessage): Record<string, unknown> {
-  // tool result messages map to a user turn carrying a tool_result content block.
-  if (m.role === 'tool') {
-    return {
-      role: 'user',
-      content: [{ type: 'tool_result', tool_use_id: m.toolCallId ?? '', content: typeof m.content === 'string' ? m.content : '' }],
-    };
-  }
-  // assistant turns that carried tool calls replay them as tool_use blocks.
+  // Tool-role messages are handled by toAnthropicMessages' batching loop and
+  // never reach here, so no tool branch is needed.
+
+  // assistant turns that carried tool calls replay them as tool_use blocks,
+  // preserving any text/image content parts that accompanied the calls.
   if (m.role === 'assistant' && m.toolCalls?.length) {
-    const content: unknown[] = [];
-    if (typeof m.content === 'string' && m.content) content.push({ type: 'text', text: m.content });
+    const content: unknown[] = [...contentBlocks(m.content)];
     for (const call of m.toolCalls) content.push({ type: 'tool_use', id: call.id, name: call.name, input: call.arguments ?? {} });
     return { role: 'assistant', content };
   }
   return {
     role: m.role === 'assistant' ? 'assistant' : 'user',
-    content: typeof m.content === 'string' ? m.content : m.content.map((part) => part.type === 'image'
-      ? { type: 'image', source: { type: 'base64', media_type: part.mimeType ?? 'image/png', data: part.imageBase64 ?? '' } }
-      : { type: 'text', text: part.text ?? '' }),
+    content: typeof m.content === 'string' ? m.content : contentBlocks(m.content),
   };
 }
 
+/** Maps text/image content — a string or a ContentPart[] — onto Anthropic content blocks. */
+function contentBlocks(content: ChatMessage['content']): unknown[] {
+  if (typeof content === 'string') return content ? [{ type: 'text', text: content }] : [];
+  return content.map((part) => part.type === 'image'
+    ? { type: 'image', source: { type: 'base64', media_type: part.mimeType ?? 'image/png', data: part.imageBase64 ?? '' } }
+    : { type: 'text', text: part.text ?? '' });
+}
+
 function textContent(content: ChatMessage['content']): string {
   if (typeof content === 'string') return content;
   if (!content) return '';
@@ -239,22 +299,3 @@ function mapStop(stop: string | undefined, hasTools: boolean): ChatResult['finis
   return 'stop';
 }
 
-function parseArgs(raw: string): unknown {
-  try {
-    return raw ? JSON.parse(raw) as unknown : {};
-  } catch {
-    return {};
-  }
-}
-
-function safeParse(line: string): unknown {
-  try {
-    return JSON.parse(line) as unknown;
-  } catch {
-    return undefined;
-  }
-}
-
-function randomId(): string {
-  return globalThis.crypto?.randomUUID?.() ?? `tool_${Date.now()}_${Math.random().toString(36).slice(2)}`;
-}

package/nugget/src/adapters/engines/base.ts

@@ -1,7 +1,7 @@
 import { AIError, fromUnknown } from '../../errors.js';
 import { fetchJson, withTimeout } from '../../transport.js';
 import type { ModelInfo, ResolvedConnection } from '../../types.js';
-import { asRecord, asString } from '../../util.js';
+import { asRecord, asString, joinUrl } from '../../util.js';
 import type { ProviderProfile } from '../profiles.js';
 
 export const DEFAULT_TIMEOUT_MS = 120_000;
@@ -17,14 +17,18 @@ export function streamError(error: unknown, timeout: { timedOut(): boolean }, pr
   return fromUnknown(error, provider);
 }
 
-/** Create a timeout/abort scope covering the full stream lifetime. */
+/**
+ * Create a timeout/abort scope covering the full stream lifetime, plus an
+ * optional idle timeout (`conn.idleTimeoutMs`) that the engine rearms via
+ * `bump()` on each received chunk.
+ */
 export function streamTimeout(conn: ResolvedConnection, signal?: AbortSignal): ReturnType<typeof withTimeout> {
-  return withTimeout(conn.timeoutMs ?? DEFAULT_TIMEOUT_MS, signal);
+  return withTimeout(conn.timeoutMs ?? DEFAULT_TIMEOUT_MS, signal, conn.idleTimeoutMs);
 }
 
 export async function listOpenModels(conn: ResolvedConnection, profile: ProviderProfile): Promise<ModelInfo[]> {
   if (!profile.listModelsPath) return [];
-  const { data } = await fetchJson(`${conn.baseUrl}${profile.listModelsPath}`, {
+  const { data } = await fetchJson(joinUrl(conn.baseUrl, profile.listModelsPath), {
     method: 'GET',
     headers: conn.headers,
     timeoutMs: conn.timeoutMs ?? DEFAULT_TIMEOUT_MS,
@@ -54,7 +58,7 @@ export async function health(conn: ResolvedConnection, profile: ProviderProfile)
   const path = profile.healthPath ?? profile.listModelsPath;
   if (!path) return { ok: true };
   try {
-    await fetchJson(`${conn.baseUrl}${path}`, {
+    await fetchJson(joinUrl(conn.baseUrl, path), {
       method: 'GET',
       headers: conn.headers,
       timeoutMs: Math.min(conn.timeoutMs ?? DEFAULT_TIMEOUT_MS, 10_000),
@@ -66,8 +70,47 @@ export async function health(conn: ResolvedConnection, profile: ProviderProfile)
   }
 }
 
-export function requireResponse(condition: unknown, message: string): asserts condition {
-  if (!condition) throw new AIError(message, { kind: 'invalid_response' });
+/**
+ * Coerce a provider's tool-call `arguments` into an object. Providers disagree
+ * on the shape: OpenAI/Anthropic stream a JSON *string*, Google and Ollama send
+ * a native object — but Ollama-compatible backends (llama.cpp and friends)
+ * sometimes send a string too. Malformed JSON degrades to `{}` so a single bad
+ * tool call surfaces as an argument-validation error rather than a stream crash.
+ */
+export function parseArgs(raw: unknown): unknown {
+  if (raw === undefined || raw === null) return {};
+  if (typeof raw !== 'string') return raw;
+  if (!raw.trim()) return {};
+  try {
+    return JSON.parse(raw) as unknown;
+  } catch {
+    return {};
+  }
+}
+
+/** JSON.parse that returns undefined instead of throwing — for per-line stream frames. */
+export function safeParse(line: string): unknown {
+  try {
+    return JSON.parse(line) as unknown;
+  } catch {
+    return undefined;
+  }
+}
+
+/** Stable-enough id for a tool call a provider didn't give one for. */
+export function randomId(): string {
+  return globalThis.crypto?.randomUUID?.() ?? `tool_${Date.now()}_${Math.random().toString(36).slice(2)}`;
+}
+
+/**
+ * A stream that ended without the provider's terminal marker (an OpenAI
+ * `finish_reason`, an Anthropic `message_delta.stop_reason`, a Google
+ * `finishReason`, an Ollama `done`) was almost certainly truncated. Every
+ * engine yields this so a dropped connection is diagnosable identically
+ * regardless of provider.
+ */
+export function streamAnomaly(reason = 'stream ended without a terminal finish reason'): { type: 'context'; kind: string; data: unknown } {
+  return { type: 'context', kind: 'stream_anomaly', data: { reason } };
 }
 
 /**

package/nugget/src/adapters/engines/google.ts

@@ -1,9 +1,9 @@
 import { AIError } from '../../errors.js';
 import { estimatedUsage } from '../../tokens.js';
 import { fetchJson, postResponse, sseLines } from '../../transport.js';
-import type { ChatMessage, ChatRequest, ChatResult, ProviderAdapter, ResolvedConnection, StreamEvent, ToolCall } from '../../types.js';
-import { asNumber, asRecord, asString, textFromMessages } from '../../util.js';
-import { DEFAULT_TIMEOUT_MS, streamError, streamTimeout } from './base.js';
+import type { ChatMessage, ChatRequest, ChatResult, ModelInfo, ProviderAdapter, ResolvedConnection, StreamEvent, ToolCall } from '../../types.js';
+import { applyProviderOptions, REASONING_BUDGET_TOKENS, asNumber, asRecord, asString, joinUrl, textFromMessages } from '../../util.js';
+import { DEFAULT_TIMEOUT_MS, randomId, safeParse, streamAnomaly, streamError, streamTimeout } from './base.js';
 
 export class GoogleAdapter implements ProviderAdapter {
   readonly provider: string;
@@ -31,12 +31,23 @@ export class GoogleAdapter implements ProviderAdapter {
     const toolCalls: ToolCall[] = [];
     const timeout = streamTimeout(conn, req.signal);
     yield { type: 'start', callId: '', provider: conn.provider, model: req.model };
+    // Gemini rejects `responseMimeType: application/json` combined with function
+    // declarations, so `body()` drops JSON mode when tools are present. Surface
+    // that downgrade instead of silently returning unstructured text.
+    if (jsonModeDowngraded(req)) {
+      yield {
+        type: 'context',
+        kind: 'json_mode_downgraded',
+        data: { reason: 'Gemini does not accept JSON response mode together with tools; the request was sent without JSON mode', provider: conn.provider },
+      };
+    }
     try {
       const streamUrl = `${conn.baseUrl}/v1beta/models/${encodeURIComponent(req.model)}:streamGenerateContent?alt=sse`;
       const res = await postResponse(streamUrl, body(req), conn.headers, timeout.signal, conn.provider);
       const contentType = res.headers.get('content-type') ?? '';
       const chunks = contentType.includes('text/event-stream') ? sseLines(res) : singleJsonLine(res);
       for await (const line of chunks) {
+        timeout.bump();
         const parsed = safeParse(line);
         const record = asRecord(parsed);
         const promptFeedback = asRecord(record?.promptFeedback);
@@ -44,6 +55,12 @@ export class GoogleAdapter implements ProviderAdapter {
         for (const part of candidateParts(record)) {
           const partText = asString(asRecord(part)?.text);
           if (partText) {
+            // Gemini marks reasoning parts with `thought: true`; route those to
+            // the reasoning channel so they don't blend into the answer text.
+            if (asRecord(part)?.thought === true) {
+              yield { type: 'reasoning', text: partText };
+              continue;
+            }
             if (firstTokenMs === null) firstTokenMs = Date.now() - started;
             text += partText;
             yield { type: 'delta', text: partText };
@@ -65,6 +82,7 @@ export class GoogleAdapter implements ProviderAdapter {
         inputTokens = asNumber(usage?.promptTokenCount) ?? inputTokens;
         outputTokens = asNumber(usage?.candidatesTokenCount) ?? outputTokens;
       }
+      if (!finish) yield streamAnomaly('stream ended without a finishReason');
       const hasTools = toolCalls.length > 0;
       yield { type: 'done', result: {
         text,
@@ -90,7 +108,7 @@ export class GoogleAdapter implements ProviderAdapter {
    */
   async health(conn: ResolvedConnection): Promise<{ ok: boolean; detail?: string }> {
     try {
-      await fetchJson(`${conn.baseUrl}/v1beta/models`, {
+      await fetchJson(joinUrl(conn.baseUrl, '/v1beta/models'), {
         method: 'GET',
         headers: conn.headers,
         timeoutMs: Math.min(conn.timeoutMs ?? DEFAULT_TIMEOUT_MS, 10_000),
@@ -101,6 +119,40 @@ export class GoogleAdapter implements ProviderAdapter {
       return { ok: false, detail: error instanceof Error ? error.message : 'Health check failed' };
     }
   }
+
+  /**
+   * Lists models from Google's `/v1beta/models`. Strips the `models/` name
+   * prefix, and maps `inputTokenLimit` → `contextWindow` and
+   * `supportedGenerationMethods` → `capabilities`.
+   */
+  async listModels(conn: ResolvedConnection): Promise<ModelInfo[]> {
+    const { data } = await fetchJson(joinUrl(conn.baseUrl, '/v1beta/models'), {
+      method: 'GET',
+      headers: conn.headers,
+      timeoutMs: Math.min(conn.timeoutMs ?? DEFAULT_TIMEOUT_MS, 15_000),
+      provider: conn.provider,
+    });
+    const rows = Array.isArray(asRecord(data)?.models) ? asRecord(data)!.models as unknown[] : [];
+    const models: ModelInfo[] = [];
+    for (const value of rows) {
+      const row = asRecord(value);
+      const name = asString(row?.name);
+      if (!name) continue;
+      const id = name.startsWith('models/') ? name.slice('models/'.length) : name;
+      const model: ModelInfo = { id, source: { provider: conn.provider, connectionId: conn.id, baseUrl: conn.baseUrl } };
+      const contextWindow = asNumber(row?.inputTokenLimit);
+      if (contextWindow !== undefined) model.contextWindow = contextWindow;
+      const methods = row?.supportedGenerationMethods;
+      if (Array.isArray(methods) && methods.every((m) => typeof m === 'string')) model.capabilities = methods as string[];
+      models.push(model);
+    }
+    return models;
+  }
+}
+
+/** True when the caller asked for JSON mode but tools force it off. */
+function jsonModeDowngraded(req: ChatRequest): boolean {
+  return req.responseFormat?.type === 'json' && !!req.tools?.length;
 }
 
 function body(req: ChatRequest): Record<string, unknown> {
@@ -126,7 +178,17 @@ function body(req: ChatRequest): Record<string, unknown> {
     if (typeof req.toolChoice === 'object') config.allowedFunctionNames = [req.toolChoice.name];
     payload.toolConfig = { functionCallingConfig: config };
   }
-  return payload;
+  if (req.reasoningEffort !== undefined) {
+    // First-class reasoning effort → Gemini's thinking budget. 0 disables
+    // thinking on models that allow it; others treat it as a minimum.
+    (payload.generationConfig as Record<string, unknown>).thinkingConfig = {
+      thinkingBudget: REASONING_BUDGET_TOKENS[req.reasoningEffort],
+    };
+  }
+  // providerOptions reaches Google-native fields: top-level `safetySettings`,
+  // `cachedContent`, and `generationConfig` extras (`thinkingConfig`,
+  // `responseModalities`, …) merged one level deep into generationConfig.
+  return applyProviderOptions(payload, req.providerOptions, ['generationConfig']);
 }
 
 function toGoogleContents(messages: ChatMessage[]): Record<string, unknown>[] {
@@ -150,28 +212,27 @@ function toGoogleContents(messages: ChatMessage[]): Record<string, unknown>[] {
 }
 
 function toGoogleContent(m: ChatMessage): Record<string, unknown> {
-  if (m.role === 'tool') {
-    return {
-      role: 'user',
-      parts: [{ functionResponse: { name: m.name ?? 'unknown', response: toolResponseObject(m) } }],
-    };
-  }
+  // Tool-role messages are handled by toGoogleContents' batching loop and never
+  // reach here, so no tool branch is needed.
   if (m.role === 'assistant' && m.toolCalls?.length) {
-    const parts: unknown[] = [];
-    if (typeof m.content === 'string' && m.content) parts.push({ text: m.content });
+    const parts: unknown[] = [...googleParts(m.content)];
     for (const call of m.toolCalls) parts.push({ functionCall: { name: call.name, args: call.arguments ?? {} } });
     return { role: 'model', parts };
   }
   return {
     role: m.role === 'assistant' ? 'model' : 'user',
-    parts: typeof m.content === 'string'
-      ? [{ text: m.content }]
-      : m.content.map((part) => part.type === 'image'
-        ? { inlineData: { mimeType: part.mimeType ?? 'image/png', data: part.imageBase64 ?? '' } }
-        : { text: part.text ?? '' }),
+    parts: googleParts(m.content),
   };
 }
 
+/** Maps text/image content — a string or a ContentPart[] — onto Google parts. */
+function googleParts(content: ChatMessage['content']): unknown[] {
+  if (typeof content === 'string') return content ? [{ text: content }] : [];
+  return content.map((part) => part.type === 'image'
+    ? { inlineData: { mimeType: part.mimeType ?? 'image/png', data: part.imageBase64 ?? '' } }
+    : { text: part.text ?? '' });
+}
+
 function textContent(content: ChatMessage['content']): string {
   if (typeof content === 'string') return content;
   if (!content) return '';
@@ -216,14 +277,3 @@ function mapFinish(finish: string | undefined, hasTools: boolean): ChatResult['f
   return 'stop';
 }
 
-function safeParse(line: string): unknown {
-  try {
-    return JSON.parse(line) as unknown;
-  } catch {
-    return undefined;
-  }
-}
-
-function randomId(): string {
-  return globalThis.crypto?.randomUUID?.() ?? `tool_${Date.now()}_${Math.random().toString(36).slice(2)}`;
-}