@nodetool-ai/runtime 0.8.1 → 0.8.2

package/dist/providers/claude-agent-provider.d.ts

@@ -1,46 +1,14 @@
-/**
- * ClaudeAgentProvider — Claude reached through the official
- * `@anthropic-ai/claude-agent-sdk` instead of an API key.
- *
- * Unlike {@link AnthropicProvider}, this provider sends no `ANTHROPIC_API_KEY`:
- * the SDK drives the bundled `claude` binary (shipped as an optional platform
- * dependency, so there is nothing to install on PATH) and lets it authenticate
- * with the machine's logged-in Claude subscription (the credentials stored
- * under `~/.claude`). The SDK streams `SDKMessage`s, which we translate into the
- * cross-provider {@link ProviderStreamItem} stream.
- *
- * `@anthropic-ai/claude-agent-sdk` is a *soft* (optional peer) dependency: it is
- * not installed by default and must be added with the package manager before
- * this provider can run. It is loaded lazily via {@link loadSdk}, so a missing
- * package surfaces as a clear install hint only when the provider is actually
- * used — the rest of the runtime, and the browser worker bundle, never need it.
- *
- * It is a *pure LLM* provider: the agentic tool loop is collapsed to a single
- * turn with every built-in tool disabled (`allowedTools: []`), filesystem
- * settings/skills disabled (`settingSources: []`), permission prompts neutered
- * (`permissionMode: "dontAsk"`), and the default agent system prompt fully
- * replaced (`systemPrompt`), so the model behaves like a plain chat completion —
- * text in, text (and thinking) out.
- *
- * Session continuity: when a `threadId` is supplied the provider routes the
- * thread through a single upstream session and, on subsequent turns, resumes it
- * (`options.resume`) and sends only the new user delta. The continuation token
- * ({@link ProviderSession}) is surfaced as a {@link ProviderSessionUpdate} and
- * persisted by the chat layer onto the assistant message; the in-memory map
- * below is only a within-process cache (the DB column is authoritative).
- */
-import type { McpSdkServerConfigWithInstance, Options, SdkMcpToolDefinition, SDKMessage } from "@anthropic-ai/claude-agent-sdk";
+import type { McpSdkServerConfigWithInstance, Options, SdkMcpToolDefinition, SDKMessage, SDKUserMessage } from "@anthropic-ai/claude-agent-sdk";
 import { BaseProvider } from "./base-provider.js";
 import { type LanguageModel, type Message, type MessageContent, type ProviderSession, type ProviderStreamItem, type ProviderSkill, type ProviderTool, type ToolCall } from "./types.js";
 import type { RunBudget, TurnBudget } from "../turn-budget.js";
 /**
  * The subset of the SDK `query` signature this provider depends on. We only
- * ever pass a string prompt and consume the result as an async iterable, so the
- * real `query` (which also accepts an async-iterable prompt and returns the
- * richer `Query`) is assignable to this. Injectable for tests.
+ * pass a string for text-only turns or a structured user message for images,
+ * and consume the result as an async iterable. Injectable for tests.
  */
 export type ClaudeQueryFn = (params: {
-    prompt: string;
+    prompt: string | AsyncIterable<SDKUserMessage>;
     options?: Options;
 }) => AsyncIterable<SDKMessage>;
 /** The SDK `createSdkMcpServer` factory. Injectable for tests. */
@@ -102,6 +70,7 @@ export declare class ClaudeAgentProvider extends BaseProvider {
         executeTool?: (toolCall: ToolCall) => Promise<string | MessageContent[]>;
         resolveMedia?: (messages: Message[]) => Promise<Message[]>;
         maxIterations?: number;
+        providedToolsOnly?: boolean;
         turnBudget?: TurnBudget | RunBudget;
         workspaceDir?: string;
         /**
@@ -113,6 +82,7 @@ export declare class ClaudeAgentProvider extends BaseProvider {
          */
         skills?: ProviderSkill[];
     }): AsyncGenerator<ProviderStreamItem>;
+    private generateSdkLoop;
     generateMessages(args: {
         messages: Message[];
         model: string;

package/dist/providers/claude-agent-provider.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"claude-agent-provider.d.ts","sourceRoot":"","sources":["../../src/providers/claude-agent-provider.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAMH,OAAO,KAAK,EACV,8BAA8B,EAC9B,OAAO,EACP,oBAAoB,EAEpB,UAAU,EAIX,MAAM,gCAAgC,CAAC;AAExC,OAAO,EACL,YAAY,EAGb,MAAM,oBAAoB,CAAC;AAE5B,OAAO,EAIL,KAAK,aAAa,EAClB,KAAK,OAAO,EACZ,KAAK,cAAc,EAEnB,KAAK,eAAe,EAEpB,KAAK,kBAAkB,EAEvB,KAAK,aAAa,EAClB,KAAK,YAAY,EACjB,KAAK,QAAQ,EACd,MAAM,YAAY,CAAC;AASpB,OAAO,KAAK,EACV,SAAS,EACT,UAAU,EAEX,MAAM,mBAAmB,CAAC;AAsC3B;;;;;GAKG;AACH,MAAM,MAAM,aAAa,GAAG,CAAC,MAAM,EAAE;IACnC,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,CAAC,EAAE,OAAO,CAAC;CACnB,KAAK,aAAa,CAAC,UAAU,CAAC,CAAC;AAEhC,kEAAkE;AAClE,MAAM,MAAM,uBAAuB,GAAG,CAAC,IAAI,EAAE;IAC3C,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,CAAC;CACpC,KAAK,8BAA8B,CAAC;AAqDrC,UAAU,0BAA0B;IAClC;;;;OAIG;IACH,OAAO,CAAC,EAAE,aAAa,CAAC;IACxB,6EAA6E;IAC7E,iBAAiB,CAAC,EAAE,uBAAuB,CAAC;CAC7C;AA8BD,qBAAa,mBAAoB,SAAQ,YAAY;IACnD,gFAAgF;IAChF,IAAa,uBAAuB,IAAI,OAAO,CAE9C;IAED,MAAM,CAAC,eAAe,IAAI,MAAM,EAAE,CAIjC;IAED,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAuB;IACvD,OAAO,CAAC,QAAQ,CAAC,yBAAyB,CAAiC;IAC3E;;;;OAIG;IACH,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAsC;IAE/D,YACE,QAAQ,GAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAM,EACtC,OAAO,GAAE,0BAA+B,EAKzC;IAED,iFAAiF;IACxE,eAAe,OAEvB;IAED;;;;;OAKG;IACY,cAAc,IAAI,OAAO,CAAC,OAAO,CAAC,CAEhD;IAED;;;;OAIG;IACY,0BAA0B,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC,CAQpE;IAED,8EAA8E;YAChE,SAAS;IAMvB,gFAAgF;YAClE,mBAAmB;IAQjC;;;;;;OAMG;IACa,YAAY,CAC1B,IAAI,EAAE,UAAU,CAAC,mBAAmB,CAAC,kBAAkB,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG;QAC7D,WAAW,CAAC,EAAE,CAAC,QAAQ,EAAE,QAAQ,KAAK,OAAO,CAAC,MAAM,GAAG,cAAc,EAAE,CAAC,CAAC;QACzE,YAAY,CAAC,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,KAAK,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC;QAC3D,aAAa,CAAC,EAAE,MAAM,CAAC;QACvB,UAAU,CAAC,EAAE,UAAU,GAAG,SAAS,CAAC;QACpC,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB;;;;;;WAMG;QACH,MAAM,CAAC,EAAE,aAAa,EAAE,CAAC;KAC1B,GACA,cAAc,CAAC,kBAAkB,CAAC,CA2LpC;IAEe,gBAAgB,CAAC,IAAI,EAAE;QACrC,QAAQ,EAAE,OAAO,EAAE,CAAC;QACpB,KAAK,EAAE,MAAM,CAAC;QACd,KAAK,CAAC,EAAE,YAAY,EAAE,CAAC;QACvB,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,CAAC;QAC5B,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,IAAI,CAAC,EAAE,MAAM,CAAC;QACd,eAAe,CAAC,EAAE,MAAM,CAAC;QACzB,gBAAgB,CAAC,EAAE,MAAM,CAAC;QAC1B,KAAK,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;QAChC,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACzB,eAAe,CAAC,EAAE,eAAe,GAAG,IAAI,CAAC;QACzC,eAAe,CAAC,EAAE,MAAM,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC;QAC3C,UAAU,CAAC,EAAE,CACX,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,KAC1B,OAAO,CAAC,MAAM,CAAC,CAAC;QACrB,MAAM,CAAC,EAAE,WAAW,CAAC;KACtB,GAAG,cAAc,CAAC,kBAAkB,CAAC,CAmBrC;IAED;;;;;OAKG;YACY,cAAc;IAgF7B,+EAA+E;YAChE,OAAO;IAoTtB,8EAA8E;IAC9E,OAAO,CAAC,gBAAgB;IAgBT,eAAe,CAAC,IAAI,EAAE;QACnC,QAAQ,EAAE,OAAO,EAAE,CAAC;QACpB,KAAK,EAAE,MAAM,CAAC;QACd,KAAK,CAAC,EAAE,YAAY,EAAE,CAAC;QACvB,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,CAAC;QAC5B,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,IAAI,CAAC,EAAE,MAAM,CAAC;QACd,eAAe,CAAC,EAAE,MAAM,CAAC;QACzB,gBAAgB,CAAC,EAAE,MAAM,CAAC;QAC1B,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACzB,eAAe,CAAC,EAAE,eAAe,GAAG,IAAI,CAAC;QACzC,UAAU,CAAC,EAAE,CACX,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,KAC1B,OAAO,CAAC,MAAM,CAAC,CAAC;QACrB,MAAM,CAAC,EAAE,WAAW,CAAC;KACtB,GAAG,OAAO,CAAC,OAAO,CAAC,CAqBnB;CACF;AAwPD,KAAK,eAAe,GAChB;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAC9B;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,CAAC;AAqCtD;;;;GAIG;AACH,wBAAgB,sBAAsB,CACpC,MAAM,EAAE,MAAM,GAAG,cAAc,EAAE,GAChC,eAAe,EAAE,CAyBnB"}
+{"version":3,"file":"claude-agent-provider.d.ts","sourceRoot":"","sources":["../../src/providers/claude-agent-provider.ts"],"names":[],"mappings":"AAsCA,OAAO,KAAK,EACV,8BAA8B,EAC9B,OAAO,EACP,oBAAoB,EAEpB,UAAU,EAGV,cAAc,EACf,MAAM,gCAAgC,CAAC;AAMxC,OAAO,EACL,YAAY,EAGb,MAAM,oBAAoB,CAAC;AAK5B,OAAO,EAIL,KAAK,aAAa,EAClB,KAAK,OAAO,EACZ,KAAK,cAAc,EAEnB,KAAK,eAAe,EAEpB,KAAK,kBAAkB,EAEvB,KAAK,aAAa,EAClB,KAAK,YAAY,EACjB,KAAK,QAAQ,EACd,MAAM,YAAY,CAAC;AASpB,OAAO,KAAK,EACV,SAAS,EACT,UAAU,EAEX,MAAM,mBAAmB,CAAC;AAyC3B;;;;GAIG;AACH,MAAM,MAAM,aAAa,GAAG,CAAC,MAAM,EAAE;IACnC,MAAM,EAAE,MAAM,GAAG,aAAa,CAAC,cAAc,CAAC,CAAC;IAC/C,OAAO,CAAC,EAAE,OAAO,CAAC;CACnB,KAAK,aAAa,CAAC,UAAU,CAAC,CAAC;AAEhC,kEAAkE;AAClE,MAAM,MAAM,uBAAuB,GAAG,CAAC,IAAI,EAAE;IAC3C,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,CAAC;CACpC,KAAK,8BAA8B,CAAC;AAqErC,UAAU,0BAA0B;IAClC;;;;OAIG;IACH,OAAO,CAAC,EAAE,aAAa,CAAC;IACxB,6EAA6E;IAC7E,iBAAiB,CAAC,EAAE,uBAAuB,CAAC;CAC7C;AA8BD,qBAAa,mBAAoB,SAAQ,YAAY;IACnD,gFAAgF;IAChF,IAAa,uBAAuB,IAAI,OAAO,CAE9C;IAED,MAAM,CAAC,eAAe,IAAI,MAAM,EAAE,CAIjC;IAED,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAuB;IACvD,OAAO,CAAC,QAAQ,CAAC,yBAAyB,CAAiC;IAC3E;;;;OAIG;IACH,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAsC;IAE/D,YACE,QAAQ,GAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAM,EACtC,OAAO,GAAE,0BAA+B,EAKzC;IAED,iFAAiF;IACxE,eAAe,OAEvB;IAED;;;;;OAKG;IACY,cAAc,IAAI,OAAO,CAAC,OAAO,CAAC,CAEhD;IAED;;;;OAIG;IACY,0BAA0B,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC,CAQpE;IAED,8EAA8E;YAChE,SAAS;IAMvB,gFAAgF;YAClE,mBAAmB;IAQjC;;;;;;OAMG;IACa,YAAY,CAC1B,IAAI,EAAE,UAAU,CAAC,mBAAmB,CAAC,kBAAkB,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG;QAC7D,WAAW,CAAC,EAAE,CAAC,QAAQ,EAAE,QAAQ,KAAK,OAAO,CAAC,MAAM,GAAG,cAAc,EAAE,CAAC,CAAC;QACzE,YAAY,CAAC,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,KAAK,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC;QAC3D,aAAa,CAAC,EAAE,MAAM,CAAC;QACvB,iBAAiB,CAAC,EAAE,OAAO,CAAC;QAC5B,UAAU,CAAC,EAAE,UAAU,GAAG,SAAS,CAAC;QACpC,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB;;;;;;WAMG;QACH,MAAM,CAAC,EAAE,aAAa,EAAE,CAAC;KAC1B,GACA,cAAc,CAAC,kBAAkB,CAAC,CAEpC;YAEc,eAAe;IA4Ld,gBAAgB,CAAC,IAAI,EAAE;QACrC,QAAQ,EAAE,OAAO,EAAE,CAAC;QACpB,KAAK,EAAE,MAAM,CAAC;QACd,KAAK,CAAC,EAAE,YAAY,EAAE,CAAC;QACvB,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,CAAC;QAC5B,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,IAAI,CAAC,EAAE,MAAM,CAAC;QACd,eAAe,CAAC,EAAE,MAAM,CAAC;QACzB,gBAAgB,CAAC,EAAE,MAAM,CAAC;QAC1B,KAAK,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;QAChC,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACzB,eAAe,CAAC,EAAE,eAAe,GAAG,IAAI,CAAC;QACzC,eAAe,CAAC,EAAE,MAAM,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC;QAC3C,UAAU,CAAC,EAAE,CACX,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,KAC1B,OAAO,CAAC,MAAM,CAAC,CAAC;QACrB,MAAM,CAAC,EAAE,WAAW,CAAC;KACtB,GAAG,cAAc,CAAC,kBAAkB,CAAC,CAqBrC;IAED;;;;;OAKG;YACY,cAAc;IAgF7B,+EAA+E;YAChE,OAAO;IA4WtB,8EAA8E;IAC9E,OAAO,CAAC,gBAAgB;IAgBT,eAAe,CAAC,IAAI,EAAE;QACnC,QAAQ,EAAE,OAAO,EAAE,CAAC;QACpB,KAAK,EAAE,MAAM,CAAC;QACd,KAAK,CAAC,EAAE,YAAY,EAAE,CAAC;QACvB,UAAU,CAAC,EAAE,MAAM,GAAG,KAAK,CAAC;QAC5B,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,IAAI,CAAC,EAAE,MAAM,CAAC;QACd,eAAe,CAAC,EAAE,MAAM,CAAC;QACzB,gBAAgB,CAAC,EAAE,MAAM,CAAC;QAC1B,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACzB,eAAe,CAAC,EAAE,eAAe,GAAG,IAAI,CAAC;QACzC,UAAU,CAAC,EAAE,CACX,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,KAC1B,OAAO,CAAC,MAAM,CAAC,CAAC;QACrB,MAAM,CAAC,EAAE,WAAW,CAAC;KACtB,GAAG,OAAO,CAAC,OAAO,CAAC,CAqBnB;CACF;AA+QD,KAAK,eAAe,GAChB;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAC9B;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,CAAC;AAqCtD;;;;GAIG;AACH,wBAAgB,sBAAsB,CACpC,MAAM,EAAE,MAAM,GAAG,cAAc,EAAE,GAChC,eAAe,EAAE,CAyBnB"}

package/dist/providers/claude-agent-provider.js

@@ -1,3 +1,5 @@
+import { getTracer } from "../telemetry.js";
+import { tracedProviderLoop, tracedToolCall, beginProviderRound, providerRoundContext, traceContentAllowed, messageTraceContent, traceTextContent } from "./loop-tracing.js";
 /**
  * ClaudeAgentProvider — Claude reached through the official
  * `@anthropic-ai/claude-agent-sdk` instead of an API key.
@@ -35,7 +37,7 @@ import * as path from "node:path";
 import { createLogger, importOptionalModule } from "@nodetool-ai/config";
 import { z } from "zod";
 import { BaseProvider, budgetStopItem, resolveTurnBudget } from "./base-provider.js";
-import { sdkNativeReplacements } from "./core-tools.js";
+import { SDK_NATIVE_TOOL_REPLACEMENTS, sdkNativeReplacements } from "./core-tools.js";
 import { isProviderMessageEvent, isProviderSessionUpdate, isProviderStop } from "./types.js";
 import { hashSystemPrompt } from "./provider-session.js";
 import { PROVIDER_IDS, isFiniteNumber, isNonEmptyString, isObjectLike, isString } from "@nodetool-ai/protocol";
@@ -63,6 +65,8 @@ const SDK_BUILTIN_TOOLS = [
     "Task",
     "TodoWrite"
 ];
+/** The built-ins a tool-using turn keeps even when it offered no counterpart. */
+const SDK_WEB_TOOLS = ["WebSearch", "WebFetch"];
 /**
  * Env vars a nested Claude subscription session leaks into its children. The
  * SDK spawns the bundled CLI as a subprocess; left in place these would make it
@@ -72,6 +76,13 @@ const SDK_BUILTIN_TOOLS = [
 const NESTED_SESSION_ENV = /^(CLAUDECODE|CLAUDE_(CODE|SESSION|ENABLE|AFTER|AUTO)_[A-Za-z0-9_]*)$/;
 /** MCP server name under which NodeTool's tools are exposed to the SDK. */
 const TOOL_SERVER_NAME = "nodetool_tools";
+/**
+ * Turns a structured-output call may take. The SDK validates the answer
+ * against the schema and, when it does not match, hands the model the errors
+ * and asks again — each retry is a turn. With one turn, the first answer that
+ * misses the schema fails the call as `error_max_turns`.
+ */
+const STRUCTURED_OUTPUT_MAX_TURNS = 3;
 const TOOL_PREFIX = `mcp__${TOOL_SERVER_NAME}__`;
 /**
  * The member expression the retired guest toolbelt was called through:
@@ -178,6 +189,9 @@ export class ClaudeAgentProvider extends BaseProvider {
      * persistable {@link ProviderMessageEvent}s.
      */
     async *generateLoop(args) {
+        return yield* tracedProviderLoop(this.provider, args.model, () => this.generateSdkLoop(args));
+    }
+    async *generateSdkLoop(args) {
         // The SDK owns the loop, so honoring the budget is this override's job:
         // reserve before the first turn, and again after every assistant turn that
         // requested tools — that is the point where the SDK will make another call.
@@ -199,17 +213,19 @@ export class ClaudeAgentProvider extends BaseProvider {
             }
             reservations.push(first);
         }
-        // Drop every NodeTool tool the SDK ships a built-in for. Those built-ins
-        // are live under bypassPermissions, so keeping the MCP copy would give one
-        // capability two surfaces and make the model pick between them.
+        // Normally prefer native equivalents to duplicate MCP surfaces. A caller
+        // restricting the loop keeps its own dispatch instead.
         const offered = args.tools ?? [];
-        const replaced = sdkNativeReplacements(offered.map((t) => t.name), args.workspaceDir);
+        const replaced = args.providedToolsOnly
+            ? new Set()
+            : sdkNativeReplacements(offered.map((t) => t.name), args.workspaceDir);
         const tools = offered.filter((t) => !replaced.has(t.name));
         if (replaced.size > 0) {
             log.debug("Using SDK built-ins in place of NodeTool tools", {
                 replaced: [...replaced]
             });
         }
+        beginProviderRound(this.provider, args.model);
         const executeTool = args.executeTool;
         // A tool dispatches either through its own `execute` or the harness
         // `executeTool`; build the MCP server when at least one route exists.
@@ -236,13 +252,11 @@ export class ClaudeAgentProvider extends BaseProvider {
                 // so a clock alone would hand them one id.
                 const toolCallId = `call_${name}_${crypto.randomUUID()}`;
                 const result = t.execute
-                    ? await t.execute(toolArgs, toolCallId)
+                    ? await tracedToolCall({ id: toolCallId, name, args: toolArgs }, () => { const execute = t.execute; if (!execute) {
+                        throw new Error("Tool dispatch was removed");
+                    } return execute(toolArgs, toolCallId); })
                     : executeTool
-                        ? await executeTool({
-                            id: toolCallId,
-                            name,
-                            args: toolArgs
-                        })
+                        ? await tracedToolCall({ id: toolCallId, name, args: toolArgs }, () => executeTool({ id: toolCallId, name, args: toolArgs }))
                         : `Tool "${name}" is not available`;
                 // A terminal tool ends the SDK loop after its result is delivered.
                 if (t.terminal)
@@ -263,7 +277,7 @@ export class ClaudeAgentProvider extends BaseProvider {
         // native skill loader discovers. Best-effort: a disk failure drops skills
         // for the turn, it does not sink it. Cleaned up in the `finally` below.
         let skillsPlugin = null;
-        if ((args.skills?.length ?? 0) > 0) {
+        if (!args.providedToolsOnly && (args.skills?.length ?? 0) > 0) {
             try {
                 skillsPlugin = await materializeSkillsPlugin(args.skills ?? []);
             }
@@ -286,6 +300,10 @@ export class ClaudeAgentProvider extends BaseProvider {
                 maxTurns: offered.length > 0 ? (args.maxIterations ?? DEFAULT_TOOL_TURNS) : 1,
                 mcp,
                 toolsOffered: offered.length > 0,
+                providedToolsOnly: args.providedToolsOnly,
+                builtinTools: args.providedToolsOnly
+                    ? []
+                    : turnBuiltins(replaced, offered.length > 0, skillsPlugin !== null),
                 cwd: args.workspaceDir,
                 skillsPlugin
             });
@@ -362,9 +380,10 @@ export class ClaudeAgentProvider extends BaseProvider {
         const resultSchema = resultTool?.inputSchema;
         const config = {
             emitMessages: false,
-            maxTurns: args.maxTurns ?? 1,
+            maxTurns: args.maxTurns ?? (resultTool ? STRUCTURED_OUTPUT_MAX_TURNS : 1),
             mcp: null,
-            toolsOffered: false
+            toolsOffered: false,
+            builtinTools: []
         };
         if (resultTool && resultSchema) {
             config.outputFormat = { type: "json_schema", schema: resultSchema };
@@ -466,8 +485,8 @@ export class ClaudeAgentProvider extends BaseProvider {
             // Do NOT load repo .claude / CLAUDE.md / skills.
             settingSources: [],
             // Run without asking for permissions: never prompt AND never deny. This
-            // lets the SDK agent use its built-in tools (WebSearch/WebFetch/Bash/…)
-            // alongside NodeTool's MCP tools. bypassPermissions requires the explicit
+            // lets the SDK agent use the built-ins `tools` keeps alongside
+            // NodeTool's MCP tools. bypassPermissions requires the explicit
             // safety flag below.
             permissionMode: "bypassPermissions",
             allowDangerouslySkipPermissions: true,
@@ -486,15 +505,35 @@ export class ClaudeAgentProvider extends BaseProvider {
             // deferred, and NodeTool defers none — it registers its handful of tools
             // in-process — so the search always comes back empty. A model that fell
             // back to it after a mis-named tool call got nothing and stalled the turn.
-            disallowedTools: plan.config.toolsOffered
-                ? ["ToolSearch"]
-                : [...SDK_BUILTIN_TOOLS, "ToolSearch"],
+            // Supplied-only loops also disable Skill and keep all supplied dispatch
+            // inside the caller's MCP boundary.
+            disallowedTools: plan.config.providedToolsOnly
+                ? [...SDK_BUILTIN_TOOLS, "Skill", "ToolSearch"]
+                : plan.config.toolsOffered
+                    ? ["ToolSearch"]
+                    : [...SDK_BUILTIN_TOOLS, "ToolSearch"],
+            tools: plan.config.builtinTools,
+            // Only the in-process NodeTool server. Without this the child also loads
+            // the account's claude.ai connectors and the user's MCP servers, and
+            // their tool schemas grow every request of the loop.
+            strictMcpConfig: true,
+            // Hide the skills bundled with the SDK. The plugin below names its own.
+            skills: [],
             includePartialMessages: true,
             // Setting env REPLACES the child env, so spread process.env minus the
             // nested-session leakage. Preserves PATH/HOME/ANTHROPIC_BASE_URL/proxies.
             env: buildChildEnv(),
             abortController
         };
+        if (args.effort !== undefined) {
+            options.effort = args.effort;
+        }
+        if (args.thinking) {
+            options.thinking =
+                args.thinking.type === "manual"
+                    ? { type: "enabled", budgetTokens: args.thinking.budgetTokens }
+                    : { type: args.thinking.type };
+        }
         if (plan.config.outputFormat) {
             options.outputFormat = plan.config.outputFormat;
         }
@@ -544,11 +583,35 @@ export class ClaudeAgentProvider extends BaseProvider {
             model: args.model,
             resume: Boolean(plan.resume)
         });
+        const tracer = getTracer();
+        const startLlm = () => {
+            const span = tracer?.startSpan(`llm.stream ${this.provider}/${args.model}`, {}, providerRoundContext());
+            span?.setAttributes({ "llm.provider": this.provider, "llm.model": args.model, "llm.request.stream": true });
+            if (traceContentAllowed()) {
+                span?.setAttribute("llm.request.messages", messageTraceContent(args.messages));
+            }
+            return span;
+        };
+        let llmSpan = startLlm();
+        let llmResponse = "";
+        let observedTurn = false;
+        const closeLlm = () => {
+            if (traceContentAllowed()) {
+                llmSpan?.setAttribute("llm.response.content", llmResponse);
+            }
+            llmSpan?.end();
+            llmSpan = undefined;
+            llmResponse = "";
+        };
         let resolvedModel = args.model;
         // When partial deltas stream, we render from them and skip the final
         // assistant message to avoid duplication; if a build omits partials we fall
         // back to the final message's content blocks.
         let streamedFromPartials = false;
+        // The CLI answers a refused login with a synthetic assistant message, then
+        // a "success" result flagged `is_error`. Hold its text back from the stream
+        // and fail the call with it instead.
+        let authFailure = null;
         // The SDK splits one API assistant turn into one frame per content block —
         // thinking, then text, then each tool_use. Emitting a message per frame
         // would hand the consumer a text-only assistant message in the middle of a
@@ -575,7 +638,7 @@ export class ClaudeAgentProvider extends BaseProvider {
             };
         };
         try {
-            for await (const msg of queryFn({ prompt: plan.prompt, options })) {
+            for await (const msg of queryFn({ prompt: sdkPrompt(plan.prompt), options })) {
                 if (msg.type === "system" && msg.subtype === "init") {
                     // Capture the session and surface it immediately so a streaming
                     // consumer can persist it onto the assistant message it creates.
@@ -596,6 +659,17 @@ export class ClaudeAgentProvider extends BaseProvider {
                     continue;
                 }
                 if (msg.type === "stream_event") {
+                    if (msg.event.type === "message_start") {
+                        if (observedTurn) {
+                            closeLlm();
+                            beginProviderRound(this.provider, args.model);
+                            llmSpan = startLlm();
+                        }
+                        observedTurn = true;
+                    }
+                    if (msg.event.type === "message_stop") {
+                        closeLlm();
+                    }
                     const captured = capturedModelFromPartial(msg);
                     if (captured)
                         resolvedModel = captured;
@@ -603,6 +677,9 @@ export class ClaudeAgentProvider extends BaseProvider {
                     if (delta?.text != null) {
                         streamedFromPartials = true;
                         plan.emitted.content = true;
+                        if (traceContentAllowed()) {
+                            llmResponse = traceTextContent(llmResponse + delta.text);
+                        }
                         yield { type: "chunk", content: delta.text, done: false };
                     }
                     else if (delta?.thinking != null) {
@@ -618,6 +695,13 @@ export class ClaudeAgentProvider extends BaseProvider {
                     continue;
                 }
                 if (msg.type === "assistant") {
+                    if (msg.error === "authentication_failed") {
+                        authFailure = finalBlocks(msg)
+                            .filter((block) => !block.thinking)
+                            .map((block) => block.content)
+                            .join("");
+                        continue;
+                    }
                     const m = msg.message;
                     if (m && isNonEmptyString(m.model))
                         resolvedModel = m.model;
@@ -625,6 +709,9 @@ export class ClaudeAgentProvider extends BaseProvider {
                     // final content blocks — kept strictly separate, never merged.
                     if (!streamedFromPartials) {
                         for (const block of finalBlocks(msg)) {
+                            if (!block.thinking && traceContentAllowed()) {
+                                llmResponse = traceTextContent(llmResponse + block.content);
+                            }
                             plan.emitted.content = true;
                             yield block.thinking
                                 ? {
@@ -682,6 +769,9 @@ export class ClaudeAgentProvider extends BaseProvider {
                     // Bill every terminal result, not just the successful ones: an
                     // errored/max-turns run still consumed (and was charged for) tokens.
                     this.trackResultUsage(msg, resolvedModel);
+                    if (authFailure !== null && msg.is_error) {
+                        throw claudeAuthError(authFailure);
+                    }
                     if (msg.subtype === "success") {
                         const flushed = flushPending();
                         if (flushed)
@@ -718,6 +808,7 @@ export class ClaudeAgentProvider extends BaseProvider {
             throw err instanceof Error ? err : new Error(String(err));
         }
         finally {
+            closeLlm();
             // Terminate the SDK query on every exit path. On normal completion the
             // query has already finished and this is a no-op; on an early `break`
             // (consumer cancelled) or a throw it is the only thing that stops the
@@ -792,8 +883,30 @@ function buildChildEnv() {
             continue;
         env[key] = value;
     }
+    // The claude.ai connectors come from the account login, not from settings.
+    // This turns them off in the child.
+    env.ENABLE_CLAUDEAI_MCP_SERVERS = "false";
     return env;
 }
+/**
+ * The built-ins one loop turn keeps: those that replace an offered NodeTool
+ * tool, the web pair when the caller offered any tool, and `Skill` when a
+ * skills plugin is loaded.
+ */
+function turnBuiltins(replaced, toolsOffered, hasSkillsPlugin) {
+    const builtins = new Set();
+    for (const name of replaced) {
+        const builtin = SDK_NATIVE_TOOL_REPLACEMENTS.get(name);
+        if (builtin)
+            builtins.add(builtin);
+    }
+    if (toolsOffered)
+        for (const name of SDK_WEB_TOOLS)
+            builtins.add(name);
+    if (hasSkillsPlugin)
+        builtins.add("Skill");
+    return [...builtins];
+}
 /**
  * A DB skill name that is safe to use as a filesystem directory. The DB
  * validates names (lowercase `a-z0-9-`), but path safety is enforced here too:
@@ -1095,6 +1208,16 @@ function jsonPropToZod(prop) {
     }
     return desc ? zt.describe(desc) : zt;
 }
+/**
+ * A refused Claude login, carrying status 401 so the provider layer classifies
+ * it as a credential failure and the UI offers to sign in again.
+ */
+function claudeAuthError(detail) {
+    const message = detail.trim() || "The Claude login was refused";
+    return Object.assign(new Error(`Claude sign-in failed (401): ${message}`), {
+        status: 401
+    });
+}
 /** Build a descriptive Error from a non-success `result` message. */
 function resultError(msg) {
     const parts = [`Claude Agent SDK query failed (${msg.subtype})`];
@@ -1138,9 +1261,14 @@ function extractSystemPrompt(messages) {
  * user input).
  */
 function buildResumeDelta(messages, checkpoint) {
-    return messages
-        .slice(checkpoint)
-        .filter((m) => m.role === "user")
+    const delta = messages.slice(checkpoint).filter((m) => m.role === "user");
+    if (hasImages(delta)) {
+        return delta.flatMap((message, index) => [
+            ...(index ? [{ type: "text", text: "\n\n" }] : []),
+            ...promptBlocks(message.content)
+        ]);
+    }
+    return delta
         .map((m) => textOf(m.content))
         .filter(Boolean)
         .join("\n\n");
@@ -1151,19 +1279,33 @@ function buildResumeDelta(messages, checkpoint) {
  * edited/branched conversation) we prime context ONCE with a single delimited
  * user message instead of rebuilding a `Human:/Assistant:` transcript — the SDK
  * cannot import external assistant turns, so this is deliberate context priming,
- * not a faithful reconstruction. Only final assistant TEXT is included
- * (thinking is stripped by {@link textOf}).
+ * not a faithful reconstruction. Text and inline images are retained, while
+ * thinking is stripped.
  */
 function buildFreshPrompt(messages) {
     const convo = messages.filter((m) => m.role !== "system");
     if (convo.length === 0)
         return "";
     if (convo.length === 1 && convo[0].role === "user") {
-        return textOf(convo[0].content);
+        return hasImages(convo) ? promptBlocks(convo[0].content) : textOf(convo[0].content);
     }
     const last = convo[convo.length - 1];
     const newTurn = last.role === "user" ? textOf(last.content) : "";
     const prior = last.role === "user" ? convo.slice(0, -1) : convo;
+    if (hasImages(convo)) {
+        const blocks = [];
+        if (prior.length) {
+            blocks.push({ type: "text", text: "<conversation_so_far>\n" });
+            for (const message of prior) {
+                blocks.push({ type: "text", text: `${message.role === "assistant" ? "Assistant" : "User"}: ` }, ...promptBlocks(message.content), { type: "text", text: "\n\n" });
+            }
+            blocks.push({ type: "text", text: "</conversation_so_far>\n\n" });
+        }
+        if (last.role === "user") {
+            blocks.push(...promptBlocks(last.content));
+        }
+        return blocks;
+    }
     const transcript = prior
         .map((m) => {
         const text = textOf(m.content);
@@ -1178,4 +1320,44 @@ function buildFreshPrompt(messages) {
         : "";
     return [primed, newTurn].filter(Boolean).join("\n\n");
 }
+function hasImages(messages) {
+    return messages.some((message) => Array.isArray(message.content) &&
+        message.content.some((part) => part.type === "image_url"));
+}
+/** Images must already be inline. This boundary never fetches caller URLs. */
+function promptBlocks(content) {
+    if (isString(content)) {
+        return [{ type: "text", text: content }];
+    }
+    const blocks = [];
+    for (const part of content ?? []) {
+        if (part.type === "text") {
+            blocks.push({ type: "text", text: part.text });
+        }
+        else if (part.type === "image_url") {
+            const image = toMcpImageBlock(part.image);
+            if (!image) {
+                throw new Error("Claude Agent image messages require inline image data. Resolve the media reference before dispatch.");
+            }
+            const mime = image.mimeType;
+            if (mime !== "image/png" && mime !== "image/jpeg" && mime !== "image/gif" && mime !== "image/webp") {
+                throw new Error(`Claude Agent image messages do not support ${mime}. Convert the image before dispatch.`);
+            }
+            blocks.push({ type: "image", source: { type: "base64", media_type: mime, data: image.data } });
+        }
+    }
+    return blocks;
+}
+function sdkPrompt(content) {
+    if (isString(content)) {
+        return content;
+    }
+    return (async function* () {
+        yield {
+            type: "user",
+            parent_tool_use_id: null,
+            message: { role: "user", content }
+        };
+    })();
+}
 //# sourceMappingURL=claude-agent-provider.js.map