@mcpmesh/sdk 2.3.0 → 2.4.0

package/dist/llm-agent.js

@@ -33,12 +33,10 @@ import { zodToJsonSchema } from "zod-to-json-schema";
 import { renderTemplate } from "./template.js";
 import { ResponseParser } from "./response-parser.js";
 import { MaxIterationsError, LLMAPIError, ToolExecutionError, } from "./errors.js";
-import { parseSSEResponse } from "./sse.js";
 import { resolveMediaInputs } from "./media/index.js";
-import { getCurrentTraceContext, getCurrentPropagatedHeaders, streamMcpTool, DEFAULT_CALL_OPTIONS, } from "./proxy.js";
-import { generateSpanId, publishTraceSpan, createTraceHeaders, injectTraceContext, } from "./tracing.js";
-import { fetchWithTimeout, isTimeoutError } from "./timeout-utils.js";
-import { getDispatcher } from "./http-pool.js";
+import { callMcpTool, streamMcpTool, DEFAULT_CALL_OPTIONS, } from "./proxy.js";
+import { isTimeoutError } from "./timeout-utils.js";
+import { envMaxIterations, sanitizeMaxIterations } from "./llm-provider.js";
 /**
  * Mesh provider that delegates to an LLM provider discovered via mesh.
  */
@@ -51,8 +49,16 @@ export class MeshDelegatedProvider {
         this.functionName = functionName;
         this.parallelToolCalls = parallelToolCalls;
     }
-    async complete(model, messages, tools, options) {
-        // Build MeshLlmRequest structure (matches Python claude_provider schema)
+    /**
+     * Build the MeshLlmRequest body shared by complete() and streamComplete().
+     *
+     * Assembles model_params (with the escape-hatch merge + typed overrides),
+     * wraps messages/tools into the MeshLlmRequest, and returns it pre-wrapped
+     * in the ``{ request }`` arguments object. Callers inject trace context /
+     * propagated headers into ``args`` afterward (per-caller — complete() uses
+     * injectTraceAndHeaders, streamComplete() lets streamMcpTool() handle it).
+     */
+    buildMeshLlmRequest(model, messages, tools, options) {
         const modelParams = {};
         // Escape-hatch merge: callers can pass vendor-specific kwargs
         // (e.g., thinking_config, output_config) via options.modelParams.
@@ -84,6 +90,18 @@ export class MeshDelegatedProvider {
         if (this.parallelToolCalls) {
             modelParams.parallel_tool_calls = true;
         }
+        // Issue #1116: forward the provider-managed loop cap. Typed field, so it
+        // takes precedence over any escape-hatch modelParams.max_iterations above.
+        if (options?.maxIterations !== undefined) {
+            modelParams.max_iterations = options.maxIterations;
+        }
+        // Issue #1112: forward the consumer-supplied output_mode override ONLY when
+        // the user explicitly set it (undefined = auto = provider's per-vendor
+        // selection; omitting keeps the provider byte-identical to today). Typed
+        // field, so it takes precedence over any escape-hatch modelParams.output_mode.
+        if (options?.outputMode !== undefined) {
+            modelParams.output_mode = options.outputMode;
+        }
         const request = {
             messages,
         };
@@ -96,152 +114,82 @@ export class MeshDelegatedProvider {
         }
         // Wrap in "request" parameter as expected by Python claude_provider
         const args = { request };
+        return { request, args };
+    }
+    async complete(model, messages, tools, options) {
+        // Build MeshLlmRequest structure (matches Python claude_provider schema).
+        // The {request} wrapper stays; callMcpTool injects trace context internally.
+        const { args } = this.buildMeshLlmRequest(model, messages, tools, options);
         // Set up timeout (default 300s to match Python SDK's stream_timeout)
         const timeoutMs = parseInt(process.env.MESH_PROVIDER_TIMEOUT_MS || "300000", 10);
-        // Tracing: propagate context to downstream provider
-        const traceCtx = getCurrentTraceContext();
-        const traceSpanId = traceCtx ? generateSpanId() : null;
-        const traceStartTime = Date.now() / 1000;
-        // Inject trace context and propagated headers into args via Rust core
-        const delegatedPropHeaders = getCurrentPropagatedHeaders();
-        if (traceCtx && traceSpanId) {
-            try {
-                const argsJson = JSON.stringify(args);
-                const headersJson = Object.keys(delegatedPropHeaders).length > 0 ? JSON.stringify(delegatedPropHeaders) : undefined;
-                const injectedJson = injectTraceContext(argsJson, traceCtx.traceId, traceSpanId, headersJson);
-                const injected = JSON.parse(injectedJson);
-                Object.assign(args, injected);
-            }
-            catch {
-                // Fallback to manual injection
-                args._trace_id = traceCtx.traceId;
-                args._parent_span = traceSpanId;
-                if (Object.keys(delegatedPropHeaders).length > 0) {
-                    args._mesh_headers = { ...delegatedPropHeaders };
-                }
-            }
-        }
-        else if (Object.keys(delegatedPropHeaders).length > 0) {
-            args._mesh_headers = { ...delegatedPropHeaders };
-        }
-        let traceSuccess = true;
-        let traceError = null;
+        // Route through the shared callMcpTool (trace injection, span publish,
+        // dispatcher pooling, job-cancel wiring all handled internally). LLM calls
+        // are expensive / non-idempotent so disable retries; LLM responses can
+        // exceed the default 10 MiB cap so raise it to effectively unbounded.
+        const callOptions = {
+            ...DEFAULT_CALL_OPTIONS,
+            timeout: timeoutMs,
+            maxAttempts: 1,
+            maxResponseSize: Number.MAX_SAFE_INTEGER,
+        };
+        // Call the mesh provider via MCP. callMcpTool throws a plain Error on
+        // failure — re-wrap into LLMAPIError to preserve the public error type.
+        let content;
         try {
-            // Call the mesh provider via MCP
-            let response;
-            try {
-                response = await fetchWithTimeout(`${this.endpoint}/mcp`, {
-                    method: "POST",
-                    headers: {
-                        "Content-Type": "application/json",
-                        "Accept": "application/json, text/event-stream",
-                        ...(traceCtx && traceSpanId ? createTraceHeaders(traceCtx.traceId, traceSpanId) : {}),
-                        ...getCurrentPropagatedHeaders(),
-                    },
-                    body: JSON.stringify({
-                        jsonrpc: "2.0",
-                        id: Date.now(),
-                        method: "tools/call",
-                        params: {
-                            name: this.functionName,
-                            arguments: args,
-                        },
-                    }),
-                    timeout: timeoutMs,
-                    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-                    dispatcher: getDispatcher(`${this.endpoint}/mcp`),
-                });
-            }
-            catch (error) {
-                if (isTimeoutError(error)) {
-                    throw new LLMAPIError(408, `Request timed out after ${timeoutMs}ms`, `mesh:${this.endpoint}`);
-                }
-                throw new LLMAPIError(0, `Fetch failed: ${error instanceof Error ? error.message : String(error)}`, `mesh:${this.endpoint}`);
-            }
-            if (!response.ok) {
-                const error = await response.text();
-                throw new LLMAPIError(response.status, error, `mesh:${this.endpoint}`);
-            }
-            // Handle SSE response from FastMCP stateless HTTP stream
-            const responseText = await response.text();
-            const result = parseSSEResponse(responseText);
-            if (result.error) {
-                throw new Error(`Mesh provider RPC error: ${result.error.message}`);
-            }
-            // Parse the MCP result content
-            const content = result.result?.content?.[0];
-            if (!content || content.type !== "text") {
-                throw new Error("Invalid response from mesh provider");
-            }
-            // Check for MCP tool execution error (isError flag in result)
-            // eslint-disable-next-line @typescript-eslint/no-explicit-any
-            if (result.result?.isError) {
-                throw new Error(`Mesh provider tool error: ${content.text}`);
-            }
-            // Parse the LLM provider response
-            // Format: { role, content, tool_calls?, _mesh_usage? }
-            const meshResponse = JSON.parse(content.text);
-            // Validate role - LLM responses should always be "assistant"
-            let validatedRole = "assistant";
-            if (meshResponse.role !== "assistant") {
-                console.warn(`[mesh.llm] Unexpected role "${meshResponse.role}" from mesh provider, defaulting to "assistant"`);
-            }
-            // Convert to OpenAI format expected by MeshLlmAgent
-            const openAiResponse = {
-                id: `mesh-${Date.now()}`,
-                object: "chat.completion",
-                created: Math.floor(Date.now() / 1000),
-                model: "mesh-delegated",
-                choices: [
-                    {
-                        index: 0,
-                        message: {
-                            role: validatedRole,
-                            content: meshResponse.content,
-                            tool_calls: meshResponse.tool_calls,
-                        },
-                        finish_reason: meshResponse.tool_calls ? "tool_calls" : "stop",
-                    },
-                ],
-                usage: meshResponse._mesh_usage
-                    ? {
-                        prompt_tokens: meshResponse._mesh_usage.prompt_tokens,
-                        completion_tokens: meshResponse._mesh_usage.completion_tokens,
-                        total_tokens: meshResponse._mesh_usage.prompt_tokens +
-                            meshResponse._mesh_usage.completion_tokens,
-                    }
-                    : undefined,
-            };
-            return openAiResponse;
+            content = await callMcpTool(this.endpoint, this.functionName, args, callOptions, "mesh-llm");
         }
         catch (err) {
-            traceSuccess = false;
-            traceError = err instanceof Error ? err.message : String(err);
-            throw err;
-        }
-        finally {
-            if (traceCtx && traceSpanId) {
-                const traceEndTime = Date.now() / 1000;
-                const traceDurationMs = (traceEndTime - traceStartTime) * 1000;
-                publishTraceSpan({
-                    traceId: traceCtx.traceId,
-                    spanId: traceSpanId,
-                    parentSpan: traceCtx.parentSpanId,
-                    functionName: "proxy_call_wrapper",
-                    startTime: traceStartTime,
-                    endTime: traceEndTime,
-                    durationMs: traceDurationMs,
-                    success: traceSuccess,
-                    error: traceError,
-                    resultType: traceSuccess ? "object" : "error",
-                    argsCount: 0,
-                    kwargsCount: 0,
-                    dependencies: [this.endpoint],
-                    injectedDependencies: 0,
-                    meshPositions: [],
-                }).catch(() => { });
-            }
-        }
+            const message = err instanceof Error ? err.message : String(err);
+            // callMcpTool catches the AbortError and re-throws a plain Error whose
+            // message is "MCP call timed out after <N>ms", so isTimeoutError (name ===
+            // "AbortError") no longer matches here — also detect the message.
+            const isTimeout = isTimeoutError(err) ||
+                (err instanceof Error && /timed out/i.test(err.message));
+            if (isTimeout) {
+                throw new LLMAPIError(408, message, `mesh:${this.endpoint}`);
+            }
+            throw new LLMAPIError(0, message, `mesh:${this.endpoint}`);
+        }
+        // The mesh provider returns a single text content item whose text is the
+        // LLM provider response JSON. callMcpTool joins/extracts it to a string.
+        if (typeof content !== "string") {
+            throw new Error("Invalid response from mesh provider");
+        }
+        // Parse the LLM provider response
+        // Format: { role, content, tool_calls?, _mesh_usage? }
+        const meshResponse = JSON.parse(content);
+        // Validate role - LLM responses should always be "assistant"
+        let validatedRole = "assistant";
+        if (meshResponse.role !== "assistant") {
+            console.warn(`[mesh.llm] Unexpected role "${meshResponse.role}" from mesh provider, defaulting to "assistant"`);
+        }
+        // Convert to OpenAI format expected by MeshLlmAgent
+        const openAiResponse = {
+            id: `mesh-${Date.now()}`,
+            object: "chat.completion",
+            created: Math.floor(Date.now() / 1000),
+            model: "mesh-delegated",
+            choices: [
+                {
+                    index: 0,
+                    message: {
+                        role: validatedRole,
+                        content: meshResponse.content,
+                        tool_calls: meshResponse.tool_calls,
+                    },
+                    finish_reason: meshResponse.tool_calls ? "tool_calls" : "stop",
+                },
+            ],
+            usage: meshResponse._mesh_usage
+                ? {
+                    prompt_tokens: meshResponse._mesh_usage.prompt_tokens,
+                    completion_tokens: meshResponse._mesh_usage.completion_tokens,
+                    total_tokens: meshResponse._mesh_usage.prompt_tokens +
+                        meshResponse._mesh_usage.completion_tokens,
+                }
+                : undefined,
+        };
+        return openAiResponse;
     }
     /**
      * Stream chunks from the mesh-delegated provider's streaming variant.
@@ -257,40 +205,8 @@ export class MeshDelegatedProvider {
      * ``ai.mcpmesh.stream`` tag opt-in (see ``MeshLlmAgent.stream()``).
      */
     async *streamComplete(model, messages, tools, options) {
-        // Build MeshLlmRequest body — same shape as complete()
-        const modelParams = {};
-        // Escape-hatch merge: callers can pass vendor-specific kwargs
-        // (e.g., thinking_config, output_config) via options.modelParams.
-        // Merged FIRST so typed fields below take precedence on collision.
-        if (options?.modelParams) {
-            Object.assign(modelParams, options.modelParams);
-        }
-        if (model && model !== "default") {
-            modelParams.model = model;
-        }
-        if (options?.maxOutputTokens)
-            modelParams.max_tokens = options.maxOutputTokens;
-        if (options?.temperature !== undefined)
-            modelParams.temperature = options.temperature;
-        if (options?.topP !== undefined)
-            modelParams.top_p = options.topP;
-        if (options?.stop)
-            modelParams.stop = options.stop;
-        if (options?.outputSchema) {
-            modelParams.output_schema = options.outputSchema.schema;
-            modelParams.output_type_name = options.outputSchema.name;
-        }
-        if (this.parallelToolCalls) {
-            modelParams.parallel_tool_calls = true;
-        }
-        const request = { messages };
-        if (Object.keys(modelParams).length > 0) {
-            request.model_params = modelParams;
-        }
-        if (tools && tools.length > 0) {
-            request.tools = tools;
-        }
-        const args = { request };
+        // Build MeshLlmRequest body — same shape as complete().
+        const { args } = this.buildMeshLlmRequest(model, messages, tools, options);
         // streamMcpTool() handles trace context injection / propagated headers /
         // dispatcher pooling internally — same path as createProxy().stream().
         // Match complete()'s env-backed timeout (MESH_PROVIDER_TIMEOUT_MS) so
@@ -314,10 +230,36 @@ export class MeshLlmAgent {
     _meta = null;
     _systemPromptOverride = null;
     _parallelLogEmitted = false;
+    // Cached output schema derived from the immutable returnSchema (Issue #459).
+    // Computed once: `null` means "not yet computed", an object holds the result
+    // (which may itself be `undefined` when conversion failed).
+    _outputSchema = null;
+    _outputSchemaSection = null;
     constructor(config) {
         this.config = config;
         this.responseParser = new ResponseParser(config.returnSchema);
     }
+    /**
+     * Build (once) the provider output schema from the immutable returnSchema.
+     * Returns `undefined` when there is no schema or conversion failed.
+     */
+    getOutputSchema() {
+        if (this._outputSchema !== null)
+            return this._outputSchema;
+        let result;
+        if (this.config.returnSchema) {
+            try {
+                const jsonSchema = zodToJsonSchema(this.config.returnSchema);
+                const schemaName = jsonSchema.title ?? "Response";
+                result = { schema: jsonSchema, name: schemaName };
+            }
+            catch {
+                // If schema conversion fails, skip
+            }
+        }
+        this._outputSchema = result;
+        return result;
+    }
     /**
      * Get metadata from the last run.
      */
@@ -337,28 +279,20 @@ export class MeshLlmAgent {
         return this._systemPromptOverride ?? this.config.systemPrompt;
     }
     /**
-     * Run the agentic loop.
+     * Build the initial LlmMessage[] shared by run() and stream():
+     * render the system prompt (+ tool schema injection), optionally append the
+     * output-schema hint, resolve media inputs, and unwind multi-turn history
+     * (attaching resolved media to the last user message).
      *
-     * @param messageInput - User message string or multi-turn message array
-     * @param context - Runtime context with tools and options
-     * @returns Parsed response (validated if schema provided)
+     * The ONLY behavioral knob is opts.includeOutputSchemaHint:
+     * - run() passes `!meshDelegated` (consumer-side schema hint when not delegated).
+     * - stream() passes `false` (always mesh-delegated; provider applies formatting).
      */
-    async run(messageInput, context) {
-        if (this.config.parallelToolCalls && !this._parallelLogEmitted) {
-            console.log("[mesh.llm] parallel tool calls enabled — tools will execute concurrently via Promise.all()");
-            this._parallelLogEmitted = true;
-        }
-        const startTime = Date.now();
-        const toolCalls = [];
-        let totalInputTokens = 0;
-        let totalOutputTokens = 0;
-        // Resolve provider
-        const provider = this.resolveProvider(context);
-        // Build initial messages
+    async buildAgentMessages(messageInput, context, opts) {
         const messages = [];
-        // Build tool definitions first (needed for schema injection)
+        // Build tool definitions first (needed for schema injection).
         // When using mesh delegation, enrich tools with endpoint URLs
-        // so the provider can execute tools directly via MCP proxies
+        // so the provider can execute tools directly via MCP proxies.
         const isMeshDelegated = !!context.meshProvider;
         const toolDefs = this.buildToolDefinitions(context.tools, isMeshDelegated);
         // Add system prompt if configured
@@ -375,7 +309,7 @@ export class MeshLlmAgent {
             // formatting via output_schema in model_params. Consumer doesn't know
             // the provider's vendor, so it must not add vendor-agnostic schema instructions.
             const outputMode = this.config.outputMode ?? "hint";
-            if (!context.meshProvider && outputMode !== "text" && this.config.returnSchema) {
+            if (opts.includeOutputSchemaHint && outputMode !== "text" && this.config.returnSchema) {
                 const outputSchemaSection = this.buildOutputSchemaSection();
                 systemContent += outputSchemaSection;
             }
@@ -426,11 +360,41 @@ export class MeshLlmAgent {
                 }
             }
         }
+        return messages;
+    }
+    /**
+     * Run the agentic loop.
+     *
+     * @param messageInput - User message string or multi-turn message array
+     * @param context - Runtime context with tools and options
+     * @returns Parsed response (validated if schema provided)
+     */
+    async run(messageInput, context) {
+        if (this.config.parallelToolCalls && !this._parallelLogEmitted) {
+            console.log("[mesh.llm] parallel tool calls enabled — tools will execute concurrently via Promise.all()");
+            this._parallelLogEmitted = true;
+        }
+        const startTime = Date.now();
+        const toolCalls = [];
+        let totalInputTokens = 0;
+        let totalOutputTokens = 0;
+        // Resolve provider
+        const provider = this.resolveProvider(context);
+        // Build tool definitions (needed for schema injection + the agentic loop).
+        // When using mesh delegation, enrich tools with endpoint URLs
+        // so the provider can execute tools directly via MCP proxies.
+        const isMeshDelegated = !!context.meshProvider;
+        const toolDefs = this.buildToolDefinitions(context.tools, isMeshDelegated);
+        // Build initial messages (system prompt + tool schema + output-schema hint
+        // + resolved media + multi-turn unwinding). run() includes the output-schema
+        // hint only when NOT mesh-delegated.
+        const messages = await this.buildAgentMessages(messageInput, context, {
+            includeOutputSchemaHint: !isMeshDelegated,
+        });
         // Get effective options (runtime options > MESH_LLM_* env > config)
-        const maxIterations = context.options?.maxIterations ??
-            (process.env.MESH_LLM_MAX_ITERATIONS
-                ? parseInt(process.env.MESH_LLM_MAX_ITERATIONS, 10)
-                : this.config.maxIterations);
+        const maxIterations = sanitizeMaxIterations(context.options?.maxIterations) ??
+            envMaxIterations() ??
+            this.config.maxIterations;
         const maxTokens = context.options?.maxOutputTokens ?? this.config.maxOutputTokens;
         const temperature = context.options?.temperature ?? this.config.temperature;
         // Determine model (mesh provider > MESH_LLM_MODEL env > config > default)
@@ -438,19 +402,8 @@ export class MeshLlmAgent {
             process.env.MESH_LLM_MODEL ??
             this.config.model ??
             this.getDefaultModel();
-        // Build output schema for provider (Issue #459) - computed once before loop
-        let outputSchema;
-        if (this.config.returnSchema) {
-            try {
-                const jsonSchema = zodToJsonSchema(this.config.returnSchema);
-                // Extract schema name from title or use generic name
-                const schemaName = jsonSchema.title ?? "Response";
-                outputSchema = { schema: jsonSchema, name: schemaName };
-            }
-            catch {
-                // If schema conversion fails, skip
-            }
-        }
+        // Build output schema for provider (Issue #459) - computed once, cached
+        const outputSchema = this.getOutputSchema();
         // Agentic loop
         let iteration = 0;
         let finalContent = "";
@@ -465,6 +418,11 @@ export class MeshLlmAgent {
                 outputSchema,
                 // Issue #1019: forward caller-supplied escape-hatch kwargs
                 modelParams: context.options?.modelParams,
+                // Issue #1116: forward the resolved provider-managed loop cap.
+                maxIterations,
+                // Issue #1112: forward the RAW (possibly-undefined) output_mode so the
+                // provider honors an explicit override; unset stays auto.
+                outputMode: this.config.outputMode,
             });
             // Track tokens
             if (response.usage) {
@@ -597,79 +555,25 @@ export class MeshLlmAgent {
         // loop. The mesh-delegated streaming provider runs its own loop on the
         // server side and emits text chunks via notifications/progress; the
         // consumer just yields each one.
-        const messages = [];
-        const isMeshDelegated = true; // by definition: we required meshProvider above
-        const toolDefs = this.buildToolDefinitions(context.tools, isMeshDelegated);
-        // System prompt with template rendering + tool schema injection.
-        // Mirrors run(): mesh-delegated path skips the output-schema hint
-        // because the provider applies vendor-specific output formatting.
-        const systemPromptTemplate = this.getSystemPrompt();
-        if (systemPromptTemplate) {
-            let systemContent = await renderTemplate(systemPromptTemplate, context.templateContext ?? {});
-            if (toolDefs.length > 0) {
-                systemContent += this.buildToolSchemaSection(toolDefs);
-            }
-            messages.push({ role: "system", content: systemContent });
-        }
-        // Resolve media items to OpenAI-compatible image_url parts
-        const mediaItems = context.options?.media;
-        let mediaParts = null;
-        if (mediaItems && mediaItems.length > 0) {
-            mediaParts = await resolveMediaInputs(mediaItems);
-        }
-        if (typeof messageInput === "string") {
-            if (mediaParts && mediaParts.length > 0) {
-                messages.push({
-                    role: "user",
-                    content: [
-                        { type: "text", text: messageInput },
-                        ...mediaParts,
-                    ],
-                });
-            }
-            else {
-                messages.push({ role: "user", content: messageInput });
-            }
-        }
-        else {
-            for (let i = 0; i < messageInput.length; i++) {
-                const msg = messageInput[i];
-                const isLastUser = mediaParts &&
-                    mediaParts.length > 0 &&
-                    msg.role === "user" &&
-                    i === messageInput.length - 1;
-                if (isLastUser) {
-                    messages.push({
-                        role: "user",
-                        content: [
-                            { type: "text", text: msg.content },
-                            ...mediaParts,
-                        ],
-                    });
-                }
-                else {
-                    messages.push({ role: msg.role, content: msg.content });
-                }
-            }
-        }
+        // Mesh-delegated by definition (we required meshProvider above).
+        const toolDefs = this.buildToolDefinitions(context.tools, true);
+        // Build initial messages (system prompt + tool schema + resolved media +
+        // multi-turn unwinding). stream() NEVER includes the output-schema hint —
+        // the provider applies vendor-specific output formatting.
+        const messages = await this.buildAgentMessages(messageInput, context, {
+            includeOutputSchemaHint: false,
+        });
         // Effective options (runtime > env > config)
+        const maxIterations = sanitizeMaxIterations(context.options?.maxIterations) ??
+            envMaxIterations() ??
+            this.config.maxIterations;
         const maxTokens = context.options?.maxOutputTokens ?? this.config.maxOutputTokens;
         const temperature = context.options?.temperature ?? this.config.temperature;
         const model = context.meshProvider?.model ??
             process.env.MESH_LLM_MODEL ??
             this.config.model ??
             this.getDefaultModel();
-        let outputSchema;
-        if (this.config.returnSchema) {
-            try {
-                const jsonSchema = zodToJsonSchema(this.config.returnSchema);
-                const schemaName = jsonSchema.title ?? "Response";
-                outputSchema = { schema: jsonSchema, name: schemaName };
-            }
-            catch {
-                // skip
-            }
-        }
+        const outputSchema = this.getOutputSchema();
         const provider = new MeshDelegatedProvider(context.meshProvider.endpoint, context.meshProvider.functionName, this.config.parallelToolCalls ?? false);
         yield* provider.streamComplete(model, messages, toolDefs.length > 0 ? toolDefs : undefined, {
             maxOutputTokens: maxTokens,
@@ -679,6 +583,11 @@ export class MeshLlmAgent {
             outputSchema,
             // Issue #1019: forward caller-supplied escape-hatch kwargs
             modelParams: context.options?.modelParams,
+            // Issue #1116: forward the resolved provider-managed loop cap.
+            maxIterations,
+            // Issue #1112: forward the RAW (possibly-undefined) output_mode so the
+            // provider honors an explicit override; unset stays auto.
+            outputMode: this.config.outputMode,
         });
     }
     /**
@@ -686,8 +595,9 @@ export class MeshLlmAgent {
      */
     createCallable(context) {
         const agent = this;
-        const callable = async (message, options) => {
-            // Handle context mode
+        // Shared "context merge vs replace" semantics used by both the buffered
+        // callable and the stream method below.
+        const mergeRunContext = (options) => {
             const contextMode = options?.contextMode ?? "merge";
             let mergedTemplateContext;
             if (contextMode === "replace" && options?.context) {
@@ -702,13 +612,14 @@ export class MeshLlmAgent {
                 // No runtime context - use base context
                 mergedTemplateContext = context.templateContext ?? {};
             }
-            // Merge options
-            const mergedContext = {
+            return {
                 ...context,
                 options: options ? { ...context.options, ...options } : context.options,
                 templateContext: mergedTemplateContext,
             };
-            return agent.run(message, mergedContext);
+        };
+        const callable = async (message, options) => {
+            return agent.run(message, mergeRunContext(options));
         };
         // Attach meta property
         Object.defineProperty(callable, "meta", {
@@ -727,23 +638,7 @@ export class MeshLlmAgent {
         // "context merge vs replace" behavior as the buffered call.
         Object.defineProperty(callable, "stream", {
             value: (message, options) => {
-                const contextMode = options?.contextMode ?? "merge";
-                let mergedTemplateContext;
-                if (contextMode === "replace" && options?.context) {
-                    mergedTemplateContext = options.context;
-                }
-                else if (options?.context) {
-                    mergedTemplateContext = { ...context.templateContext, ...options.context };
-                }
-                else {
-                    mergedTemplateContext = context.templateContext ?? {};
-                }
-                const mergedContext = {
-                    ...context,
-                    options: options ? { ...context.options, ...options } : context.options,
-                    templateContext: mergedTemplateContext,
-                };
-                return agent.stream(message, mergedContext);
+                return agent.stream(message, mergeRunContext(options));
             },
         });
         return callable;
@@ -819,17 +714,16 @@ export class MeshLlmAgent {
      * Guides the LLM to produce structured output matching the schema.
      */
     buildOutputSchemaSection() {
-        if (!this.config.returnSchema)
-            return "";
-        try {
-            const jsonSchema = zodToJsonSchema(this.config.returnSchema);
-            const schemaStr = JSON.stringify(jsonSchema, null, 2);
-            return `\n\n## Output Format\n\nYour response MUST be valid JSON matching this schema:\n\n\`\`\`json\n${schemaStr}\n\`\`\`\n\nRespond ONLY with the JSON object, no additional text.`;
-        }
-        catch {
-            // If schema conversion fails, skip injection
+        if (this._outputSchemaSection !== null)
+            return this._outputSchemaSection;
+        const cached = this.getOutputSchema();
+        if (!cached) {
+            this._outputSchemaSection = "";
             return "";
         }
+        const schemaStr = JSON.stringify(cached.schema, null, 2);
+        this._outputSchemaSection = `\n\n## Output Format\n\nYour response MUST be valid JSON matching this schema:\n\n\`\`\`json\n${schemaStr}\n\`\`\`\n\nRespond ONLY with the JSON object, no additional text.`;
+        return this._outputSchemaSection;
     }
     /**
      * Execute a tool call and record metadata.
@@ -876,129 +770,40 @@ export function createLlmToolProxy(toolInfo, description) {
     const proxy = async (args) => {
         // Set up timeout (default 30s for tool calls)
         const timeoutMs = parseInt(process.env.MESH_TOOL_TIMEOUT_MS || "30000", 10);
-        // Tracing: propagate context to downstream tool
-        const traceCtx = getCurrentTraceContext();
-        const traceSpanId = traceCtx ? generateSpanId() : null;
-        const traceStartTime = Date.now() / 1000;
-        let traceSuccess = true;
-        let traceError = null;
-        let resultType = "unknown";
-        // Build arguments with trace context injection via Rust core
-        const toolPropHeaders = getCurrentPropagatedHeaders();
-        let toolArgsWithTrace;
-        if (traceCtx && traceSpanId) {
-            try {
-                const argsJson = JSON.stringify(args);
-                const headersJson = Object.keys(toolPropHeaders).length > 0 ? JSON.stringify(toolPropHeaders) : undefined;
-                const injectedJson = injectTraceContext(argsJson, traceCtx.traceId, traceSpanId, headersJson);
-                toolArgsWithTrace = JSON.parse(injectedJson);
-            }
-            catch {
-                // Fallback to manual injection
-                toolArgsWithTrace = {
-                    ...args,
-                    _trace_id: traceCtx.traceId,
-                    _parent_span: traceSpanId,
-                    ...(Object.keys(toolPropHeaders).length > 0 ? { _mesh_headers: toolPropHeaders } : {}),
-                };
-            }
+        // Route through the shared callMcpTool (trace injection, span publish,
+        // dispatcher pooling, job-cancel wiring all handled internally). Tools may
+        // be non-idempotent so disable retries; LLM tool results can exceed the
+        // default 10 MiB cap so raise it to effectively unbounded.
+        const callOptions = {
+            ...DEFAULT_CALL_OPTIONS,
+            timeout: timeoutMs,
+            maxAttempts: 1,
+            maxResponseSize: Number.MAX_SAFE_INTEGER,
+        };
+        let result;
+        try {
+            result = await callMcpTool(toolInfo.endpoint, toolInfo.functionName, args, callOptions, "mesh-tool");
         }
-        else {
-            toolArgsWithTrace = {
-                ...args,
-                ...(Object.keys(toolPropHeaders).length > 0 ? { _mesh_headers: toolPropHeaders } : {}),
-            };
+        catch (error) {
+            // callMcpTool throws a plain Error on failure — re-wrap into the
+            // ToolExecutionError shape the agentic loop expects.
+            throw new ToolExecutionError(toolInfo.functionName, error instanceof Error ? error : new Error(String(error)));
         }
-        try {
-            // Make MCP call to the tool
-            let response;
-            try {
-                response = await fetchWithTimeout(`${toolInfo.endpoint}/mcp`, {
-                    method: "POST",
-                    headers: {
-                        "Content-Type": "application/json",
-                        "Accept": "application/json, text/event-stream",
-                        ...(traceCtx && traceSpanId ? createTraceHeaders(traceCtx.traceId, traceSpanId) : {}),
-                        ...toolPropHeaders,
-                    },
-                    body: JSON.stringify({
-                        jsonrpc: "2.0",
-                        id: Date.now(),
-                        method: "tools/call",
-                        params: {
-                            name: toolInfo.functionName,
-                            arguments: toolArgsWithTrace,
-                        },
-                    }),
-                    timeout: timeoutMs,
-                    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-                    dispatcher: getDispatcher(`${toolInfo.endpoint}/mcp`),
-                });
-            }
-            catch (error) {
-                if (isTimeoutError(error)) {
-                    throw new ToolExecutionError(toolInfo.functionName, new Error(`Tool call timed out after ${timeoutMs}ms (endpoint: ${toolInfo.endpoint})`));
-                }
-                throw new ToolExecutionError(toolInfo.functionName, error instanceof Error ? error : new Error(String(error)));
-            }
-            if (!response.ok) {
-                const errorBody = await response.text();
-                throw new ToolExecutionError(toolInfo.functionName, new Error(`Tool call failed: ${response.status} ${errorBody}`));
-            }
-            // Handle SSE response from FastMCP stateless HTTP stream
-            const responseText = await response.text();
-            const result = parseSSEResponse(responseText);
-            if (result.error) {
-                throw new Error(`Tool error: ${result.error.message}`);
-            }
-            // Parse result content
-            const content = result.result?.content?.[0];
-            if (!content) {
-                resultType = "null";
-                return null;
-            }
-            if (content.type === "text" && content.text) {
-                // Try to parse as JSON
-                try {
-                    const parsed = JSON.parse(content.text);
-                    resultType = typeof parsed;
-                    return parsed;
-                }
-                catch {
-                    resultType = "string";
-                    return content.text;
-                }
-            }
-            resultType = typeof content;
-            return content;
+        // Multi-content results are returned as structured objects as-is.
+        if (typeof result === "object") {
+            return result;
         }
-        catch (err) {
-            traceSuccess = false;
-            traceError = err instanceof Error ? err.message : String(err);
-            throw err;
+        // Empty content signals "tool returned nothing" — preserve the null result
+        // the hand-rolled path produced rather than an empty string.
+        if (result === "") {
+            return null;
         }
-        finally {
-            if (traceCtx && traceSpanId) {
-                const traceEndTime = Date.now() / 1000;
-                const traceDurationMs = (traceEndTime - traceStartTime) * 1000;
-                publishTraceSpan({
-                    traceId: traceCtx.traceId,
-                    spanId: traceSpanId,
-                    parentSpan: traceCtx.parentSpanId,
-                    functionName: "proxy_call_wrapper",
-                    startTime: traceStartTime,
-                    endTime: traceEndTime,
-                    durationMs: traceDurationMs,
-                    success: traceSuccess,
-                    error: traceError,
-                    resultType: traceSuccess ? resultType : "error",
-                    argsCount: 0,
-                    kwargsCount: 0,
-                    dependencies: [toolInfo.endpoint],
-                    injectedDependencies: 0,
-                    meshPositions: [],
-                }).catch(() => { });
-            }
+        // Parse JSON if possible, otherwise return the raw string.
+        try {
+            return JSON.parse(result);
+        }
+        catch {
+            return result;
         }
     };
     // Safely parse inputSchema - don't let malformed JSON break proxy creation