@mcpmesh/sdk 1.4.1 → 2.0.0-beta.1

package/dist/agent.js

@@ -14,6 +14,12 @@ import { startAgent, } from "@mcpmesh/core";
 import { resolveConfig, generateAgentIdSuffix, findAvailablePort } from "./config.js";
 import { enrichSchemaWithMediaTypes } from "./media-param.js";
 import { createProxy, normalizeDependency, runWithTraceContext, runWithPropagatedHeaders, PROXY_DISPATCH_META } from "./proxy.js";
+import { readJobHeaders, runWithJobContext, makeJobController, spliceJobController, } from "./inbound-job-dispatch.js";
+import { MeshJobSubmitter } from "./mesh-job-submitter.js";
+import { ClaimDispatcher } from "./claim-dispatcher.js";
+import { registerJobHelperTools } from "./jobs-helper-tools.js";
+import { registerCancelRoute } from "./jobs-cancel-route.js";
+import { clusterStrictEnabled, normalizeSchemaWithPolicy, } from "./schema-normalize.js";
 import { initTracing, generateTraceId, generateSpanId, publishTraceSpan, matchesPropagateHeader, } from "./tracing.js";
 import { buildLlmAgentSpecs, handleLlmToolsUpdated, handleLlmProviderAvailable, handleLlmProviderUnavailable, LlmToolRegistry, } from "./llm.js";
 import { llmProvider, getLlmProviderMeta } from "./llm-provider.js";
@@ -21,6 +27,7 @@ import { findAndSetBasePath } from "./template.js";
 import { getTlsOptions, getTlsConfigCached, prepareTls, cleanupTls } from "./tls-config.js";
 import { closeHttpPool } from "./http-pool.js";
 import { dispatch as poolDispatch, closePool } from "./tool-worker-pool.js";
+import { A2AClient, A2ABearer, } from "./a2a/index.js";
 /**
  * Globally-set symbol that user agent code can check to detect whether it
  * is running inside a mesh tool-isolation worker. The mesh runtime sets
@@ -116,6 +123,41 @@ export class MeshAgent {
     // In worker mode addTool() only stashes execute fns and skips all FastMCP /
     // registry / Rust core wiring (no Express port conflict, no double-register).
     _workerMode = false;
+    /**
+     * Phase 1 MeshJob substrate: per-tool ClaimHandler for `task: true`
+     * tools. Indexed by capability so the ClaimDispatcher can look up
+     * the local handler without re-traversing the tools map. Populated
+     * by addTool() at registration time; consumed by _autoStart() to
+     * spawn one dispatcher per task tool.
+     *
+     * Issue #894: also carries the per-tool retryOn whitelist so the
+     * dispatcher can pass it into `runWithJobContext` for the
+     * release-lease-on-retry-eligible-throw path.
+     */
+    _taskHandlers = new Map();
+    /**
+     * Active claim dispatchers (one per task=true capability). Started
+     * during _autoStart(); stopped during shutdown(). Empty for agents
+     * that own no task=true tools.
+     */
+    _claimDispatchers = [];
+    /**
+     * Issue #917: cache of `A2AClient` instances keyed by their
+     * `(url, skillId, auth, timeoutMs)` tuple so multiple consumer
+     * tools targeting the same backend share one outbound connection
+     * pool. Closed via `close()` on agent shutdown.
+     */
+    _a2aClients = new Map();
+    /**
+     * Issue #917: stable opaque IDs for `A2ABearer` instances used in
+     * the A2AClient cache key. Bearer fields are private so we cannot
+     * fingerprint by content (would also be a security risk — two
+     * tools with distinct literal tokens must NEVER share a cache
+     * entry). Identity-based keying is the safe default. `WeakMap`
+     * lets bearers be GC'd when the registering tool is removed.
+     */
+    _bearerIds = new WeakMap();
+    _nextBearerId = 0;
     constructor(server, config) {
         if (_isWorkerMode) {
             // Worker thread: skip ALL init. Only addTool() runs (in worker-mode
@@ -158,6 +200,168 @@ export class MeshAgent {
     addTool(def) {
         const toolName = def.name;
         const execute = def.execute;
+        // Phase 1 MeshJob substrate: validate `task: true` requires an
+        // async function. Long-running tools need a Promise-based control
+        // flow so the dispatch wrapper (Phase B) can await
+        // `MeshJob.updateProgress()` / cancellation / outbound polling.
+        // Fail loudly at `addTool` so the developer sees the misuse before
+        // the agent even tries to register with the registry.
+        //
+        // Heuristic: AsyncFunction.constructor.name === "AsyncFunction".
+        // We only flag the obvious sync case (an arrow/function literal)
+        // — any function returning a Promise will pass this check, which
+        // is the right relaxation for users who wrap their handler in a
+        // Promise factory.
+        if (def.task === true) {
+            const ctorName = execute
+                ?.constructor?.name;
+            if (ctorName !== "AsyncFunction") {
+                // We can't reliably detect Promise-returning sync functions
+                // without invoking them, but we CAN reject the unambiguous
+                // "function() { ... }" case where the developer probably
+                // forgot the `async` keyword.
+                if (ctorName === "Function") {
+                    throw new Error(`addTool({ task: true }) requires an async execute function; ` +
+                        `tool '${toolName}' has a sync execute. Mark it 'async' or ` +
+                        `remove task: true.`);
+                }
+                // Other constructor names (GeneratorFunction, etc.) are
+                // unusual; let them through with a console warning rather
+                // than blocking — the dispatch wrapper will surface any actual
+                // misuse at first invocation.
+            }
+        }
+        // Phase 1 MeshJob substrate (consumer-side validation): if the
+        // tool declares meshJobDepIndex, that index MUST be a non-negative
+        // integer pointing to a valid dependency. Catch misuse at
+        // registration so the developer doesn't see a confusing TypeError
+        // at runtime when the wrapper tries to swap the dep proxy for a
+        // submitter. Mirrors the meshJobParamIndex validation below —
+        // NaN / fractional / negative values must fail-fast here too.
+        if (def.meshJobDepIndex !== undefined) {
+            const depCount = (def.dependencies ?? []).length;
+            const v = def.meshJobDepIndex;
+            const isInt = Number.isInteger(v) && v >= 0;
+            if (!isInt) {
+                throw new Error(`addTool({ meshJobDepIndex: ${v} }) for tool '${toolName}': ` +
+                    `meshJobDepIndex must be a non-negative integer (index into ` +
+                    `dependencies[]), got: ${v}`);
+            }
+            if (v >= depCount) {
+                throw new Error(`addTool({ meshJobDepIndex: ${v} }) for tool ` +
+                    `'${toolName}' is out of range — the tool declares ${depCount} ` +
+                    `dependencies. meshJobDepIndex must be a valid index into ` +
+                    `dependencies[].`);
+            }
+        }
+        // Phase 1 MeshJob substrate (producer-side validation): if the
+        // tool declares meshJobParamIndex, that position MUST be a sane
+        // integer >= 1. Position 0 is reserved for the args payload, so
+        // the controller can only land at sig pos 1+. Without this
+        // guard, values 0 / negative / NaN / non-integer silently skip
+        // controller injection — the user's handler then sees `null`
+        // where it expected a JobController and throws a confusing
+        // `TypeError: Cannot read properties of null` at first await.
+        //
+        // Upper bound is a sanity check: > 10 almost certainly means a
+        // typo (no real producer signature has that many params).
+        if (def.meshJobParamIndex !== undefined) {
+            const v = def.meshJobParamIndex;
+            const ok = Number.isInteger(v) && v >= 1 && v <= 10;
+            if (!ok) {
+                throw new Error(`addTool({ meshJobParamIndex: ${v} }) for tool '${toolName}': ` +
+                    `meshJobParamIndex must be an integer >= 1 (position of MeshJob ` +
+                    `param after the args payload), got: ${v}`);
+            }
+        }
+        // Issue #894: validate retryOn at registration so misuse fails loud
+        // before the agent talks to the registry. Mirror Python's
+        // `mesh.decorators` validation in spirit:
+        //   - retryOn requires task: true (without the job dispatch wrapper
+        //     there's no controller to release a lease on, so the kwarg is
+        //     meaningless);
+        //   - entries must be Error constructor classes (typeof === "function").
+        // We don't filter control-flow exceptions like Python's
+        // KeyboardInterrupt / asyncio.CancelledError — JavaScript has no
+        // direct equivalent, and AbortError-style cancellation is a legitimate
+        // retry trigger for some users. They get to choose.
+        if (def.retryOn !== undefined) {
+            if (def.task !== true) {
+                throw new Error(`addTool({ retryOn }) for tool '${toolName}': retryOn is only ` +
+                    `valid with task: true; remove retryOn or set task: true.`);
+            }
+            if (!Array.isArray(def.retryOn)) {
+                throw new Error(`addTool({ retryOn }) for tool '${toolName}': retryOn must be ` +
+                    `an array of Error constructor classes (e.g., [TypeError, MyError]).`);
+            }
+            for (const entry of def.retryOn) {
+                // Must be a function that has a prototype (i.e. an actual class
+                // or a `function` declaration — not an arrow function), AND must
+                // either be Error itself or a subclass. Arrow functions have
+                // `prototype === undefined`, so `entry.prototype instanceof Error`
+                // is `false` for them — they're rejected by the second check.
+                // Without this, `err instanceof <arrow>` at dispatch time would
+                // throw `TypeError: Right-hand side of instanceof is not callable`.
+                if (typeof entry !== "function") {
+                    throw new Error(`addTool({ retryOn }) for tool '${toolName}': retryOn entries ` +
+                        `must be Error constructor classes (functions); got: ${String(entry)}`);
+                }
+                if (entry !== Error && !(entry.prototype instanceof Error)) {
+                    throw new Error(`addTool({ retryOn }) for tool '${toolName}': retryOn entries ` +
+                        `must extend Error (or be Error itself); got: ${String(entry)}`);
+                }
+            }
+        }
+        // Issue #917: validate a2aConfig at registration time so misuse fails
+        // loud BEFORE the agent talks to the registry. Match the Python
+        // `mesh.a2a_consumer` and Java `@A2AConsumer` startup-time checks.
+        let a2aClient = null;
+        if (def.a2aConfig !== undefined) {
+            const cfg = def.a2aConfig;
+            if (!cfg.url || cfg.url.trim() === "") {
+                throw new Error(`addTool({ a2aConfig }) for tool '${toolName}': url must be ` +
+                    `a non-empty string.`);
+            }
+            if (cfg.timeoutMs !== undefined) {
+                if (!Number.isFinite(cfg.timeoutMs) || cfg.timeoutMs <= 0) {
+                    throw new Error(`addTool({ a2aConfig }) for tool '${toolName}': timeoutMs ` +
+                        `must be a finite positive number (got ${cfg.timeoutMs}).`);
+                }
+            }
+            if (cfg.pollIntervalMs !== undefined) {
+                if (!Number.isFinite(cfg.pollIntervalMs) || cfg.pollIntervalMs <= 0) {
+                    throw new Error(`addTool({ a2aConfig }) for tool '${toolName}': ` +
+                        `pollIntervalMs must be a finite positive number ` +
+                        `(got ${cfg.pollIntervalMs}).`);
+                }
+            }
+            if (cfg.pollIntervalMaxMs !== undefined) {
+                if (!Number.isFinite(cfg.pollIntervalMaxMs) ||
+                    cfg.pollIntervalMaxMs <= 0) {
+                    throw new Error(`addTool({ a2aConfig }) for tool '${toolName}': ` +
+                        `pollIntervalMaxMs must be a finite positive number ` +
+                        `(got ${cfg.pollIntervalMaxMs}).`);
+                }
+            }
+            if (cfg.pollIntervalMs !== undefined &&
+                cfg.pollIntervalMaxMs !== undefined &&
+                cfg.pollIntervalMaxMs < cfg.pollIntervalMs) {
+                throw new Error(`addTool({ a2aConfig }) for tool '${toolName}': ` +
+                    `pollIntervalMaxMs (${cfg.pollIntervalMaxMs}) must be >= ` +
+                    `pollIntervalMs (${cfg.pollIntervalMs}).`);
+            }
+            if (!this._workerMode) {
+                const skillId = cfg.skillId ?? def.capability ?? toolName;
+                a2aClient = this._getOrBuildA2AClient({
+                    url: cfg.url,
+                    skillId,
+                    auth: this._buildBearerFromConfig(cfg.auth),
+                    timeoutMs: cfg.timeoutMs,
+                    pollIntervalMs: cfg.pollIntervalMs,
+                    pollIntervalMaxMs: cfg.pollIntervalMaxMs,
+                });
+            }
+        }
         // Worker mode: register the raw execute fn in the worker tool map and
         // skip FastMCP registration, dependency wiring, and metadata storage.
         // The worker entry will look up tools by name when handling dispatched
@@ -169,10 +373,51 @@ export class MeshAgent {
         // Normalize dependencies
         const normalizedDeps = (def.dependencies ?? []).map(normalizeDependency);
         const depEndpoints = normalizedDeps.map((d) => d.capability);
+        // Capture for closures — these reads must be live at invocation
+        // time (e.g. registryUrl/agentId aren't set yet at addTool time).
+        const isTaskTool = def.task === true;
+        const meshJobDepIndex = def.meshJobDepIndex;
+        const meshJobParamIndex = def.meshJobParamIndex;
+        // Issue #894: per-tool retryOn whitelist threaded into both
+        // dispatch paths (inbound HTTP wrapper below + ClaimHandler
+        // registered in this.taskHandlers). Captured here so the closure
+        // sees a stable reference even if def is mutated post-registration.
+        const retryOn = def.retryOn;
+        // Phase 1 MeshJob substrate: when a job-bound tool exists AND the
+        // user explicitly opted into worker isolation via env, log a single
+        // warning at registration time. The wrapper force-disables
+        // isolation for job-bound tools because controllers + the
+        // AsyncLocalStorage / Rust task-local job context don't cross the
+        // worker_threads boundary cleanly. Without this log the
+        // force-disable was silent — users who set MCP_MESH_TOOL_ISOLATION
+        // expected it to apply to every tool.
+        const isJobBoundForLog = isTaskTool || meshJobDepIndex !== undefined;
+        const isolationEnvSet = typeof process.env.MCP_MESH_TOOL_ISOLATION === "string" &&
+            process.env.MCP_MESH_TOOL_ISOLATION.toLowerCase() !== "false";
+        if (isJobBoundForLog && isolationEnvSet) {
+            console.warn(`[mesh-tool] '${toolName}' has ` +
+                (isTaskTool ? "task: true" : `meshJobDepIndex: ${meshJobDepIndex}`) +
+                `; worker isolation is disabled for job-bound tools ` +
+                `(controllers/AsyncLocalStorage don't cross worker boundaries). ` +
+                `Set 'task: true' explicitly if you intend a producer.`);
+        }
         // Create wrapper that injects dependencies positionally and handles tracing
         const wrappedExecute = async (args) => {
             // Build positional deps array using composite keys (toolName:dep_index)
-            const depsArray = normalizedDeps.map((_, depIndex) => this.resolvedDeps.get(`${toolName}:dep_${depIndex}`) ?? null);
+            // Phase 1 MeshJob substrate (consumer-side): if meshJobDepIndex is
+            // set, swap the McpMeshTool proxy at that slot for a
+            // MeshJobSubmitter targeting that dep's capability. We bind the
+            // submitter to the live registryUrl/agentId so it can submit
+            // jobs without needing access to the agent instance.
+            const depsArray = normalizedDeps.map((dep, depIndex) => {
+                if (depIndex === meshJobDepIndex) {
+                    // Build the submitter lazily per call so we always pick
+                    // up the current registryUrl (test harnesses sometimes
+                    // mutate it between calls).
+                    return new MeshJobSubmitter(dep.capability, this.agentId, this.config.registryUrl);
+                }
+                return this.resolvedDeps.get(`${toolName}:dep_${depIndex}`) ?? null;
+            });
             const injectedCount = depsArray.filter((d) => d !== null).length;
             // Extract trace context from arguments (injected by upstream proxy)
             // This is the fallback mechanism since fastmcp doesn't expose HTTP headers
@@ -219,7 +464,26 @@ export class MeshAgent {
             // Python implementation in _mcp_mesh/shared/tool_executor.py.
             // Default ON; set MCP_MESH_TOOL_ISOLATION=false to revert to inline
             // execution on the main loop (legacy behavior).
-            const isolationEnabled = (process.env.MCP_MESH_TOOL_ISOLATION ?? "true").toLowerCase() !== "false";
+            //
+            // Phase 1 MeshJob substrate: force-disable isolation for tools
+            // that bind to a JobController or MeshJobSubmitter. The
+            // controller/submitter wrap napi-rs handles plus
+            // AsyncLocalStorage state that cannot be cleanly serialised
+            // across the worker_threads boundary. Running inline on the
+            // main loop is the right trade — task=true tools are
+            // long-running by definition and benefit less from isolation
+            // (their wall-clock time is dominated by the user's `await`s,
+            // not CPU bursts that block the event loop).
+            // Issue #917: A2A consumer tools force-disable isolation along
+            // with job-bound tools. The framework-injected `A2AClient` wraps
+            // an undici dispatcher handle that cannot be cleanly serialised
+            // across the worker_threads boundary; running inline keeps the
+            // cached client + connection pool intact across calls.
+            const isA2aBound = a2aClient !== null;
+            const isJobBound = isTaskTool || meshJobDepIndex !== undefined;
+            const isolationEnabled = !isJobBound &&
+                !isA2aBound &&
+                (process.env.MCP_MESH_TOOL_ISOLATION ?? "true").toLowerCase() !== "false";
             try {
                 let result;
                 if (isolationEnabled) {
@@ -257,16 +521,66 @@ export class MeshAgent {
                 }
                 else {
                     // Legacy inline execution on the main thread. Preserved as a clean
-                    // fallback for users who explicitly opt out of isolation.
+                    // fallback for users who explicitly opt out of isolation, AND used
+                    // unconditionally for job-bound tools (see isJobBound above).
+                    //
+                    // Phase 1 MeshJob substrate: when this tool is task=true and the
+                    // inbound headers carry X-Mesh-Job-Id, build a JobController,
+                    // splice it into the call args at meshJobParamIndex, and run the
+                    // user function inside both the JS-side ALS (CURRENT_JOB) and the
+                    // Rust-side task-local (withJobAsync) so cancel-registry binding
+                    // + outbound header injection work transparently.
                     result = await runWithTraceContext(traceContext, async () => {
                         return await runWithPropagatedHeaders(propagatedHeaders, async () => {
+                            if (isTaskTool) {
+                                const [jobId, deadlineSecs] = readJobHeaders(propagatedHeaders);
+                                let controller = null;
+                                if (jobId && this.config.registryUrl && this.agentId) {
+                                    try {
+                                        controller = makeJobController(jobId, this.agentId, this.config.registryUrl);
+                                    }
+                                    catch (err) {
+                                        // Don't silently fall back to a regular tool call —
+                                        // a `task: true` tool that needs a controller will
+                                        // misbehave (return a dict instead of completing the
+                                        // row, leaving the registry's job stuck in `working`
+                                        // until lease expiry). Surface the failure so the
+                                        // outer FastMCP handler reports it AND the inbound
+                                        // wrapper's catch (or its caller) can fail-fast.
+                                        console.error(`[mesh-jobs] makeJobController failed for tool ` +
+                                            `'${toolName}' job=${jobId} agent=${this.agentId} ` +
+                                            `registry=${this.config.registryUrl}:`, err);
+                                        throw err;
+                                    }
+                                }
+                                // Build the call args, splicing the controller (or null) at
+                                // meshJobParamIndex if specified. Position 0 is `args`; deps
+                                // begin at position 1. The MeshJob slot is orthogonal —
+                                // when meshJobParamIndex skips a position, deps shift past
+                                // it (caller's signature must reflect that).
+                                const callArgs = spliceJobController(cleanArgs, depsArray, controller, meshJobParamIndex);
+                                // Issue #917: append the framework-cached A2AClient as
+                                // the trailing positional arg when this tool declares
+                                // a2aConfig. Mirrors the producer-side JobController
+                                // splice — A2AClient never participates in the
+                                // ordered-deps math, it always lands last.
+                                if (a2aClient !== null) {
+                                    callArgs.push(a2aClient);
+                                }
+                                return await runWithJobContext(jobId, deadlineSecs, controller, () => Promise.resolve(execute(...callArgs)), retryOn);
+                            }
+                            if (a2aClient !== null) {
+                                return await execute(cleanArgs, ...depsArray, a2aClient);
+                            }
                             return await execute(cleanArgs, ...depsArray);
                         });
                     });
                 }
-                // Auto-serialize non-string results (like Python SDK does)
-                // This allows users to return natural types (numbers, objects, arrays)
-                // without manually calling JSON.stringify() or String()
+                // Auto-serialize non-string results (like Python SDK does).
+                // NOTE: structuredContent removed in #917 — FastMCP TS rejects it via
+                // strict zod schema (ContentResultZodSchema.strict()) even though
+                // the field is part of the MCP spec. Re-enable when FastMCP TS
+                // upstream accepts the field. Tracked in #925.
                 if (typeof result === "string") {
                     return result;
                 }
@@ -274,6 +588,10 @@ export class MeshAgent {
                     return "";
                 }
                 else {
+                    // Return JSON-stringified text only — every consumer parses
+                    // content[0].text back into an object anyway. FastMCP TS will
+                    // auto-build {content: [{type: "text", text: <string>}]} from
+                    // this bare string return, satisfying its strict schema.
                     return JSON.stringify(result);
                 }
             }
@@ -323,17 +641,66 @@ export class MeshAgent {
             parameters: parametersWithPassthrough,
             execute: wrappedExecute,
         });
+        // Phase 1 MeshJob substrate: register a ClaimHandler for this
+        // tool so the per-capability ClaimDispatcher (spawned in
+        // _autoStart) can dispatch claimed jobs to the same execute fn
+        // — without going through FastMCP's HTTP transport. The handler
+        // builds the same callArgs shape the inbound wrapper does, but
+        // gets the controller passed in directly (no header parsing
+        // needed) and bypasses FastMCP's tool-call serialisation.
+        if (isTaskTool) {
+            const capability = def.capability ?? toolName;
+            const handler = async (payload, controller) => {
+                const liveDeps = normalizedDeps.map((dep, depIndex) => {
+                    if (depIndex === meshJobDepIndex) {
+                        return new MeshJobSubmitter(dep.capability, this.agentId, this.config.registryUrl);
+                    }
+                    return this.resolvedDeps.get(`${toolName}:dep_${depIndex}`) ?? null;
+                });
+                const callArgs = spliceJobController(payload, liveDeps, controller, meshJobParamIndex);
+                // Issue #917: A2A consumer tools dispatched via the claim
+                // path get the same trailing A2AClient argument as the
+                // inbound HTTP path.
+                if (a2aClient !== null) {
+                    callArgs.push(a2aClient);
+                }
+                return await execute(...callArgs);
+            };
+            this._taskHandlers.set(capability, { handler, retryOn });
+        }
         // Store mesh metadata with JSON Schema for LLM tool resolution
         const inputSchema = this.convertZodToJsonSchema(def.parameters);
         enrichSchemaWithMediaTypes(inputSchema);
+        // Issue #547: extract output schema if user supplied one. Zod cannot
+        // infer return types from the handler signature, so this is opt-in.
+        let outputSchemaRaw;
+        if (def.outputSchema) {
+            outputSchemaRaw = this.convertZodToJsonSchema(def.outputSchema);
+        }
         this.tools.set(toolName, {
             capability: def.capability ?? toolName,
             version: def.version ?? "1.0.0",
             tags: def.tags ?? [],
             description: def.description ?? "",
             inputSchema: JSON.stringify(inputSchema),
+            outputSchemaRaw,
+            // Issue #547 Phase 4: per-tool override (default true = current behavior).
+            outputSchemaStrict: def.outputSchemaStrict !== false,
             dependencies: normalizedDeps,
             dependencyKwargs: def.dependencyKwargs,
+            // Phase 1 MeshJob substrate: stamp producer's long-running flag
+            // so the heartbeat pipeline ships it to the registry. Consumers
+            // read this to decide between job semantics and a regular
+            // tools/call.
+            task: def.task === true,
+            meshJobParamIndex: def.meshJobParamIndex,
+            meshJobDepIndex: def.meshJobDepIndex,
+            // Issue #917: A2A consumer marker so heartbeat-build appends the
+            // surrounding agent name to the tag list before shipping to the
+            // registry. Captured here at addTool time so a downstream rename
+            // of `this.config.name` doesn't desync the registered tag.
+            a2aConsumer: def.a2aConfig !== undefined,
+            a2aAgentName: def.a2aConfig !== undefined ? this.config.name : undefined,
         });
         return this;
     }
@@ -390,11 +757,78 @@ export class MeshAgent {
         }
         return this;
     }
+    /**
+     * Issue #917: build an `A2ABearer` (or undefined) from the
+     * user-friendly auth config supported on `MeshA2AConfig.auth`. The
+     * config can be either an `{ token, tokenEnv }` shorthand object OR
+     * a pre-built `A2ABearer` instance the user constructed manually.
+     *
+     * Tightened to `instanceof A2ABearer` so a stray `{ token,
+     * authorizationHeader: () => ... }` object cannot duck-type its way
+     * past A2ABearer's validation (which catches blank tokens and
+     * mutually-exclusive `token`/`tokenEnv`).
+     */
+    _buildBearerFromConfig(auth) {
+        if (auth === undefined)
+            return undefined;
+        if (auth instanceof A2ABearer)
+            return auth;
+        return new A2ABearer(auth);
+    }
+    /**
+     * Issue #917: cache `A2AClient` instances by their config tuple so
+     * multiple consumer tools targeting the same backend share one
+     * outbound connection pool. Auth instances participate in the cache
+     * key by reference (same `A2ABearer` ref → same client); two
+     * separately-constructed bearers — even ones holding identical
+     * tokens — get separate clients. Identity-based keying is the safe
+     * default: A2ABearer's private fields make content-fingerprinting
+     * impossible from outside, and a content-derived key risks leaking
+     * tool-A's bearer onto tool-B's outbound traffic.
+     */
+    _bearerCacheKey(bearer) {
+        if (!bearer)
+            return "none";
+        // A2AClientConfig.auth permits a raw A2ABearerConfig too, but the
+        // call site below always normalises via `_buildBearerFromConfig`
+        // first, so in practice we only ever see real A2ABearer instances.
+        // Defensively pass-through the config-shape case as a content-free
+        // fallback key — never collide with the bearer-id namespace.
+        if (!(bearer instanceof A2ABearer))
+            return "raw-config";
+        let id = this._bearerIds.get(bearer);
+        if (id === undefined) {
+            id = `bearer-${this._nextBearerId++}`;
+            this._bearerIds.set(bearer, id);
+        }
+        return id;
+    }
+    _getOrBuildA2AClient(config) {
+        const key = [
+            config.url,
+            config.skillId,
+            this._bearerCacheKey(config.auth),
+            config.timeoutMs ?? "default",
+            config.pollIntervalMs ?? "default",
+            config.pollIntervalMaxMs ?? "default",
+        ].join("|");
+        const existing = this._a2aClients.get(key);
+        if (existing)
+            return existing;
+        const client = new A2AClient(config);
+        this._a2aClients.set(key, client);
+        return client;
+    }
     /**
      * Convert Zod schema to JSON Schema.
      */
     convertZodToJsonSchema(schema) {
-        return zodToJsonSchema(schema, { $refStrategy: "none" });
+        // $refStrategy: "root" preserves $ref + definitions for recursive Zod
+        // schemas (e.g. z.lazy(...)). With "none", zod-to-json-schema can't expand
+        // the cycle and falls back to {} (empty), which erases the recursion from
+        // the canonical hash. Non-recursive shapes are unchanged because they have
+        // no references to inline.
+        return zodToJsonSchema(schema, { $refStrategy: "root" });
     }
     /**
      * Internal: Start the agent (called by auto-start mechanism).
@@ -491,11 +925,90 @@ export class MeshAgent {
         }
         // 2. Register LLM tools from LlmToolRegistry
         this.registerLlmTools();
+        // 2.5 Phase 1 MeshJob substrate: register the three framework
+        // helper tools (`__mesh_job_status`/`_result`/`_cancel`) on
+        // every TS agent regardless of whether it owns task=true tools.
+        // Mirrors Python's JobsHelperToolsStep. Skipped when there's no
+        // registry URL — the helpers can't function without it.
+        this.registerJobsHelperTools();
+        // 2.6 Phase 1 MeshJob substrate: mount POST /jobs/:job_id/cancel
+        // on FastMCP's underlying Hono app so the registry's cancel
+        // forwarder can fire the in-process cancel token. Best-effort —
+        // failures here are logged, not fatal. When this agent owns
+        // task: true tools and the route fails to register, escalate to
+        // a second console.error so the operator can't miss the
+        // cancel-mid-flight regression in logs.
+        if (this.config.registryUrl) {
+            const cancelRouteOk = registerCancelRoute(this.server);
+            if (!cancelRouteOk && this._taskHandlers.size > 0) {
+                console.error(`[mesh-jobs] agent ${this.agentId} owns ${this._taskHandlers.size} ` +
+                    `task: true tool(s) but the cancel route failed to register. ` +
+                    `Cancel requests for in-flight jobs will fall through to ` +
+                    `lease expiry — see the prior [mesh-jobs] error for the cause.`);
+            }
+        }
         // 3. Start heartbeat to registry via Rust core
         await this.startHeartbeat();
+        // 3.5 Phase 1 MeshJob substrate: spawn one ClaimDispatcher per
+        // task=true tool so the agent can poll the registry's
+        // /jobs/claim and dispatch claimed work locally. Started after
+        // heartbeat so the registry already knows this replica when the
+        // first claim arrives (eliminates the "claim before
+        // registration" race).
+        this.startClaimDispatchers();
         // 4. Install signal handlers for graceful shutdown
         this.installSignalHandlers();
     }
+    /**
+     * Phase 1 MeshJob substrate: register the three framework helper
+     * tools on the FastMCP server AND in the agent's tool catalog so
+     * the heartbeat ships them to the registry as visible capabilities.
+     */
+    registerJobsHelperTools() {
+        if (!this.config.registryUrl) {
+            return;
+        }
+        let helpers;
+        try {
+            helpers = registerJobHelperTools(this.server, this.config.registryUrl);
+        }
+        catch (err) {
+            console.warn("[mesh-jobs] failed to register job helper tools:", err);
+            return;
+        }
+        for (const [name, meta] of helpers.entries()) {
+            // Don't overwrite a user-defined tool with the same name.
+            if (this.tools.has(name))
+                continue;
+            this.tools.set(name, {
+                capability: meta.capability,
+                version: meta.version,
+                tags: meta.tags,
+                description: meta.description,
+                inputSchema: meta.inputSchema,
+                outputSchemaStrict: true,
+                dependencies: [],
+                dependencyKwargs: undefined,
+                task: meta.task,
+            });
+        }
+    }
+    /**
+     * Phase 1 MeshJob substrate: spawn ClaimDispatchers for every
+     * task=true tool registered. Skipped if no registry URL or no task
+     * handlers are present.
+     */
+    startClaimDispatchers() {
+        if (!this.config.registryUrl)
+            return;
+        if (this._taskHandlers.size === 0)
+            return;
+        for (const [capability, entry] of this._taskHandlers.entries()) {
+            const dispatcher = new ClaimDispatcher(capability, this.agentId, this.config.registryUrl, entry.handler, entry.retryOn);
+            dispatcher.start();
+            this._claimDispatchers.push(dispatcher);
+        }
+    }
     /**
      * Register LLM tools from LlmToolRegistry.
      * This adds tool metadata for LLM tools created via mesh.llm().
@@ -561,6 +1074,9 @@ export class MeshAgent {
     async startHeartbeat() {
         // Get LLM tool registry for llmFilter/llmProvider
         const llmRegistry = LlmToolRegistry.getInstance();
+        // Issue #547 Phase 4: read cluster strict knob once; per-tool override
+        // is read inside the loop below.
+        const clusterStrict = clusterStrictEnabled();
         // Build the agent spec for Rust core
         const tools = Array.from(this.tools.entries()).map(([name, meta]) => {
             // Check if this tool has LLM config
@@ -581,21 +1097,89 @@ export class MeshAgent {
                     tags: llmConfig.provider.tags ?? [],
                 });
             }
+            // Issue #547 / Phase 4: normalize via Rust core and apply verdict policy.
+            // Throws on (effective) BLOCK to refuse agent startup; demoted BLOCKs
+            // and WARNs are logged loudly and shipped in schemaWarnings.
+            const toolStrict = meta.outputSchemaStrict !== false;
+            let inputSchemaCanonical;
+            let inputSchemaHash;
+            let outputSchemaCanonical;
+            let outputSchemaHash;
+            const combinedWarnings = [];
+            if (meta.inputSchema) {
+                let inputRaw;
+                try {
+                    inputRaw = JSON.parse(meta.inputSchema);
+                }
+                catch {
+                    // shouldn't happen, but fall through without normalizing
+                }
+                if (inputRaw) {
+                    const r = normalizeSchemaWithPolicy(inputRaw, `tool '${name}' input`, clusterStrict, toolStrict);
+                    inputSchemaCanonical = r.canonicalJson ?? undefined;
+                    inputSchemaHash = r.hash ?? undefined;
+                    combinedWarnings.push(...r.warnings);
+                }
+            }
+            let outputSchemaJson;
+            if (meta.outputSchemaRaw) {
+                outputSchemaJson = JSON.stringify(meta.outputSchemaRaw);
+                const r = normalizeSchemaWithPolicy(meta.outputSchemaRaw, `tool '${name}' output`, clusterStrict, toolStrict);
+                outputSchemaCanonical = r.canonicalJson ?? undefined;
+                outputSchemaHash = r.hash ?? undefined;
+                combinedWarnings.push(...r.warnings);
+            }
+            // Issue #917: when this tool was registered with a2aConfig,
+            // append the consumer agent's name to the tag list (defensive
+            // copy — never mutate meta.tags). Skips when the agent has
+            // no name (consumer-only / nameless agent) or when the tag
+            // already appears, mirrors Java's
+            // MeshToolRegistry.injectConsumerNameTags semantics.
+            let effectiveTags = meta.tags;
+            if (meta.a2aConsumer) {
+                const agentName = meta.a2aAgentName;
+                if (agentName &&
+                    agentName.trim() !== "" &&
+                    !meta.tags.includes(agentName)) {
+                    effectiveTags = [...meta.tags, agentName];
+                }
+            }
             return {
                 functionName: name,
                 capability: meta.capability,
                 version: meta.version,
-                tags: meta.tags,
+                tags: effectiveTags,
                 description: meta.description,
                 // Pass dependencies to Rust core for registry resolution
                 // Note: tags may contain nested arrays for OR alternatives (TagSpec[])
                 // Serialize to JSON for Rust binding - preserves nested structure
-                dependencies: meta.dependencies.map((dep) => ({
-                    capability: dep.capability,
-                    tags: JSON.stringify(dep.tags ?? []),
-                    version: dep.version,
-                })),
+                dependencies: meta.dependencies.map((dep) => {
+                    // Issue #547: normalize per-dep expectedSchemaRaw. There's no
+                    // per-tool override on the consumer side (override is producer-
+                    // side); we still apply cluster strict so WARN→BLOCK works.
+                    let expectedCanonical;
+                    let expectedHash;
+                    if (dep.expectedSchemaRaw) {
+                        const r = normalizeSchemaWithPolicy(dep.expectedSchemaRaw, `dependency on '${dep.capability}'`, clusterStrict, true);
+                        expectedCanonical = r.canonicalJson ?? undefined;
+                        expectedHash = r.hash ?? undefined;
+                    }
+                    return {
+                        capability: dep.capability,
+                        tags: JSON.stringify(dep.tags ?? []),
+                        version: dep.version,
+                        expectedSchemaCanonical: expectedCanonical,
+                        expectedSchemaHash: expectedHash,
+                        matchMode: dep.matchMode,
+                    };
+                }),
                 inputSchema: meta.inputSchema,
+                outputSchema: outputSchemaJson,
+                inputSchemaCanonical,
+                inputSchemaHash,
+                outputSchemaCanonical,
+                outputSchemaHash,
+                schemaWarnings: combinedWarnings.length > 0 ? combinedWarnings : undefined,
                 // LLM filter/provider as JSON strings (matches Python format)
                 llmFilter,
                 llmProvider,
@@ -833,6 +1417,27 @@ export class MeshAgent {
      * Shutdown the agent gracefully.
      */
     async shutdown() {
+        // Phase 1 MeshJob substrate: stop claim dispatchers first so
+        // they don't pull a fresh job mid-shutdown.
+        for (const d of this._claimDispatchers) {
+            try {
+                await d.stop();
+            }
+            catch (err) {
+                console.warn(`[mesh-jobs] error stopping claim dispatcher:`, err);
+            }
+        }
+        this._claimDispatchers = [];
+        // Issue #917: mark all cached A2AClients closed so any in-flight
+        // user code raises cleanly instead of reusing a torn-down instance.
+        // Close in parallel so one slow client doesn't block the others —
+        // the undici Agent pool is shared via closeHttpPool() below.
+        const closePromises = Array.from(this._a2aClients.values()).map((client) => client.close().catch((err) => {
+            console.warn("[mesh-a2a] Error closing A2AClient:", err);
+            return null;
+        }));
+        await Promise.allSettled(closePromises);
+        this._a2aClients.clear();
         try {
             await closeHttpPool();
         }