@theokit/sdk-handoff 0.1.1 → 0.1.3

package/dist/index.d.ts

@@ -1,7 +1,7 @@
 import { SDKAgent, Plugin } from '@theokit/sdk';
 import { ZodType } from 'zod';
-import { H as HandoffDescriptor, a as HandoffOptions } from './handoff-D-Ujv-lA.js';
-export { b as HandoffLoopError, c as HandoffNameCollisionError, d as HandoffPairLoopError, e as HandoffReceiverDisposedError, f as HandoffSelfReferenceError } from './handoff-D-Ujv-lA.js';
+import { H as HandoffDescriptor, a as HandoffOptions } from './handoff-D7malWe_.js';
+export { b as HandoffLoopError, c as HandoffNameCollisionError, d as HandoffPairLoopError, e as HandoffReceiverDisposedError, f as HandoffSelfReferenceError } from './handoff-D7malWe_.js';
 
 /**
  * Public `Handoff` class — factory for handoff descriptors (Adoption
@@ -9,7 +9,8 @@ export { b as HandoffLoopError, c as HandoffNameCollisionError, d as HandoffPair
  *
  * Usage:
  *
- *   import { Agent, Handoff } from "@theokit/sdk";
+ *   import { Agent } from "@theokit/sdk";
+ *   import { Handoff } from "@theokit/sdk-handoff";
  *
  *   const billing = await Agent.create({
  *     name: "billing",
@@ -32,53 +33,167 @@ export { b as HandoffLoopError, c as HandoffNameCollisionError, d as HandoffPair
  * @public
  */
 
-/** Recommended system-prompt prefix for senders (D215 / EC-13). */
+/**
+ * Prose to prepend to a SENDING agent's `systemPrompt` so the model knows the `transfer_to_*` tools
+ * exist and what they mean.
+ *
+ * ```ts
+ * systemPrompt: `${RECOMMENDED_HANDOFF_PROMPT_PREFIX}\n\nYou triage support requests.`
+ * ```
+ *
+ * Nothing applies it for you — neither {@link Handoff.create} nor {@link Handoff.asPlugin} touches
+ * the system prompt, so omitting it is legal and usually shows up as a model that never transfers.
+ * Only the sender needs it; the receiver is unaware it was handed a conversation.
+ *
+ * @public
+ */
 declare const RECOMMENDED_HANDOFF_PROMPT_PREFIX: string;
+/**
+ * Peer-to-peer delegation: one agent hands the conversation to another and stops.
+ *
+ * A handoff is not a subagent call. A subagent runs, returns a result, and the parent continues; a
+ * handoff TRANSFERS the turn — the target answers the user directly and the source does not resume.
+ * Reach for it when the right responder is a different agent (billing, escalation, a specialist),
+ * and for a tool-shaped "go find this out and come back", use `agents` / `Tool.create` instead.
+ *
+ * Two entry points, and they are not interchangeable:
+ *
+ * - {@link Handoff.create} builds one descriptor, for `Agent.create({ handoffs: [...] })`. A bare
+ *   `SDKAgent` in that array is auto-wrapped, so call this explicitly only to customise (input
+ *   schema, filter, callback).
+ * - {@link Handoff.asPlugin} installs a whole set as a plugin, which is the SDK 2.x+ shape and what
+ *   the README's migration section points at.
+ *
+ * A namespace class: `new Handoff()` is a compile error, matching `Agent.create` / `Tool.create`.
+ */
 declare class Handoff {
     private constructor();
     /**
-     * Build a `HandoffDescriptor` for a target agent. Wrap with custom options
-     * (filter / inputType / callback / whitelist / etc); pass to
-     * `Agent.create({ handoffs: [...] })`.
+     * Describe one handoff target. Pass the result inside `Handoff.asPlugin({ targets })` or
+     * `Agent.create({ handoffs })`.
+     *
+     * ```ts
+     * Handoff.create(billing, { toolName: "escalate_billing" })
+     * ```
+     *
+     * A bare `SDKAgent` in either array is auto-wrapped with empty options, so call this explicitly
+     * only to customise — see {@link HandoffOptions}, and note that `tools` there is currently
+     * ignored.
      *
-     * Raw `SDKAgent` instances in `handoffs[]` are auto-wrapped by the runtime
-     * — call `Handoff.create()` explicitly only when you need to customize.
+     * The tool the model sees is named `transfer_to_<slug>`, where the slug comes from the target's
+     * `name` (falling back to its `agentId`, then to `"anonymous"`) with a leading `agent-` stripped,
+     * every run of characters OUTSIDE `[A-Za-z0-9_-]` folded to a single `_`, leading and trailing
+     * `_` trimmed, and a 64-char truncation. Hyphens and underscores are preserved, so `"billing EU"`
+     * and `"billing (EU)"` both become `billing_EU` while `"billing-EU"` stays distinct. Two targets
+     * whose names collapse to the same slug are NOT caught here — the collision is raised later, when
+     * the set is normalised.
+     *
+     * Throws `ConfigurationError` with `code: "handoff_target_required"` for a null/undefined target,
+     * and `code: "handoff_target_invalid"` for anything without a `send` method. It validates the
+     * target only, never the options.
      */
     static create<TInput extends ZodType = ZodType>(target: SDKAgent, options?: HandoffOptions<TInput>): HandoffDescriptor<TInput>;
     /**
-     * Plugin-based wiring (SDK 2.x preferred). Wraps `targets` in synthetic
-     * `transfer_to_<receiver>` tools and registers them via `ctx.registerTool`
-     * at agent init time.
+     * Expose each target as a `transfer_to_<receiver>` tool on the host agent — the SDK 2.x way to
+     * wire handoffs.
+     *
+     * ```ts
+     * const support = await Agent.create({
+     *   name: "support",
+     *   systemPrompt: `${RECOMMENDED_HANDOFF_PROMPT_PREFIX}\n\nYou answer support requests.`,
+     *   plugins: [Handoff.asPlugin({ parentAgentId: "support", targets: [billing] })],
+     * });
+     * ```
+     *
+     * Pass `parentAgentId` — it defaults to `"anonymous"`, and it is what self-reference detection
+     * and the chain trace compare against, so leaving it out weakens both. `maxHandoffDepth` defaults
+     * to 5.
+     *
+     * Four behaviours that are easy to be surprised by:
      *
-     * Replaces the legacy `Agent.create({ handoffs: [...] })` option (which is
-     * still supported as a transitional convenience while sdk-handoff is
-     * installed — the framework lazy-imports the tool-injector at runtime).
+     * - **`maxHandoffDepth: 0`, or an empty `targets`, registers NOTHING** and returns a plugin that
+     *   silently does nothing. There is no error and no warning; the model simply never sees a
+     *   transfer tool.
+     * - **Registration is awaited.** `register()` returns a promise that settles once the tools are
+     *   registered, and the plugin manager awaits it — so the transfer tools exist before the first
+     *   `send()`, and a failure in the lazy import or in target validation reaches the caller.
+     *   Before #355 it returned immediately and both of those were untrue.
+     * - **The receiver gets the user's LAST message, not the whole conversation.** The tool handler
+     *   forwards the supervisor's transcript, from which the dispatcher takes the most recent user
+     *   turn (#354 — before that it forwarded nothing, and the receiver was sent the literal string
+     *   `` `(Handoff from <sender> — no prior user message in history.)` ``). That placeholder is
+     *   still what a receiver gets when there genuinely is no prior user turn. Anything beyond the
+     *   last question has to be in the target's own system prompt, or you drive the handoff yourself
+     *   with {@link handoffTo}, which passes an explicit message through.
+     * - **The handoff tool never throws at the caller.** Every failure — loop detected, depth
+     *   exceeded, disposed receiver, `isEnabled` false, input that fails `inputType` — is caught
+     *   inside the tool handler and returned to the MODEL as
+     *   `{"ok":false,"error":"<ErrorName>","message":"…"}`. The exported error classes are real, but
+     *   in this wiring they never reach your `try`/`catch`; watch the tool results instead.
      *
-     * @example
-     *   const support = await Agent.create({
-     *     name: "support",
-     *     plugins: [Handoff.asPlugin({ parentAgentId: "support", targets: [billing] })],
-     *   });
+     * A self-referencing target and two targets resolving to the same tool name are both rejected —
+     * from `register`, which the plugin manager awaits — so `HandoffSelfReferenceError` and
+     * `HandoffNameCollisionError` reject the `Agent.create` you can `catch` around (#355; they used
+     * to arrive as an unhandled rejection instead, leaving an agent silently without handoff tools).
      */
     static asPlugin(opts: AsPluginOptions): Plugin;
 }
 /**
- * Options for `Handoff.asPlugin()`. `parentAgentId` defaults to `"anonymous"`;
- * pass the host agent's `name` for correct loop detection in chains.
+ * Options for {@link Handoff.asPlugin}.
+ *
+ * @public
  */
 interface AsPluginOptions {
+    /**
+     * Agents this one may transfer to. A bare `SDKAgent` is auto-wrapped; use
+     * {@link Handoff.create} for a customised entry.
+     *
+     * An EMPTY array registers no tools at all and produces a plugin that does nothing — silently.
+     */
     readonly targets: ReadonlyArray<SDKAgent | HandoffDescriptor>;
+    /**
+     * Identity of the HOST agent, as it will appear in the chain trace. Default `"anonymous"`.
+     *
+     * Self-reference detection compares `target.agentId` against this exact string, so a default
+     * `"anonymous"` means an agent listing itself among `targets` is NOT caught, and the pair-loop
+     * guard is the only thing left between you and a recursion.
+     */
     readonly parentAgentId?: string;
+    /**
+     * Maximum hops in one chain before `HandoffLoopError`. Default 5. `0` disables handoffs entirely
+     * rather than allowing zero hops.
+     *
+     * The counter is created FRESH for each tool invocation, so it bounds one dispatch, not the whole
+     * `send()` — cross-tool depth accumulation is not implemented. In practice a single invocation
+     * makes one hop, so this rarely fires; the pair guard (same sender → same receiver twice) is what
+     * actually catches ping-pong.
+     */
     readonly maxHandoffDepth?: number;
 }
 /**
- * Imperative escape hatch (D225). Useful for tests / programmatic flows
- * that need deterministic handoff without LLM routing.
+ * Hand `message` to `target` right now and return its reply text — no LLM routing, no tool call.
+ *
+ * ```ts
+ * const reply = await handoffTo(triage, billing, "refund for order 42");
+ * ```
+ *
+ * Use it for tests and for flows where YOU decide the destination. Unlike the plugin wiring, this
+ * passes the message through verbatim, so the receiver actually sees what the user said.
  *
- * NOTE: this is a STANDALONE helper rather than a method on `SDKAgent`
- * to avoid invasive refactor of the agent class. Behavior is identical
- * to invoking the corresponding synthetic tool would be.
+ * It THROWS, where the tool-based path swallows: a disposed receiver raises
+ * `HandoffReceiverDisposedError`, `isEnabled: false` raises a plain `Error`, and an
+ * `inputType` that rejects raises a plain `Error` wrapping the Zod message. Depth is fixed at 5 and
+ * the chain state is fresh per call, so `HandoffLoopError` is unreachable here and only the
+ * same-pair guard can fire — within a single call, never across calls.
+ *
+ * A receiver that does not finish cleanly does not throw either: you get the sentinel string
+ * `` `(Handoff target <id> returned status=<status>)` `` as the reply. Check for it if the
+ * distinction matters.
+ *
+ * Standalone rather than a method on `SDKAgent` so the agent class need not know about handoffs.
+ *
+ * @public
  */
 declare function handoffTo(sender: SDKAgent, target: SDKAgent, message: string, options?: HandoffOptions): Promise<string>;
 
-export { type AsPluginOptions, Handoff, HandoffDescriptor, HandoffOptions, RECOMMENDED_HANDOFF_PROMPT_PREFIX, handoffTo };
+export { type AsPluginOptions, Handoff, type HandoffDescriptor, type HandoffOptions, RECOMMENDED_HANDOFF_PROMPT_PREFIX, handoffTo };

package/dist/index.js

@@ -1,17 +1,33 @@
 import { createRequire } from 'module';
 import { z, toJSONSchema } from 'zod';
-import { Plugin } from '@theokit/sdk';
+import { ConfigurationError, Plugin } from '@theokit/sdk';
 
 var __defProp = Object.defineProperty;
 var __getOwnPropNames = Object.getOwnPropertyNames;
-var __esm = (fn, res) => function __init() {
-  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+var __esm = (fn, res, err) => function __init() {
+  if (err) throw err[0];
+  try {
+    return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+  } catch (e) {
+    throw err = [e], e;
+  }
 };
 var __export = (target, all) => {
   for (var name in all)
     __defProp(target, name, { get: all[name], enumerable: true });
 };
 
+// src/internal/slugify-agent-name.ts
+function slugifyAgentName(candidate) {
+  return candidate.slice(0, MAX_INPUT_LENGTH).replace(/^agent-/i, "").replace(/[^a-zA-Z0-9_-]+/g, "_").replace(/^_+|_+$/g, "").slice(0, 64) || "anonymous";
+}
+var MAX_INPUT_LENGTH;
+var init_slugify_agent_name = __esm({
+  "src/internal/slugify-agent-name.ts"() {
+    MAX_INPUT_LENGTH = 1024;
+  }
+});
+
 // src/types/handoff.ts
 var HandoffLoopError, HandoffPairLoopError, HandoffSelfReferenceError, HandoffReceiverDisposedError, HandoffNameCollisionError;
 var init_handoff = __esm({
@@ -225,11 +241,15 @@ function extractUserText(content) {
   const text = content.filter((c) => c?.type === "text").map((c) => c.text).join("\n");
   return text.length > 0 ? text : void 0;
 }
+function userTextOf(entry) {
+  const m = entry;
+  if (m?.type === "user" && m.message?.role === "user") return extractUserText(m.message.content);
+  if (m?.role === "user") return extractUserText(m.content);
+  return void 0;
+}
 function extractLastUserMessage(history, senderAgentId) {
   for (let i = history.messages.length - 1; i >= 0; i -= 1) {
-    const m = history.messages[i];
-    if (m?.type !== "user" || m.message?.role !== "user") continue;
-    const text = extractUserText(m.message.content);
+    const text = userTextOf(history.messages[i]);
     if (text !== void 0) return text;
   }
   return `(Handoff from ${senderAgentId} \u2014 no prior user message in history.)`;
@@ -262,7 +282,11 @@ async function dispatchHandoff(args) {
     toolName: descriptor.resolvedToolName
   });
   try {
-    const run = await receiver.send(lastUserMessage);
+    const toolAllowlist = descriptor.options.tools;
+    const run = await receiver.send(
+      lastUserMessage,
+      toolAllowlist !== void 0 ? { activeTools: [...toolAllowlist] } : {}
+    );
     const result = await run.wait();
     const reply = buildReply(result, receiver.agentId);
     return {
@@ -341,10 +365,7 @@ function autoWrap(agent) {
 }
 function resolveTargetName(agent) {
   const candidate = agent.name ?? agent.agentId ?? "anonymous";
-  return slugify(candidate);
-}
-function slugify(input) {
-  return input.replace(/^agent-/i, "").replace(/[^a-zA-Z0-9_-]+/g, "_").replace(/^_+|_+$/g, "").slice(0, 64) || "anonymous";
+  return slugifyAgentName(candidate);
 }
 function buildHandoffTool(parentAgentId, descriptor, maxHandoffDepth) {
   const description = descriptor.options.toolDescription ?? `Transfer the conversation to the ${descriptor.target.agentId} agent. Use this when the user's request matches their specialty.`;
@@ -356,7 +377,7 @@ function buildHandoffTool(parentAgentId, descriptor, maxHandoffDepth) {
     name: descriptor.resolvedToolName,
     description,
     inputSchema,
-    handler: async (input) => {
+    handler: async (input, ctx) => {
       const chainState = createChainState(parentAgentId, maxHandoffDepth);
       try {
         const { reply, result } = await dispatchHandoff({
@@ -364,8 +385,12 @@ function buildHandoffTool(parentAgentId, descriptor, maxHandoffDepth) {
           senderAgentId: parentAgentId,
           chainState,
           rawInputJson: input,
-          history: { messages: [] }
-          // v1: history replay deferred
+          // #354 — the supervisor's transcript, which the SDK hands every tool handler as
+          // `ctx.messages`. This used to be `{ messages: [] }` with the note "v1: history replay
+          // deferred", so the dispatcher found no user message and sent the receiver the
+          // placeholder instead of the question — and `inputFilter`, the documented redaction
+          // hook, was handed an empty transcript to redact.
+          history: { messages: ctx?.messages ?? [] }
         });
         return JSON.stringify({
           ok: true,
@@ -388,9 +413,13 @@ var init_tool_injector = __esm({
     init_handoff();
     init_dispatcher();
     init_registry();
+    init_slugify_agent_name();
     init_to_json_schema();
   }
 });
+
+// src/handoff.ts
+init_slugify_agent_name();
 var RECOMMENDED_HANDOFF_PROMPT_PREFIX = `
 You can transfer the conversation to other specialist agents when their
 expertise matches the user's request. Invoke the appropriate
@@ -401,19 +430,39 @@ var Handoff = class {
   constructor() {
   }
   /**
-   * Build a `HandoffDescriptor` for a target agent. Wrap with custom options
-   * (filter / inputType / callback / whitelist / etc); pass to
-   * `Agent.create({ handoffs: [...] })`.
+   * Describe one handoff target. Pass the result inside `Handoff.asPlugin({ targets })` or
+   * `Agent.create({ handoffs })`.
+   *
+   * ```ts
+   * Handoff.create(billing, { toolName: "escalate_billing" })
+   * ```
+   *
+   * A bare `SDKAgent` in either array is auto-wrapped with empty options, so call this explicitly
+   * only to customise — see {@link HandoffOptions}, and note that `tools` there is currently
+   * ignored.
+   *
+   * The tool the model sees is named `transfer_to_<slug>`, where the slug comes from the target's
+   * `name` (falling back to its `agentId`, then to `"anonymous"`) with a leading `agent-` stripped,
+   * every run of characters OUTSIDE `[A-Za-z0-9_-]` folded to a single `_`, leading and trailing
+   * `_` trimmed, and a 64-char truncation. Hyphens and underscores are preserved, so `"billing EU"`
+   * and `"billing (EU)"` both become `billing_EU` while `"billing-EU"` stays distinct. Two targets
+   * whose names collapse to the same slug are NOT caught here — the collision is raised later, when
+   * the set is normalised.
    *
-   * Raw `SDKAgent` instances in `handoffs[]` are auto-wrapped by the runtime
-   * — call `Handoff.create()` explicitly only when you need to customize.
+   * Throws `ConfigurationError` with `code: "handoff_target_required"` for a null/undefined target,
+   * and `code: "handoff_target_invalid"` for anything without a `send` method. It validates the
+   * target only, never the options.
    */
   static create(target, options = {}) {
     if (target === void 0 || target === null) {
-      throw new Error("Handoff.create: target agent is required");
+      throw new ConfigurationError("Handoff.create: target agent is required", {
+        code: "handoff_target_required"
+      });
     }
     if (typeof target.send !== "function") {
-      throw new Error("Handoff.create: target must be an SDKAgent instance");
+      throw new ConfigurationError("Handoff.create: target must be an SDKAgent instance", {
+        code: "handoff_target_invalid"
+      });
     }
     const resolvedToolName = options.toolName ?? `transfer_to_${slugifyName(target)}`;
     return {
@@ -423,19 +472,47 @@ var Handoff = class {
     };
   }
   /**
-   * Plugin-based wiring (SDK 2.x preferred). Wraps `targets` in synthetic
-   * `transfer_to_<receiver>` tools and registers them via `ctx.registerTool`
-   * at agent init time.
+   * Expose each target as a `transfer_to_<receiver>` tool on the host agent — the SDK 2.x way to
+   * wire handoffs.
+   *
+   * ```ts
+   * const support = await Agent.create({
+   *   name: "support",
+   *   systemPrompt: `${RECOMMENDED_HANDOFF_PROMPT_PREFIX}\n\nYou answer support requests.`,
+   *   plugins: [Handoff.asPlugin({ parentAgentId: "support", targets: [billing] })],
+   * });
+   * ```
+   *
+   * Pass `parentAgentId` — it defaults to `"anonymous"`, and it is what self-reference detection
+   * and the chain trace compare against, so leaving it out weakens both. `maxHandoffDepth` defaults
+   * to 5.
+   *
+   * Four behaviours that are easy to be surprised by:
    *
-   * Replaces the legacy `Agent.create({ handoffs: [...] })` option (which is
-   * still supported as a transitional convenience while sdk-handoff is
-   * installed — the framework lazy-imports the tool-injector at runtime).
+   * - **`maxHandoffDepth: 0`, or an empty `targets`, registers NOTHING** and returns a plugin that
+   *   silently does nothing. There is no error and no warning; the model simply never sees a
+   *   transfer tool.
+   * - **Registration is awaited.** `register()` returns a promise that settles once the tools are
+   *   registered, and the plugin manager awaits it — so the transfer tools exist before the first
+   *   `send()`, and a failure in the lazy import or in target validation reaches the caller.
+   *   Before #355 it returned immediately and both of those were untrue.
+   * - **The receiver gets the user's LAST message, not the whole conversation.** The tool handler
+   *   forwards the supervisor's transcript, from which the dispatcher takes the most recent user
+   *   turn (#354 — before that it forwarded nothing, and the receiver was sent the literal string
+   *   `` `(Handoff from <sender> — no prior user message in history.)` ``). That placeholder is
+   *   still what a receiver gets when there genuinely is no prior user turn. Anything beyond the
+   *   last question has to be in the target's own system prompt, or you drive the handoff yourself
+   *   with {@link handoffTo}, which passes an explicit message through.
+   * - **The handoff tool never throws at the caller.** Every failure — loop detected, depth
+   *   exceeded, disposed receiver, `isEnabled` false, input that fails `inputType` — is caught
+   *   inside the tool handler and returned to the MODEL as
+   *   `{"ok":false,"error":"<ErrorName>","message":"…"}`. The exported error classes are real, but
+   *   in this wiring they never reach your `try`/`catch`; watch the tool results instead.
    *
-   * @example
-   *   const support = await Agent.create({
-   *     name: "support",
-   *     plugins: [Handoff.asPlugin({ parentAgentId: "support", targets: [billing] })],
-   *   });
+   * A self-referencing target and two targets resolving to the same tool name are both rejected —
+   * from `register`, which the plugin manager awaits — so `HandoffSelfReferenceError` and
+   * `HandoffNameCollisionError` reject the `Agent.create` you can `catch` around (#355; they used
+   * to arrive as an unhandled rejection instead, leaving an agent silently without handoff tools).
    */
   static asPlugin(opts) {
     const parent = opts.parentAgentId ?? "anonymous";
@@ -445,22 +522,29 @@ var Handoff = class {
       name: `handoff-${parent}`,
       version: "1.0.0",
       kind: "general",
-      register(ctx) {
+      // #355 — `async`, and the promise is RETURNED. The plugin contract types `register` as
+      // `(ctx) => void | Promise<void>` and the manager awaits it, so returning it is all this
+      // needed. It used to run an unawaited async IIFE and return immediately: the tools appeared a
+      // module-load later, so whether they existed for the first `send()` depended on timing no
+      // caller controls, and `normalizeHandoffs`' validation errors became unhandled rejections
+      // that could not be caught around `Agent.create`.
+      //
+      // The import stays lazy — deferring it until `register` runs is what keeps the cold path
+      // lean, and that never required leaving it unawaited INSIDE `register`.
+      async register(ctx) {
         if (maxDepth === 0 || targets.length === 0) return;
-        void (async () => {
-          const { normalizeHandoffs: normalizeHandoffs2, buildHandoffTool: buildHandoffTool2 } = await Promise.resolve().then(() => (init_tool_injector(), tool_injector_exports));
-          const normalized = normalizeHandoffs2(parent, targets);
-          for (const { descriptor } of normalized) {
-            ctx.registerTool(buildHandoffTool2(parent, descriptor, maxDepth));
-          }
-        })();
+        const { normalizeHandoffs: normalizeHandoffs2, buildHandoffTool: buildHandoffTool2 } = await Promise.resolve().then(() => (init_tool_injector(), tool_injector_exports));
+        const normalized = normalizeHandoffs2(parent, targets);
+        for (const { descriptor } of normalized) {
+          ctx.registerTool(buildHandoffTool2(parent, descriptor, maxDepth));
+        }
       }
     });
   }
 };
 function slugifyName(agent) {
   const candidate = agent.name ?? agent.agentId ?? "anonymous";
-  return candidate.replace(/^agent-/i, "").replace(/[^a-zA-Z0-9_-]+/g, "_").replace(/^_+|_+$/g, "").slice(0, 64) || "anonymous";
+  return slugifyAgentName(candidate);
 }
 async function handoffTo(sender, target, message, options = {}) {
   const descriptor = Handoff.create(target, options);