Feat/agent stop start

.github/workflows/publish-cli.yaml

@@ -88,6 +88,18 @@ jobs:
         working-directory: cli
         run: bun run build
 
+      - name: Verify built binary reports the package version
+        if: steps.gate.outputs.publish == 'true'
+        working-directory: cli
+        run: |
+          EXPECTED="${{ steps.versions.outputs.package }}"
+          ACTUAL=$(node dist/ranch.mjs --version)
+          if [ "$ACTUAL" != "$EXPECTED" ]; then
+            echo "::error::Built binary reports v$ACTUAL but package.json is v$EXPECTED — refusing to publish a stale dist."
+            exit 1
+          fi
+          echo "Built binary reports v$ACTUAL ✓"
+
       - name: Pack (dry run, sanity check)
         if: steps.gate.outputs.publish == 'true'
         working-directory: cli

admin/slices/setup/api/data/repositories/api/sdk.gen.ts

@@ -80,6 +80,8 @@ import type {
   AgentControllerDemoteAdminData,
   AgentControllerPromoteAdminData,
   AgentControllerRestartData,
+  AgentControllerStopData,
+  AgentControllerStartData,
   RestartByTemplateData,
   FileControllerListData,
   FileControllerReadData,
@@ -98,6 +100,7 @@ import type {
   ResetBridleTranscriptResponse,
   GetBridleTranscriptData,
   GetBridleTranscriptResponse,
+  ArchiveBridleTranscriptData,
   SkillControllerFindAllData,
   SkillControllerCreateData,
   SkillControllerListSourcesData,
@@ -1433,6 +1436,38 @@ export class AgentsService {
     });
   }
 
+  /**
+   * Stop an agent without deleting it: cancels its workflow and deletes its pod to free cluster CPU/memory, then marks it `stopped`. Use this to free a slot so another agent can start. Bring it back with POST :id/start. Admin or Owner.
+   */
+  public static agentControllerStop<ThrowOnError extends boolean = false>(
+    options: Options<AgentControllerStopData, ThrowOnError>,
+  ) {
+    return (options.client ?? _heyApiClient).post<
+      unknown,
+      unknown,
+      ThrowOnError
+    >({
+      url: "/agents/{id}/stop",
+      ...options,
+    });
+  }
+
+  /**
+   * Start a stopped agent: deploys a fresh pod and reattaches the runtime. Inverse of POST :id/stop. Admin or Owner.
+   */
+  public static agentControllerStart<ThrowOnError extends boolean = false>(
+    options: Options<AgentControllerStartData, ThrowOnError>,
+  ) {
+    return (options.client ?? _heyApiClient).post<
+      unknown,
+      unknown,
+      ThrowOnError
+    >({
+      url: "/agents/{id}/start",
+      ...options,
+    });
+  }
+
   /**
    * Restart every agent that uses this template. Pulls latest template-owned files into each agent and redeploys, preserving runtime state. Concurrency capped at 5 to avoid overwhelming the cluster. Admin or Owner.
    */
@@ -1692,6 +1727,22 @@ export class BridleService {
       ...options,
     });
   }
+
+  /**
+   * Archive the persisted chat transcript for an agent/channel — the live JSONL is moved to a timestamped sibling (`bridle:<channel>.<iso-ts>.archived.jsonl`) and the live slot starts empty. Used by the embed's "New chat" action when the visitor wants a clean slate but we still want the prior conversation for admin/audit. No-op (returns `{}`) when there's nothing to archive.
+   */
+  public static archiveBridleTranscript<ThrowOnError extends boolean = false>(
+    options: Options<ArchiveBridleTranscriptData, ThrowOnError>,
+  ) {
+    return (options.client ?? _heyApiClient).post<
+      unknown,
+      unknown,
+      ThrowOnError
+    >({
+      url: "/api/agent/{agentId}/transcript/archive",
+      ...options,
+    });
+  }
 }
 
 export class SkillsService {

admin/slices/setup/api/data/repositories/api/types.gen.ts

@@ -1965,6 +1965,32 @@ export type AgentControllerRestartResponses = {
   201: unknown;
 };
 
+export type AgentControllerStopData = {
+  body?: never;
+  path: {
+    id: string;
+  };
+  query?: never;
+  url: "/agents/{id}/stop";
+};
+
+export type AgentControllerStopResponses = {
+  201: unknown;
+};
+
+export type AgentControllerStartData = {
+  body?: never;
+  path: {
+    id: string;
+  };
+  query?: never;
+  url: "/agents/{id}/start";
+};
+
+export type AgentControllerStartResponses = {
+  201: unknown;
+};
+
 export type RestartByTemplateData = {
   body?: never;
   path: {
@@ -2175,6 +2201,24 @@ export type GetBridleTranscriptResponses = {
 export type GetBridleTranscriptResponse =
   GetBridleTranscriptResponses[keyof GetBridleTranscriptResponses];
 
+export type ArchiveBridleTranscriptData = {
+  body?: never;
+  path: {
+    agentId: string;
+  };
+  query?: {
+    /**
+     * Session channel — defaults to "admin".
+     */
+    channel?: string;
+  };
+  url: "/api/agent/{agentId}/transcript/archive";
+};
+
+export type ArchiveBridleTranscriptResponses = {
+  200: unknown;
+};
+
 export type SkillControllerFindAllData = {
   body?: never;
   path?: never;

api/src/slices/bridle/bridle.controller.ts

@@ -80,7 +80,7 @@ export class BridleController {
 
     return new Promise((resolve) => {
       const timeout = setTimeout(() => {
-        this.hub.unregisterClient(clientId);
+        this.hub.unregisterClient(clientId, agentId);
         resolve({
           text: chunks.join('') || 'Timeout: no response from agent',
           messageId: '',
@@ -95,7 +95,7 @@ export class BridleController {
           const event = data as Record<string, unknown>;
           if (event.type === 'message' || event.type === 'stream_end') {
             clearTimeout(timeout);
-            this.hub.unregisterClient(clientId);
+            this.hub.unregisterClient(clientId, agentId);
             resolve({
               text: event.text ?? chunks.join(''),
               messageId: event.messageId,

api/src/slices/bridle/data/bridle.gateway.ts

@@ -36,9 +36,18 @@ export class BridleGateway extends IBridleGateway {
   /** Agent connections: agentId → send function */
   private agents = new Map<string, (data: unknown) => void>();
 
-  /** Browser clients: clientId → { agentId, send } */
+  /**
+   * Browser clients keyed by `${clientId}\u0000${agentId}`. Keying by the pair
+   * (not clientId alone) lets ONE user hold several concurrent conversations —
+   * e.g. a multi-slot dashboard chatting with N agents on N sockets — without
+   * later sockets overwriting earlier ones (they share clientId='admin'/sub).
+   */
   private clients = new Map<string, IBridleClientData>();
 
+  private clientKey(clientId: string, agentId: string): string {
+    return `${clientId}\u0000${agentId}`;
+  }
+
   /** Pending sync requests awaiting agent ack: requestId → pending */
   private pendingSyncs = new Map<string, IPendingSync>();
 
@@ -97,7 +106,8 @@ export class BridleGateway extends IBridleGateway {
     isAdmin: boolean,
     prompt?: string,
   ): void {
-    this.clients.set(clientId, {
+    this.clients.set(this.clientKey(clientId, agentId), {
+      clientId,
       agentId,
       send,
       isAdmin,
@@ -108,10 +118,10 @@ export class BridleGateway extends IBridleGateway {
     );
   }
 
-  unregisterClient(clientId: string): void {
-    this.clients.delete(clientId);
+  unregisterClient(clientId: string, agentId: string): void {
+    this.clients.delete(this.clientKey(clientId, agentId));
     this.logger.log(
-      `Browser client unregistered: ${clientId} (total: ${this.clients.size})`,
+      `Browser client unregistered: ${clientId} agentId=${agentId} (total: ${this.clients.size})`,
     );
   }
 
@@ -126,7 +136,7 @@ export class BridleGateway extends IBridleGateway {
       this.logger.warn(
         `Cannot send to agent — not connected (agentId=${agentId})`,
       );
-      this.sendToClient(clientId, {
+      this.sendToClient(clientId, agentId, {
         type: 'message',
         text: 'Agent is not connected. Please try again later.',
         parts: [
@@ -141,7 +151,7 @@ export class BridleGateway extends IBridleGateway {
       return;
     }
 
-    const client = this.clients.get(clientId);
+    const client = this.clients.get(this.clientKey(clientId, agentId));
     agentSend({
       type: 'message',
       clientId,
@@ -152,8 +162,8 @@ export class BridleGateway extends IBridleGateway {
     });
   }
 
-  sendToClient(clientId: string, data: unknown): void {
-    const client = this.clients.get(clientId);
+  sendToClient(clientId: string, agentId: string, data: unknown): void {
+    const client = this.clients.get(this.clientKey(clientId, agentId));
     if (client) {
       client.send(data);
     }
@@ -163,8 +173,8 @@ export class BridleGateway extends IBridleGateway {
     const clientId = data.clientId;
     if (!clientId) return;
 
-    const client = this.clients.get(clientId);
-    if (client && client.agentId === agentId) {
+    const client = this.clients.get(this.clientKey(clientId, agentId));
+    if (client) {
       client.send(data);
     }
   }

api/src/slices/bridle/domain/bridle.gateway.ts

@@ -30,8 +30,8 @@ export abstract class IBridleGateway {
     text: string,
     parts: BridlePart[],
   ): void;
-  /** Send an event to a specific browser client */
-  abstract sendToClient(clientId: string, data: unknown): void;
+  /** Send an event to a specific browser client (scoped to clientId + agentId) */
+  abstract sendToClient(clientId: string, agentId: string, data: unknown): void;
   /** Register a browser client for a specific agent */
   abstract registerClient(
     clientId: string,
@@ -42,8 +42,8 @@ export abstract class IBridleGateway {
      * agent on every message in this session. */
     prompt?: string,
   ): void;
-  /** Unregister a browser client */
-  abstract unregisterClient(clientId: string): void;
+  /** Unregister a browser client (scoped to clientId + agentId) */
+  abstract unregisterClient(clientId: string, agentId: string): void;
   /** Register an agent connection for a specific agent */
   abstract registerAgent(agentId: string, send: (data: unknown) => void): void;
   /** Unregister an agent connection for a specific agent */

api/src/slices/bridle/domain/bridle.types.ts

@@ -113,6 +113,7 @@ export interface IBridleAgentHealthData {
 
 /** Registered client metadata */
 export interface IBridleClientData {
+  clientId: string;
   agentId: string;
   send: (data: unknown) => void;
   isAdmin: boolean;

api/src/slices/bridle/handlers/bridleClientWs.handler.ts

@@ -192,9 +192,12 @@ export class BridleClientWsHandler
 
   handleDisconnect(client: Socket) {
     const clientId = client.data?.clientId as string | undefined;
-    if (clientId) {
-      this.hub.unregisterClient(clientId);
-      this.logger.log(`Browser disconnected: clientId=${clientId}`);
+    const agentId = client.data?.agentId as string | undefined;
+    if (clientId && agentId) {
+      this.hub.unregisterClient(clientId, agentId);
+      this.logger.log(
+        `Browser disconnected: clientId=${clientId} agentId=${agentId}`,
+      );
     }
   }
 

api/src/slices/workflow/data/agent-workflow.manifest.ts

@@ -220,8 +220,16 @@ function buildAgentPod(
           // running browser_play / Chromium). Guaranteed (requests == limits
           // at 2 CPU / 2Gi) made schedules fail on small Hetzner nodes —
           // even one agent didn't fit alongside the platform pods.
+          //
+          // CPU request is 100m, not 500m: idle agents actually use ~10-20m,
+          // so 500m × ~14 agents reserved 100% of an 8-vCPU cx43 node while it
+          // ran at 8% — new agents stuck Pending on "Insufficient cpu" despite
+          // a near-idle node. The scheduler packs by requests, not usage, so
+          // an honest low floor is what lets agents fit. Bursting is unaffected
+          // (limits still come from i.cpu). Memory floor stays 512Mi — that's
+          // a real idle footprint and becomes the next ceiling (~28/node).
           resources: {
-            requests: { cpu: '500m', memory: '512Mi' },
+            requests: { cpu: '100m', memory: '512Mi' },
             limits: { cpu: i.cpu, memory: i.memory },
           },
           ports: [{ containerPort: 3000 }],