fix(gastown): dispatch circuit breaker — per-bead attempt cap, exponential backoff, town-level breaker (#1653)

cloudflare-gastown/src/db/tables/beads.table.ts

@@ -50,6 +50,8 @@ export const BeadRecord = z.object({
       }
     })
     .pipe(z.record(z.string(), z.any())), // z.any() needed for Rpc.Serializable compatibility
+  dispatch_attempts: z.number().int().default(0),
+  last_dispatch_attempt_at: z.string().nullable().default(null),
   created_by: z.string().nullable(),
   created_at: z.string(),
   updated_at: z.string(),
@@ -129,10 +131,20 @@ export function createTableBeads(): string {
     created_by: `text`,
     created_at: `text not null`,
     updated_at: `text not null`,
+    dispatch_attempts: `integer not null default 0`,
+    last_dispatch_attempt_at: `text`,
     closed_at: `text`,
   });
 }
 
+/** Idempotent ALTER statements for existing databases. */
+export function migrateBeads(): string[] {
+  return [
+    `ALTER TABLE beads ADD COLUMN dispatch_attempts integer not null default 0`,
+    `ALTER TABLE beads ADD COLUMN last_dispatch_attempt_at text`,
+  ];
+}
+
 export function getIndexesBeads(): string[] {
   return [
     `CREATE INDEX IF NOT EXISTS idx_beads_type_status ON ${beads}(${beads.columns.type}, ${beads.columns.status})`,

cloudflare-gastown/src/dos/Town.do.ts

@@ -1555,6 +1555,25 @@ export class TownDO extends DurableObject<Env> {
       switch (action) {
         case 'RESTART':
         case 'RESTART_WITH_BACKOFF': {
+          // Fix 4 (#1653): if the hooked bead has exhausted its dispatch
+          // attempts, fail it instead of restarting — prevents infinite loops.
+          const restartBeadId = snapshotHookedBeadId ?? targetAgent?.current_hook_bead_id;
+          if (restartBeadId) {
+            const restartBead = beadOps.getBead(this.sql, restartBeadId);
+            if (restartBead && restartBead.dispatch_attempts >= scheduling.MAX_DISPATCH_ATTEMPTS) {
+              beadOps.updateBeadStatus(this.sql, restartBeadId, 'failed', input.agent_id);
+              if (targetAgent?.current_hook_bead_id === restartBeadId) {
+                if (targetAgent.status === 'working' || targetAgent.status === 'stalled') {
+                  dispatch
+                    .stopAgentInContainer(this.env, this.townId, targetAgentId)
+                    .catch(() => {});
+                }
+                agents.unhookBead(this.sql, targetAgentId);
+              }
+              break;
+            }
+          }
+
           // Stop the agent in the container, reset to idle so the
           // scheduler picks it up again on the next alarm cycle.
           if (targetAgent?.status === 'working' || targetAgent?.status === 'stalled') {

cloudflare-gastown/src/dos/town/actions.ts

@@ -517,6 +517,19 @@ export function applyAction(ctx: ApplyActionContext, action: Action): (() => Pro
         `,
         [agentId]
       );
+      // Track dispatch attempts on the bead itself (not just the agent).
+      // The bead counter is never reset by hookBead, preventing the
+      // infinite retry loop (#1653).
+      query(
+        sql,
+        /* sql */ `
+          UPDATE ${beads}
+          SET ${beads.columns.dispatch_attempts} = ${beads.columns.dispatch_attempts} + 1,
+              ${beads.columns.last_dispatch_attempt_at} = ?
+          WHERE ${beads.bead_id} = ?
+        `,
+        [now(), beadId]
+      );
       beadOps.updateBeadStatus(sql, beadId, 'in_progress', agentId);
 
       const capturedAgentId = agentId;

cloudflare-gastown/src/dos/town/agents.ts

@@ -282,13 +282,16 @@ export function hookBead(sql: SqlStorage, agentId: string, beadId: string): void
     unhookBead(sql, stale.bead_id);
   }
 
+  // Do NOT reset dispatch_attempts here — per-bead dispatch tracking
+  // lives on the beads table now (beads.dispatch_attempts). Resetting
+  // the agent counter on every hook was the root cause of the infinite
+  // retry loop (#1653).
   query(
     sql,
     /* sql */ `
       UPDATE ${agent_metadata}
       SET ${agent_metadata.columns.current_hook_bead_id} = ?,
           ${agent_metadata.columns.status} = 'idle',
-          ${agent_metadata.columns.dispatch_attempts} = 0,
           ${agent_metadata.columns.last_activity_at} = ?,
           ${agent_metadata.columns.agent_status_message} = NULL,
           ${agent_metadata.columns.agent_status_updated_at} = NULL

cloudflare-gastown/src/dos/town/beads.ts

@@ -4,7 +4,13 @@
  */
 
 import { z } from 'zod';
-import { beads, BeadRecord, createTableBeads, getIndexesBeads } from '../../db/tables/beads.table';
+import {
+  beads,
+  BeadRecord,
+  createTableBeads,
+  getIndexesBeads,
+  migrateBeads,
+} from '../../db/tables/beads.table';
 import {
   bead_events,
   BeadEventRecord,
@@ -65,7 +71,7 @@ export function initBeadTables(sql: SqlStorage): void {
   dropCheckConstraints(sql);
 
   // Migrations: add columns to existing tables (idempotent)
-  for (const stmt of [...migrateConvoyMetadata(), ...migrateAgentMetadata()]) {
+  for (const stmt of [...migrateBeads(), ...migrateConvoyMetadata(), ...migrateAgentMetadata()]) {
     try {
       query(sql, stmt, []);
     } catch {