Heartbeats and a run watchdog

README.md

@@ -59,6 +59,7 @@ needs to be useful:
 | 📋 **Task board** | A Trello-style Kanban board per workspace: create tickets, assign them to agents, drag them across *Backlog → Todo → In progress → In review → Done*. A ticket that lands in *Todo* starts its agent. Every ticket works in its own git worktree of the workspace's repository (a cloned URL or a local folder), on its own branch, so agents running side by side never get in each other's way. When the agent on a **Coding** ticket is done, Godmode pushes its branch and opens a pull request — merged pull requests close their ticket. Reply to a ticket to ask for changes; the pull request updates. Archive finished tickets (one by one, or a whole *Done* column) to keep the board focused — they keep everything and come back in one click. Tickets carry a **priority**, a **due date** and **labels** (filter by them; urgent ones start first and the agent is told the deadline). Every ticket keeps a **timeline** — each delivery with its full result, your feedback, the agent's notes, blocks and pull requests — and shows what it **cost** and how long the agent worked. Review in one click (**Approve** / **Request changes** / **Reopen**); A run that fails on a hiccup (an API error, a crash, a restart) is **tried again on its own** — twice at most — so work gets done without you; *Blocked* is left for what needs you, and says why — the agent needs something, the run kept failing — and offers the matching way on. A ticket whose agent set itself a follow-up shows *Waiting — continues tomorrow at 10:00* instead of landing in review. A ticket can **wait for** others (*Reconcile October* waits for *#3 Get the bank statements*): it stays in Todo, says so, and starts by itself once they are delivered — with their results in its brief. Big tickets split into **parts**: a lead hands them to its team (or you add them), its ticket waits until they're delivered and then continues with their results — approve the whole and its parts are done too. |
 | ⚡ **Automations** | Work starts when it happens: on a schedule (“weekdays at 08:00”), when something happens in a connected app (a new email, a Slack message, a calendar event, a Notion update), when a condition you describe comes true (“competitor pricing changes”), or when a webhook is called. Describe it in one sentence and Godmode sets it up. |
 | ⏰ **Follow-ups** | When a task needs waiting — a reply to an email, a delivery, a build, office hours — the agent sets itself a time and picks the chat up again on its own, with all the context, like a coworker who says “I'll check back tomorrow at 10”. You see when it comes back and can continue now, move it or cancel it. |
+| **Heartbeats & watchdog** | Agents with real work keep moving between your prompts. Switch on an agent's **heartbeat** (every 15 minutes to once a day, optionally only in working hours on weekdays) and it wakes on its own rhythm: it picks up its tickets that stand still, tries a failed one again, starts one that never did — each in the ticket's own chat, told exactly what changed since it last worked on it — and runs its standing **checklist**. Every beat ends with a trail on the agent's *Heartbeat* tab; a beat with nothing to do is quiet and costs nothing. The **watchdog** looks at every run while it works: a run that shows no sign of life, or calls the same tool with the same result over and over, is stopped with a precise report of where it stood. Board tickets try again once on their own; anything else is escalated to you. |
 | 🤝 **Delegation** | Agents hand tasks to peer agents or spawn short-lived subagents. The Godmode agent can list, check, create and configure all agents. Follow a handoff both ways: open the teammate's chat from the handoff card, and jump back to the chat that asked from there. |
 | 🧰 **Claude Code & MCP** | Build your team from the terminal: connect Claude Code — or Cursor, Codex, Claude Desktop, any app that speaks MCP — and say *“Create a Godmode agent that…”*. One click in **Settings → Claude Code & MCP** adds Godmode to Claude Code for every project. The app gets Godmode's own management tools (agents, automations, tasks, runs, VMs) through a key of its own, with full or look-only access; scripts reach the same tools with `godmode call`. Passwords, 2FA codes and settings stay out of its reach. |
 | 🏢 **Team & org chart** | Give every agent a role (*Bookkeeper*, *Research analyst*) and a lead; agents are told who does what and who they report to, so work goes to the right one. The **org chart** shows the whole team with what each is doing right now. Status tells the truth — *Working in 2 chats*, *Queued*, *Needs your answer*, *Last run failed* (one click opens the run) — and each agent's page shows **what's on its plate**: live work, waiting chats, board tickets, follow-ups and its next automations. Duplicate an agent in one click; messages from automations, the board or other agents are labelled as such, never shown as yours. Or start a **whole team** in one click — *Back office*, *Marketing*, *Sales*, *Product & QA*: a lead and its reports, wired up, with their schedules. |
@@ -161,6 +162,10 @@ needs to be useful:
     <td><img src="docs/screenshots/followup-waiting.png" alt="An agent waiting for a signed contract, set to continue tomorrow at 09:00" /><br /><sub><b>Follow-ups</b> — the agent sets itself a time to continue; move it, cancel it or continue now</sub></td>
     <td><img src="docs/screenshots/followup.png" alt="The agent picked the chat up again at the time it set and finished the task" /><br /><sub><b>Back on it</b> — at that time the agent picks the chat up again, with all the context</sub></td>
   </tr>
+  <tr>
+    <td><img src="docs/screenshots/heartbeat.png" alt="An agent's Heartbeat tab: on, beating every hour in working hours, with its board" /><br /><sub><b>Heartbeat</b> — the agent wakes on its own rhythm to move its tickets forward and run its checklist</sub></td>
+    <td><img src="docs/screenshots/heartbeat-beats.png" alt="Recent beats and the watchdog's interventions with their reports" /><br /><sub><b>Beats & watchdog</b> — every beat leaves a trail; stalled or looping runs are stopped with a precise report</sub></td>
+  </tr>
   <tr>
     <td><img src="docs/screenshots/pause.png" alt="A paused chat with a Continue button above the message box" /><br /><sub><b>Pause & continue</b> — the agent finishes its step, stands still and picks the work up where it stopped</sub></td>
     <td><img src="docs/screenshots/pause-limit.png" alt="A chat waiting for Claude's session limit to reset, set to continue by itself" /><br /><sub><b>Usage limit</b> — the work waits for the reset and continues by itself, with what you wrote meanwhile</sub></td>
@@ -507,6 +512,7 @@ See [CONTRIBUTING.md](CONTRIBUTING.md).
 - [x] Working folders, Claude Code slash commands, live browser preview in chat, desktop auto-update
 - [x] Computer use: share a window (background control via Cua Driver), a display, every monitor or a browser tab
 - [x] Automations: schedules, app events (Composio triggers), plain-language conditions and webhooks
+- [x] Heartbeats: agents wake on their own rhythm to move their tickets forward and run a checklist; a watchdog stops stalled and looping runs
 - [x] Agents working in a dedicated macOS VM (Tart / Virtualization.framework): shell, files and screen, assigned per agent, chat or workspace
 - [x] API tools: any API with a key (Nano Banana, OpenAI, ElevenLabs…) for agents, global / workspace / agent
 - [x] SSH servers: agents run commands, edit files and copy files on remote machines — password or key, sudo, pinned host keys

apps/desktop/src/components/agents/detail/chats-tab.tsx

@@ -2,7 +2,7 @@ import { useEffect, useState, type ComponentType } from "react";
 import { Link } from "react-router";
 import { motion } from "motion/react";
 import { formatDistanceToNowStrict } from "date-fns";
-import { Archive, ArchiveRestore, MessageSquare, MessagesSquare, Moon, Pin, Plug, Plus, Search, Share2, SquareKanban, Trash2, Workflow } from "lucide-react";
+import { Archive, ArchiveRestore, HeartPulse, MessageSquare, MessagesSquare, Moon, Pin, Plug, Plus, Search, Share2, SquareKanban, Trash2, Workflow } from "lucide-react";
 import type { Agent, Conversation, ConversationOrigin } from "@godmode/shared";
 import { errorMessage } from "@/lib/api";
 import { useArchivedConversations, useConversations } from "@/lib/hooks";
@@ -31,6 +31,7 @@ const ORIGIN: Record<ConversationOrigin, { icon: ComponentType<{ className?: str
   telegram: { icon: PLATFORMS.telegram.glyph, label: "Telegram" },
   teams: { icon: PLATFORMS.teams.glyph, label: "Microsoft Teams" },
   task: { icon: SquareKanban, label: "Task" },
+  heartbeat: { icon: HeartPulse, label: "Heartbeat" },
 };
 
 function useDebounced<T>(value: T, ms = 250): T {