forge-heartbeat

Detects a dispatched agent gone silent mid-run from events.jsonl. Use during multi-agent runs or when a run feels quiet — silent agent, stalled run, heartbeat check, timeout.

SKILL.md

GitHub

forge-heartbeat — stall/silence watchdog

Flags a dispatched Boss that went quiet mid-work — crashed, stuck in a loop, or waiting on something nobody is watching. Read-only: it never touches a live agent, it only reads the real events already logged.

When to use

  • During a multi-agent /forge run — periodically, or whenever the run "feels quiet" (no new dashboard activity for a while).
  • When the owner asks "is anything stuck?" or "did that agent die?"
  • As a future forge-doctor check — same read-only, evidence-only pattern as forge-doctor/forge-verify; can be wired in later without changing this tool's contract.

How to run

node .claude/forge-bin/forge-heartbeat.cjs check <run_id>
  # print stalled agents (or "no stalled agents"), exit 0 clean / 1 stalled

node .claude/forge-bin/forge-heartbeat.cjs check <run_id> --window 15   # override the silence window (minutes)
node .claude/forge-bin/forge-heartbeat.cjs check <run_id> --root <projectRoot>
node .claude/forge-bin/forge-heartbeat.cjs check <run_id> --json        # full machine-readable result

What STALLED means

  • An agent is tracked once it has logged agent_started or subagent_started.
  • An agent is finished — and never flagged, no matter how old its last event is — once it has logged agent_completed, subagent_completed, agent_failed, or subagent_failed anywhere in its history. Silence after a real completion/failure is expected, not a stall.
  • An unfinished, tracked agent is STALLED when it has been silent for longer than the window (default 10 minutes) since its last real event.
  • Within stalled agents, NEVER PROGRESSED is a stronger flag: the agent's start event is also its only event — it never logged a single progress/output event before going quiet.
  • Events with no resolvable timestamp are never treated as "just happened" or invented a fake time for — they're tracked but excluded from the last-seen calculation, so a stall is never hidden or invented by a missing timestamp.

Honesty

  • Reads only real events from <run>/events.jsonl — it never pings a live agent process and never guesses at agent state.
  • Silence is a flag to CHECK the agent, not to auto-kill it. This tool never cancels, restarts, or reassigns work — it surfaces a signal for the Lead (or the owner) to look at. A stalled agent might simply be doing something slow and legitimate; treat the flag as "go look", not "it failed".
  • A malformed line in events.jsonl is skipped, never crashes the check, and is reported in the malformed count.