Handsfree · fleet control · Herdr ≥ 0.7.1

Hark

When your agents need a word.

Handsfree supervision for the whole herd. When agents in Herdr block, swarm, or wait on you, Hark speaks the ask and takes your answer by voice—so the fleet keeps moving while you’re away from the keyboard.

Handsfree Mode A Herd · fleet · multi-session R2/R3 always confirm

Handsfree · herd topology

You speak to the Hark handsfree agent; it uses cloud STT/TTS and talks to coding agents inside Herdr. You! operator · mic voice tts Hark handsfree · Mode A outside Herdr Services cloud STT cloud TTS xAI · … api watch · answer Herdr sessions · panes · agents herdr.dev ↗ Agent blocked? Agent working Agent done? local machine · multi-session OK
You ↔ Hark ↔ services · Hark → agents in Herdr
Hark, the herald agents sing: “Human input, please, we bring.” Blocked in Herdr, questions rise; Hark relays your voice replies.
<2s
Blocked → TTS start
0
Stale double-sends
N
Sessions · one feed
R0–R3
Risk-aware confirm

Handsfree · herd control

One voice. Many agents. No keyboard required.

A swarm in Herdr will pause on lockfiles, menus, and permissions. You can’t babysit every pane. Hark is the handsfree control plane: listen when the fleet needs you, speak once, deliver to the right target.

01

Fleet awareness

One watch feed across local and remote sessions. Blocked, done, and ambient prompts—tagged by pane and risk.

02

Handsfree dialogue

TTS asks, STT answers, radio partials with HOLD. Mic mutes while speaking; record starts when you talk.

03

Safe herd routing

Fingerprint + revision refuse stale panes. R2/R3 always confirm. The LLM never invents the delivery target.


Usage

Run the skill. The agent sets up the rest.

You don’t need to wire Mode A by hand. Point a capable coding agent at this repo, invoke the hark (or handsfree) skill, and let it arm watch, speech, and safety. You stay on voice.

  1. Install / open the repo Clone or point the agent at ultradyn/hark.
  2. Run the skill /hark or /handsfree (or load skill/hark/SKILL.md).
  3. Agent bootstraps Doctor, Herdr sessions, watch Monitor, ambient wake, TTS status of any already-blocked panes.
  4. You supervise by voice When the herd blocks, it asks; you answer; it delivers. Multi-session queue, one target at a time.
you · one shot
# In Claude / Grok / Pi / OpenCode (with Monitor)…
/hark
# or: use the hark skill · handsfree alias
agent: running hark doctor…
agent: arming watch --for-monitor…
agent: ambient listening · say hey hark
agent.blocked work/w3:p2 R2
▸ speaks the question · listens · answers
delivered · fleet continues

Orchestrators

Needs a Monitor tool. Works with the herd you already run.

Mode A expects a coding CLI/agent that can run shell tools and keep a long-lived Monitor (or equivalent) on hark watch --for-monitor. Without that, blocks won’t interrupt the session.

CLI / agent Monitor Notes
Claude Code Native Skill + Monitor on watch feed; full Mode A loop
Grok Build / Grok CLI Native Same pattern; skill install + persistent monitor
Pi Plugin Requires a Monitor-capable plugin / hook path
OpenCode Plugin Requires a Monitor-capable plugin / extension
Codex / others If Monitor exists Any agent with shell + long-lived watch can drive Hark

Herdr hosts the worker agents. Your orchestrator (outside Herdr) runs the skill and speaks for the fleet.


The loop

From blocked to unblocked without a screen.

Mode A keeps the orchestrator outside Herdr: local mic, speakers, and a skill that never invents your answers.

Agent blocked

Monitor line with event_id, pane, risk.

Speak the ask

hark tts --listen or hark ask—mic mutes while Eve talks.

You reply

Speech-gated capture; start cue only when you speak.

Bound deliver

hark answer <event_id> with stale-safe checks.

hark · mode a
$ hark doctor
overall: OK · ambient armed · xAI OAuth
$ hark watch --for-monitor
agent.blocked default/w7:p1 risk=R2
$ hark ask --confirm always "Approve mesh peer?"
▸ mic muted · Eve speaking
▸ speech open · record cue
transcript: "option one, chris-ws"
$ hark answer 19f5… --keys 1 enter
delivered · pane working

Safety first

The verse is playful. Routing is not.

Pane text is untrusted. Humans stay in the loop. Permissions and destructive actions always require confirmation.

Risk Meaning Confirm
R0 Informational / trivial Never forced
R1 Ordinary free-text / soft choices When unsure
R2 Permissions, network, edits Always
R3 Destructive / secrets / prod Always

Install

One-line install. Then run /hark.

The installer puts the hark CLI on your PATH and copies skills to ~/.claude/skills. Prefer letting the agent run doctor + Mode A after /hark. Ambient wake: say hey hark, hey herald, or your custom trigger.

Install method
skills · npx
$ npx skills add ultradyn/hark -g -y
# agent skills only · still need hark CLI (bash install or uv) for Mode A
# then in agent: /hark

Handsfree when the herd calls

Blocked agents call.
Hark answers.

Fleet control by voice. Human-in-the-loop, without hands on the keyboard.