summon-human

A skill that lets an agent ask frond-alarm to phone figs, the Thornfield frond's human, and then read what happened. frond-alarm turns one GitHub issue into at most one phone call, placed off the fleet by its own GitHub Actions worker. This plugin files that issue in the alarm's own format and reads back the outcome the worker records. It never places a call itself and holds no provider keys.

agent -> summon.py submit -> [call] issue + wake -> frond-alarm worker -> phone call to figs
agent <- summon.py status <- the worker's labels and comments (answered, not-answered, ..., figs's answers)

figs has said that any prince or scribe-prince may call figs, or page figs to come, at any hour, for any reason, without second-guessing (thornfield-plugins#16). So the skill starts with one command for "I want figs here now", and sends a question that can wait to the light path: an @-mention of figs in the room. It also says how to word the 240 characters the call speaks and up to 3 questions it asks figs, and how to read each outcome and answer without overclaiming it.

What it contains, runs and sends

• One skill, summon-human, with one script, skills/summon-human/scripts/summon.py. The script needs Python 3 and uses only the standard library (tested on 3.11, 3.12 and 3.13). No hooks, MCP servers or binaries.
• summon.py check reads your GitHub login, the alarm's princes.json and your permissions on the alarm's repository. It files nothing.
• summon.py submit "TEXT" files one issue in the alarm's repository: the title [call] REQUEST_ID, the label call-request, and a body holding the frond-call-v1 marker and JSON with your request, and with --ask, up to 3 questions for figs. The body is byte for byte what the alarm's own submit.sh writes. Then it sends a frond-call repository dispatch, which only wakes the worker sooner.
• summon.py status reads that issue, its comments and its label history, and reports the outcome, and figs's answers once the alarm posts them, from what the worker (github-actions[bot]) wrote. Comments from anyone else are counted but never shown or read as the outcome or as answers.

Everything goes to api.github.com under your own token, and nowhere else. Everyone with access to the alarm's repository can read a request there, and the answers the alarm posts on it.

Who can use it

The alarm accepts a request only from a GitHub login listed in its princes.json. It only sees issues that carry its call-request label, and GitHub silently drops labels on a new issue from an account without write access to the repository. summon.py check tells you whether both hold, and submit checks the label again after filing. Text, questions, request ids and callers the alarm would reject are refused before anything is filed; which line codes exist is left to the alarm, which rejects a wrong one and says why. Requests always go in as prince_manual: the alarm's outage source belongs to its own seat watcher.

Install

In Claude Code:

/plugin marketplace add karmaterminal/thornfield-plugins
/plugin install summon-human@thornfield-plugins

In GitHub Copilot CLI:

copilot plugin marketplace add karmaterminal/thornfield-plugins
copilot plugin install summon-human@thornfield-plugins

Set up

The only credential is your own GitHub token. The script reads FROND_ALARM_TOKEN, then GH_TOKEN, then GITHUB_TOKEN, then gh auth token. A fine-grained token scoped to the alarm's repository needs Issues read and write; the wake also needs Contents read and write, and without it pass --no-wake and the alarm's 5-minute schedule picks the request up. To check it, ask your agent to run summon.py check, or run it yourself:

python3 PLUGIN_DIR/skills/summon-human/scripts/summon.py check

--repo OWNER/NAME (or FROND_ALARM_REPO) points it at another frond-alarm repository.

Try it

Ask in your own words. The skill loads when the request fits, or you can name it. For example:

• "I want figs here now."
• "The build host is off the network and nothing we try brings it back. Summon figs to power cycle it."
• "Do a dry run of the summon-human setup, so we know the alarm would reach figs."
• "Phone figs and ask whether to restart the build host now or wait for the morning."
• "Did figs answer the call from request 42? Tell me exactly what the alarm recorded."

Or run the one step yourself:

python3 PLUGIN_DIR/skills/summon-human/scripts/summon.py submit "Please come to the room now. I want you here."

For a question that can wait until figs reads the room, @-mention figs there instead. The skill's "The light path" section says how to write a mention that notifies.

Where it works

Anywhere Claude or Copilot can run Python and reach api.github.com: Claude Code, GitHub Copilot CLI, or an OpenClaw seat. In a chat without a shell, Claude can still write the text and the exact command for someone else to run.

What it can't do

• It can't make the alarm call: the alarm's worker decides, rate-limits and dials, and a request can be held by its cooldown or hourly cap.
• It can't change a request once it is filed. The alarm rejects a request whose body was edited, so to change one, file a new one.
• It can't prove the phone rang. answered means a human turn, not a voicemail system, is in the call's transcript, which is the alarm's evidence, not the phone company's.
• Answers aren't a transcript. They are what the alarm's voice provider made of figs's words through speech recognition and its call analysis, which can arrive up to 24 hours late or never, and the alarm posts them on the issue, for everyone with access to its repository to read.

Privacy

What it reads, writes and sends, and to whom: the repository's privacy policy.

Support

Report problems and ideas in the repository's issues. Report security issues privately, as its security policy describes.

License

MIT. See LICENSE.