oo-publish - SKILL.md Agent Skill

name: oo-publish description: Use when the user wants to publish, share, ship, or release their agent so others can subscribe. Triggers on phrases like "publish my agent", "share my skills", "ship my agent", "announce my agent", "make my setup subscribable". allowed-tools: Bash, Read, Edit

Publish over the relay

Curate ~/.co/agent.json and run one command. The relay receives a signed ANNOUNCE that carries the alias, bio, and every skill marked publish: true (with its full SKILL.md body inlined). Skills marked publish: false (or missing the flag) never leave the machine.

You orchestrate; co announce does the crypto and networking.

Prerequisites

ls ~/.co/keys/agent.key ~/.co/agent.json >/dev/null \
  || { echo "Run the oo-init skill first (or co setup directly)."; exit 1; }
co announce --help >/dev/null \
  || { echo "MISSING: pip install -U connectonion"; exit 1; }

Phase A — Surface current state

Read ~/.co/agent.json. Tell the user:

Alias and bio (flag if bio still says "Edit ~/.co/agent.json to customize.").
How many skills are in the library, and how many are already publish: true.

cat ~/.co/agent.json

If the bio is a placeholder, block until the user supplies a real one — a vague bio is the most common reason a profile gets ignored.

Phase B — Pick what to publish

Group agent.json.skills by purpose (writing / shipping / social / reviewers) and ask the user which clusters to publish. First-time profiles should be 5–15 curated items, not the whole library.

~/.co/agent.json is a plain JSON file. Use the Edit tool (or $EDITOR) to flip "publish": false → "publish": true on the chosen entries and leave the rest alone. No script needed — it's a one-character edit per skill, and keeping it manual means the user sees exactly what's going out.

Phase C — Dry-run, then announce

co announce --dry-run is the validator. It reads ~/.co/agent.json, inlines every publish: true SKILL.md, signs the payload, and prints what would be sent — without sending. If a description is missing or a SKILL.md is gone, it surfaces there. Run it, scan the list, then send:

co announce --dry-run    # inspect the signed payload + skill list
co announce              # send it

co announce does:

Reads ~/.co/agent.json.
Filters skills to publish: true, inlines each SKILL.md body.
Builds {alias, bio, version, skills:[{name, description, body}]} and signs the whole ANNOUNCE with the Ed25519 key at ~/.co/keys/agent.key.
WebSocket → wss://oo.openonion.ai/ws/announce. Relay verifies the signature and persists the profile + bodies to its database.

Phase D — Confirm

✓ Published <alias> (<address>).

Friends can subscribe with the oo-subscribe skill:
  "subscribe to <alias>"   or   "subscribe to <0xaddress>"

Updates

To re-publish after editing skills or the bio: bump version in ~/.co/agent.json for meaningful changes, then run co announce again. Each announce overwrites the relay-side profile.

Anti-patterns

Don't mkdir bundle/, mktemp, or write a publish dir. There is no bundle directory. ~/.co/ is the source of truth; co announce reads bodies from ~/.co/skills/<name>/SKILL.md directly.
Don't sign in bash. co announce signs once over the whole message. Don't add a second signature on a per-body basis.
Don't gh pr create. Publishing is relay-based pub/sub, not a PR flow.
Don't modify ~/.co/skills/ to curate. Curation lives in agent.json.skills[].publish. The library stays intact.
Don't publish every skill on first run. 5–15 carefully chosen items beats 30 random ones.
Don't wrap agent.json edits in a python script. It's a plain JSON file; the user (or the Edit tool) flips publish flags directly. co announce --dry-run does the validation — don't re-implement it in bash or python.