new-agent

name	new-agent
description	Scaffold a new Verona agent in the user-agents dir (default ~/.verona/agents/<name>/) with SOUL.md, tasks/, memory/INDEX.md, and a valid agent.toml. Use when the user wants to create a new agent.

/verona:new-agent

Create a new Verona agent in the user-agents dir (default ~/.verona/agents/<name>/, override via $VERONA_AGENTS_DIR).

DO NOT write to agents/examples/ in the Verona source repo. That dir holds read-only canonical templates. The user's own agents live separately so git pull of Verona never touches them.

If the user wants to start from a bundled template, run verona agents init <name> --template <template> (CLI does the copy + name rewrite). Use this skill for from-scratch authoring or for guided customization beyond what init produces.

Inputs

1. name (kebab-case) — agent directory name and [agent].name. Must match ^[a-z][a-z0-9-]*$.
2. description — one sentence. What does this agent do?
3. adapter — one of claude-cli (default, subscription), anthropic-api, openai, openrouter. Recommend claude-cli unless the user has a specific reason.
4. default_effort — low | medium | high | max. Recommend medium.
5. soul_traits — 3–5 short traits (tone, voice, opinions). You'll expand these into a 200–500 word SOUL.md.
6. initial_tasks — cron tasks only by default. For each, ask: id, schedule (cron expression or every Nm/h/d), prompt summary. Don't ask about on_message — Slack/inbound replies just resume the prior session and need no task definition. Only suggest an on_message = true task if the user explicitly wants a strict per-message protocol (e.g. a triage prompt that must run on every reply).
7. slack — optional. If yes, ask for the channel (e.g. #research-feed).

Resolve target dir

target_dir = $VERONA_AGENTS_DIR/<name>/    # if env var set
           = ~/.verona/agents/<name>/      # otherwise

Refuse to overwrite if target_dir already exists. Suggest a different name.

Steps

1. Confirm all inputs. Show a one-paragraph summary back to the user before writing anything.
2. Create <target_dir>/ with:
   • agent.toml — schema in src/config/schema.ts
   • SOUL.md — 200–500 words; OpenClaw style (specificity > generality, real opinions over safe positions)
   • tasks/<task-id>.md for each declared task — title + body explaining the task's goal, allowed tools, what to write to memory
   • memory/INDEX.md — initial routing table (mirror the structure of agents/examples/*/memory/INDEX.md)
3. Run verona agents add <target_dir> to register it in the state dir. Skip this step if verona doctor reports the state dir is missing — point the user at verona init first.
4. Print a short next-steps list: register adapter API keys (if non-claude-cli), verona connectors add slack (if slack), verona schedule run <name>:<task> for a smoke test.

Tool allowlist guidance

claude -p denies any tool not listed in allowed_tools. Pick the smallest set that covers the task. Common combinations:

Use case	allowed_tools
Smoke test / minimal	["Read", "Write"]
Research / news / monitoring	["Read", "Write", "WebFetch", "WebSearch"]
Document / code editing on memory	["Read", "Write", "Edit"]
Anything that needs shell	add "Bash(curl:*)" (or specific subcommands) — never bare "Bash" for an always-running agent

Do NOT include WebFetch without WebSearch for a research-style task. WebFetch reads a known URL; WebSearch finds pages by topic. Most research workflows need both. The bundled researcher template includes both.

If the user asks "why didn't the agent search the web," the most likely cause is that WebSearch is missing from allowed_tools — check that first.

Things to NOT do

• Don't write into agents/examples/ — that's the source repo's read-only template tree.
• Don't put real tokens, API keys, or Slack channel IDs in committed files.
• Don't skip agent.toml validation — let verona agents add fail loudly if you got the schema wrong.
• Don't write to state/ directly. Use the CLI.
• Don't co-author commits.

After scaffolding

If the user asks for a smoke test, suggest:

verona schedule run <name>:<first-task-id>
verona logs <name> --latest

If the agent uses Slack:

verona connectors add slack    # interactive token paste
verona connectors test slack --destination '#their-channel'