fountainlisted

# Fountain from Hermes You have `fountain_*` tools. Each call talks to a Fountain instance on the user's behalf; the work runs **there**, in a sandbox Fountain provisions, not on this machine. ## Mental model | Primitive | What it is | |---|---| | **Agent** | a named runtime config: runtime (claude/codex/gemini/opencode), model, system prompt, an environment (packages, repos, secrets, network policy), optional skills and MCP servers | | **Conversation** | one running instance of an agent in a fresh sandbox; multi-turn; the sandbox keeps state between turns until it is terminated or reclaimed | | **Turn** | one prompt and everything the agent did in response | | **Vault** | a bag of secret overrides picked per conversation (wins over the environment on key collision) | The user chooses *which* agent by name; you do not configure agents from here. ## Workflow 1. `fountain_agents` — once per session, or when the user names an agent you have not seen. Match by name; pass the name (or id) to `fountain_run`. 2. `fountain_run(agent, prompt)` — write the prompt as if briefing a contractor with **no other context**: the goal, the repo/paths if the environment clones one, what "done" looks like, and what to report back. It provisions a sandbox (typically 30–90 s) and waits for the answer. 3. Read `output`. If `done` is `false`, the turn is still running: call `fountain_wait(conversation_id)` — repeat until `done`. Do not re-run the task. 4. Follow-ups go to the