Models
The bot runs its agents on models that Pi can reach.
You choose them in roundtable.config.ts:
export default { // ... model: env("MODEL"), thinking: "medium", judge: { model: "anthropic/claude-haiku-4-5", threshold: 0.6 }, delegation: { model: "anthropic/claude-haiku-4-5", thinking: "medium" }, agentDir: "./data/pi",} satisfies RoundtableConfig;Only model is required.
The <provider>/<id> format
Section titled “The <provider>/<id> format”Every model setting is written <provider>/<id>, for example anthropic/claude-sonnet-5-5.
The part before the first slash is the provider, and the rest is the model id.
Anything else stops the start and names the key:
config model: expected <provider>/<id>, got "<value>". Write it like anthropic/claude-sonnet-5-5.The same check applies to judge.model and delegation.model, under their own names.
model and thinking
Section titled “model and thinking”model sets the agents’ model.
An agent follows it unless that agent has a model of its own.
thinking is one of off, minimal, low, medium, high, or xhigh, and defaults to medium.
By default an agent’s thinking setting is auto: before each turn the judge reads the message and picks low, medium, high, or xhigh.
If the judge is unsure, the agent keeps the previous turn’s level, or uses thinking as the fallback for its first turn.
The judge is a small model that answers short questions for the bot:
- how much thinking a turn needs,
- whether your reply to a held action is a go-ahead, and
- which members of a group should answer a message.
| Key | Default | Meaning |
|---|---|---|
judge.model |
the value of model |
The model that answers those questions. |
judge.threshold |
0.6 |
A number from 0 to 1. How sure the judge must be before its answer is used for a thinking level or a go-ahead. |
A cheaper model is enough here.
If the judge’s model is not available, the judge fails with the judge's model <provider>/<id> is not available; log in to its provider or change judge.model.
delegation
Section titled “delegation”An agent can hand research to a background worker with delegate_task.
The worker starts a fresh session with web search and page reading as its only tools.
Its report comes back to the agent’s channel.
| Key | Default | Meaning |
|---|---|---|
delegation.model |
the value of model |
The model the worker runs on. |
delegation.thinking |
medium |
The worker’s thinking level. |
The worker loads pi-web-access, a peer dependency of pi-roundtable.
A model and thinking level per agent
Section titled “A model and thinking level per agent”An agent has its own model and thinking level only when you set them.
Ask the coordinator, which changes them with agent_update:
Put Researcher on openai-codex/gpt-6-sol and set its thinking to high.agent_update takes a model as <provider>/<model>, or default to follow the assistant again.
A model the bot cannot run is refused, and the answer lists the usable ones.
Thinking takes one of the six levels or default; default means auto, the judge’s choice.
A fixed level skips the judge for that agent.
The change applies from the agent’s next turn, and its conversation stays.
See Create agents.
Model login
Section titled “Model login”The bot needs a login for the provider of model.
roundtable doctor checks for one, and accepts either of these:
- The provider’s API key in
.env, such asANTHROPIC_API_KEYforanthropicorOPENAI_API_KEYforopenai. - A login made with Pi, stored in
auth.jsoninside the agent directory.
If neither exists, the check fails:
✗ model login: no login for anthropic, the provider of anthropic/claude-sonnet-5-5. Set the provider's API key in .env (ANTHROPIC_API_KEY for anthropic, OPENAI_API_KEY for openai), or sign in with Pi so ./data/pi/auth.json holds it.When it passes, it says where the login comes from, for example anthropic login from <source>.
doctor reads the login without creating the agent directory or auth.json.
agentDir
Section titled “agentDir”agentDir is Pi’s agent directory, where the model logins live.
It defaults to <dataDir>/pi.
When the bot starts it sets PI_CODING_AGENT_DIR to this directory for the Pi packages it loads.
Sign in with Pi using this directory, or point agentDir at the directory that already holds your login.