Skip to content

v0.7.2

Choose a tag to compare

@github-actions github-actions released this 13 Nov 06:30
· 27 commits to main since this release

The bot sees what its children are doing and can free a stuck one, chooses each child's model, and pots login sets up everything in one run.

Added

  • The bot sees what a working child is doing now. Wherever a working child is described (the task tool's list, the answer to send, the children lines in wakes, pots children), indented lines say which tool it is in and for how long, the start of its command, its last output line and how old it is, and whether a message to it is still waiting: a child takes messages only between its steps.
  • task interrupt ends only a child's current command. The child sees that its tool call was aborted, takes the messages that waited, and goes on. stop still ends the child.

Changed

  • pots login adds all you need in one run. After each sign-in, key or local server it says what is ready, or why not, and comes back to its menu; "Done" leaves and lists the models and judges pi can use, as pots login --list does. A judge's key, such as TypeSafe's for Jev, is tried at once, as pots init tries it.
  • pots init asks again when getting the AI fails. A failed sign-in, a model or local server that does not answer, or Ctrl+C in one of its prompts takes you back to "How should your bot get its AI?" instead of ending pots init.
  • The setup page for AI agents (docs/for-agents.md) has the agent explain what you sign in to, and why, before it gives you pots login: the model the bot thinks with, usually the subscription you already use, and a judge such as Jev, which decides which group messages wake the bot. You add both in one run.
  • The bot chooses each child's model. When it starts a child, the task tool takes a model (provider/modelId) and any of pi's thinking levels, from off to max. A child it starts without them has the bot's model, also after a change in bot.json, and thinking xhigh. A child keeps the model chosen for it across changes of bot.json and restarts. A model that is unknown or not signed in on this computer is refused at once, with the reason, instead of failing at the child's first request. The prompt tells the bot that a code mode script lists the signed-in models, that it talks at medium thinking while children work at xhigh or max, and that which model does which work is the person's to say, kept in its AGENTS.md. A new bot asks about that at its first start.
  • childModel and childThinking in bot.json are no longer used. A bot.json that still has them loads as before, and the gateway's output says that they are ignored. Children started before the upgrade have the bot's model.
  • Children report only when they end, and the bot knows it. Its prompt says that for long work it can set itself a wake (pots wake) to look in on a child now and then and tell the person briefly how it is going.

Fixed

  • An empty API key is no longer saved. Enter at an empty key prompt in pots login or pots init saved an empty key over the one pi had, and pi then counted the provider as not signed in: pots login ended with "signed in, but pi lists no model or judge for this login". Now the prompt asks again, and Ctrl+C goes back without changing the key.
  • For contributors: npm test no longer depends on the environment it runs in. POTS_* variables, a desktop (DISPLAY) and your home folder no longer leak into the tests, and the sign-in tests can no longer open a real browser. A test of memory passes that sometimes failed under load is fixed in the code: the log line "memory pass got no answer" is written after the pass has saved its progress.

What's Changed

  • Replace the demo video with a real Slack recording by @Crokily in #41
  • Update the demo video with captions and Slack windows by @Crokily in #42
  • Tests: independent of the caller's environment, no real browser, and no memory race by @Crokily in #47
  • pots login adds all you need in one run, and an empty key is asked again by @Crokily in #46
  • Show what a working child does now, and let the main agent interrupt its current command by @Crokily in #44
  • Let the main agent choose each child's model by @Crokily in #45

Full Changelog: v0.7.1...v0.7.2