Skip to main content

Pi release notes v0.1.0

v0.1.0 is the first version and has not been published yet; what follows is what it already does.

New Features​

  • Built-in Pi project: a coding agent driven by the pi agent SDK; you assign tasks in the conversation and the process and result stay in the transcript.
  • --agent plan then build: plan only researches and proposes, and once you approve it the same session switches to build.
  • --effort reasoning effort: seven levels from off to max, defaulting to high, with unsupported levels converging automatically.

Added​

  • Added the built-in Pi project, a coding agent driven by the pi agent SDK.
  • Added task assignment in the conversation: the agent reads and writes files in the working directory, runs commands, and leaves both process and result in the transcript.
  • Added a reasoning-effort parameter (--effort) accepting off / minimal / low / medium / high / xhigh / max, defaulting to high.
  • Added an unsupported --effort level converging to the nearest level the model supports, instead of failing.
  • Added a permission-policy parameter (--permission-policy) accepting read-only / workspace / auto.
  • Added read-only hiding the write tools, workspace (the default) asking for approval on every write, and auto running without asking.
  • Added an agent-mode parameter (--agent) accepting build and plan.
  • Added plan researching and proposing only, then asking you to approve the plan before the same session switches to build or keeps the plan unchanged.
  • Added sending /compact or /summarize — or clicking the context usage in the conversation — to compact the current turn.
  • Added replacing the older conversation with a model-written summary instead of prompting the model with it, and saying so when there is nothing to compact.
  • Added tool calls, reasoning, context compaction and failures as collapsible detail rows, while the answer itself stays flat in the conversation body.