Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.selftune.dev/llms.txt

Use this file to discover all available pages before exploring further.

​
Status: Experimental

selftune supports Pi as an experimental tracked platform. Pi can feed session evidence into selftune through extension hooks and batch ingest, and selftune can run judge and optimizer workflows through pi -p.

​
Setup

Initialize with the Pi agent type, or let selftune init auto-detect Pi if the runtime is already installed: 
selftune init --agent pi
selftune pi install

​
Hooks

selftune pi install writes extension hook scripts under ~/.pi/extensions/selftune/ for:
  • message
  • tool_call
  • tool_result
  • session_shutdown

​
LLM-backed workflows

Pi can run selftune’s LLM-backed workflows:
  • Judge and validation calls use pi -p
  • Optimizer agents are inlined into the system prompt because Pi does not have a native subagent flag
  • selftune runs these calls in ephemeral mode (--no-session) so validation traffic does not pollute normal Pi session history

​
What gets tracked

selftune can capture the same core evidence types from Pi sessions as other supported runtimes:
  • Skill invocations — when a skill is triggered and how it performs
  • Execution facts — outcomes, grades, and quality signals
  • Improvement signals — low-rated or failed interactions that feed evolution proposals

​
Syncing data

After using Pi, sync your session data to the selftune cloud: 
selftune sync
Then check your skill status: 
selftune status

​
Limitations

  • Pi does not expose a native subagent flag, so selftune emulates optimizer agents by inlining their instructions into the system prompt
  • This integration is still experimental and less battle-tested than Claude Code

​
See also