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

Codex support is in early development, but it is not ingest-only anymore. selftune can capture Codex sessions through hooks, a wrapper, or batch ingest, run judge and optimizer workflows through codex exec, and use experimental runtime replay validation during selftune evolve.

​
Real-time hooks

selftune codex install
This writes selftune-managed entries to ~/.codex/hooks.json for SessionStart, PreToolUse, PostToolUse, and Stop.

​
Batch ingest

selftune ingest codex
Reads Codex session artifacts and imports them into selftune’s database.

​
Wrapper mode

Wrap a Codex invocation to capture session data in real-time: 
selftune ingest wrap-codex -- <codex args>
This runs Codex with the provided arguments while capturing session data for selftune.

​
LLM-backed workflows

Codex can run selftune’s judge, eval, and optimizer paths.
  • Judge and validation calls use codex exec
  • Specialized optimizer agents are bundled by selftune and inlined into the prompt because Codex has no native --agent flag

​
Limitations

  • Codex has no native --agent flag, so optimizer agents are emulated by prompt inlining
  • Capture quality depends on whether you installed hooks or used the wrapper; plain batch ingest is still post-session
  • Runtime replay validation is still experimental and depends on codex exec --json
  • This path is still experimental and less battle-tested than Claude Code