Guidelines
How to actually use Fettle, and why every fix is safe by construction.
The safe loop
- 01DiagnoseFettle turns trace data into metrics + findings, each routed to fix / coach / patch.
- 02IsolateCodex edits a throwaway copy of the agent — never your live code.
- 03GateA benchmark (a machine-checkable definition of “fixed”) decides merge vs rollback.
- 04Merge or roll backOnly --apply on a passing gate touches the real agent. Everything is reversible.
Commands
fettle checkHealth-check every agent Fettle knows about and list what needs attention.fettle fix <id>Repair bugs — stale tool schemas, broken calls, retry storms.fettle coach <id>Refactor to a better professional workflow and explain the change.fettle patch <id>Close security holes — prompt-injection exposure, leaked secrets.fettle watchArm the guardian: maintain idle agents automatically on an interval.Flags
--dry-runShow the exact plan Fettle would hand to Codex, without running anything. Zero cost.
--applyMerge the fix into the real agent — but only if the benchmark gate passes.
--bypass-sandboxRun Codex without its OS sandbox. Use when the host sandbox is unavailable (e.g. it clashes with antivirus on Windows).
A typical run
fettle checkfettle fix support-bot --dry-run # inspect the planfettle fix support-bot --apply # fix, gate, merge