Skip to main content
Accurate for Stack 0.1.0 and later unless a section cites a newer release. See Stack Changelog.
Stack runs GEPA prompt optimization locally and surfaces the loop in the cockpit, so you can optimize a prompt against your own task and inspect every step as a receipt.

​
What GEPA does

GEPA iteratively proposes and evaluates prompt variants against a task’s training and held-out sets, keeping changes that improve the score. Stack drives the optimizer service, runs the rollouts, and records the result as a StackEval packet you can replay.

​
Run a built-in task

Stack ships StackEval GEPA tasks you can run end-to-end:
TaskWhat it optimizes
banking77-local-gepaintent classification on the Banking77 set
crafter-local-gepaan agent policy prompt on Crafter
Prepare and launch a task from the cockpit harness: 
cd ~/Documents/GitHub/stack
./bin/stackeval harness prepare banking77-local-gepa --preset smoke
This brings up stackd and the Stack TUI, scaffolds a StackEval packet (initial_prompt.txt, metadata.json, trace pointers), and starts the optimization loop. Progress, candidate scores, and the final improved prompt are written to the packet’s trace directory and shown in the cockpit.

​
Inspect the run

Every optimizer run is grounded in stackd receipts: 
./bin/stackeval harness status   --capture-pane -o "$STACKEVAL_PACKET/harness.debug.json"
./bin/stackeval harness export-thread --packet-dir "$STACKEVAL_PACKET" -o "$STACKEVAL_PACKET/harness.export.json"
The export is the auditable record of what the optimizer tried, what scored, and the prompt it shipped.

​
Hosted optimizers

The same loop can run on Synth’s hosted optimizer service when you are signed in. See the hosted-gepa Codex skill for connecting a Stack session to hosted runs.