GEPA

[gepa] + [proposer] in gepa.toml. Shared TOML: Optimizers. Container routes: Containers.

[proposer]
code
[proposer]backend = "codex_app_server"execution_mode = "local_process"model = "gpt-5.4-nano"auth_mode = "api_key"api_key_env = "OPENAI_API_KEY"copy_host_auth = falsesandbox_mode = "workspace-write"approval_policy = "never"timeout_seconds = 900reasoning_effort = "medium"
FieldTypeNotes
backendstringmust be codex_app_server
execution_modestringlocal_process
modelstring
auth_modestringapi_key
api_key_envstringenv var for Codex key
copy_host_authboolfalse for API-key cookbooks
sandbox_modestringworkspace-write
approval_policystringnever
timeout_secondsinteger
reasoning_effortstringoptional
[gepa]
code
[gepa]max_generations = 3proposals_per_generation = 6minibatch_size = 40max_total_rollouts = 4096max_train_rollouts = 520max_heldout_rollouts = 220rollout_submission_mode = "sync"max_cost_usd = 0.0
FieldTypeNotes
max_generationsinteger
proposals_per_generationinteger
minibatch_sizeinteger
minibatch_accept_marginfloatdefault 0
max_total_rolloutsinteger
max_train_rolloutsintegeroptional
max_heldout_rolloutsintegeroptional
rollout_submission_modesync | async
rollout_poll_interval_msintegerasync
rollout_async_timeout_secondsintegerasync
rollout_chunk_sizeintegeroptional
frontier_typestring
acceptance_criterionstring
max_cost_usdfloat0 = off
max_time_secondsintegeroptional stop
no_improvement_generationsintegeroptional stop
code
[gepa.pipeline]mode = "sync_serial"[gepa.pipeline.workers]rollout = 20max_in_flight_candidates = 6[gepa.pipeline.adaptive_rollout_concurrency]enabled = trueinitial = 50min = 1max = 120
gepa.pipeline.mode
sync_serial | async_pipelined
gepa.pipeline.workers.rollout
integer
gepa.pipeline.workers.max_in_flight_candidates
integer
gepa.pipeline.adaptive_rollout_concurrency.*
enabled, initial, min, max, increase_step, decrease_step, increase_after_successes
Env overrides
EnvOverrides
SYNTH_OPTIMIZERS_GEPA_PIPELINE_MODEgepa.pipeline.mode
SYNTH_OPTIMIZERS_GEPA_ROLLOUT_SUBMISSION_MODEgepa.rollout_submission_mode
SYNTH_OPTIMIZERS_GEPA_ROLLOUT_CHUNK_SIZEgepa.rollout_chunk_size
SYNTH_OPTIMIZERS_GEPA_ROLLOUT_HTTP_RETRIESrollout HTTP retries
OPENAI_API_KEYproposer (when proposer.api_key_env=OPENAI_API_KEY)
Commands
code
bash run_fresh_gepa.sh --profile smokesynth-optimizers gepa run --config gepa.toml