Repository Issue Activity (beta)

#61 Parallelization and Concurrency: Async vs Multithread GEPA

#448 Bugs Identified Post 0.1.4

#116 Abstract GEPAEngine.run() to have discrete steps that can be run by an external controller

#440 Refiner score is credited to the unrefined candidate

#438 Decoupling and parallelizing rollouts + reflections

#97 Reflective dataset contains potentially redundant data for ReAct agent

#409 Code review findings — main @ f919db0a (2026-07-24)

#424 Evaluation cache conflates trainset and valset ids: valset scores are silently served from trainset rollouts

#428 Full-val candidate evaluation incorrectly forces capture_traces=True

#390 Truncated reflection LM output (missing code fence) is silently adopted as a candidate prompt — should format failures be handled on the reflection side as in #265?

#427 Verify evals on Papers with Code

#121 Dealing with VLM

#419 Project Submission

#417 omni(meta-harness): Multi-objective search for meta-harness + Codex integration

#359 Add first-class vLLM / Hugging Face model loading support

#408 Reference: GEPA public API surface & ecosystem usage (stability tiers for future refactors)

#381 Add a run_dir admission manifest for candidate acceptance/rejection audits

#395 ComBEE §3.3: dynamic batch size controller (adaptive reflection_minibatch_size)

#367 When is the next release planned?

#388 Qwen3-8B Reproducible recipe

#329 Design: Unified Reflection LM Pool — sessions, agents, and ComBEE as first-class reflection LMs

#316 Cached eval still counts toward metric budget

#265 feat: format validation as first-class metric with automatic failure penalization

#379 GEPA doesn't remember rejected proposals — re-sampling the same parent repeats the same failed mutation

#267 feat: StructuredCandidate — typed sectioned prompt with append semantics