Skip to content

v0.7.7-alpha

This alpha release packages the changes landed after v0.7.6-alpha.

Use it if you need faster ingestion and UI paths, refreshed agent backend defaults, and clearer operator behavior around environment checks and reset flows.

Highlights

  • Faster MAP-Elites ingest and archive paths. This release adds query/index work for ingestion, archive lookups, and UI-facing job scans.
  • Cursor pagination for large datasets. The API now exposes /api/v1/jobs/page, /api/v1/commits/page, and /api/v1/archive/records/page, and the Streamlit UI uses those endpoints for page-by-page navigation.
  • Better commit browsing. The commits API and UI now support server-side text filtering instead of loading the full dataset into the browser session.
  • Refreshed Kilocode default backend. The default planning/coding backend uses the Kilocode CLI as kilo, with explicit support for WORKER_KILOCODE_AGENT, model, and variant settings.
  • More predictable Codex CLI worker runs. The Codex backend now runs in a stricter non-interactive mode with isolated CODEX_HOME handling and better error capture.
  • Stronger local examples and tests. The circle-packing example and local eval flow now work without a required submodule checkout, and the test suite covers the new pagination, backend, snapshot, UI, and ingestion behavior.

Upgrade Notes

Kilocode CLI defaults

  • If you rely on the default planning/coding backend, make sure kilo is on PATH.
  • If your environment still sets WORKER_KILOCODE_BIN=kilocode, switch it to kilo.
  • WORKER_KILOCODE_MODE still works as a legacy alias, but new setups should use WORKER_KILOCODE_AGENT.

Environment and preflight behavior

  • OPENAI_API_KEY and OPENAI_BASE_URL now also accept the aliases LORELEY_LLM_API_KEY and LORELEY_LLM_BASE_URL.
  • Scheduler and worker preflight checks now skip the OpenAI key requirement when the configuration uses local-hash embeddings and does not need trajectory summarization.

API and UI paging

  • For large job, commit, and archive datasets, prefer the new cursor-based /page endpoints.
  • The older list endpoints still exist, but the UI no longer depends on offset-only pagination for these views.

Reset behavior

  • uv run loreley reset-db --yes now clears the Redis namespace for the current experiment in addition to recreating the database schema.
  • If you intentionally keep multiple experiments, use separate Postgres databases. EXPERIMENT_ID remains the namespace for processes and external resources inside one database, not a multi-experiment switch for a shared database.

Validation

  • CI continues to run uv sync --locked --all-extras followed by uv run pytest.
  • Local validation for this release candidate: 318 passed in 6.96s.
  • Performance improvements across ingestion, archive queries, and UI data access.
  • New cursor-paginated API/UI flows for jobs, commits, and archive records.
  • Default backend/docs refresh for Kilocode CLI (kilo) and related worker settings.
  • Safer operator behavior for preflight checks, reset flows, and local example evaluation.