loreley 0.5.0a0

Whole-repository Quality-Diversity optimization for real git codebases.

Loreley

Whole-repository Quality-Diversity optimization for real git codebases.

Loreley is an automated Quality-Diversity optimization system that evolves entire git repositories, not just single files or scripts. It continuously samples promising commits, asks external agents to plan and implement changes, evaluates them, and archives the best-performing and most diverse variants for later reuse.

Loreley is inspired by systems such as AlphaEvolve and open-source efforts like OpenEvolve, but is designed from day one for whole-repo evolution, a learned behaviour space, and a production-grade distributed loop.

---

Challenges → core ideas

Challenge in real repositories	Loreley core idea
Single-file evolution cannot express cross-module refactors and production changes	Whole-repo evolution
Hand-crafted behaviour descriptors do not generalise across projects	Learned behaviour space
Demo-style pipelines do not scale to distributed, long-running operation	Production-grade distributed loop

---

Methodology

Loreley treats software evolution as quality-diversity search over the commit graph of a real repository, guided by a learned behaviour space and driven by a production-grade distributed loop. Instead of using LLMs as one-shot patch generators, it organises planning, editing, evaluation, and archiving into a repeatable system that can safely explore improvements while remaining auditable (git), testable (evaluator), and operable (scheduler + workers).

---

Core ideas

1) Whole-repo evolution

Whole-repo evolution makes the git commit the fundamental unit of search. This solves the practical limitation of single-file optimisation: real improvements often require changing multiple modules, updating configs and build scripts, and keeping tests and tooling intact.

Repository-scale evolution has been demonstrated in the literature (for example, SATLUTION), but many repository-scale loops are champion-based and rulebase-driven: a single “current best” becomes the next baseline, and extensive human-authored rules are used to keep the agent on track. This design can limit diversity and makes quality-diversity methods difficult to realise.

Loreley is designed to be QD-native at repository scale:

• it keeps a MAP-Elites archive of multiple elites across behavioural niches (not a single champion line),
• it samples from those niches as inspirations for new jobs,
• and it uses evaluator gates + repository semantics as the primary source of constraints, minimising dependence on domain-specific rulebases.

2) Learned behaviour space

Quality-diversity methods require a behaviour space. Hand-crafted behaviour descriptors (file counts, line deltas, test counts, etc.) are brittle and often project-specific.

Loreley derives behaviour descriptors from repo-state code embeddings (file-level embeddings cached by git blob SHA and aggregated into a commit vector), optionally reduced with PCA. Summary embeddings remain available as an optional utility, but are not used for MAP-Elites behaviour descriptors in repo-state mode.

Under similar fitness, the archive can preserve structurally different improvements (refactors vs micro-optimisations vs feature shifts) as distinct behavioural niches, enabling exploration without collapsing to a single style of change.

3) Production-grade distributed loop

Production-grade evolution requires more than an agent loop: it needs distributed execution, resource controls, and persistent traceability.

Loreley runs a long-lived loop with:

• a scheduler that ingests completed jobs, samples base commits, and enqueues new jobs,
• a Redis/Dramatiq worker fleet that runs planning/coding/evaluation per job,
• a PostgreSQL-backed store for experiments, commits, metrics, and archive state,
• explicit lifecycle controls (max unfinished jobs, optional total job caps, seed population, best-candidate branch export).

You can run a long optimisation campaign on a repository, scaling workers horizontally, while keeping the evolution process reproducible and observable.

---

System overview

At a high level, Loreley sits between your git repository, a pool of LLM-based agents, and a MAP-Elites archive:

flowchart LR
  repo["Git repository<br/>(target project)"]
  sched["Scheduler<br/>(EvolutionScheduler)"]
  queue["Redis / Dramatiq<br/>(job queue)"]
  w1["Evolution worker 1"]
  wN["Evolution worker N"]
  db[("PostgreSQL<br/>(experiments + metrics)")]
  archive["MAP-Elites archive<br/>(learned behaviour space)"]

  repo --> sched
  sched -->|enqueue evolution jobs| queue
  queue --> w1
  queue --> wN

  w1 -->|checkout + push commits| repo
  wN -->|checkout + push commits| repo

  w1 --> db
  wN --> db

  db --> archive
  archive -->|sample base commits| sched

• Scheduler: keeps the experiment in sync with the repository, ingests completed jobs, samples new base commits from the MAP-Elites archive, and enqueues evolution jobs.
• Workers: for each job, check out a base commit, call external planning/coding/evaluation agents, create a new commit, and persist metrics.
• Archive: stores a diverse set of high-performing commits in a learned behaviour space, which the scheduler uses to inspire the next round of jobs.

---

Evolution loop

The core evolution loop connects the scheduler, workers, git repository, evaluation plugins, and the MAP-Elites archive:

sequenceDiagram
  participant S as Scheduler
  participant Q as Redis/Dramatiq
  participant W as Evolution worker
  participant G as Git repo
  participant L as LLM agents
  participant E as Evaluator
  participant DB as Postgres
  participant M as MAP-Elites archive

  S->>DB: ingest completed jobs<br/>(update metrics)
  S->>M: update archive from metrics
  S->>M: sample promising base commits
  M-->>S: elite commit candidates
  S->>Q: enqueue evolution jobs

  Q->>W: dispatch evolution job
  W->>G: checkout base commit<br/>(create job branch)
  W->>L: plan + implement changes
  W->>G: apply edits and commit
  W->>E: run tests / evaluation
  E-->>W: metrics
  W->>DB: persist results

This loop can run indefinitely against a long-lived repository, gradually populating the MAP-Elites archive with diverse, high-quality commits.

---

Architecture at a glance

Loreley is organised into a few main areas:

• Configuration (loreley.config) – a single Settings object (pydantic-settings) centralises environment-driven configuration for logging, database, Redis/Dramatiq, scheduler, worker repositories, and MAP-Elites knobs.
• Database (loreley.db) – SQLAlchemy engine/session helpers plus ORM models for repositories, experiments, commits, metrics, evolution jobs, and archive state.
• Experiments (loreley.core.experiments) – helpers for normalising the target git worktree into a Repository, deriving an Experiment from MAP-Elites and evaluator settings, and reusing the same experiment across scheduler runs.
• MAP-Elites core (loreley.core.map_elites) – repository preprocessing/chunking utilities, repo-state code embeddings (file cache), dimensionality reduction, archive management, sampling, and snapshot persistence via MapElitesManager (summary embeddings are optional utilities).
• Worker pipeline (loreley.core.worker) – worktree lifecycle, planning, coding, evaluation, evolution commits, commit summaries, and job persistence used by Dramatiq actors.
• Tasks (loreley.tasks) – Redis broker helpers and the run_evolution_job(job_id) Dramatiq actor that runs the evolution worker.
• Scheduler (loreley.scheduler) – EvolutionScheduler ingests completed jobs into MAP-Elites, dispatches pending jobs, maintains a seed population (when a root commit is configured), and can create a best-fitness branch when an experiment reaches its job cap.
• Operational scripts (script/run_scheduler.py, script/run_worker.py) – CLI shims that wire up Loguru/Rich logging, settings, the Redis broker, and the scheduler/worker entrypoints.
• Docs (docs/) – focused guides for configuration, scheduler behaviour, worker operations, and the MAP-Elites pipeline.

Module-level documentation lives under docs/loreley/** and docs/script/**. The rendered site is built with MkDocs into site/.

---

Requirements & tooling

• Python 3.11+
• uv for dependency management
• PostgreSQL and Redis (Dramatiq broker)
• Git (worktrees, LFS optional)
• Access to configured external planning/coding/evaluation agents

---

Quick start

1. Clone and install

git clone <YOUR_FORK_OR_ORIGIN_URL> loreley
cd loreley
uv sync          # install dependencies from pyproject.toml / uv.lock

If you already have an environment, you can pin dependencies without creating a workspace:

uv sync --no-workspace

2. Configure

All runtime settings come from environment variables consumed by loreley.config.Settings. Common examples:

• Core app
  • APP_NAME
  • APP_ENV (environment name, for example development / staging / production)
  • LOG_LEVEL (Loguru and Dramatiq log level)
• Database
  • DATABASE_URL
  • or individual DB_SCHEME, DB_HOST, DB_PORT, DB_USER, DB_PASSWORD, DB_NAME, DB_POOL_SIZE, DB_MAX_OVERFLOW, DB_POOL_TIMEOUT, DB_ECHO
• Task queue / Dramatiq
  • TASKS_REDIS_URL or (TASKS_REDIS_HOST, TASKS_REDIS_PORT, TASKS_REDIS_DB, TASKS_REDIS_PASSWORD)
  • TASKS_REDIS_NAMESPACE, TASKS_QUEUE_NAME
  • TASKS_WORKER_MAX_RETRIES, TASKS_WORKER_TIME_LIMIT_SECONDS
• Scheduler
  • SCHEDULER_REPO_ROOT
  • SCHEDULER_POLL_INTERVAL_SECONDS
  • SCHEDULER_MAX_UNFINISHED_JOBS, SCHEDULER_MAX_TOTAL_JOBS
  • SCHEDULER_SCHEDULE_BATCH_SIZE, SCHEDULER_DISPATCH_BATCH_SIZE, SCHEDULER_INGEST_BATCH_SIZE
• Worker repository
  • WORKER_REPO_REMOTE_URL, WORKER_REPO_BRANCH, WORKER_REPO_WORKTREE
  • WORKER_REPO_ENABLE_LFS, WORKER_REPO_FETCH_DEPTH, WORKER_REPO_JOB_BRANCH_PREFIX, WORKER_REPO_JOB_BRANCH_TTL_HOURS
• Worker planning / coding
  • WORKER_PLANNING_CODEX_* and WORKER_CODING_CODEX_* options configuring the default Codex-based planning/coding backends
  • WORKER_PLANNING_BACKEND and WORKER_CODING_BACKEND optional dotted paths to custom agent backends that implement the same protocol as the default Codex CLI backend
• Evaluator
  • WORKER_EVALUATOR_PLUGIN, WORKER_EVALUATOR_PYTHON_PATHS
  • WORKER_EVALUATOR_TIMEOUT_SECONDS, WORKER_EVALUATOR_MAX_METRICS
• Evolution objective
  • WORKER_EVOLUTION_GLOBAL_GOAL – plain-language global objective shared across jobs
• MAP-Elites
  • MAPELITES_EXPERIMENT_ROOT_COMMIT – optional experiment root commit used for seeding
  • MAPELITES_EMBEDDING_MODE (currently only repo_state)
  • MAPELITES_FILE_EMBEDDING_CACHE_BACKEND (db (default) or memory)
  • MAPELITES_REPO_STATE_MAX_FILES (optional cap for repo-state file enumeration)
  • MAPELITES_PREPROCESS_*, MAPELITES_CHUNK_*, MAPELITES_CODE_EMBEDDING_*
  • MAPELITES_DIMENSION_REDUCTION_*, MAPELITES_FEATURE_*, MAPELITES_ARCHIVE_*, MAPELITES_FITNESS_*, MAPELITES_SAMPLER_*, MAPELITES_SEED_POPULATION_SIZE
  • MAPELITES_SUMMARY_*, MAPELITES_SUMMARY_EMBEDDING_* (optional summary embedding utilities; not used for MAP-Elites behaviour descriptors in repo-state mode)

See docs/loreley/config.md for the exhaustive list.

---

Running Loreley

Scheduler loop

The scheduler drives ingestion, scheduling, dispatch, seeding, and archive maintenance:

uv run python script/run_scheduler.py        # continuous loop
uv run python script/run_scheduler.py --once # single tick

# or invoke the module directly
uv run python -m loreley.scheduler.main        # continuous loop
uv run python -m loreley.scheduler.main --once # single tick

See docs/script/run_scheduler.md and docs/loreley/scheduler/main.md for details on scheduler behaviour and configuration.

Worker process

A worker process consumes jobs from Dramatiq, applies planning/coding/evaluation, and pushes results back into the database:

uv run python script/run_worker.py

The worker configures Loguru/Rich logging, initialises the Redis broker defined in loreley.tasks.broker, imports loreley.tasks.workers, and launches a single-threaded Dramatiq worker bound to TASKS_QUEUE_NAME.

See docs/script/run_worker.md, docs/loreley/core/worker/evolution.md, and the other worker module docs under docs/loreley/core/worker/ for deeper operational guidance.

---

Documentation map

The full documentation lives under docs/ and is rendered into site/ via MkDocs. Useful entry points:

• docs/index.md – high-level overview and navigation.
• docs/loreley/config.md – global settings and environment variables.
• docs/loreley/db/ – database engine/sessions and ORM models.
• docs/loreley/core/map-elites/ – MAP-Elites pipeline (repo-state repository embeddings, preprocessing/chunking helpers, dimensionality reduction, sampler, snapshots, and optional summary embeddings).
• docs/loreley/core/worker/ – planning, coding, evaluator, evolution loop, commit summaries, job store, and worker repository.
• docs/loreley/scheduler/main.md – scheduler internals and configuration.
• docs/loreley/tasks/ – Redis broker and Dramatiq actors.
• docs/loreley/api.md – UI API (FastAPI, read-only JSON).
• docs/loreley/ui.md – Streamlit UI dashboard (read-only observability).
• docs/script/ – CLI wrappers for the scheduler and worker.

---

Project layout

• loreley/ – core services (config, db, core/map_elites, core/worker, scheduler, tasks).
• script/ – CLI shims (run_scheduler.py, run_worker.py).
• docs/ – module-level docs under docs/loreley and docs/script, rendered into site/.
• pyproject.toml, uv.lock – dependency definitions for uv.
• examples/ – self-contained optimisation examples used for testing and demos.

---

Examples

• examples/circle-packing – a geometric optimisation benchmark based on the classical circle packing problem. The example defines:

  • a small standalone problem repository in examples/circle-packing with a solution.py that exposes pack_circles(n: int = 26) and returns n non-overlapping circles inside the unit square; the default instance uses n = 26 and the objective is to maximise the sum of radii subject to the non-overlap and boundary constraints;
  • a companion evaluation environment in examples/circle_packing_env with an evaluate.py plugin that checks geometric validity (no overlap, inside bounds) and reports sum_radii (main objective), packing_density, and num_circles.

  You can wire this into the worker by pointing WORKER_EVALUATOR_PYTHON_PATHS at examples/circle_packing_env and setting WORKER_EVALUATOR_PLUGIN=evaluate:plugin, then letting MAP-Elites evolve better packings.

  The best evolved result for this example is available on the evolution/best/b11e0430 branch of the examples/circle-packing repository.

---

License

See LICENSE for details.