Architecture

Crucible is a single Rust binary. Everything it needs, including the dashboard, lives in the same process. Data is persisted to SQLite in the configured data_dir, so the daemon can be stopped and restarted without losing state.

Pieces of the binary

Module	Responsibility
Git manager	Clones tracked repositories, enumerates commits on matching branches, and builds each revision on demand using the engine’s configured build command.
UCI adapter	Talks to engine binaries over the standard UCI protocol.
Match runner	Plays games, enforces time control, and streams events for SPRT bookkeeping and training-data collection.
SPRT	Computes the log-likelihood ratio after every game, applies the configured bounds, and reports H0Accepted, H1Accepted, or Inconclusive.
Scheduler	Prioritises work across engines and branches. See Scheduling.
Bisect runner	Drives regression hunts across a range of commits. See Regression hunts.
Gate runner	Runs release gauntlets against external engine binaries. See Release gates.
Training writer	Writes JSONL training exports from self-play and regression matches.
Storage	SQLite schema, queries, migrations, and the claim-and-recover logic that keeps workers consistent across restarts.
Web server	Axum router serving the JSON API and the embedded HTML dashboard.
TUI	Ratatui-based terminal UI, used either alongside the daemon or attached later.

The testing loop

1. Fetch. The scheduler polls every tracked engine’s remote for new commits matching its configured branches.
2. Build. When a new commit is picked up, a worker checks it out and runs the engine’s build command in the cloned repo.
3. Schedule. The scheduler decides which pair of revisions to test next, using the priorities described in Scheduling.
4. Test. A worker plays a match between the two revisions, streaming games into SQLite as they finish.
5. Analyse. SPRT is evaluated after every game. When it accepts a hypothesis, or the match hits max_games, the worker records the Elo estimate, error, and LOS, and releases the slot.
6. Repeat. The daemon never stops. New commits are picked up on the next poll, and the Timeline rolls forward.

Storage layout

Everything Crucible persists lives under data_dir:

<data_dir>/
  crucible.db           # SQLite database
  repos/<engine>/       # clones of tracked repositories
  gates/*.json          # release gate summaries
  training/.../*.jsonl  # exported training data

The database holds engines, revisions, jobs, games, bisect sessions, training run metadata, and gate runs. Cancelled or interrupted jobs are re-queued when the daemon restarts, so nothing gets stuck in a stale “running” state.