Expand description
abstractcore-console: the AbstractCore configuration console.
Architecture (the proven gateway-console shape): the UI thread owns
all signals; one worker thread owns every file read and every
abstractcore subprocess. Commands cross via mpsc, results come
back as closures posted through WakeHandle (the engine’s
live-data law).
Re-exports§
pub use screens::catalog;pub use screens::catalog;pub use screens::display_available;pub use screens::display_from;pub use screens::engines;pub use screens::engines;pub use screens::schedule_job_poll;pub use screens::spawn_worker;pub use screens::system_open;pub use screens::Access;pub use screens::JobPoll;pub use screens::OpenError;pub use screens::Opener;pub use screens::Remote;pub use screens::ScreenCmd;pub use screens::ScreensCtx;pub use screens::ScreensOptions;pub use screens::ScreensStore;pub use transport::CliTransport;pub use transport::ConsoleTransport;pub use transport::InstallLocation;pub use transport::ServerAction;pub use transport::TransportCaps;pub use transport::TransportError;pub use transport::TransportErrorKind;
Modules§
- cli
- The
abstractcoreCLI subprocess client — the console’s only lane into Python-derived views (and, in M2, into every coupled write). - config
- Reading
abstractcore.jsonand folding it into a REDACTED display model. No writes live here (M1 is read-only; M2’s write paths will re-read fresh at write time — the model below is display currency). - models
- Model-id classification for the pickers. Discovery gives NAMES
ONLY (
config models/test-providercarry no type metadata — probed 2026-07-25), so class filtering is a name heuristic: embedding models are strongly conventional (they exist to be found by name — embed, minilm, bge, e5, gte, …) while generative model names are unbounded. The pickers therefore filter INTO the embedding class or OUT of it, never pretend finer classes (vision-capable vs text cannot be told apart by name), and every consumer keeps a free-typing lane so a heuristic miss can never block a valid configuration. - probes
- The M3 test vocabulary: probe specs (what to test) and pure verdict
folds (what the evidence means), mirroring
writes.rs’s split — builders + folds are pure and unit-tested; the worker owns the subprocess/socket mechanics. - schema
- The DISPLAY schema: the 17 config sections and their 73 scalar fields, with dataclass defaults and the validation Python actually enforces. Source: docs/config-surface-inventory.md §2 (every entry cites abstractcore/config/manager.py at the probe date).
- screens
- The shared Models and Engines screens (contract H) — one
implementation, mounted by both the core console (over
CliTransport) and the gateway console (over its own HTTP transport). - store
- App state: one struct of signals owned by the UI thread, plus the
typed rows parsed out of
abstractcore config … --jsonpayloads. Nothing here does I/O. - transport
- The seam between the shared Models / Engines screens and whatever answers them (contract H).
- ui
- UI root: one PageHost shell over the 8 screens, two modes — browse
(free tabs) and wizard (a gated linear walk over the CLI wizard’s
phases). Screens are plain component functions; durable UI state
(selections, wizard step) lives in
UiStatecreated at the root, so screen remounts on tab switches lose nothing (PageHost disposes page scopes on switch by design). - worker
- The worker thread: owns every file read/write and every
abstractcoresubprocess. The UI thread owns all signals; this thread never touches them directly — it posts closures throughWakeHandle, which run on the UI thread in the next frame’s user phase. - writes
- The write vocabulary: how every editable surface writes, and how every write VERIFIES. Grounded in live-probed CLI behavior (docs/backlog/proposed/0001_write_lane_design.md):
Constants§
Functions§
- run_cli
- CLI entry — returns the process exit code.