Skip to content

Command Reference

The full command surface of er (equivalently python -m executor.cli). Everything here is grounded in er --help and each subcommand’s --help. There are no flags beyond the ones documented below.

er [-h] [--db DB] {rig,list,show,preflight,provision,teardown,status} …

er is executor.cli:main — installed as a console script by pip install -e ., or run as python -m executor.cli.

Option Description
-h, --help show help and exit (available on the top-level command and every subcommand).
--db DB path to the recipe DB (tensor-engine-knowledge/engines.db). Overrides ER_RECIPES_DB and the fallback path. Not needed for rig.

The DB is resolved in priority order: --dbER_RECIPES_DB → the fallback ../../readouts/tensor-engine-knowledge/engines.db relative to the repo. See Getting Started.

er rig

Show the detected rig fingerprint. Reads only the live hardware (nvidia-smi + environment) — no recipe DB required. Prints GPU, VRAM, sm, driver, CUDA runtime/toolkit, OS surface, WSL2 version, and the semantic compat-band that recipes resolve against.

No options beyond -h/--help.

er list [--all]

List recipes from the DB as a table of KIND, BACKEND, V (verified flag), and SLUG.

Flag Description
--all include non-executable recipes too. By default list shows only executable recipes.

A recipe whose backend is null (a modifier) shows (inherits) in the BACKEND column.

er show <slug>

Show one recipe in full: kind / backend / axis / verified / executable, plus its baselines, constraints, artifacts (with the index-only count), compensators, and cross-KB techniques. This is a read of the immutable knowledge artifact — no side effects.

Positional Description
slug the recipe slug (from er list).

An unknown slug prints no recipe: <slug> on stderr and exits non-zero.

er preflight <slug>

Resolve the recipe against the live rig and run the capability/constraint checks, the resolvability floor, and compute the instance_idno side effects. Prints each check with a status glyph (ok / HALT / warn / .... / note), a count of note-only and deferred checks, the reproducible: verdict, and a final PREFLIGHT OK or PREFLIGHT HALT. See Reading the check statuses.

Positional Description
slug the recipe slug.

Exit codes: 0 on PREFLIGHT OK; 2 on PREFLIGHT HALT (one or more blocking halts).

er provision <slug> [--execute] [--model MODEL] [--from FROM] [--port PORT]

Plan (dry-run) or, with --execute, materialize → launch → measure. Dry-run is the default and touches nothing — it prints every step and its compensator. --execute is the only flag that touches the rig.

Positional / Flag Description
slug recipe slug to provision (see er list).
--execute run for real. Without it, dry-run plan only — the only flag that touches the rig.
--model MODEL path to the model (.gguf) to serve/measure. Required with --execute.
--from FROM reuse an existing engine dir already on the rig (skip the download/vendor step).
--port PORT 127.0.0.1 port for the launched server. Default 8080.

Live execute is implemented for the native-win-compile backend on the launchable-server kind. Other backend/kind combinations print a clean BLOCKED: … in the plan and a structured exit on --execute rather than running.

Exit codes: 0 on a clean dry-run or a READY execute; 2 on an ANDON halt during execute (the run stops, compensators fire newest-first, and the instance is never marked ready). Missing --model under --execute, or an unimplemented provider, exits non-zero with a one-line message.

er teardown <instance>

Roll back a provisioned instance. Compensators run newest-first and each stop is identity-verified (only a PID that is alive and whose exe lives under our instance dir is killed). The ledger state is honest — rolled-back only when the server is verifiably stopped or gone; stop-failed (instance dir kept) when liveness/identity is inconclusive.

Positional Description
instance instance id to roll back (see er status).

An unknown instance id prints no ledger entry for <instance> and returns without error.

er status

Show provisioned instances — the persisted ledger. Each row is instance-id state recipe, plus pid=… :port and measured tok/s for a running server. With nothing provisioned, prints no provisioned instances.

No options beyond -h/--help.

Code Meaning
0 success — command completed cleanly (clean preflight, clean dry-run, READY execute, normal read).
1 error — a structured one-line message on stderr (e.g. a corrupt-ledger RuntimeError, a missing recipe/DB). Never a raw traceback.
2 ANDON halt — a blocking defect: PREFLIGHT HALT, or an ANDON during --execute (compensators fire, instance not marked ready).

The 2 exit is the load-bearing one: it distinguishes the pipeline stopped on a defect from an ordinary error. A RuntimeError out of any command (notably the corrupt-ledger guard) is caught and surfaced as er: error: <message> on stderr with exit 1, never as a stack trace.