Local intelligence & extraction
These commands are the Edge Agent's on-device intelligence: a governed model-serving runtime, federated cognition that lets a device answer inside a disconnected pocket, federated-semantics scoring, and document extraction. All of it runs without a round trip to a cell.
s9 model-serving
The Model Serving family, verify, register, load, and run models locally,
under a governed runtime with a local audit journal. Signing (mint,
mint-approval) runs on an operator workstation; a deployed device never holds a
mint key.
Inspect and run
| Subcommand | Description | Key flags |
|---|---|---|
status | Runtime status, profile, and loaded models. | --json |
list-models | List registered models on this host. | --json |
inventory | Capability inventory for orchestration / planner consumers. | --json |
config-show | Show resolved deployment profile + state-dir paths. | --json |
load <MODEL_ID> / unload <MODEL_ID> | Force-load / unload a registered model. | --json |
infer | One-shot inference call against the local IPC server. | --model, --task, --input (- = stdin), --max-tokens, --label |
embed | Generate embeddings for inputs from a file (or stdin). | --model, --input, --output, --label |
audit-tail | Tail the local audit journal. | --limit, --since <RFC3339>, --json |
Import and verify
| Subcommand | Description | Key flags |
|---|---|---|
verify-bundle <PATH> | Verify a bundle on disk without registering it. | --json |
import <PATH> | Verify and register a bundle from a local directory. | --json |
mint <MODEL_DIR> | Sign a model directory into a bundle a device can import (operator workstation only). | --manifest, --key, --key-id, --trust-root-out |
mint-approval | Sign an approval releasing a model held pending by a manual-approval posture (operator workstation only). | --model-id, --key, --key-id, --out, --reason |
apply-approval | Verify and apply a signed approval, releasing a pending model. | --path <PATH> |
Operator controls
| Subcommand | Description | Key flags |
|---|---|---|
quarantine <MODEL_ID> | Stop serving a model until released. | --reason (required), --confirm (for non-TTY) |
release <MODEL_ID> | Release a quarantined model. | --reason |
revoke <MODEL_ID> | Local emergency revocation. Backend revocations propagate via the heartbeat path independently. | --reason (required), --confirm |
The family-level flags select where the runtime keeps state and how strictly it
verifies: --state-dir, --trust-roots, and --deployment-profile
(dev, standard-edge, or tactical-edge).
# Import a couriered, signed model bundle, then run it
s9 model-serving import /media/usb/models/mymodel/
s9 model-serving infer --model mymodel --task chat --input - --max-tokens 512
# Quarantine a model pending investigation
s9 model-serving quarantine mymodel --reason "under review" --confirm
s9 cognition
The Cognition family, this device as a leaf in a federated reasoning tree that becomes a sub-conductor, and then a conductor, as the cloud goes dark. Partials it produces while disconnected are spooled and ride home later.
| Subcommand | Description | Key flags |
|---|---|---|
submit | Run a request through the leaf and emit the signed partial + provenance. | --request <PATH>, --utilization |
partial | Show a produced partial and its provenance by request id. | --request-id, --json |
forward | Ride spooled partials home over the transport: load the persisted partials and submit each. | --operational-endpoint, --utilization, --json |
prereduce | Verify, dedup, and combine peer partials into one sealed partial to ride home. | --peers <DIR>, --allow <ENTITY> (repeatable), --json |
conduct | Answer a question inside a disconnected pocket: admit + verify peer partials, then run the reduce locally to produce an actionable answer. | --peers, --allow, --org-conductor-reachable, --json |
inventory | Models, capacity, band-load, and reachability for this device. | --utilization, --cognition-enabled, --json |
status | Leaf / sub-conductor / conductor role and current band zone. | --utilization, --org-conductor-reachable, --holds-prereducer-lease, --json |
Peer partials are admitted fail-closed: a partial whose producer is not in the
--allow set is rejected. Family-level flags set the identity and ceiling for
this leaf: --state-dir, --entity-id, and --ceiling (the highest
classification label this device may run a request over).
# Produce a partial into the local spool while disconnected
s9 cognition submit --request ./request.json
# Answer locally inside a disconnected pocket, admitting two trusted peers
s9 cognition conduct --peers /media/mesh/peers \
--allow entity:dev-2 --allow entity:dev-3 \
--org-conductor-reachable false
# When a link returns, drain the spool home
s9 cognition forward
s9 federated-semantics
The Federated Semantics family, this device as a delegated leaf chart-holder. It scores observations against admitted projections and classifies text, model-free, at the edge.
| Subcommand | Description | Key flags |
|---|---|---|
list | List admitted projections (current generation per key). | --store <PATH> |
verify | Verify a projection JSON file: schema, structure, content hash, and (when present) signature. | --input, --expected-key-id, --allow-unsigned |
score | Score an observation against an admitted projection (off-projection distance + local intrinsic dimension, grounded to the nearest anchor). | --store, --subject-area, --input, --classification-level, --k |
classify | Zero-shot classify text into candidate labels by embedding cosine (model-free; the argmax label is marked *). | --input, --label (comma-separated), --dim |
s9 federated-semantics classify --input "cannot log in" --label billing,bug,praise
s9 federated-semantics score --store ./proj.redb --subject-area support --input "disk full"
s9 extract
Extract text from a document: plaintext, CSV, HTML, Markdown, or PDF.
| Argument / flag | Description |
|---|---|
<FILE> | Path to the file to extract text from. |
--output <FMT> | text (default) or json. |
--chunk-size <N> | Chunk the output into pieces of this size (characters). |
--chunk-strategy <S> | paragraph (default), sentence, or fixed_size. |
--chunk-overlap <N> | Character overlap between fixed-size chunks (default 100). |
s9 extract report.pdf --output json --chunk-strategy paragraph
s9 extract-entities
Extract typed entities from a document using ontology-guided prompts.
| Argument / flag | Description |
|---|---|
<FILE> | Path to the file to extract entities from. |
--ontology-types <TYPES> | Comma-separated type names, each optionally with properties, e.g. Patient(name,age,diagnosis),Device(serial,model). |
--chunk-size <N> | Max chunk size in characters (default 1000). |
--output <FMT> | text or json (default). |
s9 extract-entities intake.pdf --ontology-types "Patient(name,age),Device(serial,model)"
Related topics
- Agent Platform · Agent Runtime, the runtime that consumes local models.
- Data Platform · Governance, the classification and enforcement model the ceilings here reference.
- The on-device agent, the chat and mission commands that reason with these models.