semiont-cli

name: semiont-cli description: Help users accomplish knowledge work tasks using the Semiont CLI disable-model-invocation: false user-invocable: true allowed-tools: Bash, Read, Write, Glob, Grep

You are helping a user work with a Semiont knowledge base using the semiont CLI.

Semiont is a knowledge system where humans and AI agents collaborate as peers. Documents (resources) are stored in a knowledge base and enriched with W3C Web Annotations — highlights, comments, tags, and entity references. The CLI is the primary tool for this work. All commands talk to a running Semiont backend via a cached auth token.

Authentication

Before any API command will work, the user must be logged in:

semiont login --bus http://localhost:4000 --user alice@example.com
semiont login --bus https://api.acme.com   # interactive password prompt
semiont login --refresh --bus https://api.acme.com

Tokens are cached at $XDG_STATE_HOME/semiont/auth/<bus-slug>.json and are valid for 24 hours. If a command fails with an auth error, prompt the user to run semiont login first.

All API commands accept --bus <url> (short: -b) to target a specific backend. Falls back to $SEMIONT_BUS.

Primary workflow

The core pipeline is: mark → gather → match → bind. Detect entity references in a document, assemble context around each one, search the KB for a match, and link the annotation to its target. When no match exists, yield --delegate generates a new resource and binds to it.

# 1. Detect entity references (AI-assisted)
semiont mark "$RESOURCE_ID" --delegate --motivation linking \
  --entity-type Location --entity-type Person

# 2. For each unresolved annotation, find candidates
semiont match "$RESOURCE_ID" "$ANN_ID"

# 3. Bind to the best match
semiont bind "$RESOURCE_ID" "$ANN_ID" "$TARGET_ID"

# 4. Or generate a new resource if no good match exists
semiont yield --delegate --resource "$RESOURCE_ID" --annotation "$ANN_ID" \
  --storage-uri file://generated/output.md

The semiont-wiki skill runs this pipeline end-to-end as a TypeScript script using @semiont/sdk — the canonical implementation of the Canonicalize mentions archetype. The SDK loop composes browse → gather → match → bind / yield.fromAnnotation cleanly; the CLI commands above are useful for ad-hoc testing of individual steps.

browse — Inspect the knowledge base

semiont browse resources
semiont browse resources --search "Paris"
semiont browse resources --entity-type Location --limit 20
semiont browse resource <resourceId>
semiont browse resource <resourceId> --annotations
semiont browse resource <resourceId> --references
semiont browse annotation <resourceId> <annotationId>
semiont browse references <resourceId>
semiont browse events <resourceId>
semiont browse history <resourceId> <annotationId>
semiont browse entity-types
semiont browse files
semiont browse files docs
semiont browse files docs --sort mtime
semiont browse files --sort annotationCount

browse files lists a project directory, merging live filesystem entries with KB metadata. Each entry is marked tracked: true/false. Dotfiles and .semiont/ are excluded. Paths that escape the project root are rejected. --sort accepts name (default), mtime, annotationCount.

semiont browse files | jq '.entries[] | select(.tracked) | .name'

gather — Fetch context for downstream workers

Use gather instead of browse when feeding data into automation or pipelines.

semiont gather resource <resourceId>
semiont gather annotation <resourceId> <annotationId>

mark — Create an annotation

Delegate mode is the primary path for bulk annotation — an AI worker scans the document and creates annotations automatically. Manual mode is for one-off corrections or additions.

# Delegate — AI worker detects and creates annotations
semiont mark <resourceId> --delegate --motivation highlighting
semiont mark <resourceId> --delegate --motivation linking --entity-type Person --entity-type Place
semiont mark <resourceId> --delegate --motivation tagging --schema-id science --category Biology

# Manual — specify selector and body directly
semiont mark <resourceId> --motivation highlighting --quote "key phrase"
semiont mark <resourceId> --motivation commenting --quote "phrase" --body-text "my comment"
semiont mark <resourceId> --motivation linking --quote "Paris" --link <targetResourceId>

Motivations: highlighting, commenting, linking, tagging, assessing, describing.

match / bind — Resolve a reference

semiont match <resourceId> <annotationId>
semiont match <resourceId> <annotationId> --user-hint "look for papers about Paris"
semiont bind <resourceId> <annotationId> <targetResourceId>

# Typical pipeline
TARGET=$(semiont match <resourceId> <annotationId> --quiet | jq -r '.[0]["@id"]')
semiont bind <resourceId> <annotationId> "$TARGET"

listen — Stream domain events

semiont listen
semiont listen resource <resourceId>
semiont listen | jq .type

yield — Upload or generate a resource

# Upload a local file
semiont yield --upload ./paper.pdf
semiont yield --upload ./paper.pdf --name "My Paper"
semiont yield --upload ./a.md --upload ./b.md

# Generate a new resource from an annotation's context
semiont yield --delegate \
  --resource <resourceId> \
  --annotation <annotationId> \
  --storage-uri file://generated/output.md

For corpus-wide ingest — declaring the KB's entity-type vocabulary via frame.addEntityTypes and then yielding many files in one run — use the SDK-based semiont-ingest skill instead. The CLI's per-file yield is right for one-off uploads; for systematic corpus loading the SDK version is canonical.

beckon — Direct another participant's attention

Sends a focus signal to a named participant connected to the backend. Ephemeral — dropped if the participant is not connected.

semiont beckon <resourceId>
semiont beckon <resourceId> <annotationId>

Guidance for the AI assistant

• Resolve IDs before acting. If the user refers to a resource by name, use semiont browse resources --search "<name>" to find the ID first.
• Check auth early. If a command fails unexpectedly, check whether the token is missing or expired (semiont login).
• Compose with jq. All commands emit JSON to stdout. Suggest jq pipelines for filtering, extracting IDs, or chaining commands.
• browse files vs browse resources. browse files shows what is on disk (tracked or not); browse resources shows only what is in the KB. Use browse files when the user wants to see their project directory or find untracked files.
• Delegate before manual for mark. Suggest --delegate when the user wants to annotate a whole document. Manual mode is for targeted corrections.
• The gather → match → bind pipeline is the standard pattern for resolving unresolved reference annotations — the Canonicalize mentions archetype, building Layer #3 of the layered data model. Walk the user through the three steps in order, or point them to the SDK-based semiont-wiki skill for full automation. For repeated / automated use, the SDK version is canonical; the CLI is for ad-hoc testing.
• beckon coordinates attention between participants, not navigation within the app. It is useful for directing a human reviewer's attention to a specific annotation from a script or agent.