Skip to content

cli.md

Latest commit

 

History

History
471 lines (344 loc) · 12.7 KB

cli.md

File metadata and controls

471 lines (344 loc) · 12.7 KB

CommonGround CLI Reference

This reference describes the current cg command surface implemented by CommonGround/cli.py. It is a minimal open-source reference for users and integration authors. Use cg <command> --help as the most precise local argument reference.

Option Surfaces

cg no longer has root-level shared auth flags. Put connection and auth flags on the command that actually uses them:

  • operator commands, such as cg setup project *, cg local run, cg service run, and cg admission run, expose only local process, database, setup-artifact, host, port, and token-file artifact options.
  • join command cg agent join accepts a server URL, join code, --base-url, --admin-base-url, and --config. It redeems a join code and writes local config/profile/token files; it does not require an existing profile, AgentCredential, or Admin bearer token.
  • admin commands cg admission invite create and cg profile ensure-agent accept --admin-base-url, --admin-auth-token, --admin-auth-token-file, and --config. cg profile ensure-agent also accepts --base-url and --profile because it writes or refreshes a local destination profile.
  • cg_service commands accept --base-url, --profile, --auth-token, --auth-token-file, and --config. Read/projection commands that do not have their own actor field may also accept --caller-project-id and --caller-agent-id.

Admin Service bearer flags are not CommonGround Service auth, and AgentCredential flags are not Admin Service auth. Use cg <command> --help for the exact option surface.

For service commands that already name the actor, the CLI infers caller identity:

  • cg dispatch uses --requested-by.
  • cg turn resume uses --requested-by.
  • cg agent drain and cg agent resume use --requested-by, or --agent-id when omitted.
  • cg provision spawn uses --requested-by.
  • cg report work-memory uses --agent-id.
  • cg worker once and cg worker loop use --profile, or --project-id with --agent-id.
  • cg worker claim next and cg worker claim run use --agent-id.
  • cg smoke pair uses --from.
  • Claim-file worker commands infer identity from the claim file where possible.

Claim-file worker commands still need a matching AgentCredential; the claim file identifies the claim and actor, but it is not a standalone bearer token.

Output Envelope

One-shot CLI commands write one JSON envelope to stdout.

Successful commands:

{
  "ok": true,
  "result": {}
}

Handled failures:

{
  "ok": false,
  "error": {
    "code": "invalid_arguments",
    "message": "human-readable detail",
    "status": 2
  }
}

cg local run, cg service run, and cg admission run are long-running local operator entrypoints. They do not emit the one-shot JSON envelope.

Local Setup

Seed a local project namespace:

PG_DSN=postgresql://USER:PASSWORD@HOST:PORT/DBNAME \
cg setup project seed --default-local

Inspect local project readiness:

PG_DSN=postgresql://USER:PASSWORD@HOST:PORT/DBNAME \
cg setup project status --default-local

This checks setup artifacts and bootstrap readiness. It is not a live CommonGround service health check.

Write local CLI client config:

Single-port local bundle:

cg setup project client-config --default-local \
  --base-url http://127.0.0.1:8000 \
  --admin-base-url http://127.0.0.1:8000

Separated services:

cg setup project client-config --default-local \
  --base-url http://127.0.0.1:8000 \
  --admin-base-url http://127.0.0.1:8001

Run the CommonGround Service:

PG_DSN=postgresql://USER:PASSWORD@HOST:PORT/DBNAME \
cg service run

Run the single-port local first-run bundle:

PG_DSN=postgresql://USER:PASSWORD@HOST:PORT/DBNAME \
cg local run --project-id cg-demo --host 0.0.0.0 --port 8000

Use 127.0.0.1 for same-machine local testing. Use 0.0.0.0 only when another machine must reach this server, and protect the host and network accordingly.

cg local run serves /v3r1 and /admin/v1 from one uvicorn process and one port. This combines the local runtime shape only; /v3r1 still uses AgentCredential and claim fencing, while /admin/v1 still owns product-layer admission and join policy.

Run the reference Admin Service admission API:

PG_DSN=postgresql://USER:PASSWORD@HOST:PORT/DBNAME \
cg admission run --project-id cg-demo

cg local run, cg service run, and cg admission run require the server-ready CLI package:

uv tool install 'commonground-kernel[server]'

If those dependencies are missing, the CLI reports missing_extra.

Agent Onboarding

Create a scoped join invite through the Admin Service:

cg admission invite create \
  --project-id cg-demo \
  --agent-id worker-1 \
  --join-base-url http://10.0.0.10:8000

By default this creates a single-use, 24-hour conversation-worker invite with profile kind byoa.conversation_worker.v1 and runtime kind manual.shell.v1. The result includes a copyable canonical command:

cg agent join http://10.0.0.10:8000 cgjoin_abc123

Redeem the join code on the Agent machine:

cg agent join http://10.0.0.10:8000 cgjoin_abc123

cg agent join writes base_url, admin_base_url, a local profile, and an AgentCredential token file with 0600 permissions. It does not require or print the Admin Service bearer token, and its receipt does not print the AgentCredential secret.

Profiles

Profiles let the CLI request and store an Agent credential through the Admin Service.

cg profile ensure-agent \
  --profile cg-demo/reporter \
  --project-id cg-demo \
  --requested-agent-id reporter \
  --profile-kind byoa.work_memory_reporter.v1 \
  --runtime-kind local.cli.v1 \
  --display-name "Local Reporter"

A profile command needs Admin Service bearer auth through --admin-auth-token, --admin-auth-token-file, or config. The resulting Agent credential token is stored in a local token file and referenced from the CLI config.

For an invited conversation-worker profile, pass the invitation code through a file:

cg profile ensure-agent \
  --profile cg-demo/worker-1 \
  --project-id cg-demo \
  --requested-agent-id worker-1 \
  --profile-kind byoa.conversation_worker.v1 \
  --runtime-kind external-runtime.v1 \
  --display-name "Worker 1" \
  --invitation-code-file ./invite-code.txt

--invitation-code is also available for local throwaway demos, but --invitation-code-file is preferred because it avoids shell history exposure.

For new conversation-worker onboarding, prefer cg admission invite create plus cg agent join. cg profile ensure-agent remains useful for work-memory reporter profiles and lower-level Admin Service testing.

Dispatch

Dispatch a root Turn:

cg dispatch \
  --project-id cg-demo \
  --requested-by requester \
  --target-agent worker \
  --turn-kind turn.conversation.v1 \
  --request-id req-001 \
  --payload-json '{"prompt":"Summarize the repo status"}'

--request-id and --dispatch-key are birth-time causality and idempotency anchors. At least one is required. If only one is provided, the CLI mirrors it to the other.

Payload input modes:

  • --payload-json '<json>'
  • --payload-file payload.json
  • --payload-stdin

Turns

Fetch a Turn snapshot:

cg turn get --project-id cg-demo --turn-id <turn_id>

Fetch semantic context:

cg turn context \
  --project-id cg-demo \
  --turn-id <turn_id> \
  --after-turn-seq 0 \
  --limit 100

Wait for a Turn to close:

cg turn wait \
  --project-id cg-demo \
  --turn-id <turn_id> \
  --timeout-seconds 60 \
  --poll-interval-ms 500

Resume a suspended Turn:

cg turn resume \
  --project-id cg-demo \
  --turn-id <turn_id> \
  --requested-by requester

Agents

Fetch an Agent snapshot:

cg agent get --project-id cg-demo --agent-id worker

Drain an Agent so it stops accepting new Turns:

cg agent drain \
  --project-id cg-demo \
  --agent-id worker \
  --requested-by operator

Resume accepting new Turns:

cg agent resume \
  --project-id cg-demo \
  --agent-id worker \
  --requested-by operator

Project Observation

List Agents through the projection API:

cg project agent list \
  --project-id cg-demo \
  --accepts-work-only \
  --limit 100

List canonical Turn offers:

cg project offer list \
  --project-id cg-demo \
  --turn-kind turn.conversation.v1

Get a single Turn offer:

cg project offer get \
  --project-id cg-demo \
  --turn-kind turn.conversation.v1 \
  --agent-id worker

List Turns:

cg project turn list \
  --project-id cg-demo \
  --target-agent-id worker \
  --state closed \
  --limit 100

Get Turn lineage:

cg project turn lineage \
  --project-id cg-demo \
  --turn-id <turn_id>

Fetch the project feed:

cg project feed \
  --project-id cg-demo \
  --after-ledger-seq 0 \
  --limit 100

Worker Lifecycle

Run the generic shell worker adapter once:

cg worker once \
  --profile cg-demo/worker \
  --command ./worker-bin --flag

Run the same adapter in a loop:

cg worker loop \
  --profile cg-demo/worker \
  --command ./worker-bin --flag

cg worker once and cg worker loop claim work, write context to CG_CONTEXT_FILE, renew the claim while the child command runs, and finish or suspend the Turn from CG_FINAL_FILE or CG_SUSPEND_FILE. These adapter paths do not place the active claim token in stdout or the child environment.

Claim the next available Turn:

cg worker claim next \
  --project-id cg-demo \
  --agent-id worker \
  --claim-out-file claim.json

Claim a Turn, keep the lease renewed, and run a child command:

cg worker claim run \
  --project-id cg-demo \
  --agent-id worker \
  --claim-out-file claim.json \
  --context-out-file context.json \
  -- ./worker-bin --flag

Renew a claim:

cg worker claim renew --claim-file claim.json

The lower-level cg worker claim * commands are different from cg worker once/loop: the claim file and cg worker claim run child environment contain active claim authority. Do not log, paste, commit, or pass those claim artifacts to untrusted child processes.

Append a semantic record under the active claim:

cg worker claim append \
  --claim-file claim.json \
  --payload-file progress.json \
  --role progress

Dispatch a child Turn from a parent claim:

cg worker claim dispatch-child \
  --claim-file claim.json \
  --requested-by worker \
  --target-agent reviewer \
  --payload-file child-payload.json \
  --dispatch-key child-001

Suspend or finish the active Turn:

cg worker claim suspend \
  --claim-file claim.json \
  --reason blocked \
  --note "Waiting for external input"

cg worker claim finish \
  --claim-file claim.json \
  --outcome succeeded \
  --payload-file final.json \
  --final-record-role deliverable

Smoke

Run a pair collaboration smoke test:

cg smoke pair \
  --from cg-demo/requester \
  --to worker

The command checks target offer discovery, dispatches a turn.conversation.v1, waits for terminal state, fetches context, and returns the terminal payload.

Provisioning

Dispatch a provision request to a provisioner Agent:

cg provision spawn \
  --project-id cg-demo \
  --requested-by requester \
  --provisioner-agent provisioner \
  --role researcher \
  --request-id provision-001

This uses Turn kind turn.provision.agent.spawn.v1. Role discoverability is projected from canonical Turn offers, not from kernel authority.

Work-Memory Reports

Submit a local work-memory manifest:

cg report work-memory \
  --profile cg-demo/reporter \
  --project-id cg-demo \
  --agent-id reporter \
  --manifest-file work-memory.json \
  --request-id report-001

Bootstrap or refresh the reporting Agent profile first with the Admin Service:

cg profile ensure-agent \
  --profile cg-demo/reporter \
  --project-id cg-demo \
  --requested-agent-id reporter \
  --profile-kind byoa.work_memory_reporter.v1 \
  --runtime-kind local.cli.v1 \
  --display-name "Local Reporter"

The manifest must be a JSON object and must not contain meta.

CLI vs Integration API

Use the CLI for local operation, debugging, scripted demos, and operator workflows. It provides stable command-level envelopes and profile bootstrap conveniences.

Use the Python clients or HTTP API for runtime integrations:

  • CommonGround.agent_client.http_client.HttpAgentClient covers Agent, Turn, claim, dispatch, credential, and work-memory operations.
  • CommonGround.projection_client.http_client.ProjectionHttpClient covers read-only project projection routes.

The CLI is not the kernel authority. It is a client over the same HTTP service and Admin Service surfaces used by integrations.