arp-jarvis-daemon

Daemon core + HTTP server for managing local ARP runtime instances (spawn/stop/list) and routing run requests to them.

Implements the ARP Daemon API v1 and can run fully standalone.

This single CLI has two modes:

• Server mode (default): arp-jarvis-daemon (or arp-jarvis-daemon serve) starts the HTTP server.
• Client mode: arp-jarvis-daemon <subcommand> talks to a running daemon server over HTTP.

Install

From PyPI:

pipx install arp-jarvis-daemon

python3 -m pip install arp-jarvis-daemon

From local repo source (with dev dependencies):

python3 -m pip install -e ".[dev]"

The local editable installation is for when you are making changes to this daemon repo, and has dependencies like pyright installed.

Run the daemon server

arp-jarvis-daemon --host 127.0.0.1 --port 8082

Or alternatively:

arp-jarvis-daemon serve --host 127.0.0.1 --port 8082

OpenAPI docs (FastAPI) are available at http://127.0.0.1:8082/docs.

Data directory layout

By default, the daemon stores state/artifacts under ~/.arp/daemon (override with --data-dir):

• Runtime profiles (safe list): <data-dir>/runtime_profiles.json
• Instance registry (persisted): <data-dir>/instances.json
• Cached venvs (managed runtimes): <data-dir>/venvs/<runtime_profile>/...
• Run artifacts: <data-dir>/runs/<run_id>/... (includes trace.jsonl)

Pip/index policy for managed runtimes (optional)

arp-jarvis-daemon --pip-index-url https://pypi.org/simple
arp-jarvis-daemon --pip-no-index
arp-jarvis-daemon --pip-extra-index-url https://<private-index>/simple --pip-trusted-host <private-index-host>
arp-jarvis-daemon --pip-upgrade-pip

These flags only apply in server mode (they affect how the daemon runs pip install for managed runtimes).

Client mode quickstart (profiles → instances → runs)

Terminal A: start the daemon server.

Terminal B: point the CLI client at the server:

export ARP_DAEMON_URL=http://127.0.0.1:8082
arp-jarvis-daemon runtime-profiles list

Runtime profiles (safe list)

The daemon only creates managed runtime instances from runtime profiles (safe list). This is specifically designed with security in mind: we do not want jarvis daemon to become a remote code execution engine.

Profiles are stored in <data-dir>/runtime_profiles.json (by default ~/.arp/daemon/runtime_profiles.json).

A sample runtime profile JSON file is provided in jarvis-runtime-profile.json.

Profiles must include an exec config under extensions["arp.jarvis.exec"] (namespaced extension key).

pip_spec must be pinned (one of):

• pkg==version
• local path (/abs/path, ./rel/path, ../rel/path, file:///...)
• name @ file:///... (recommended for local dev)

Example request payload (pip driver, recommended): jarvis-runtime-profile.json

{
  "runtime_name": "arp-jarvis-runtime",
  "extensions": {
    "arp.jarvis.exec": {
      "driver": "pip",
      "pip_spec": "arp-jarvis-runtime==v0.1.0", // if you want to use package from local repo of a runtime, do something like "your-runtime-package @ file:///path/to/your/runtime"
      "entrypoint": "arp-jarvis-runtime", // Or your runtime's CLI starting point.
      "entrypoint_args": ["serve"]
    }
  }
}

Tip: for repo-local dev, run python3 scripts/bootstrap_jarvis_local.py to upsert a working profile that points at your local arp/JARVIS_Runtime checkout.

CLI:

arp-jarvis-daemon runtime-profiles upsert jarvis-local --request-json ./jarvis-runtime-profile.json
arp-jarvis-daemon runtime-profiles list
arp-jarvis-daemon runtime-profiles delete jarvis-local

Managed instances (create/list/delete)

These commands talk to a running daemon server. Set --daemon-url (or ARP_DAEMON_URL) if not using the default http://127.0.0.1:8082.

Create instances from a runtime profile:

arp-jarvis-daemon start --runtime-profile jarvis-local --count 1 --tool-registry-url http://127.0.0.1:8000

List instances:

arp-jarvis-daemon list

Delete an instance (managed: stop process + remove record; external: deregister only):

arp-jarvis-daemon delete <instance_id>

Overrides for managed instances:

• --tool-registry-url is passed to the runtime as ARP_TOOL_REGISTRY_URL
• --env KEY=VALUE and --arg ... are passed through to the runtime process

Managed runtimes are installed into cached venvs under <data-dir>/venvs/<runtime_profile>/.

Unmanaged instances (register/unregister)

Register an already-running runtime endpoint (daemon will route runs to it, but will not start/stop it):

arp-jarvis-daemon register --runtime-api-endpoint http://127.0.0.1:43120

Runs (submit/status/result/trace)

Submit a run request (JSON file containing a RunRequest), routed by runtime profile:

A sample run-request.json is provided in this repo as a reference:

{
  "input": { "goal": "Say hello" },
  "runtime_selector": { "runtime_profile": "jarvis-local" }
}

arp-jarvis-daemon run --request-json ./run-request.json
arp-jarvis-daemon status <run_id>
arp-jarvis-daemon result <run_id>
arp-jarvis-daemon trace <run_id>

To target a specific instance, use runtime_selector.instance_id instead of runtime_profile.

Traces and run artifacts

The daemon stores run artifacts under <data-dir>/runs/<run_id>/ (by default <data-dir> is ~/.arp/daemon).

• Trace JSONL: <data-dir>/runs/<run_id>/trace.jsonl
• API: GET /v1/runs/{run_id}/trace returns trace_uri and (when available) parsed events.

Dev scripts

• Start server with repo-local data dir: scripts/dev_server.sh
• Bootstrap local runtime profile + create instance: scripts/bootstrap_jarvis_local.py
• Submit a run and poll for result: scripts/submit_run.py

End-to-end (3 terminals):

• Terminal A: arp-jarvis-tool-registry
• Terminal B: bash scripts/dev_server.sh
• Terminal C: python3 scripts/bootstrap_jarvis_local.py --tool-registry-url http://127.0.0.1:8000 && python3 scripts/submit_run.py --goal "Say hello" --runtime-profile jarvis-local

Daemon API endpoints (selected):

• GET /v1/health
• GET /v1/version
• GET /v1/admin/runtime-profiles
• PUT /v1/admin/runtime-profiles/{runtime_profile}
• DELETE /v1/admin/runtime-profiles/{runtime_profile}
• GET /v1/instances
• POST /v1/instances
• POST /v1/instances:register
• DELETE /v1/instances/{instance_id}
• GET /v1/runs
• POST /v1/runs
• GET /v1/runs/{run_id}
• GET /v1/runs/{run_id}/result
• GET /v1/runs/{run_id}/trace