agx

Use when working on the AGX CLI or its cloud-backed task/project workflows. Covers task lifecycle commands, project assignment, repo attachment with local analysis, daemon/board flows, and the wake-work-sleep execution model.

AGX

AGX is a cloud-backed task orchestrator with a local CLI. Use this skill when changing AGX itself or when you need the exact CLI flows and mental model.

Model

AGX agents are stateless between runs. Continuity comes from persisted API state:

WAKE -> load state -> WORK -> persist state -> SLEEP

Treat the CLI as a client for the cloud API plus local runtime helpers like daemon, board, and repo analysis.

Core CLI flows

agx init
agx -p "explain this code"
agx new "Build a REST API with auth"
agx run <task-id-or-slug>
agx -a -p "Build a REST API with auth"

Tasks

agx new "<goal>"
agx run <task-id-or-slug>
agx status [task-id-or-slug]
agx complete <task-id>
agx retry <task-id-or-slug> [--from <stage>]
agx deps <task> [--depends-on <task> ... | --clear]
agx task ls
agx task logs <id>
agx task stop <id>

Task identifiers usually accept UUIDs, slugs, or cached #N references.

Projects and repos

Use projects to scope task ownership and attach repo context.

agx project list
agx project get <id-or-slug>
agx project create --name <name>
agx project update <id-or-slug> [flags]
agx project delete <id-or-slug>
agx project assign <project> --task <task>
agx project unassign --task <task>
agx new "goal" --project <project-id-or-slug>

Attach a repo from the current directory:

agx repo add . --project <project-id-or-slug>

What agx repo add does:

  • resolves the repo root from the supplied path
  • analyzes local signals like package.json, lockfiles, scripts, key config files, and README.md
  • generates repo notes
  • appends the repo entry to the target project through the cloud API

Optional flags:

agx repo add . --project <project> --name <display-name>
agx repo add . --project <project> --notes "extra notes"
agx repo add . --project <project> --json

Scheduling

Manage recurring prompt jobs that run on a cron schedule via agx-cloud.

agx schedule ls [--state active|paused|stopped] [--project <id>]
agx schedule create "name" --cadence "every day" --prompt "do the thing"
agx schedule create "name" --cadence "0 9 * * 1" --prompt-file ./tasks/weekly.md
agx schedule get <id>
agx schedule update <id> --cadence "every 2 hours" --name "new name"
agx schedule pause <id>
agx schedule resume <id>
agx schedule rm <id>
agx schedule runs <id>
agx schedule cancel <id>
  • --cadence accepts natural language ("every day", "every 2 hours") or cron expressions ("0 9 * * 1")
  • --prompt-file reads the prompt from a file (useful for long prompts)
  • Provider and model default to null; the runtime resolves the environment default
  • --overlap skip|queue|allow controls what happens when a run is still active
  • --catch-up fire_once|replay_all|skip controls behavior for missed runs

Daemon and board

agx board start
agx board stop
agx board show
agx daemon start
agx daemon stop
agx daemon status

Use these when validating end-to-end queue execution or dashboard behavior.

Providers

agx claude [args]
agx codex [args]
agx gemini [args]
agx ollama [args]

Aliases:

  • claude -> c
  • codex -> x
  • gemini -> g
  • ollama -> o

Flags that matter

-p, --prompt
-P, --provider
-m, --model
-a, --autonomous
-y, --yolo
--swarm

Notes:

  • -a is the full autonomous path: create task, start daemon, and work until done.
  • -y skips confirmations during execution and is typically paired with autonomous runs.
  • Project-aware task creation should prefer --project <id-or-slug>.

When editing AGX

  • Keep CLI help text, README command examples, and command implementations aligned.
  • Prefer wiring new behavior through existing cloudRequest and project-resolution helpers instead of creating parallel API clients.
  • For project/repo mutations, preserve existing repo entries when patching repos.