agp is the command-line client for the daemon. It talks to agentd over the
localhost REST API, so the daemon must be running. Everything here is also
doable from the dashboard.
agp <group> <command> [options]
| Command | Description |
|---|---|
task add <title> |
Add a main-orchestrated task. -w/--workspace <repo|portfolio|scratch> (default repo), -r/--repo (repo target, default current Git root), --repo-root (portfolio root), --provider <claude|codex>, -p/--prompt, -f/--prompt-file, -m/--model, -e/--effort <low|medium|high|xhigh|max|ultra> (Codex default high), -P/--priority <0-4> (default 2), -b/--blocked-by <id>, -v/--verify <cmd>. |
task ls |
List tasks. -s/--status <status>, --ready (only queued tasks with no open blockers). |
task show <id> |
Full task detail as JSON. |
task update <id> |
Update fields: -s/--status, -P/--priority, --provider <claude|codex>, -m/--model, -e/--effort, -b/--blocked-by <taskId|none>, --result <summary>. Provider changes are rejected while the task has a live agent. -b records a sequencing constraint — the task stays out of --ready until that blocker is done — and -b none clears it; a self-block, a cycle, or an unknown blocker is rejected. |
task claim <id> |
Atomically claim a queued task. |
task cancel <id> |
Close a task from any state (kills its live worker/reviewer). --rm-worktree also removes the worktree (uncommitted work is lost). Approved unpublished Human-publishes work is retained; destructive cleanup requires --rm-worktree --discard-unpublished. Reports any tasks left dangling by blocked_by. |
task resume <id> |
Reopen a done/cancelled task in place. -p/--prompt or -f/--prompt-file supplies changed requirements. Preserves a resumable same-provider session and safely falls back to a fresh session with the archived handoff. |
task diff <id> |
Show the diff on the task's branch. --stat for stat + commits only. |
Spawn an independent adversarial reviewer for a task in review. -m/--model
overrides the reviewer model (a slug matching the reviewer provider; for Claude
it defaults to CC_REVIEWER_MODEL, then the task model for an existing Claude
workflow, otherwise opus; a Codex model is never passed to Claude).
-p/--provider <claude|codex> runs a cross-model reviewer (otherwise the default
is Claude, or the CC_REVIEWER_PROVIDER / CC_REVIEWER_VARIETY policy), and
-e/--effort sets a Codex reviewer's reasoning effort. Watch it with
agp agent peek <id>.
| Command | Description |
|---|---|
agent spawn -t <id> |
Spawn a Claude Code or Codex worker. --provider, -m/--model, and Codex -e/--effort override the task; --fresh forces a clean session. Resume is same-provider only. |
agent ls |
List agents. -a/--all includes dead agents. |
agent kill <id> |
Kill an agent's tmux window. --requeue puts its task back in the queue; --rm-worktree removes the worktree. |
agent peek <id> |
Print the agent's visible terminal output. -n/--lines <n> (default 50). |
agent send <id> <text…> |
Send a message into an agent's interactive session. |
agent session <id> |
Show provider, session ID, transcript path (when available), working directory, and a copyable detached resume command. Codex commands include Command Center's isolated CODEX_HOME; plain codex resume <id> searches the wrong session store. |
Attach to an agent's tmux window interactively (detach with Ctrl-b d). Uses
switch-client when you're already inside tmux, otherwise attach.
Spawn the Claude orchestrator main agent. -m/--model (default fable, or
CC_MAIN_MODEL; override with CC_MAIN_MODEL=opus / --model opus). One live
main agent at a time. Dashboard/CLI/MCP task creation uses an explicit workspace
and is delivered here before any worker is spawned.
| Command | Description |
|---|---|
scheduler status |
Show scheduler state (default subcommand). |
scheduler on / scheduler off |
Enable / disable autonomous spawning for direct legacy and cron tasks (the kill switch). Explicit workspace tasks remain owned by Claude main. |
scheduler set |
--max <n>, --limit <n> (daily spawn budget), --verify-concurrency <n> (verify_cmd runs allowed at once, fleet-wide), --stall <minutes>, --hours <"22-6"|"always">, --auto-review <on|off>, --escalate <minutes>. |
See configuration.md for what each knob does and its default.
| Command | Description |
|---|---|
cron add <name> -s <cron> |
Create a template. --provider <claude|codex>, -p/--prompt or -f/--prompt-file (required), -r/--repo, --title, -m/--model, Codex -e/--effort (default high), -P/--priority, -v/--verify, --no-pr (no PR; auto-completes without a merge — for report/doc crons), --no-review (skip the adversarial reviewer; complete on worker finish — implies --no-pr). |
cron ls |
List crons with enabled state, schedule, and next/last run. |
cron enable <idOrName> / cron disable <idOrName> |
Toggle a cron. |
cron run <idOrName> |
Enqueue this cron's task immediately. |
cron rm <idOrName> |
Delete a cron. |
dream setup creates the dreaming cron disabled (-s/--schedule default
0 4 * * *, -m/--model default opus, -r/--repo default commandcenter).
Test it with agp cron run dreaming; enable it with agp cron enable dreaming.
| Command | Description |
|---|---|
memory ls |
Recent memories. -n/--limit (default 20). |
memory search <query…> |
Full-text search. |
memory add <text> |
Store a memory. -t/--tags <comma,separated>. |
memory rm <id> |
Delete a memory. |
Recent platform events. -n/--limit <n> (default 30).
Rebuild (npm run build:all) and restart the daemon's tmux window in place, then
health-check it — fixes the stale-daemon warning after a source change.
--main also respawns the main agent so it picks up new MCP tools.
agp codex setup generates the isolated Command Center Codex profile and
static lifecycle hooks, then prints the one-time login and /hooks trust steps.
agp codex doctor checks the CLI, profile parse, isolated login, MCP build,
hook bridge build, and the hook policy/JSON contract. It never bypasses hook
trust or the Codex sandbox; a missing runtime SessionStart is surfaced by the
watchdog instead of being reported as a healthy worker.