Mac CLI

The mac CLI

What this skill owns, and what it does not

It owns the CLI's shape: which object a verb hangs off, which verbs mean something other than their name, what the default views hide, and the traps that have cost a wrong command against a live fleet.

It does not own: architecture decisions (those are docs/adr/), how to diagnose a failed task, how to decompose work, or repository conventions (CLAUDE.md, CONTRIBUTING.md). If a question is "what should this task be", this is the wrong document.

Stating the boundary is worth the four lines: a skill that quietly grows into everything stops being read, and the reader cannot tell what it is responsible for being right about.

Before acting after a compaction or handoff

If this conversation has been compacted, summarised, or handed over, reread this skill and the live state before any mutating command — do not act on a summary's recollection of either.

This is not hypothetical caution. In one session a task was reopened from a summary that did not record that its work had already merged, and an agent immediately claimed it and began re-implementing a merged module. The summary was accurate about what had happened; it just did not carry the one fact that made the action wrong.

Cheap checks, in order: mac task show <id> for current state, and gh pr list --search "<task_id>" --state all for work that already landed.

mac <object> <verb> [args]. Everything is an object with verbs, and the objects that matter day to day are project, task and agent. Everything else lives under admin.

Non-interactive stdout defaults to JSON and disables pagers. Interactive terminals default to compact text. --json explicitly selects JSON in any position: mac task list --json and mac --json task list are the same command.

If you are registered as an agent, send your own heartbeat

An interactive session that registers itself as a runner is a real agent on a real host, and nothing will report your liveness for you. The hub stamps only resources.virtual agents (operator, hub-reviewer) — constructs whose liveness the hub itself constitutes. Vouching for a session on someone's laptop would describe a closed lid as a healthy worker, so it deliberately does not.

If you do not heartbeat, the stale sweep marks you offline. Observed 2026-08-20: a session sat offline for 33 minutes while actively working, and the operator's mac agent list said it was gone.

mac agent heartbeat <agent_id> --status idle --health-status healthy

Send one when you register, and again at natural checkpoints in a long session — after landing a change, before a long-running gate.

The id is not the name. The name keeps hyphens, the id substitutes underscores: name claude-session-yowza, id agent_claude_session_yowza. mac agent show takes the id and answers agent not found for the name, which reads like the agent is gone rather than misaddressed. mac --json agent list gives you both.

A hold survives a heartbeat, and should: a held session reports itself alive without becoming eligible for dispatch.

The traps

These are not hypothetical. Each one cost a wrong command against a live fleet.

Everything unfamiliar is under admin. The top level was deliberately reduced to the four objects above. dispatch, human, memory, machine, fleet, client, openshell, secret and the rest all moved. mac dispatch prints a redirect rather than working, so if a command "used to exist", try mac admin <thing> before assuming it was removed.

A paused project is activated, not resumed. mac project resume does not exist. Agents use the opposite pair: mac agent hold / mac agent resume. So the verb depends on the object, and guessing from the other one is wrong.

mac agent install is not a verb. mac agent is the fleet worker: register, list, hold, resume, heartbeat. The harness installer is mac admin plugin install. Guessing mac agent install from a design doc that said "agent install" sends you to a usage error.

mac agent update cannot change status. It takes --capabilities, --add-capability, --remove-capability, --instance-kind, --owner, --visibility. To clear a draining agent, PUT {"status": "idle"} to /agents/{id} on the hub.

mac project pause only reaches REGISTERED projects. A project name that came from task metadata rather than mac project create returns Not Found, so pausing "the fleet" by walking the project list silently misses work. To stop dispatch fleet-wide, hold the agents.

Pausing a project does not drain it. In-flight tasks keep running. A deploy started at that moment fails with active work attached while release compensation ran. Wait until no agent reports current_task_id.

delete is usually a rename of something gentler. mac project delete is unregister; with --force it sets tasks.project = NULL rather than destroying tasks. Read the help before assuming a delete is destructive — or that it is not.

Lists are scoped by default

list shows what you can still act on, not the whole table. Both defaults exist because the unscoped view was unreadable on a real hub.

mac task list                  active work only (17 rows on the live hub)
mac task list --all-states     every task, terminal included (507)
mac task list --state=cancelled   one state, unchanged

mac project list               projects with live work OR a registration (10)
mac project list --all         every project, including derived ghosts (19)

When to pass the flag. --all-states / --all are for auditing and archaeology — counting what a project ever held, finding a task you cancelled last week, verifying a prune. For anything about what the fleet is doing now, the default is the answer, and a flag that adds 490 cancelled rows is actively misleading.

What project list --all reveals. A project with no project_id is DERIVED: it exists only because some task carries that string. mac project show, pause, activate and unregister all return Not Found for one, and nothing can be dispatched to it. It is a label, not an object. That is why it is hidden unless it holds live work.

How the project is inferred

mac task create with no --project infers one from the working directory: the name of the repository, resolved through git rev-parse --git-common-dir, which every linked worktree shares.

This matters because CLAUDE.md mandates a worktree per agent. Inferring from --show-toplevel instead — which is what this used to do — takes the basename of the WORKTREE, so filing from /tmp/mac-dev created a project called mac-dev. Six such ghosts were live on the hub, each holding one or two cancelled tasks. If you see a project named after a branch, that is what it is.

Pass --project '' for none, or --project NAME to override.

Finding things

mac help                      the object list
mac <object> help             verbs for one object
mac admin help                everything that moved under admin
mac <object> <verb> --help    arguments

mac task create help is intercepted and prints help; it does not file a task called "help".

The commands that come up most

mac task list --state=open          mac task show <id>
mac task ready --limit 10           mac task why-unclaimed <id>
mac task create "title" --description-file=f.txt
mac task reopen <id>                recovery: terminal/stuck -> open
mac agent list                      mac agent hold/resume <id>
mac project pause/activate <name>
mac admin human register <username>
mac admin dispatch submit <file>    literate-ai execution requests
mac admin judgement status          process-quality daemon last report
mac admin judgement run             run one judgement cycle now
mac admin events news --follow      significant task and agent activity
mac admin login --local-console     hub-local enrollment without SSH
mac admin plugin install --scope global
mac admin plugin install --scope repo --repo PATH
mac admin plugin status
mac admin plugin uninstall

mac admin events news is the human-facing fleet feed: task creation, claim and state changes plus meaningful agent lifecycle/status changes. Add --follow to subscribe continuously, --project NAME to restrict task activity to one project, or --json for newline-delimited records in follow mode.

mac admin login --local-console is only for a shell on the hub. It asks the running API service for a new scoped, independently revocable credential over its kernel-authenticated Unix socket; it never reads the shared hub admin token. The ordinary scope boundary is read,write,dispatch. Any broader scope requires the API service's OS account or root plus --allow-elevated; a configured supplementary group grants ordinary enrollment only. Remote clients continue to use mac admin login --ssh ... or a registered fleet SSH route. Use mac admin login renew --local-console to rotate a direct local-console profile through the same socket; the new bearer is validated before replacing the local credential.

Sequence tasks before releasing them

Both mac task create and mac task update accept --dependencies. Update replaces the dependency list; pass --dependencies "" to remove it. Editable-state and cycle checks still apply. A running task is not editable.

File gating tasks first and create dependents with their dependencies already set. If you need to assemble a batch before sequencing it, create it with --no-dispatch, update the dependencies, inspect them, then release the tasks. Under the default all_success join policy, only completed dependencies release their dependents; failed or cancelled dependencies do not.

What happens if you skip this. On 2026-08-20 eleven tasks were filed flat in one session. Within the hour four were reviewing, four failed, one running — including an unsupervised attempt at a 5,565-line Hermes removal that should have been gated behind an ADR and behind resolving a vendored directory the deploy still required. Deleting that code first would have made the fleet undeployable. Two more tasks were being worked by agents while the same defects were being fixed by hand in open PRs, because nothing connected the filing to the fix.

Two habits that follow:

• If you are also fixing the thing yourself, say so in the task at creation, or do not file it. A detailed task plus a parallel fix is duplicated work, and the agent that claims it cannot know.
• An investigation task and its implementation are two tasks with an edge, not one task. File the investigation, then create the implementation depending on it.

Priority is inverted, and 0 is the bottom

priority is an integer sorted DESCENDING — the allocator sorts on -int(task.priority). Higher is more urgent. The convention on the live hub is:

P0  ->  priority 100      the top of the open queue
escalations above it      900, 1000, 2000 (used sparingly)

So "make this P0" means --priority 100. mac task update <id> --priority 0 does the exact opposite of what it reads like: it sends the task to the back of the queue, behind everything. There is at least one P0:-titled task on the hub sitting at priority 0 for this reason.

mac task update <id> --priority N is a targeted field write. It preserves description, metadata, dependencies, capabilities, max_attempts, state and project — verified across 18 tasks. You do not need to re-send the description to change a priority, and you should not: passing --description-file rewrites the whole field.

If you are working the ledger, you are a participant — register and claim

This applies to YOU, an interactive session, not only to loop-mode workers.

An interactive session with hub credentials is a fleet participant whether or not it is registered. If it is not registered, its work is invisible: no agent row, no claim, no owner_agent_id, no attempt_count. Other agents cannot see that the task is being worked, so nothing stops one of them starting the same task. That is the same claim-exclusivity failure that produced five divergent implementations of one task and eleven duplicate PRs on 2026-08-19 — an unregistered session is one more way to cause it.

This was found by audit: a session on a host that was not any fleet node fixed and merged two tasks (PR #478, PR #465), and hours later both still read state=open, owner_agent_id=None, attempt_count=0.

So:

mac admin machine register <host> --machine-id machine_<host>
mac agent register machine_<host> <session-name> --agent-id agent_<name>
mac agent hold <agent_id> --reason "Interactive session; do not auto-dispatch"

An ordinary claim respects agent and task holds. For an explicitly authorized local host implementation, record the single-task execution-boundary exception with mac task break-glass <task_id> <agent_id> --reason "...", then claim and start that exact task. The authorization expires and is consumed once; it does not release other work. Inspect mac task break-glass-list <task_id> to confirm it. Normal fleet execution uses the ordinary allocator, without this exception.

mac task claim <task_id> <agent_id>
mac task start <task_id> <agent_id>

Register held, and understand what the hold does. An interactive session must never be a dispatch target — it cannot answer an assignment. mac agent hold <id> prevents that. agent_operator is already held for precisely the failure to avoid: "claimed task_1b76d5ed and held it 6h silently without executing." A claim you are not actively executing is worse than no claim, because it stops an agent that would have done the work.

Claim before, not after. Claiming retroactively does not work, and the state machine is what stops you: from open the only legal moves are blocked, cancelled, claimed, failed, needs_input, waiting — completed is not among them. Completion is reachable only through claimed -> running -> needs_review, i.e. through the fleet's own review loop. Repository completion still requires canonical integration proof, including for force-complete. Do not substitute an operator acceptance record for that proof. If a held task is superseded by another implementation task, use the explicit superseded disposition and replacement task so the ledger preserves why no separate execution is needed.

Releasing a hold hands the task out immediately. mac task release clears a --no-dispatch hold (use it — do not hand-edit metadata.no_dispatch). But if the task is already satisfied by merged work, an agent will claim it within seconds and start re-implementing it. Confirm the work is genuinely outstanding before releasing, and close it first if it is not.

Before you work a task, check whether it already has a PR

A retried task currently files a NEW pull request on each attempt rather than updating the existing one. On 2026-08-19 the open queue held 23 PRs covering 12 distinct pieces of work; one task had emitted five, by two different agents, with five divergent implementations (+598 to +2285 lines).

Two consequences for anyone working a task:

• Search for the task id before starting. PR titles carry it, so gh pr list --search "<task_id>" finds prior attempts.
• Finding an open PR for your task does not mean the work is done. It means an earlier attempt got that far. Read it before writing a second one.

Until the idempotency fix lands, this is manual. It is the single largest source of wasted fleet capacity observed to date.

Reading agent state

mac agent show <name> often fails for an agent that mac agent list displays. show resolves by id, and an agent registered before the current id convention has an opaque uuid that matches neither its name nor the id the code expects. hub-reviewer is the live example: it is agent_da539e43b3e147ffb580ebdd9f7cde6b, while the constant in services.py says agent_hub-reviewer, and both mac agent show hub-reviewer and mac agent show agent_hub-reviewer return "agent not found". Get the real id from mac agent list --json.

Virtual agents flap. hub-reviewer is registered by the hub itself so hub-side contract verification has an identity to sign with. It has no worker process, so nothing heartbeats for it, and health_status is only whatever was last written. It moves between idle, offline and degraded with no underlying fault. Do not diagnose it as a broken worker; it is not a worker.

why-unclaimed may name no reason at all. For a task that is genuinely eligible and simply not being dispatched, it prints attempt counts and nothing else. attempt_count: 0 with idle capable agents and no dispatch hold is a real state, and it means the fault is in dispatch rather than in anything about the task. Do not read a bare why-unclaimed as "nothing is wrong".

Requirements: capabilities versus hardware

Capabilities are set membership over a DECLARED vocabulary — agents advertise python, testing, review. Host facts are NOT capabilities: os and cpu_arch are probed into resources.hardware, and no agent will ever advertise linux. Asking for one as a capability produces a task that is created and never claimed.

WRONG   --capabilities linux
RIGHT   --hardware '{"os": ["linux"], "cpu_arch": ["x86_64"]}'

mac task preflight answers whether the fleet could ever claim a task, and names this mapping error specifically. Run it before filing anything with requirements; a task that cannot be satisfied is accepted and queued, and then waits forever rather than failing. Use mac task why-unclaimed <id> for one that is already filed and sitting.

Where mac is talking to

Resolution order: --db (direct Postgres, prints a redacted banner), then --hub-url, then $MAC_API_URL/$MAC_URL/$MAC_HUB_URL, then --fleet via ~/.mac/fleets.yaml. Nothing configured is an error, never a silent default.

Not every ControlPlane method is wrapped for hub mode. "X is not yet supported in hub mode" means the RemoteDispatch wrapper is missing, not that the feature is absent — the hub route usually exists.