capabilities/skills/commander/SKILL.md

name, description, version

name	description	version
Commanders	Use when the user asks about a commander — what one is, how to start one, why one is not firing, how much one costs — or when they describe work that should keep happening while they are away ("watch X and do Y", "every time Z, run this"). Carries what a commander is, how the pieces actually fit, and how to help somebody arrive at one that works. Open the Officer platform skill too if the answer needs an endpoint.	1

What a commander is

An agent that keeps working when nobody is watching. It waits for something, acts on it, and reports.

Mechanically it is an ordinary chat session — same harness, same transcript, same socket — working in a directory that belongs to it: $DATA_PATH/<email>/commanders/<id>/, created before its first turn, holding a CLAUDE.md written at creation and a state/ for what it has done.

The directory IS the commander. The conversation is how you talk to it. That is worth stating early because it has two consequences people assume otherwise about. A commander is not a special kind of agent with special powers — it is a conversation in a folder, plus a loop somebody started from inside it. And it outlives any particular conversation: everything it knows is on disk, so a session that dies can be replaced in the same directory and the commander carries on.

Your CLAUDE.md is your own file. Extend it as the job is settled and whenever what you watch or what you do about it changes. It is read at the start of every session, so what you write there is what the next conversation knows.

The three pieces, and which one people get wrong

1. The conversation. Where you and it agree what the job is. This is the part to spend time on.

2. The wait. A shell loop, started as a tracked background job from inside that conversation. It polls, and when its condition is true it acts. The wait happens in the shell — not by the model waking up to check. A model-side tick re-reads the whole conversation to learn nothing, and costs a turn to do it.

3. The action. The loop cannot do judgement — it cannot read a plan, weigh a situation, or write a report. So when it fires it spawns a headless agent (claude -p) with a prompt telling it to read a procedure and follow it. That agent does the work and exits.

The piece people get wrong is 3. They expect the commander itself to wake up and act. It does not: it started a loop, and the loop starts somebody else.

Never arm on the first turn

Not a guideline. The first exchange is for establishing what to watch and what to do about it, and a watcher armed before that is a watcher armed on a guess — which fails silently in both directions: it never fires, or it fires constantly, and either way it is found out days later.

There used to be a mechanical interlock behind this, and it is gone: the paths a commander records itself at could not be resolved on turn one, so their absence WAS the signal. They are in your CLAUDE.md before the first turn now, written when your directory was created — so nothing stops you arming early except knowing better. The rule is the whole guard now rather than a backstop to it.

When you started yourself

Officer can start you with a button instead of a prompt. When it does, everything you have been told is a name and a REFERENCE PATH — the person has not said what they want yet, and they are waiting for you to ask. That is the whole reason the button exists: writing a good opening prompt is the hard part, and asking somebody to do it before they have talked to you gets it done badly or not at all.

Turn one is orientation, not questions. Read the REFERENCE PATH in your CLAUDE.md — that is what you were pointed at, and it is not where you live. Your own directory holds your CLAUDE.md and an empty state/; there is nothing to learn there and listing it teaches you nothing. What orientation means depends entirely on what the reference turns out to be — a repository, a folder somebody drops files into, a mailbox, a library of documents, a machine that runs one script. List it, read whatever explains it, and form a view. You are not looking for anything in particular; you are trying to stop being a stranger before you ask somebody to explain their work to you.

Your name is the brief. Whoever started you chose it before anything else and it is usually the only thing they have said — "Platform Deploys", "Invoice Inbox", "Backup Watch". Read this directory looking for what that name refers to, and let it settle what is in scope. A name pointing at deploys means the question you are working towards is about deploys; the fact that the repository is also mid-refactor is not your business and asking about it wastes the turn.

You are looking for something that HAPPENS AGAIN, not for what people here are doing. This is the trap, and a codebase sets it perfectly: a TODO.md, a list of open items and a week of commits are all right there, and reading them produces a confident, useless question about which piece of work matters most. That is a colleague's question. Yours is narrower — what event recurs here, and what does somebody do by hand every time it does? Current priorities are not a trigger. Nothing you can only do once is a commander.

Whatever that job turns out to be, look for whether it has been done here by hand already. It usually has, and the shape of it is on disk — a folder of previous runs, a procedure somebody wrote down, a script with the steps already in order, a naming convention nobody documented. Finding that is worth more than any question you could ask, because it turns "how should this work" into "is this still how you want it", and the second question gets a much better answer.

When they name a system, read its skill before you reason about it. Your skills index lists what is available here by description; the bodies load when you open one. So a system that Officer knows about — itself included — has its knowledge somewhere better than your guess, and reaching for it is one tool call. Do that instead of inferring how something works from its directory listing, and instead of asking them to explain something already written down. If there is no skill for what they named, say so plainly and ask; that is a real answer, and inventing a mechanism to fill the gap is not.

What you say next depends on what you found, and there are two cases.

The job is already written down here. A procedure, a folder of previous runs, a script with the steps in order — enough that you could do it yourself without guessing. Then do not interview anybody. Say what you understood, in your own words and not as a summary of files, and ask them to confirm or correct it. Reading it back is the whole check: they find out in one exchange whether you understood, and you get the correction while it is still cheap.

It is not written down, or only partly. An empty directory, a folder somebody drops things into, a job that lives in their head. Then ask what you should be watching for, and ask only that. Wait for the answer, follow it where it goes, and settle it before you ask what to do when it fires — the questions further down are the ones to work through, one at a time.

The honest test between the two: could I write the procedure from what is here, or would I be guessing? Guessing counts as not written down.

Both cases end the same way, and this is the part that is easy to leave out. Once it is agreed, do the first run yourself, and then arm. A question left hanging for the owner to turn into action is not a finished intake — they came here to stop doing this by hand.

You never cause the thing you are watching for. Not while setting up, not on a first run, not as a favour. If a backlog is sitting there and clearing it would trigger you, that is the owner's to start — say it is waiting, and leave it. The temptation is real and it is always polite: the work is genuinely there, and offering is genuinely helpful-seeming. It is still the wrong side of the line, and a first run you triggered yourself proves nothing about whether the watching works.

"Do the first run" means running the action against something that has ALREADY happened, so the action is known to work before it runs unattended. It never means manufacturing an occasion to run.

No formatting in that opening. No bullets, no headings, no backticks, no commit hashes, tag names, file paths or table names. If a specific belongs anywhere, it belongs in the OPTIONS — when you offer choices, they carry the detail and the question carries none. Those options were the good part of the message above; the preamble is what needed deleting.

Then work through the questions below, one at a time, in that order. Follow the answers where they go — this is a conversation and the interesting parts are the corrections. Do not move on from a trigger that cannot be reduced to a shell test; that is the one question where "roughly" is not good enough.

Before you arm anything, write the specification down and show it. Trigger, action, baseline, what must never happen twice, what a report contains. In their words, not a config dump. Two things come out of this: they catch the misunderstanding while it is still cheap, and you end up with the procedure file you were going to need anyway. Ask them to confirm it. Then, and only then, arm — and the rule below still holds, so it is not on this turn.

Helping somebody shape one

Most of the value is here, and it is a conversation rather than a form. Four questions, in this order:

What is the trigger, exactly? Not "when there is a new release" — what file, tag, row or output changes, and what command reveals it. If the trigger cannot be reduced to a shell test that exits 0 or 1, there is no commander yet, only an intention. Push until it can.

What happens when it fires? If the answer needs judgement, the action is a spawned agent and it needs a written procedure to follow. Say that out loud — a procedure in a file, versioned, is what makes the same the same job repeatable next month, and reviewable by somebody else. A prompt typed into a dialog is not.

What must never happen twice? This is the question nobody volunteers and it is the expensive one. A trigger that stays true after a failed action fires again, and again. Ask what the cost of a second run is: if it is money, or a duplicate message, or a half-applied change, the action must claim its work before starting and refuse a second attempt. One attempt per unit of work, enforced by something the loop cannot undo — a file created with mkdir, not a flag it sets itself.

What does "new" mean — and this one you must supply, not ask for. "Watch for new X" is not a condition, whatever X is — a tag, a file in a folder, a row, an unread message, an alert. A watcher armed today sees everything that already exists, and unless something says otherwise it will treat all of it as news and act on every item in turn. Nobody says "and ignore the ones already there", because to a person that is obviously implied. It is not implied to a shell loop.

So a watcher needs a baseline, and there are only two honest kinds:

• A record of what has been handled. Best where acting is expensive or must not repeat: the loop skips anything with a record and creates one before acting. This also survives a restart, which a remembered baseline does not.
• A snapshot taken at arming. Fine where the history is genuinely irrelevant — "anything that lands from now on". Cheap, and it forgets everything the moment the loop is restarted, so say that out loud.

Choose one and say which, in the condition itself. A watcher with neither looks identical to a working one until the moment it fires on a backlog.

What does "done" look like to you? A report is worth more than a success flag, and the parts worth reading are the ones that are hard to reconstruct: what it skipped, what it could not verify, what it left outstanding. Ask for those explicitly or you will get a summary of what went right.

Writing the loop, if you are the one writing it

The background-job dialog already sends the load-bearing mechanics — use the harness's own background mechanism, keep the wait in the shell, put the test into the shell condition, and ask first if the condition is not completely clear rather than arming a guess. That last one matters more than it reads: a wrong condition fails silently, either never firing or firing constantly, and it is found out days later. Arming late costs a minute.

Four details the mechanics do not carry, each of which was learned by getting it wrong:

• OFFICER_ROOT is empty in the sidecar's environment. Any path built from it resolves to /data/..., so anything the loop writes to record its own progress lands where nothing reads it — and a watcher whose record is invisible to itself has no baseline, so it acts on the same work for ever. Set it explicitly in the loop.

• Keep stderr out of a --output-format=stream-json pipe. The CLI prints warnings there — a connectors notice, among others — and one of them landing in the stream corrupts the first jq parse. Redirect it to its own reader.

• And filter that connectors notice before you narrate stderr. On this machine ANTHROPIC_BASE_URL points at a local proxy which holds the real credential, so ANTHROPIC_API_KEY is set to a token for that proxy — and the CLI sees a key set, assumes the worst, and warns on every single spawn:

  ⚠ claude.ai connectors are disabled because ANTHROPIC_API_KEY or another auth source is set…
  

  Nothing is wrong and nothing here uses connectors. But a loop that reports stderr whenever it is non-empty will report this every time, which trains whoever reads the job panel to skim the one channel that exists to carry real failures. Drop it and report what is left:

  if grep -qv 'claude.ai connectors' "$errlog" 2>/dev/null; then
    say "stderr: $(grep -v 'claude.ai connectors' "$errlog" | tr '\n' ' ' | cut -c1-200)"
  fi
  

  Filter by the known line, never by suppressing stderr wholesale. The reason to keep the channel is the failure you have not seen yet.

• jq --unbuffered, or the job panel shows nothing for as long as the work takes and reads as hung. A background job that produces no output is indistinguishable from a dead one.

• The claim goes before the spawn, not inside the action. If the action claims its own work, an action that dies before claiming has claimed nothing and will run again.

Telling the owner, without a second channel

A loop writes to its job panel, and nobody reads a job panel. So a watcher that fires while the owner is asleep has done its work and told no one, which from their side is indistinguishable from not firing.

Say two things in the CONVERSATION, and only two:

• When it fires. One line: what appeared, and that you are handling it. Nothing else — the detail is in the job panel by definition, and repeating it in chat means both places are half-read.
• When the work ends. A short report: what was done, what was skipped, what went wrong and what the owner has to do about it. "Five servers, nothing to report" is a complete message. So is "four went, one needs a console and here is why".

Nothing in between. A progress commentary in chat is a notification per step, which is how somebody turns notifications off.

How, with nothing new

You cannot post into your own conversation from a shell loop — you only speak when a turn runs. But you can cause a turn, and both halves of that already exist:

1. The child writes its report to a file, named for its own session id, somewhere under your state directory. Same machine, same user, no authentication to arrange.
2. The loop inserts a row into nudges for your own transcript id, over the local Postgres socket, with a message telling you what happened and where the file is. The nudger ticks every 60 seconds, wakes this conversation, and you read the file and write the report yourself.

So the shape is: the loop knows the FACTS and puts them in a file; you provide the JUDGEMENT and put it in the chat. It costs one turn per event, which is exactly what buys a summary rather than a pasted log tail.

Three things about that mechanism, each of which will bite otherwise:

• nudges holds one row per (user, session), deliberately. A second schedule for the same conversation produces doubled prompts, so overwrite your pending nudge rather than queueing behind it. Two events in the same minute become one turn covering both, which is the better message anyway.
• The nudge message is a PROMPT you will be handed, not a notification the owner sees. Write it to yourself: what fired, where the file is, and that you are to report briefly. What the owner reads is whatever you say on that turn.
• Up to 60 seconds late, because the nudger polls on a minute. Fine for anything a watcher does; if something genuinely needs to be instant, it is not a watcher.

Once you have spoken, the ordinary chat notification path takes it to the owner's devices — so the phone tap lands in this conversation, with the job panel one step away. Nothing to build for that either.

Things that are true and surprising

• Its cost line understates. The recorded cost counts the commander's OWN turns — one row per result that passes through the platform's chat socket. A spawned agent's cost never touches that socket, so a commander that spent several dollars through its children can honestly display cents. Do not quote the number as total spend.
• A spawned agent cannot be found from its parent. Nothing links them: the child's transcript looks exactly like any other session, and entry types that appear to mark one (queue-operation, ai-title, last-prompt) are the harness's own and appear in both. If a link is wanted, the loop must print or record the child's session id itself — there is no way to recover it afterwards.
• A child transcript is not permanent. It is readable while it lives and for a while after. Two from a previous day were gone. Anything worth keeping must be captured while it runs.
• Reach what you are watching by the shortest path available. A public hostname for something sitting on the same machine goes out and back through whatever fronts it, and shared-IP rate limiting there refuses connections for long stretches — presenting as an access-rights error, which it is not. Minutes of detection latency become hours. The Officer-specific instance of this is in the Deploying Officer skill.
• The platform may not know a watcher exists. Its bookkeeping engages only for the fire behaviour that hands back and re-arms. Started another way, the loop runs perfectly and the row still reads unarmed, so the UI shows no watcher and has nothing to disarm.

When something is not firing

In this order, because each rules out the one below:

1. Is the loop alive? It is a tracked background job — look at the job, not at the agent.
2. Is its condition ever true? Run the test by hand, in the same shell form the loop uses.
3. Can it reach what it polls? An access-rights error on a repository or an endpoint that plainly exists is usually a rate-limited public path, not a credential.
4. Has it already claimed this work? A refusal to act twice looks identical to a failure to notice.
5. Only then the model, the prompt or the procedure.

Where to look next

For endpoints, permissions or why something is refused, open the Officer platform skill. For a deploy — the worked example of everything above, with a real procedure and a real fleet behind it — open the Deploying Officer skill.