lab-edison: cited answers from the Edison platform¶

Edison is a research platform, run by Edison Scientific. It answers questions from the published literature with citations, runs code on a dataset you upload, and in Kosmos has a research agent that works a whole goal over many rounds. This skill drives it from your machine through edison-cli, a command that ships with the plugin. No cluster is involved. It is the one skill here that talks only to a cloud service.

You ask for it by name¶

The skill never loads on its own. The other three trigger from what you say; this one waits.

/lab-edison what is known about ...

Anything else you would put to a researcher goes after that same command: "Has anyone already done this experiment?", "Analyse this table and tell me what it shows", "Set up my Edison key on this machine."

Every run spends credits from your own account. That is why nothing here starts without you.

Your key, and where it goes¶

You need an Edison API key. It lives in one file on your own machine, ~/.claude/lab/edison.env:

mkdir -p ~/.claude/lab
printf 'export EDISON_PLATFORM_API_KEY=PASTE-YOUR-EDISON-KEY-HERE\n' \
  > ~/.claude/lab/edison.env
chmod 600 ~/.claude/lab/edison.env

Then open the file in your editor and paste your key over the placeholder. You do not have to remember any of this: when something is missing, the check the skill runs first prints those exact commands back to you. templates/edison.env is the same file with the reasoning in its comments. The variable name is the one the client reads and no other, and the chmod leaves the file readable by you alone.

The key sits beside personal.md, never inside it. An agent reads personal.md into the chat, so a key written there would be copied into every transcript you ever record.

Before it does anything, the skill checks that file. If it is missing, still holds the placeholder, or is readable by other people, the agent stops and says which check failed. It offers to write the file for you, with the placeholder in place, and then hands it back to you. It will never ask you to type or paste your key into the chat.

What happens before it spends¶

You see the job and the query. The skill names the job it picked and prints your question in full, before the task goes out.

The costly runs stop and ask: a deep literature run, a data analysis run, and any batch of more than three tasks. A follow-up rides on the run you just read, instead of paying again for the same background.

No prices appear anywhere in this repo. They change, and a stale number is worse than none at all. Check the balance on the platform when the cost matters.

A run takes minutes, and the id is the receipt¶

Edison runs are slow enough that nobody should sit and watch one. The skill sends the task, prints the task id straight away, and then checks on it. You get your terminal back.

That id outlives the session that made it. If your laptop sleeps, the window closes, or the agent stops early, the run carries on and finishes on the platform. So a run is never really lost: the skill can list the tasks on your account, newest first, and find the one you started. It offers to do that before asking the same question twice, because asking twice means paying twice for an answer you already own. It can also cancel a run you started by mistake, which is worth knowing before a long data analysis job.

Kosmos, for a goal rather than a question¶

Kosmos is the platform's heavyweight agent, and the reason to come here when what you have is bigger than one query. It is not a task you send off. You pick a persona, make a project inside it, and talk to it there. Kosmos splits your goal into many smaller tasks and works through them in rounds until it decides it is done. One real project ran thirteen of them.

That shape is also what makes it cost. A run pays for every task it fans out to, so there is no flat price to quote, and the skill will not start one unless you ask for that run in plain words.

Everything it does around a run is free. It drafts the goal with you, checks your data is in a shape Kosmos can use, and shows you a run you already started, task by task, without opening a browser.

Names, not long ids¶

The platform gives every persona and project a long random id. You do not have to keep track of them. Say the name, and the skill finds the id for you. If nothing matches that name, it stops and says so. If more than one thing matches, it stops and lists them all with their ids. It never guesses.

Names it has looked up are kept in a small file next to your key. That file holds names and ids and nothing else. Anything that spends money, or removes something, skips it and asks the platform again. So a name that has since moved to a different project cannot point the skill at the wrong one.

Deleting a project is the one step here you cannot undo. The skill prints the project, its name and how many runs it holds, and asks whether the run history goes with it. Nothing is removed until you say yes.

One question, one job¶

For a single question the skill picks one of the platform's one-shot jobs and routes to it.

Ask in plain words and the skill picks for you. Say which one you want and it uses that.

Where your data sits¶

Your own machine, by default, because that is where your key belongs. Data on your laptop uploads and runs with nothing to arrange. Data on a cluster comes down first and uploads from here, so your key stays off shared hardware. You can run Edison on arc, but only after you put the key there yourself, in your own session. The skill never copies your key anywhere: not to a cluster, not into a job script, not onto a command line.

What it needs, and why the first run waits¶

You need pixi on your machine. It is the only extra tool this skill asks for, and no other skill here asks for it.

Nothing else has to be installed. The first Edison command builds the environment it runs in, and dozens of packages download before your question is even sent. It can look hung. It is not.

That environment lives inside the plugin's own folder, not in any project of yours. An update puts the new version of the plugin in a fresh folder, so the build runs once more after each update. It is quicker then, because the downloads are already on your disk.