Use the `osd` command to manage workbench Screens, sessions, and terminals. Open or split panes, find and read conversations, inspect agent terminals, and send messages to sessions or running Claude Code and Codex instances.
Use for anything about the workbench's own sessions, Screens, terminals and settings — opening a Screen, splitting it into several sessions or terminals, closing one; reading another conversation or the Claude Code / Codex in a terminal and messaging it; or acting on an "[OSD session]", "[OSD terminal]" or "[OSD screen]" reference the user pasted; or a setting the task needs changed (terminals, remote access, approval mode, proxy). All of it goes through the `osd` command.
Run score Not tested
See how to installReview the instructions the agent will receive before you install.
osd is on your PATH and already connected to this workbench. Three things
are reachable through it:
osd screen ls shows all of it at once, with the ids every other command
takes. osd session find <words> finds a conversation by what was said in it.
osd screen new --sessions 6 --name "Six" # a new Screen, split evenly among 6 new conversations
osd screen new ses_a ses_b --terminals 1 # existing conversations and a terminal, side by side
osd screen new --terminals 3 --run claude # three terminals, each starting Claude Code
osd screen add g12 --new # one more pane: a new conversation (or a SESSION_ID, or --terminal)
osd screen add --here --new # a new conversation right of you, on the Screen you are on
osd screen show g12 # bring it up in the user's window
osd screen close g12 # close it; the conversations are kept
New conversations and terminals start in your folder unless you pass
--directory DIR or --project NAME. A new Screen appears in the user's
tab bar right away, and the user stays on the Screen they were looking at.
| You have | Read it | Send to it |
|---|---|---|
a session ses_… | osd session show <id> --last 2 (its last two turns) | osd session send <id> "<message>" --wait |
| a terminal running Claude Code or Codex | osd term read <id> (--turns 3 for more) | osd term send <id> "<message>" --wait |
| a terminal running a plain shell | osd term read <id> (last 40 lines; --lines N) | osd term send <id> "<command>". This runs the command. |
osd session wait <id> waits for a turn someone else started and prints its
answer; osd session status <id> says whether one is running.
Read before you write: a message lands in the middle of someone else's work.
osd term read on an agent says whether it is working; a message sent then
waits in its queue and does not interrupt.
The other agent does not know who you are. Write every message so it stands alone: who you are, what you need, what to send back.
I'm the agent in the "Fold proteins" session. I need the path of the AlphaFold outputs you produced today. Reply with the path only.
When you tell the user about another conversation's files, give full paths. Their folder is not yours.
osd
says to stop and ask the user. Tell the user what you have done and ask
whether to go on.terminals are off means "Allow terminals" is off: ask for
remote.terminal=on (see Settings below).osd settings lists the app's settings and their values. You change none
yourself: you ask the user, once, for everything the task needs, with why.
osd settings set remote.terminal=on --reason "To run Claude Code in a terminal next to me."
The user sees each change in a dialog and approves all, some or none; the command waits and prints what was applied. A decline is final: do not ask again unless the user brings it up. Ask only for what the task needs. API keys and sign-ins are not settings here; the user enters them in Settings.
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Control native desktop apps and desktop browser windows through their accessibility elements, then read back the state to check what happened. Useful for work that requires a visible app and cannot be done through a shell or page-level browser automation.
Reviews analysis code for known scientific error patterns across physics, earth science, biology, chemistry, social science, and bioprocessing. It reports structured findings such as unit mismatches or unsuitable statistical methods; a clean review does not prove the code is correct.
Inspect large data files before reading their contents into an agent’s context. The skill returns a compact pointer with schema, shape, samples, and key values using bounded memory, so the agent can select specific rows, columns, or ranges to read.
Run compute jobs that need cloud CPUs, GPUs, or elastic capacity on Modal. The skill writes a function in the workspace, runs it with the user's authenticated Modal CLI, and saves outputs and run details for later traceability.
A minimal test skill that says hello. Use it to check whether a skill loads and whether the skill system is responding as expected.
Create or review charts, tables, and figures for papers using a consistent publication style. The skill checks color, layout, and readability at final size, and requires static exports for paper delivery; interactive Plotly or HTML can support exploration.