nc://coder

Coder: run a command inside your own remote Coder workspace, and list or read those workspaces. A Nodecode rewrite of perqin/hermes-plugin-coder by perqin.

The original replaced the agent's own terminal with a Coder workspace, so every command ran there instead of on the machine you sit at. A cell has no terminal to replace, so the same reach is three tools the model calls:

Ask "what is the build saying in my dev workspace?", "run the tests in data and tell me what failed", "is my workspace agent up?" — the model picks the workspace with coder_workspaces, and runs the command with coder_exec.

Get it

Tell the agent "add the coder cell", or run nodecode add nc://coder.

Set it up

You need a Coder deployment you can reach over HTTPS (yours, or your organisation's) and an existing workspace in it. This cell does not create workspaces; it uses one.

1. A session token. In Coder: your avatar, Account, Security, Session token, and copy it. Put it in an environment variable, e.g. CODER_API_KEY, and set token_env to that name; or save it in a file and set token_file to the path.
2. The deployment's URL, scheme included and nothing after the host: https://coder.example.com.
3. A workspace name, e.g. my-workspace: the one coder_exec runs in when a call names none, and the one to check first with coder_workspaces.

Optional: forward_env lists environment variable names whose values are copied into the shell for one command (never the session token itself), startup_timeout bounds how long a stopped workspace may take to start, and timeout bounds one command.

Use it

coder_exec {"command": "git -C ~/app status --short"}
coder_exec {"command": "pytest -q", "workspace": "data", "timeout": 600}

A stopped workspace is started first, which can take minutes; the answer says so with started. The command runs as bash -lc inside the workspace, so it sees that workspace's files, tools and network, and its exit code is the workspace shell's own. The command rides the PTY's query string, so keep it under a few thousand characters: for anything longer, write a script inside the workspace and run the script.

The tools are the coder package's functions, so the same thing from eval:

(coder:run-command (nlk:contract-decode (nlk:contract 'nodecode-coder::command-run)
                                         (nlk:decode-json "{\"command\":\"ls\"}")))