Skills for Herdr.
Starts a Herdr tab in the right workspace, with an agent already primed on a task. The task can be a Linear or GitHub ticket, a chat thread or a plain description.
Ask for example: "start a new task for ABC-123" or "new tab in Support: look into the failing nightly export".
The skill reads your workspaces from ~/my-workspace.toml. If the file is missing, the skill offers to create it on first use. It lists your current Herdr workspaces, asks what each one is for, and writes the file after you approve it.
To write it by hand, copy the example. It documents every field.
Some agent kinds need launch steps that Herdr cannot do alone. Each has a harness file named <kind>.md. The skill looks in two places, and uses the first file it finds:
references/harnesses/in the skill, which shipscodex.mdandopencode.md~/.config/herdr-new-task/harnesses/, for your own kinds
- Workspaces match by label, not folder, because several workspaces can share a folder.
- Claude gets its prompt as a launch argument. Claude Code wraps a long typed-in paste in
<pasted_content>tags, and the agent then treats it as text from someone else. - Claude starts with
--permission-mode auto, so it does not stall on permission prompts in a tab nobody watches. Codex--approve-for-medoes the same for Codex.
- Herdr, with the skill running inside a Herdr pane
jq- For tickets: the Linear MCP server, or the
ghCLI for GitHub issues