Add Claude Code, split connectors into triggers, workflows, and actions, and rework the UI

[imsobear]

Adds Claude Code as an agent, splits what a connector offers into triggers, workflows, and actions, and reworks the app UI around shorter forms and denser lists. Also updates the docs and site to the new story.

Agents and runners

• Claude Code is a third agent next to Codex and Cursor Agent.
• A loop's folder is now relative to the runner's home (code/web), and each runner resolves it. Folder loops no longer have to run on Loopable's host. Older absolute paths still run there, and a migration turns paths under a home folder into relative ones.
• The Runners page shows the install line with the join command.

Triggers, workflows, actions

• Each connector registers what can start a loop (Starts when), ready-made jobs on those triggers (Workflows), and where an answer can go (Send to). Trigger ids are the old workflow ids, so saved loops and polls are unchanged.
• Custom loop: pick any trigger and write your own prompt.
• GitHub: "Review pull requests" and "Reply to issues". Plan and Implement are gone (the code-writing path is still in place, unused, and covered by tests).
• Gmail: "An email arrives" is a trigger, and "Email this account" sends the answer to the connected inbox, in the same thread when the run started from an email. Gmail now asks for gmail.send, so existing connections need reconnecting to send.
• "Keep in Inbox" shows as Inbox only rather than under Schedule.
• Slack connections are named @bot · workspace.

UI

• Shared page header, form sections, and list rows; cards no longer pad their tops twice.
• Loop page: Settings and Tasks tabs as links (?tab=tasks), on/off and Check/Run in the header, items that were not run folded into one line with a Run on each.
• Loop form: When / Agent / Answer, the folder with the agent, no team notes (a migration moves existing notes onto the prompt), and fixed trigger details folded away.
• New loop: Custom loop and Schedule first; connector pages list their loops; Inbox asks an agent from a dialog; product logos for connectors and agents.
• Inbox prompts finish as done instead of prepared (with a migration for earlier ones). Hydration warnings from folds and relative times are gone.

Docs and site

• README, narrative, concepts, connectors, and deploy describe the shared-agent story, triggers/workflows/actions, custom loops, Gmail, and relative folders.
• The site is rebuilt around the problem, a shared agent, and running on your own machine, with Claude Code, custom loops, and Gmail in the loop list. Not redeployed from this PR.

Releasing

• CI (.github/workflows/ci.yml) runs typecheck, tests, and both builds on every PR and on main. Publishing also runs the tests now, and pnpm is pinned.
• pnpm release patch|minor|major publishes only from a clean main that matches GitHub: it runs every check, bumps, commits, tags vX.Y.Z, publishes, and pushes. --dry-run checks without changing anything.
• Runners report their version and a protocol number. Folder jobs only go to runners that understand relative folders, a loop with only old runners says to upgrade, and the Runners page flags old runners. Older runners report nothing and count as protocol 1.

Testing

• pnpm typecheck and pnpm test (364 tests) pass.
• Checked each page in a running pnpm dev, at desktop and phone width.

Migrations

0025_prompt_tasks_done, 0026_guidance_into_prompt, 0027_relative_folders, 0028_runner_version run when the dispatcher starts. Upgrade runners on other machines too (npm install -g loopable-cli): an old runner keeps working, but no longer takes folder jobs.

🤖 Generated with Claude Code