Claude Code and Codex
Hand a post to a coding agent. Work on it, on a post's page, opens Claude Code or Codex with the post written as the prompt, in the repository of its GitHub issue. Nothing runs until you press Enter in the agent.
Start an agent on a post#
Open the dialog#
On a post's page in the dashboard, click Work on it. The dialog shows the Prompt, which you can edit, and, where your workspace takes issues in GitHub repositories, the Repository the agent will work in.
Select where the agent opens#
Under Open in, select a tile. Each one names the agent and where it opens (see the list below).
Open the agent#
Click the button under the tiles, such as Open Claude Code or Open ChatGPT. Your browser opens the agent's link, and the agent shows the prompt in its input box and waits. Read the prompt there, then press Enter.
Swingback sends nothing itself: your browser opens the agent, and the agent runs under your own Claude or ChatGPT account, which counts it like anything you type there. The dialog remembers your choice of tile while the tab stays open.
Work on it is there for every member of your team, on every plan, on any approved post. It doesn't show on a post that's held for review, rejected, in the trash or merged.
What the button opens#
- On the web (Claude Code): a session on claude.ai/code, in a new tab. You need a Claude account with Claude Code on the web, your GitHub connected there, and an environment for the repository.
- Desktop (Claude Code): a new session in Claude's desktop app, where you choose the folder or the environment.
- Terminal (Claude Code): a terminal on your computer, in a clone of the repository where you've run
claudebefore, or else in your home folder. Claude Code warns that the prompt came from a link. It handles these links only after you've sent it a first prompt in an interactive session. - VS Code (Claude Code): the Claude Code panel in VS Code, with the extension installed and the project open.
- ChatGPT (Codex): a new chat in the ChatGPT desktop app, in the workspace whose Git remote matches the repository. The repository must be one of the app's workspaces.
- Other: Copy prompt copies the prompt, to paste into any other agent.
What the prompt contains#
- The post's number and its address in the dashboard.
- Its title and up to 2,000 characters of its text, as its author wrote them, marked as a description of a need and never as instructions. Never the author's name, email or company.
- For a post that not everyone can see: an instruction to describe the change in the agent's own words, and to copy nothing from the post into commits or the pull request.
- Its GitHub and Linear issues, five at most, each with its title and address, for the
agent to read before it starts, since your team may have written more there. Also the words that close them from a
pull request: for GitHub, such as
Fixes acme/web#123; for Linear, such asFixes ENG-123, through Linear's own GitHub integration. - On Pro, where the MCP server is available: an instruction to read the post's comments there and, once the pull request is open, to leave an internal note on the post with its link. Never to change the post's status.
A prompt is at most 5,000 characters, the most every agent's link accepts. Longer text is cut; on Pro, the agent can read the whole post on the MCP server.
The repository#
The dialog picks the repository of the post's first GitHub issue, or else the only repository your workspace takes issues in. If there are several, select one, or select None: the agent then works wherever it opens. A Linear team doesn't name a repository.
When the work is done#
The agent opens a pull request for you to review. When you merge it, it closes the issue it names, and Swingback updates the post as your connection's settings say (see GitHub and Linear). The post's followers hear about it then, not while the agent works.
