Work start

Put the session in the right worktree with the full picture of what is asked, and stop before implementing. Run every wt and gt command from the repo the user is in; wt acts on the clone, so any worktree of it works.

Steps

1. Parse the target. pr <n> (or a PR URL) is a pull request; a bare number is an issue. Done when you hold one number and its kind. Without a number, ask for one.

2. Read it.

   • Issue: df-ti <n>. When gt issue view <n> --json kind --jq .kind prints pull, the number is a pull request: switch to the PR branch of every later step.
   • PR: df-tpr <n> prints its state, base and head branch, body, comments, inline review comments, and review bodies. Read each issue its body lists under resolves: or Related to.

   When the issue or PR has an image you cannot fetch, ask the user to paste it. Done when every ask, every unresolved review comment, and every linked issue has been read.

3. Find existing work. List the worktrees with wt list --format=json and the PRs with gt pr list --state all -L 200 --json number,state,head,title,body.

   • Issue: a candidate is a worktree or PR whose branch starts with jd-<n> (either jd-<n>- or jd-<n>/), or a PR whose body resolves #<n>.
   • PR: the candidate is the worktree whose branch equals the PR's head branch.

   wt list --format=json prints an object (schema, repo, items, …), not an array; the worktrees are under items: jq -c '.items[] | select(.branch == "<branch>") | {branch, path: .worktree.path}'. On PR #925 a script that iterated the top level crashed with 'str' object has no attribute 'get'.