[JacobStephens2]

Summary

Continuation (ADR-0002): if an Issue branch is already on origin with no PR, or with an open PR, the Run picks it up instead of starting over.

 run
   pre-flight: Origin match, issue open, git identity
-  base = checked-out branch (detached HEAD → error)
-  worktree: new issue-<n> from origin/<base>
-  fresh prompt
+  branch::select                      # git ls-remote + gh pr list --state all
+    highest used Issue branch, on origin, PR absent/open → Continuation
+    none used                                           → Fresh issue-<n>
+    PR merged/closed or branch gone                     → error (numbered branches: #7)
+    local copy of the chosen branch ≠ origin's         → error, nothing created
+  base = open PR's base, else the checked-out branch (note on stderr when they differ)
+  pre-flight on base: exists on origin, local copy (if any) not ahead
+  Fresh        → worktree -b <branch> origin/<base>, fresh prompt
+  Continuation → worktree -B <branch> origin/<branch>, continuation prompt
+                 ("Update PR <url>" form with an open PR, else "Create a pull request")
   session → push → check PR

src/
├── branch.rs     # new: Issue branch selection, Base branch, local-branch check
├── preflight.rs  # Base branch checks split out to run after selection
├── github.rs     # + gh pr list; PrState enum; shared gh JSON helper
├── prompt.rs     # + continuation prompt
├── worktree.rs   # + continue_existing
└── run.rs        # wires it together
tests/
├── continuation.rs   # new
├── fakes/gh          # + pr list (head: prefix search), call recording
└── support/mod.rs    # + origin_has_branch, github_has_pr, launch_checks_out, gh_calls, first_prompt

Tested seams

• The thirdshift CLI (the one seam agreed in the spec Spec: thirdshift, turn a GitHub issue into a ready-for-review PR unattended #2): tests run the compiled binary against real git and fake gh/claude, and check stdout, stderr, the exit code, origin's branches, the fake GitHub state, and the prompt claude received.

Evidence

• Before (main): thirdshift <issue 7> with issue-7 already on origin creates issue-7 fresh from main. The push is rejected (non-fast-forward) and the run exits 1.
  After: cargo test: 12 in tests/continuation.rs + 11 in tests/fresh_run.rs + 10 in tests/preflight.rs, all passing:

  an_issue_branch_with_no_pr_is_continued                       # commits: Continue ← Earlier ← Initial
  without_a_pr_the_continuation_prompt_asks_for_one              # exact create-PR form
  an_issue_branch_with_an_open_pr_is_continued_and_that_pr_is_the_result
  with_an_open_pr_the_continuation_prompt_asks_to_update_it      # "Update PR …/pull/1 …"
  the_open_prs_base_is_the_base_branch_whatever_is_checked_out   # develop checked out → main + stderr note
  without_an_open_pr_the_checked_out_branch_stays_the_base_branch
  an_open_pr_supplies_the_base_branch_even_on_a_detached_head
  a_local_issue_branch_that_differs_from_origin_stops_the_run    # no session, worktree, temp dir or log
  a_local_issue_branch_missing_from_origin_stops_the_run
  a_local_issue_branch_matching_origin_is_continued
  a_failed_runs_branch_is_continued_from_its_failure_commit
  selection_asks_github_once_for_the_pr_history                  # gh calls == [issue view, pr list, pr view]
  another_issues_branch_and_pr_do_not_count_as_this_issues       # issue-70 ≠ issue-7 (fresh_run.rs)
  

  I checked that the local-branch tests aren't vacuous: with the local-branch check disabled, both "stops the run" tests fail.

Merge Danger

Door: two-way

The change adds no stored state or migrations. A Continuation only adds commits and pushes; it never rebases or force-pushes. The one destructive step (worktree add -B, then the existing branch -D at cleanup) runs only after selection has checked that the local Issue branch matches origin's.

Blast Radius: runs

Every Run now makes a gh pr list call before the worktree exists. A gh failure there now stops a run that used to be fresh. A merged or closed PR on issue-<n> now errors instead of starting over on issue-<n> (that case gets numbered branches in #7).

Unaddressed findings

Standards

• S5, duplicated text between the fresh and continuation prompts: each prompt stays one literal that matches the spec's text line for line, so it can be checked against Spec: thirdshift, turn a GitHub issue into a ready-for-review PR unattended #2 at a glance. The two differ in layout (single vs blank-line separated), so sharing a tail would split both.
• S8, Worktree::add takes raw git flags (-b/-B): private, two callers, and the flags say exactly what git does. An enum would only rename them.
• S2 (part), the second match on Selection in run.rs for worktree + prompt: the Base branch logic moved onto Selection (the domain rule from ADR-0002). Building the worktree and prompt stays in run.rs, which is where the Run's steps live.
• S3 (part), an IssueBranch type: the repeated format! is gone (everything goes through branch_name). A dedicated type would only wrap two private functions used in one module.

Spec

• P1, the Base branch pre-flight checks (exists on origin, local not ahead): resolved by merging main (Pre-flight checks #4). Pre-flight is split: Origin match, issue open and git identity run first; the Base branch checks run after selection, on the base selection picked, and still before anything is created. A detached HEAD no longer blocks a Continuation whose open PR supplies the Base branch, and a PR base with no local copy skips the "ahead" check.
• P2, no test counts git ls-remote calls: git is real in the harness, and counting its network calls would mean tracing git internals. Selection makes exactly one ls-remote call by construction; the gh side is asserted.
• P3, a local issue-<n> with no origin copy is an error: kept on purpose. User story 16 says local-only commits are never destroyed, and the run would otherwise delete that branch at cleanup.
• P4, a merged/closed PR or a vanished branch is an error: a placeholder until Numbered Issue branches after a merged or closed PR #7 (numbered Issue branches, blocked by this issue). Reusing issue-<n> there would reopen finished work, which ADR-0002 forbids.
• P5, fork PRs are filtered out (isCrossRepository): a fork's issue-7 is not this repository's Issue branch. The filter costs nothing and avoids a wrong Continuation.
• P6, the prompt names <base>, and the local <base> may be stale or missing when the PR's base isn't checked out: the prompt wording is the spec's verbatim, and the same staleness applies to fresh runs. Fixing it means either creating or moving the user's local branches or changing the spec's prompt; worth deciding in the spec, not here. origin/<base> is fetched so the agent can use it.
• P7, gh pr list --search head: relies on GitHub's search index, which can lag: the alternative, listing every PR without --search, grows with the repo's whole PR history. The lag is seconds, and the only case it affects is a PR merged moments before the Run, which matters for Numbered Issue branches after a merged or closed PR #7. Revisit there if needed.

Closes #6

🤖 Generated with Claude Code