[JacobStephens2]

Summary

When the highest-numbered Issue branch's PR is merged or closed, the Run now starts fresh on the next number instead of stopping with "numbered Issue branches are not supported yet".

 select(issue)
   highest = max number seen on origin branches ∪ PR heads   # one ls-remote, one pr list
   pr = newest PR for highest
   match (branch on origin?, pr)
+    (_, merged | closed)  → Fresh on issue-<n>-branch-<highest+1>   # even if the branch was deleted
     (yes, none | open)    → Continuation
-    otherwise             → error "not supported yet"
+    (no, open)            → error "has an open PR but is gone from origin"

A Fresh numbered branch gets the fresh prompt, is created from the Base branch, and must not already exist locally, the same as issue-<n>.

Tested seam: the thirdshift CLI (tests/numbered_branches.rs). The tests cover merged → -branch-2, closed → -branch-2, deleted merged issue-7 → -branch-2, open PR on -branch-2 → Continuation, merged -branch-2 → -branch-3 (including when it was deleted), fresh prompt, PR head and "Closes #7", and a local -branch-2 stopping the run.

Evidence

• Before: issue-7 with a MERGED PR → exit 1: issue-7 can't be continued (its PR …/pull/1 is merged), and numbered Issue branches are not supported yet (7 of 8 new tests failed).
  After: the Run implements on issue-7-branch-2 from main and prints …/pull/2. cargo test: all suites pass (9/9 in numbered_branches).

Merge Danger

Door: two-way

Selection logic only; reverting restores the old error.

Blast Radius: small

Only affects Runs on issues whose latest Issue branch PR is merged or closed. Those Runs used to fail during pre-flight.

Unaddressed findings

Standards

• Shared parameterised AGENT_COMMITS_AND_OPENS_PR in tests/support (Duplicated Code): refactoring fresh_run.rs is outside this issue. The local helper needs the branch as a parameter.
• Helper for the repeated ["Add feature", "Initial commit"] assertion (Duplicated Code): the testing guidance asks for expected values spelled out from the scenario, and a literal log reads better than a helper.
• Scenario::last_pr() helper for gh_state()["prs"]…last() (Message Chain): it has one use, so a helper would be speculative.

Spec

• New error when the highest branch has an open PR but is gone from origin (not in spec): it replaces the old catch-all error for a case GitHub rarely produces, because deleting a head branch closes its PR. Flagged here so you know about it.
• The newest PR on a branch decides, and that isn't documented in CONTEXT.md or ADR-0002: this behaviour already existed from Continuation of an existing Issue branch #6, so documenting it is a separate domain-docs decision.
• A Fresh numbered run bases on the checked-out branch, not the merged PR's base: this is correct under the CONTEXT.md definition of Base branch, so nothing to change.

Closes #7

🤖 Generated with Claude Code