[JacobStephens2]

Summary

Running scripts/release.sh [--review] <version> again carries on from the first step not yet done. It keeps main's refusals (#157) and --review for a fresh run.

 main
   fetch_main
+  find_earlier_run          # newest PR from release-<version>, else the branch on origin
+  stop_if_tagged            # after an earlier run: tag on the merged bump → "already tagged", exit 0;
+                            #   tag elsewhere on origin → refused
+  case PR state
+    none, no branch → commit_bump (check_can_release, bump in a worktree)
+                      write_summary (--review: n → "nothing was pushed")
+                      push the branch, open_pr, wait_and_merge
+    none, branch    → write_summary (--review: n → "no PR was opened"), open_pr, wait_and_merge
+    OPEN            → wait_and_merge
+    MERGED          → use the PR's merge commit
+    CLOSED          → refused: reopen it or delete the branch
   push the tag on the merge commit

On a resume, check_can_release is skipped because the first run already passed it, and after the merge the version is no longer higher than main's. With no earlier run, an existing tag is still refused by check_can_release, in main's order and wording.

The fake gh's pr list now takes --head (exact match) and reports headRefOid and mergeCommit. The README's Releasing section gains a paragraph on resuming.

Closes #158

Evidence

• Before: against the old script, 6 of the first 7 resume tests failed. The seventh (a tag with no bump PR) hung waiting on CI.
  After: cargo test --test release passes 35 of 35, main's refusal and review tests included. The full cargo test passes, and cargo clippy --all-targets and shellcheck scripts/release.sh are clean.

branch pushed, no PR               → PR opened on the existing head, merged, merge commit tagged
branch pushed, no PR, --review n   → "stopped at the review, so no PR was opened", origin unchanged
PR open                            → still one PR, head unchanged, merged, tagged
merged, no tag                     → the merge commit is tagged, no second merge
tag on the merged bump             → "v0.2.0 is already tagged…", exit 0, origin refs and gh state unchanged
  (lightweight and annotated)
tag elsewhere / PR open / no PR    → "refusing to release 0.2.0: v0.2.0 already exists on origin…", exit 1, origin untouched
PR closed unmerged                 → refused, origin untouched

Merge Danger

Door: two-way

Only the maintainer's release script, its test fakes and the README change. Nothing ships in the binary.

Blast Radius: release-tooling

A wrong resume decision could tag the wrong commit. The tag is only ever pushed onto the merge commit of the release-<version> PR, and an existing tag anywhere else on origin is refused.

Unaddressed findings

Standards

• Steps return by setting globals (pr_state, url, head, merge, summary, diff), which is a Data Clump and gives the steps a hidden order. This matches how the script already shares version/tag/branch, and bash has no clean way to return several values.
• "earlier run" and "resuming" reuse CONTEXT.md's Run and Resume. release.sh is outside the factory's domain, and those are the plain words for what it does.
• The jq string is copied between release.sh and the fake (PR_LINE_OF_NEWEST). This follows the fake's existing WORKFLOW_RUN_LINES pattern. The fake fails loudly on any other --jq, so the two can't drift silently.
• The resume tests repeat the push_bump / open_bump_pr / merge_bump_pr setup. Each test spells out the interrupted state it sets up. A combined helper would hide which steps the earlier run got through.
• format!("release-{version}") is built in three test helpers. It's a small, fixed naming rule. A helper for it would save little.

Spec

• A PR closed without merging is refused, which the spec didn't ask for. Kept, because silently opening a new PR on a branch the maintainer closed seemed worse than stopping with a hint.
• A resumed branch is trusted without checking that it is a bump on current main. Only an earlier run creates that branch. Its PR is merged with a merge commit and CI runs on it, so a branch based on an older main still tags what CI tested.
• Only the newest PR from the branch is read, so an older merged PR behind a newer closed one is refused. A second PR for the same version only appears if someone opens it by hand. Refusing is the safe outcome.
• A resume doesn't refuse a local v<version> tag. The script pushes the tag from the merge commit's SHA, never from a local tag, so a local tag can't change what gets tagged. The docs now say "elsewhere on origin".
• A pushed branch with no PR gets a new summary, and with --review is asked about again. The PR needs a summary, and the first run's summary was never saved anywhere.
• The branch history has an empty "thirdshift: failed run" commit. Removing it would mean rewriting history that is already pushed to the open PR.

🤖 Generated with Claude Code