[JacobStephens2]

Summary

scripts/release.sh <version> goes from a version number to a pushed v<version> tag, which starts the existing dist release workflow.

release.sh 0.4.0
  fetch origin/main → temporary detached worktree   # maintainer's checkout untouched
  bump version in Cargo.toml + Cargo.lock only → commit → push release-0.4.0
  body = "## Summary" + <!-- release-summary:start --> generated notes <!-- release-summary:end -->
         + "## Version diff" (```diff written by the script)
  gh pr create → "release: opened <url>"
  poll actions/runs?head_sha=<head> until every workflow run completed → "release: waiting for CI on <url>"
    any not success/skipped/neutral → "CI failed on <url> (<names>) …" exit 1, PR left open, no tag
  gh pr merge --merge --match-head-commit <head> → "release: merged <url> as <sha>"
  git push origin <merge>:refs/tags/v0.4.0 → "release: tagged v0.4.0 on <sha> and pushed the tag"

It waits on workflow runs rather than check runs because release.yml also runs on pull_request, and its build jobs only appear after plan finishes. The generated notes cover the PRs merged since the last v* tag (previous_tag_name), with target_commitish set to the origin/main commit.

 tests/fakes/gh
+  gh api repos/<repo>/releases/generate-notes   # notes from merged PRs' mergeCommit reachability
+  gh api repos/<repo>/actions/runs?head_sha=…   # served from the sha's check runs
+  a merged PR records mergeCommit; {owner}/{repo} is accepted as the fake's repo
 tests/release.rs                                # black box vs a bare origin, user git config kept out
 README.md  ## Releasing                         # rewritten around the one command

Evidence

• Before: there was no release script. Releasing was the three manual README steps.
  After: cargo test --test release passes 6 tests:

  • full release: the diff is only version in both files; the merge commit's parents are (old main, bump head); v0.2.0 on origin points at that merge commit
  • the PR body's marked summary holds the notes since v0.1.0 (and not the PR from before the tag), followed by the version diff
  • progress lines appear on stderr in order and stdout stays empty
  • a red check exits non-zero, leaves the PR OPEN, doesn't move main and pushes no tag
  • the maintainer's wip branch, local edits, untracked files, unpushed commit, branch list and worktree list are all unchanged, and none of them leak into the release
  • a failure to generate the notes stops the script before any PR is opened

• The full cargo test, cargo clippy --all-targets -D warnings and shellcheck scripts/release.sh are clean.

• Real stderr from the test run:

  release: bumping the version to 0.2.0 on release-0.2.0 from main at c6d618f
  release: opened https://github.com/JacobStephens2/thirdshift/pull/3
  release: waiting for CI on https://github.com/JacobStephens2/thirdshift/pull/3
  release: merged https://github.com/JacobStephens2/thirdshift/pull/3 as dd818b7
  release: tagged v0.2.0 on dd818b7 and pushed the tag
  

Unaddressed findings

Spec

• RELEASE_POLL_SECONDS is an extra knob. It is kept as the test seam for the poll interval and documented in the script header. A maintainer never needs to set it.
• previous_tag_name comes from local tags (git describe). Fetching main auto-follows tags, so local tags normally mirror origin. Fetching origin's tags into the maintainer's repo risks clobbering their local tags. The refusals ticket is a better place to revisit this.
• {owner}/{repo} is ambiguous in a clone with several remotes and no gh repo set-default. This clone has a single origin. Deriving GH_REPO from the origin URL would need a GitHub-shaped origin in the tests. Left for when that case comes up.
• The script could read the merge commit from gh pr view --json mergeCommit instead of git. The reviewer marked this "not a bug". Finding the merge on origin/main whose second parent is the head is exact, and it only reads from git.

Standards

• Marker comments are speculative. The ticket asks for "the marker that the later agent summary (Release script: agent-written summary in the bump PR #159) and the Release page step will use", so they stay.
• The fake's jq expression is coupled to the script's. That is deliberate. The fake already matches exact --jq strings so an unexpected call fails loudly.
• isolated()/git() duplicate the helpers in site_deploy.rs and support/mod.rs. This follows the existing per-file prior art. The name isolated stays because, unlike without_user_git_config, it also sets the git identity.
• The setup (new → ci_reports(GREEN) → run) repeats across tests. The repeated Arrange step keeps each test readable on its own. The duplicated headRefOid lookup was extracted into pr_head.

Merge Danger

Door: two-way

This only adds a maintainer script, a test and extensions to the fake. Nothing ships in the binary, because the crate's include leaves scripts/ out. Running the script is one-way: the pushed tag publishes to crates.io. Merging this PR runs nothing.

Blast Radius: tooling

The fake gh changes are additive. The {owner}/{repo} acceptance and the shared -f parsing touch paths the existing tests use, and the full suite passes.

Closes #156

🤖 Generated with Claude Code