[JacobStephens2]

Summary

The release script now has claude write the bump PR's summary. If the agent fails, it falls back to GitHub's generated notes.

 pr_body <base> <tag>
-  notes = gh generate-notes
+  diff  = version_diff <base>
+  release_summary <base> <tag> <diff>
+    input = <diff> + "### #N: title\n\nbody" for each PR merged since the last v* tag
+    prompt file + input | claude -p --tools '' --strict-mcp-config
+    if it failed or printed nothing
+      warn on stderr
+      "_The agent summary was unavailable, so this is GitHub's generated notes._"
+      gh generate-notes
   ## Summary  <!-- release-summary:start --> … <!-- release-summary:end -->
   ## Version diff

 scripts/
 ├── release.sh
+└── release-summary.md   # the agent's prompt, with the emphasis rules and the 0.3.0 summary as the model
 tests/fakes/
 ├── claude               # + plain print mode: records stdin, prints only the script's output
 └── gh                   # + `pr view <number>` and `--jq .<field>`

• The PRs merged since the last tag come from the Merge pull request #N merge commits in <last tag>..origin/main. The bump PR isn't merged yet when this runs, so it is never among them.
• The README's Releasing section describes the agent summary and the fallback.

Evidence

• Before: on main, the summary section holds GitHub's generated notes and claude is never called.
  After: cargo test --test release passes all 9 tests, including 4 new ones:

the_agent_gets_the_prompt_file_then_the_version_diff_and_the_prs_merged_since_the_last_tag
  stdin == prompt file + version diff + "#2: Add the frobnicator" + body + "#3: Tidy the site"
  (the PR merged before v0.1.0 and the bump PR are left out)
the_agent_runs_in_print_mode_with_tools_disabled         -p, --tools '', --strict-mcp-config
the_pr_body_has_the_agents_summary_in_a_marked_section_then_the_version_diff
when_the_agent_fails_or_prints_nothing_the_summary_is_the_generated_notes
  exit 1 / blank output → a stderr warning, the note, then the generated notes; v0.2.0 is still tagged

• The full cargo test --no-fail-fast passes. In one earlier run, merge_run's ctrl_c_after_the_merge_… hit its 10-second "agent never started" deadline while the machine was under heavy load. It passed on the re-run, and it doesn't touch the release script.
• shellcheck scripts/release.sh and cargo clippy --all-targets -- -D warnings are clean.

Merge Danger

Door: two-way

This only changes the maintainer's release script and its tests. Nothing ships in the binary. The tag push it leads to was already one-way.

Blast Radius: release notes

At worst, a poor agent summary ends up in a bump PR body, and the maintainer can edit it. Releasing now needs claude on PATH. If claude is missing or fails, the script falls back to the generated notes rather than stopping.

Unaddressed findings

Standards

• Two gh pr view calls per PR (title, then body): a single call would need a jq expression that joins a multi-line body. Separate calls stay simple, and a release has only a few PRs.
• The fake gh builds a second field map for --jq: pr_fields deliberately leaves out title and body, which pr list doesn't return, so the --jq path adds them locally.
• The argv test checks exact flags: "tools disabled" is a safety property the spec names, so asserting the flags is the point of the test.

Spec

• The prompt says the summary is also for the Release page: that is where Cut a release with one command: bump PR, agent-written summary, merge and tag #150 puts the summary. The wording tells the agent who reads it and doesn't implement the Release page hook.
• --strict-mcp-config beyond "tools disabled": --tools '' only disables built-in tools, so this flag also keeps MCP tools out, as Cut a release with one command: bump PR, agent-written summary, merge and tag #150 intends ("it can't take actions").
• Squash-merged or rebase-merged PRs are missed: this repo merges with merge commits, and the release script itself merges that way.
• A failure to collect the input stops the release instead of falling back: the spec's fallback covers agent failures. A gh failure while reading PRs also means the generated notes would likely fail.
• An untested assumption that "not logged in" exits non-zero: I couldn't check the real logged-out claude in this headless session. The whitespace-only check also catches an empty result.
• No test with a previous bump PR inside the range: that bump's merge commit carries the previous tag, so it sits outside <last tag>..main by construction.

Closes #159

🤖 Generated with Claude Code