Squash merge discards commit bodies (squash_merge_commit_message: BLANK) — intentional?

[yuanchen8911]

Summary

The repo merges with squash_merge_commit_title: PR_TITLE and squash_merge_commit_message: BLANK, so the commit that lands on main is the PR title plus trailers and nothing else — every branch commit body is discarded. The rationale for a change therefore exists only in the PR description, never in git log.

Raising this as a question rather than a defect: it may well be deliberate.

Evidence

$ gh api repos/NVIDIA/aicr --jq '{squash_merge_commit_title, squash_merge_commit_message}'
{"squash_merge_commit_message":"BLANK","squash_merge_commit_title":"PR_TITLE"}

Recent merges confirm it. #2183, #2187 and #2195 each landed on main with a body consisting solely of Signed-off-by:. The full commit message written on the branch — in #2183's case around 170 lines describing the fail-closed behaviour and its trade-offs — was dropped at merge.

Trailers do survive, so DCO sign-off is unaffected.

Consequences

• git log on main carries titles only. Reading why a change was made requires GitHub, a network connection, and the PR still being accessible. git log, git blame -w, git bisect and offline archaeology all lose the reasoning.
• Effort spent on commit bodies is discarded. CONTRIBUTING.md teaches contributors to write a body explaining "the problem being solved and why this approach was chosen"; that text never reaches main. PR docs: fix the branch-hygiene push rule — append commits mid-review #2201 (merged 2026-08-16) added a note documenting this, but the underlying setting is the real question.
• The PR title becomes the entire durable record, which raises the stakes on title quality considerably.

The question

Is BLANK intentional?

• If yes — reasonable, since PR bodies here can be long and some maintainers dislike that in git log. Then the guidance should say plainly that the PR title is the durable artifact and commit bodies serve reviewers only. PR docs: fix the branch-hygiene push rule — append commits mid-review #2201 already did this for the agent rules and CONTRIBUTING.md when it merged on 2026-08-16.
• If no — squash_merge_commit_message: PR_BODY would preserve the reasoning in history permanently, at the cost of longer commit messages on main. COMMIT_MESSAGES is the other option, though it composes from branch commits rather than the curated PR description.

This is an admin-only repo setting, so it needs a maintainer decision either way.

Related

• PR docs: fix the branch-hygiene push rule — append commits mid-review #2201 (merged 2026-08-16) documents the current behaviour in the branch-hygiene rules and CONTRIBUTING.md. The documentation half is done; only the maintainer decision on the setting is outstanding.
• Issue Adopt a force-push guard that holds in shared clones (bare --force-with-lease fails open) #2203 covers the separate force-push guard question from the same review.

[yuanchen8911]

Closing — the setting is correct for this repo and should stay as-is.

Answer

squash_merge_commit_message: BLANK is the right choice here, for two reasons specific to AICR rather than any general convention:

1. PR_BODY would be worse than nothing, given our template. .github/PULL_REQUEST_TEMPLATE.md is 66 lines across eight sections with 25 checkboxes — Type of Change, Component(s) Affected, Risk Assessment, Checklist. Under PR_BODY, every commit on main would carry that scaffold into git log, git show, and git blame:

- [ ] Bug fix (non-breaking change that fixes an issue)
- [x] New feature (non-breaking change that adds functionality)
- [ ] Breaking change ...

PR_BODY pays off when PR descriptions are prose. Ours are structurally not that. COMMIT_MESSAGES is worse still — it concatenates branch commits, including review-round fixups.

2. Durable rationale already has an in-tree home. docs/design/ holds 17 ADRs (002 through 018). Design reasoning that must survive offline, outlive GitHub, and stay greppable belongs there by existing convention — and does. What lives only in a PR body is tactical review context (why this ordering, why this test approach), which is genuinely ephemeral.

So the concern in this issue is real but narrower than it first appeared: what's lost is reviewer-facing context, not architectural rationale.

Prior art, for anyone reading later

BLANK is not the industry default. The most common deliberate choice is PR_TITLE + PR_BODY; BLANK is a well-established minority pattern typical of conventional-commits repos and those with heavy templates. GitHub's own default is COMMIT_OR_PR_TITLE + COMMIT_MESSAGES, which almost nobody keeps. The case for BLANK here is situational, not conventional.

Already covered

The consequence — that the PR title is the entire permanent record — is documented as a consequence of this setting, not as style preference:

• AGENTS.md:570 — "squash-merge-only, so GitHub composes the commit on main from the PR title; by default only the title and trailers reach main. Durable wording belongs in the PR title."
• CONTRIBUTING.md:394-396 — same, explicitly.

No documentation change needed.

Follow-up

One idea worth its own issue rather than a note here: with the title as the only durable record, a CI check enforcing conventional-commit format on PR titles would be cheap insurance. No such workflow exists in .github/workflows/ today. Filing separately; not urgent.

Reopen if there's context I'm missing — a past decision, a plan to slim the template, or an org convention that argues otherwise.