Investigation is done, and I've changed nothing. PR 2's branch hasn't moved (dc16fb4822 on both sides), so no rebase was needed, and origin/muskan/squash-redo-final-block-inheritance..HEAD is empty. Here is the pre-edit brief.

Problem

PR 1 and PR 2 link a saved commit only through its last paragraph, but with simplified rules:

• Reader. ParseLastParagraphCheckpoints doesn't use git's trailer grammar. It ignores continuations, Key : value, and mixed Signed-off-by: blocks.
• Fallback. Under core.commentChar=auto or cleanup whitespace/verbatim/scissors, readTrailerRules falls back to scanning the whole message. In those setups a mention in the body can still steer attach and --generate (#2255).
• Prepare. It reads the draft as #-commented with default cleanup. It never asks whether git opens the editor, and it ignores commit.status, git's scissors cut and git's auto pick.
• Writers. They find "git's comment block" by looking for the first # line. appendTrailerLine classifies the last paragraph on trimmed lines and with a hardcoded # (#2256).
• commit-msg. Emptiness hardcodes #.

Root cause

All of these come from one fact: there is no model of how git treats comments. Evidence: commit_trailers.go:26 (readTrailerRules, the only config read, with no commit.status and no auto resolution), trailers.go appendTrailerLine/finalTrailerBlock, manual_commit_hooks.go:302 (hasUserContent), and :3065/:3085 (first # line).

Proposed change

I'll carve the source components 1 to 10 as listed. Each step is test-first, and each commit is small:

1. trailers: git's grammar (isTrailerBlock, lastTrailerBlock, finalTrailerBlock, ParseAllCheckpointsFromFinalTrailerBlock, the new trailerLineRe), the comment and scissors helpers, AutoCommentPrefix with AutoRule, and HasOPFApplied through lastTrailerBlock(msg, "", true). It removes ParseLastParagraphCheckpoints and ParseUnindentedCheckpoints. I'll port the source's trailers_test.go cases first.
2. appendTrailerLine fix from #2259 (closes #2256).
3. telemetry.GitVersion exported.
4. strategy: one gitCommentConfig(ctx) (a plain read for now: git config --get-regexp once, plus git --version only under auto), commentRules, gitOpensEditor, messageCommentRules and preparedFinalTrailers. Then stampedTrailer, squash/redo "already present" and the amend preserve check go through them. addCheckpointTrailerAboveComments with trailerLine/blockEndsCommit serves every writer and replaces addInheritedCheckpointTrailer. addCheckpointTrailerWithComment takes the prefix and uses EditorCommentBlockStart, and keeps the current "trailer above" layout and wording.
5. LinkedCheckpoints: final block with no prefix first, then with trailerPrefix only if the message names a checkpoint. It removes trailerRules, readTrailerRules and the whole-message fallback, and keeps linkedCheckpointsReader's at-most-once read.
6. commit-msg: hasUserContent(message, cfg, edited), explanationPrefix, skipped sign-offs, and the scissors cut.
7. Test hygiene: gitHookEnv sets GIT_EDITOR=true, and the -m simulations append GIT_EDITOR=:.
8. The real-git matrix merged with PR 2's squash and redo rows, plus starts, startsBefore251, aborts, links and gitAutoSkipsTrailingComments.
9. Docs: the reader, squash/redo and commit-msg paragraphs, with PR 1's fallback wording dropped.

Not ported: WithCommentConfig, the snapshot and withGitDir (PR 5), the rules.edited && rules.stripped gating (PR 4), and the 52ae5ec22 layout.

Decisions

1. Read-only views. I recommend keeping them config-free: ParseAllCheckpointsFromFinalTrailerBlock(msg, "#"), then the first ID anywhere. Signatures stay as they are (no ctx), and listings make no extra git config calls until PR 5. The tradeoff: under ; a view can show an ID that LinkedCheckpoints wouldn't link. That's PR 1's existing contract.

2. Explanation wording. It stays "trailer above", so trailerExplanation matches the text.

3. Prepare config reads. One hook can call git config up to about four times: squash, redo, amend and stamping, plus git --version under auto. Post-commit and commit-msg read at most once. The PR description will say this until PR 5.

4. This one needs your call: it touches PR 4's area. Without PR 4's gating, the editor path always writes the explanation. Under auto with a cleanup that keeps comments, git picks ;, so the explanation's ; lines are recorded inside the trailer paragraph. The trailer grammar reads recorded commits with #, so it won't skip them, and the commit ends up unlinked. That breaks the "never unlink" contract for the source rows "template under auto with whitespace / scissors cleanup (and -v)". This is my reading of the code; the matrix test will confirm it before any fix.

   • Recommended: in PR 3, write the explanation only when the grammar will skip it (rules.stripped || rules.prefix == rules.trailerPrefix), and otherwise use addCheckpointTrailerAboveComments. PR 4 then narrows the condition to rules.stripped.
   • Alternative: leave those rows out of PR 3 and add them in PR 4.

   Either way, PR 3 drops the source's two "explanation/prompt never reach the commit" assertions. Those belong to PR 4: under whitespace/verbatim with #, PR 3 still records the explanation, but the commit stays linked.

Local git is 2.56, so the startsBefore251 branch only runs through unit tests here.

Do you approve the change, with decision 1 as recommended and decision 4 as recommended or the alternative?

Read-only views are the commands that only show which checkpoint a commit is linked to. They never write into a checkpoint. In this branch they are:

• entire explain <commit> and explain --commit: show a commit's checkpoint (explain.go:2829).
• The branch listing in explain: one checkpoint per commit (explain.go:2467).
• explain --checkpoint <id>: finds the commits that carry that ID (explain.go:1696).
• Export: explain_export.go:205.
• The head checkpoint flags: head_checkpoint_flags.go:55.
• The API reader, which resolves a commit's checkpoint for the web view (checkpoint_api_reader.go:375).

All of them call CommitCheckpoint or CommitCheckpoints.

The opposite group is the writers. Post-commit condenses into a checkpoint, attach joins one, and explain --generate writes a summary into one. They call LinkedCheckpoints, which reads git config.

Why the split matters. A writer that picks the wrong checkpoint corrupts data: it puts your work into someone else's record. So writers have to be exact, and they read your real comment settings. A view that picks wrong only shows the wrong ID on screen. That's why views get two shortcuts:

• They don't read git config. They assume #, so a listing of 200 commits costs no git config calls.
• If the last trailer block names nothing, they show the first ID anywhere in the message. A squash commit often carries its squashed commits' trailers in the body, and showing one of them beats showing nothing.

Where they can disagree. Say your repo sets core.commentChar=; and a commit ends like this:

With ; as the comment character, git keeps #42 is the issue as text. That makes the last paragraph not a trailer block, so LinkedCheckpoints says the commit links nothing. A view reads with #, skips that line, and shows abc123def456. The difference is only on screen; nothing gets written.

Decision 1 asks whether PR 3 should keep that shortcut, now using git's grammar instead of the lenient reader, or make views read config too. Reading config would remove the disagreement, but every listing would pay a git config call until PR 5 adds the snapshot. I recommend keeping the shortcut.

Before I start editing, I need two answers:

1. Is "okay" your approval of the whole pre-edit brief, with decision 1 as recommended (read-only views stay config-free)?
2. Decision 4 is still open: under auto with a cleanup that keeps comment lines, the explanation can leave the commit unlinked. Which do you want?
   • Recommended: PR 3 writes the explanation only when git's trailer grammar will skip it. PR 4 later narrows that condition.
   • Alternative: PR 3 leaves out the matrix rows for "template under auto with whitespace / scissors cleanup (and -v)", and PR 4 adds them.

Here's decision 4 with a concrete example.

Background. When you run git commit and the editor opens, Entire adds two things to the message:

That's the trailer, plus a short note to you (the "explanation") written as comment lines. Normally git deletes all comment lines when you save, so only the trailer remains. Even when a setting tells git to keep comment lines, Entire still finds the trailer, because when it reads a commit it skips lines starting with #.

The problem case. With core.commentChar=auto, git chooses the comment character itself. If your message already has a line starting with # (say "#42 is the issue"), git switches to ;. Entire then writes its note with ;, so git will recognize it as a comment. Add a setting that keeps comments (commit.cleanup=whitespace), and the saved commit ends like this:

When Entire reads a saved commit, it only skips # lines, not ;. So it sees a trailer followed by two lines of ordinary text, decides "this isn't a trailer block", and the commit isn't linked. Your work loses its checkpoint.

The two options:

• Recommended: in that situation, Entire skips the note and writes only the trailer. The commit stays linked. PR 4 is already changing when the note gets written, so this is a small early piece of that.
• Alternative: PR 3 doesn't touch the note and doesn't test this case. The bug stays until PR 4 lands.

I recommend the first option, because PR 3's job is "follow git's real rules without unlinking anything", and leaving the bug in would break that.

It's an unusual setup (auto plus a cleanup setting that keeps comments), but it's real.

PR 4 is "Prompt out of commits". The note Entire writes in the editor can quote your last prompt to the agent:

Normally git deletes those # lines, so the prompt never reaches the commit. But with a setting that keeps comment lines (commit.cleanup=whitespace, verbatim or scissors), git saves them, and your prompt to the agent ends up permanently in the commit history. Once pushed, anyone can read it.

PR 4 makes two changes:

1. The note is written only when git will delete it. In setups that keep comment lines, Entire writes just the trailer, with no note. That's the rules.edited && rules.stripped check from the source (97529a64c).
2. The "Last Prompt" line goes away entirely, even in normal setups. This is a product choice rather than a bug fix: people lose a hint in the editor, but the prompt can never leak. PR 4's description says so explicitly, so maintainers can push back.

How it connects to decision 4. Change 1 is the general rule: write the note only when git deletes it. The case I raised is a narrower one inside it: when the note would also leave the commit unlinked. So the recommended option puts that narrow piece into PR 3, and PR 4 widens it into the general rule. With the alternative, PR 3 leaves the note alone and PR 4 fixes both cases at once.

I need to correct what I told you about decision 4, and the correction changes the decision.

What I checked. I temporarily added the excluded rows to the matrix and ran them on PR 3 (the file is restored now).

• The source's three rows pass on git 2.56. Newer git skips the template's trailing # paragraph and picks #, so the note is skipped. Those rows only fail on git older than 2.51.
• A row with a # line git does count, like a #42 Subject first line, fails on any git version. Git picks ;, Entire writes its note with ;, and the commit is recorded as:
  The final trailer block reads as empty, so the commit is unlinked.

What I got wrong. I called this an existing bug that would just wait for PR 4. It's actually new in PR 3. Today main links this commit, and so do PR 1 and PR 2: under core.commentChar=auto, PR 1's fallback scans the whole message and finds the trailer. PR 3 removes that fallback, so this commit goes from linked to unlinked. That breaks the stack's "never unlink" contract between PR 3 and PR 4.

Options:

1. Recommended: fix it in PR 3 with the narrow guard. Write the note only when git's trailer grammar will skip it (rules.stripped || rules.prefix == rules.trailerPrefix), and otherwise add just the trailer. PR 4 narrows the condition to rules.stripped. Then PR 3 never unlinks anything, and these two rows go into the matrix.
2. Keep the gap: PR 3 temporarily unlinks commits in this setup (auto, a cleanup that keeps comments, and a counted # line), and its description says so.

Everything else is in place and passing locally: grammar, config read, prepare, readers, commit-msg and the matrix. Docs and the commits come after your answer. Which option do you want?

While mise run check runs, my plan for the commits: two local commits, because the reader removal and its strategy callers have to land together to build.

1. fix: link commits through git's final trailer block and comment config: code and tests.
2. docs: describe git's trailer grammar and comment config.

The check covers the final tree, which is commit 2's tree; commit 1 differs from it only in docs.

I'll commit as soon as mise run check finishes. I'll report the result first, including any of the known load-sensitive timeouts and their reruns.

mise run check is running again. The first run stopped on a single lint finding (maintidx on the matrix test), which I've fixed. I'll commit once it passes.

The check is running again and retries while another session holds the lint lock. I'll commit once it passes.

PR 3 is implemented and committed locally on muskan/git-config-fidelity; nothing is pushed. git log origin/muskan/squash-redo-final-block-inheritance..HEAD shows only my two commits:

• 78427c4bd6 fix: link commits through git's final trailer block and comment config
• b2e9dc365d docs: describe git's trailer grammar and comment config

Verification

• mise run check: lint clean. The full test run failed on three timeouts, in packages this branch doesn't touch:

  • TestProbeConfigDir_AsksClaude (5.00s) and TestDiscoverAndRegister_ContinuesAfterRegistrationError (10s) are on your known-flaky list.
  • TestResolveClaudeConfigDir_PrefersClaudesAnswer (5.00s) isn't on the list. It's in the same package as the first and timed out at the same 5.00s.

  Run alone, all three pass. The canary didn't run inside the check because go test failed first, so I ran it separately and it passed (53 + 4).

• Matrix: TestTrailerAuthority_RealGitCommits passes all rows, including the two new #-subject rows for the decision-4 fix. Both failed before the guard.

• Lint suppression: the merged matrix test needed //nolint:maintidx, with a reason, following checkpoint/remote/util_test.go.

• Old git: git here is 2.56, so the matrix's pre-2.51 expectations didn't run. The pre-2.51 auto rule is covered by unit tests only.

Three things to know

• Line with text after the ID: Entire-Checkpoint: <id> (some note) no longer links. The grammar only accepts a line that is exactly the trailer, the same rule RemoveCheckpointTrailers already uses. Read-only views still show it through their first-ID fallback.
• Commit message: the commit body says "Fixes #2255, #2256". On merge, GitHub closes those issues from the commit as well as from the PR description. I'll drop the line from the commit if you'd rather not.
• Ported test wording: the "explanation never reaches the commit" assertion is skipped only for the whitespace, scissors and verbatim rows. The note's wording there is the same as PR 2 ("trailer above").

Draft PR title

Read checkpoint trailers by git's rules

Draft PR description

Publish plan (not run; needs your approval)

1. Run mise run lint, then git push -u origin muskan/git-config-fidelity.
2. You, in the browser: on #2938, click the stack icon, then "Add to stack". Pick muskan/git-config-fidelity and create the PR with the title and body above. Then send me the PR number.
3. Run entire trail link 2635 --repo gh/entireio/cli --branch muskan/git-config-fidelity --title "Read checkpoint trailers by git's rules" unsandboxed. Check that it reuses the stacked PR. If it opens a draft "shadow" PR based on main, I stop and ask before closing it.
4. Check entire trail show 2635 --repo gh/entireio/cli --json to confirm only trail #2635 holds the branch, and that the PR's file list is only these 21 files.
5. Run entire trail finding list cli/<n> --freshness any, again about 10 minutes later. Fix real findings and re-check.
6. Comment on #2253: pr 3 of the split is up: #<n>, git config fidelity. same trail #2635.

Do you approve the title, the description and this plan?

Here is the draft PR description:

The proposed title is Read checkpoint trailers by git's rules.

Do you want any changes before you approve it?

Lint came back clean (0 issues), and muskan/git-config-fidelity is pushed to origin with upstream set.

Step 2 is yours, in the browser:

1. Open #2938 (https://github.com/entireio/cli/pull/2938).
2. Click the stack icon in the header, then "Add to stack".
3. Pick muskan/git-config-fidelity as the head branch.
4. Use the title Read checkpoint trailers by git's rules and paste the approved description.
5. Create the PR and send me its number.

Once I have the number, I'll run entire trail link and the remaining steps.

Two. #2923 (PR 1) and #2938 (PR 2) are open. PR 3's branch is pushed, but its PR doesn't exist until you create it from #2938's "Add to stack" (step 2). Once you do, there will be three, and I'll link it to trail #2635.

No, I haven't linked anything for PR 3. There's no PR for muskan/git-config-fidelity yet, so nothing exists to link. Only the branch is pushed.

If I ran entire trail link now, it would probably open a draft "shadow" PR based on main instead of joining the stack, which is the problem step 2 avoids. So the order is:

1. You create the PR from #2938 (stack icon, then "Add to stack", head muskan/git-config-fidelity).
2. Send me the number, and I'll run entire trail link 2635 ... --title "Read checkpoint trailers by git's rules".

If you'd rather I link first and accept a possible shadow PR, tell me and I'll do it that way.

That's not how the plan you gave me works, and I can't create the PR myself either way:

• Your rule: I never create PRs directly (no gh pr create). The only way a PR appears from my side is entire trail link, which makes Entire open its mirrored PR.
• The pasted plan says to create the PR from #2938's "Add to stack" before trail link. Then trail link reuses that stacked PR. If I link first, Entire will probably open its own PR based on main, and the plan calls that a "shadow" PR to stop and ask about.

Your order might still work, if GitHub's "Add to stack" can take an existing PR and change its base to muskan/squash-redo-final-block-inheritance. I can't check that from here. If it can't, we'd have to close the shadow PR and redo it.

Two ways forward:

1. As planned (safer): you click "Add to stack" on #2938 and create the PR, then send me the number and I link it.
2. Your order: I run entire trail link now and send you the PR it opens. You add it to the stack, and if GitHub won't let you, we close it and fall back to option 1.

Which do you want?

I'm stopping here because this changes something you asked for. The installed CLI (0.12.2-nightly.202610242156) has no --title on entire trail link. The current flags are --base, --branch, --branch-action, --idempotency-key, --json, --status, --project and --repo.

Your instructions say to always pass --title. Without it, the Change gets named after the branch (muskan/git-config-fidelity), and the CLI can't rename it later.

Options:

1. Link without a title: the Change is named muskan/git-config-fidelity. You might be able to rename it on entire.io if the web UI allows it; I can't check that.
2. Update the CLI first and retry: a newer build may have --title back. You'd need to update it (for example, via your usual install path).
3. Link from the entire.io UI: add the branch to trail #2635 there, if the UI lets you set a title.

Which do you want? I'd go with 2 if updating is quick, otherwise 3.

PR 3 is up as #2944 (https://github.com/entireio/cli/pull/2944) and linked to trail #2635 as Change cli/1717. Only trail #2635 holds the branch, and it now has three Changes: 1706, 1714 and 1717. I've set the approved title and description, keeping Entire's trail-link block at the top.

Two things still need you on GitHub:

1. Add #2944 to the stack on top of #2938. It's based on main right now, so it shows 37 files; once it's on muskan/squash-redo-final-block-inheritance, it should show only PR 3's 21 files.
2. Mark it ready for review. Entire opened it as a draft, while #2923 and #2938 aren't drafts. I can run gh pr ready 2944 if you want; I didn't, because the plan doesn't cover it.

Once it's stacked, I'll check the file count. Then I'll run the findings check (entire trail finding list cli/1717 --freshness any) now and again about 10 minutes later, and post the approved comment on #2253: pr 3 of the split is up: #2944, git config fidelity. same trail #2635.

The CLI can't rename a Change. entire trail update --title changes the whole trail's title ("fix(checkpoints): link a commit only through its last paragraph"), not cli/1717's, and trail link has no title flag in this build. That's why cli/1706 and cli/1714 are also still named after their branches. A Change rename would have to happen in the entire.io UI, if it offers one.

For the PR title, #2923 and #2938 use the conventional-commit style, so I'd match them:

fix(checkpoints): read checkpoint trailers by git's rules

Should I set #2944 to that title? Also, for the Change titles (1717, and 1706 and 1714 if you want them consistent), are you able to rename them in the entire.io UI, or should I leave them as the branch names?

The 37-file view is because PR 2's branch was rebased onto a newer main, and PR 3 was still sitting on its old commits.

Done so far:

• #2944 base: now muskan/squash-redo-final-block-inheritance.
• #2944 title: now fix(checkpoints): read checkpoint trailers by git's rules.
• It's no longer a draft. I didn't change that, so GitHub or Entire probably flipped it when the base changed.
• Local rebase: my two commits are rebased onto the new PR 2 tip (11602d3287). It applied cleanly, and the range is only my two commits (ae87542815, 042f85a66d).

Next, and it needs your approval: force-push. Once mise run check (running now) passes, I'd run mise run lint, then git push --force-with-lease origin muskan/git-config-fidelity. That's what brings #2944 down to its own 21 files. Can I force-push once the check passes?

The "Add to stack" button may also appear now that the base is PR 2's branch.

Approval noted. I'll force-push as soon as mise run check finishes and passes; it's still running.

#2944 is updated: GitHub now shows only PR 3's 21 files, on top of PR 2's current tip (11602d3287), and it isn't a draft.

• Before the push: lint passed on the rebased tree (0 issues). The full mise run check is still running in the background. No test had failed when I pushed, and I'll report its final result when it finishes.
• Findings: cli/1717 has none yet. The review runners just got the new push, so I'll check again in about 10 minutes.

Still open:

• Stack button: see whether "Add to stack" shows up on #2938 or #2944 now that the base is right.
• Change titles: cli/1706, 1714 and 1717 are still named after their branches. The CLI can't rename them, so that's an entire.io UI step if you want it.
• Comment on #2253: once findings are clear, I'll post the approved comment: pr 3 of the split is up: #2944, git config fidelity. same trail #2635.

Here is the draft comment for #2253:

Should I post it as written?

About the 21 files: that's PR 3's own change, not leftovers from PR 2. Before the rebase, GitHub showed 37 because it was also counting PR 1, PR 2 and the newer main. PR 3 touches these 21 files:

• Trailer grammar: trailers/trailers.go and its test.
• Config read, readers, prepare and placement: strategy/commit_trailers.go, manual_commit_hooks.go, manual_commit_provenance.go and manual_commit_redo.go.
• Strategy tests: commit_trailers_test, manual_commit_test, manual_commit_provenance_test, manual_commit_squash_test, manual_commit_redo_test, and global_test (the GIT_EDITOR isolation).
• Exported GitVersion: telemetry/detached.go.
• Integration tests: checkpoint_trailer_real_git_test.go (the matrix), checkpoint_trailer_authority_test.go and testenv.go (GIT_EDITOR hygiene).
• Comment wording only (last paragraph → final trailer block): attach.go, attach_test.go, checkpoint_api_reader.go and explain.go.
• Docs: docs/architecture/sessions-and-checkpoints.md.

That's about +1,600 / −430 lines. Most of it is tests and the ported grammar.

The comment is posted on #2253: https://github.com/entireio/cli/pull/2253#issuecomment-6469259759

Still pending: the background mise run check result and the second findings check on cli/1717 in about 10 minutes. I'll report both.

PR 3 is published as #2944 (Change cli/1717 on trail #2635) and has no findings so far. The rebased branch passed its checks except for two timeouts that pass when run alone.

Post-rebase check:

• Lint: 0 issues on the rebased tree.
• Two timeouts in the full run: both are the Claude config-dir probe timing out, not the trailer code. Run alone, both pass.
  • TestResolveClaudeConfigDir_PrefersClaudesAnswer (5.00s) timed out again; it also did before the rebase.
  • TestResume_AsksClaudeForItsConfigDir (6.33s) is new to this run, and isn't on your known-flaky list either.
• Canary: passes (53 + 4). It ran separately because the failed go test step stopped the gate before it.

Findings: none on cli/1717, checked twice. The first check came right after the force-push and the second a few minutes later. That second check was earlier than the planned 10 minutes, so I'll check once more if you want.

Left for you:

• Stack link: see whether "Add to stack" now shows on #2938 or #2944.
• Change names: if you want proper names on cli/1706, 1714 and 1717, that's an entire.io UI step. The CLI can't rename a Change.

The re-run is in progress, and I'll tell you how it ends. Once it passes, only the reviewer approval is left on #2944. PRs 1 and 2 are waiting on approval the same way.

PR 4 is scoped, and I'm waiting on your approval before I create the branch or edit anything. The CI re-run on #2944 is still going, and I'll report it when it ends.

Pre-edit brief: PR 4, "keep the prompt out of commits"

Problem. With an editor commit and commit.cleanup set to whitespace, verbatim or scissors, PR 3 still writes Entire's explanation comment, including # Last Prompt: <the user's prompt>. Those modes keep comment lines, so git records both into history. PR 3's real-git matrix marks those three rows keepsComments and skips the "no prompt in the commit" assertion for them.

Root cause.

• The gate: prepare's default branch (manual_commit_hooks.go:607) writes the explanation whenever rules.stripped || rules.prefix == rules.trailerPrefix. The second half is true in every keep mode that uses #.
• What the hooks can't see: a mode passed on the command line (git commit --cleanup=whitespace) never reaches the hooks. They only read config (gitCommentConfig), so no gate can be exact in that case.

Proposed change (stacked on PR 3, about 5 files):

1. Tests first.
   • In checkpoint_trailer_real_git_test.go, drop keepsComments so every row asserts no explanation and no prompt in the recorded message. The three keep-mode rows should then fail on PR 3.
   • Add one row for git commit --cleanup=whitespace given on the command line. It asserts the commit links and records no prompt. A small recordsExplanation flag allows the explanation in that row only.
   • Update the unit tests at manual_commit_test.go:860-885 and :1018 to assert no Last Prompt line.
2. Gate: write the explanation only when rules.edited && rules.stripped, which is the source's commit 97529a64c. Otherwise prepare uses addCheckpointTrailerAboveComments with no comment. This also covers the auto case PR 3 guarded.
   • I said earlier that PR 4 would narrow the condition to rules.stripped. The source also requires edited, and I'd follow it.
   • That extra condition only skips writing a comment nobody sees: no editor opens, and git strips the comment anyway.
3. Drop the Last Prompt: line from the editor comment. Of the two changes, this is the product choice the plan says the PR description must state, so maintainers can push back.
   • The reason is the --cleanup gap: the most that can then leak is the generic explanation, never the user's prompt.
   • The prompt still shows in the terminal confirmation when you commit with -m from a terminal. That output goes to the screen, not into the commit.
   • addCheckpointTrailerWithComment loses its prompt parameter.
4. Docs: update the Placement bullet in sessions-and-checkpoints.md to say the explanation is written only when git opens the editor and its cleanup drops comments, and that it no longer quotes the prompt.

Tradeoff: people who liked seeing their last prompt in the editor lose it. Each change could also ship alone: the gate without dropping the line, or the reverse.

Setup: work new cli "keep the prompt out of commits" --stack-on git-config-fidelity. This branches from PR 3, joins trail #2635 and takes over this Rex tab.

Approve all four, or just the gate without dropping the line?

CI on #2944 is green now: the canary passed on the re-run, and so did the test summary job. Entire Gates still shows its earlier fail. It last ran before the re-run, so its Checks gate may stay red until something triggers a new run, such as an approval or a push. Its Approvals gate needs a reviewer either way.

PR 4 is still waiting on your answer to the brief above: approve all four, or just the gate without dropping the line?

This work has one trail, #2635, holding three Changes. PR 4 isn't on it yet: it has no push and no Change.

Trail #2635, "fix(checkpoints): link a commit only through its last paragraph" (open, Linear ENT-3643): https://entire.com/gh/entireio/-/trails/2635

1. cli/1706, PR 1: #2923 "link a commit only through its last paragraph", based on main.
2. cli/1714, PR 2: #2938 "inherit squash and redo trailers from last paragraphs", stacked on PR 1.
3. cli/1717, PR 3: #2944 "read checkpoint trailers by git's rules", stacked on PR 2.
4. PR 4: in progress on muskan/keep-the-prompt-out-of-commits. It becomes a Change once it's pushed and linked.

PRs opened today (Oct 28, your time):

• #2944 (open): PR 3 above, trail #2635, Change cli/1717. Done in this session.
• #2938 (open): PR 2 above, trail #2635, Change cli/1714.
• #2939 (merged): "Limit review finding snippets to files in the reviewed change", trail #2658, Change cli/1715. It merged, but trail #2658 still shows open, so check it for follow-ups before closing it.
• #2936 (open): "ask for the push policy on the first terminal push", trail #2630 "Hold checkpoints back before push", Change cli/1712.

#2923 (PR 1) opened yesterday evening, Oct 27, alongside #2925 and #2921 on the checkpoint-push work.

PR 4 status: the fix is in. Before it, the updated real-git matrix failed on exactly the four expected rows: the three keep-mode configs and the new --cleanup=whitespace row. I'm now running the trailer tests plus the strategy and trailers packages against the fix, and I'll report the result, then do the docs bullet and mise run check.

PR 4 is implemented and tested, and it's ready to commit once you approve the messages below. Nothing is committed or pushed yet.

What changed (4 files, +41/−78):

• Gate: prepare writes the explanation comment only when git opens the editor and its cleanup drops comment lines. Otherwise the trailer goes in the same place with no comment.
• No prompt: the Last Prompt: line is gone from that comment, and the writer no longer takes a prompt. The prompt still shows in the terminal question you get when committing with -m from a terminal, which never goes into the commit.
• Docs: a new "Explanation" bullet in sessions-and-checkpoints.md, next to the Placement bullet.

Tests:

• Matrix: the real-git matrix now checks every row for "no prompt in the commit", and every row except the command-line one for "no explanation". The new row is git commit --cleanup=whitespace. Before the fix exactly four rows failed: the three cleanup modes set in config and the new row. After the fix the matrix passes, and so do the strategy and trailers packages.
• Full check: mise run check passed format and lint. Tests failed only on TestFetchCheckpointRefFrom_ConfiguredHonorsFetchTimeout (new to our flaky list) and TestExternalCommand_SigintReachesPlugin (known flaky). Both pass alone. The canary passes (53 + 4).
• Review lessons: I checked the diff against General.md and cli.md, and none of the lessons apply.

One limit to know: a cleanup mode given on the command line (git commit --cleanup=whitespace) still records the explanation, because the hooks can't see that flag. It no longer records the prompt. The new test row and the docs bullet both say so.

Commit drafts, for your approval:

1. Code and tests:

1. Docs:

After you approve, I'll commit both and run mise run lint. Pushing and linking it as a Change on trail #2635 will each need your go-ahead. work publish will print its plan for that.

The four cleanup reviews (reuse, simplification, efficiency, altitude) are running on PR 4's uncommitted diff. I'll triage what they find and show you any cleanups before changing code. The commit drafts are on hold until then.

The altitude review is back with no findings. The reuse, simplification and efficiency reviews are still running.

Reuse found nothing either. The efficiency review found one real cleanup, and I've checked it in the code. I'm still waiting on the simplification review.

All four reviews are back. One cleanup is worth doing and one small test nit is optional. I'll wait for your approval before editing either.

1. Read the last prompt only when the terminal question shows it (efficiency). This is the cleanup I'd do.

• Problem: prepare calls trailerPromptContext on every commit (manual_commit_hooks.go:560). That function reads the session's prompt.txt (getLastPrompt), then collapses and truncates the text.
• Why this diff causes it: the editor comment no longer uses the prompt, so its only remaining use is the "Last prompt:" detail. That detail appears only for -m/-F commits made from a terminal while commit_linking is not always. Every other commit, including every editor commit, now reads that file for nothing.
• Cleanup:
  • trailerPromptContext returns only the agent type and commit_linking.
  • The read and formatting move into a small displayLastPrompt(ctx, sessions), called inside that terminal branch just before askConfirmTTY.
• Scope: about 15 lines in one function and its one caller. No behaviour changes, and the read also drops out of the resolve_session_metadata timing span.

2. Drop the NotContains(t, result, "Prompt:") check in TestAddCheckpointTrailerWithComment_HasComment (optional nit).

• Problem: the writer no longer takes a prompt, so that check can't fail.
• Cleanup: delete it. The real-git matrix's "no prompt in the commit" check on every row is what guards the behaviour.

Skipped:

• Named method for the gate: the altitude review suggested something like explanationDropped() for rules.edited && rules.stripped. Skipped because the condition has only one user, as that review itself said.
• Hard-coded prompt in the test: the matrix checks for "change mine.txt", which is tied to how authorityAgentTurn builds its prompt. Skipped because there's no shared helper to reuse, and the coupling is local to the test file.
• settings.Load on every commit: it runs on every path too, but that cost existed before this diff, so it's out of scope.
• Gate placement: the gate stays at its one caller rather than moving into commentRules or the writer. Both checks in it are needed: commit.cleanup=strip drops comments even when no editor opens, so stripped alone isn't enough.

Apply 1 and 2? After that I'll rerun the strategy tests and the trailer matrix, then come back to the commit drafts.

Both cleanups are in, and the tests and lint pass.

• Prompt read moved: prepare now reads and formats the last prompt only in the terminal branch, through the new displayLastPrompt, just before askConfirmTTY. trailerPromptContext returns only the agent type and commit_linking, so other commits no longer read prompt.txt.
• Check removed: the dead NotContains("Prompt:") check is gone from TestAddCheckpointTrailerWithComment_HasComment.

Verification:

• Strategy package: passes.
• Integration tests: the trailer-authority matrix and the related integration tests pass.
• Lint: mise run lint passes on the current tree.
• Diff: now 4 files, +53/−93.

Not covered by tests: no test checks the "Last prompt:" detail itself, because it only shows in an interactive terminal question. The move doesn't change what it displays.

Here are the two commit drafts again, unchanged:

Approve these and I'll commit. Pushing and linking it as a Change on trail #2635 will each still need your go-ahead.

Claude Code can't use your Claude subscription, usually because a payment is past due · Pay or renew at https://claude.ai/settings/billing (if you just paid, try again shortly), or use an Anthropic API key instead