Split Commit Into Stack

Turn one large commit into a stack where each PR is a coherent story that compiles and passes on its own. The work is mostly analysis: the mechanics are five commands, but choosing the cut and proving it holds is where the value is.

Most of the cost of a big PR is that a reviewer cannot tell which changes belong to which decision. A split earns its keep when it separates decisions, not when it merely reduces line count.

Step 0: Measure before proposing anything

Measure the whole branch against its base, not the tip against its parent: a branch that already holds several commits is split as one diff, and HEAD~1 would silently measure, inspect and later reset only the last of them.

base="$(git merge-base "$(gt parent)" HEAD)"   # the branch's base, whatever the commit count
git --no-pager log --oneline "$base"..HEAD     # several commits that already match the concerns: see Graphite notes
git --no-pager diff --numstat "$base" HEAD | awk '{i+=$1;d+=$2; printf "+%-6d -%-6d %s\n",$1,$2,$3} END {printf "TOTAL +%d -%d\n",i,d}'
git --no-pager diff --name-status "$base" HEAD

Every later step reads "$base" from this shell, so run them in the same one or set it again.

Two numbers change the conversation before any splitting is discussed:

• Production vs test insertions. A 3,500-line diff where 1,000 lines are one new test file is not a 3,500-line review. Report the production-only figure; sometimes it dissolves the problem and no split is warranted.
• Per-concern totals. Group files by concern (below) and total each. If one group is 85% of the diff and is atomic, splitting the remaining 15% is the entire available win
  • say so plainly rather than proposing a heroic three-way split.
• Whether one file serves both concerns. Two or three such files is normal and Step 4 usually dissolves them; a dozen means the concerns are not actually separable and the honest answer is to leave the commit whole.

Use --name-status and not just --numstat: added (A) and deleted (D) files change what the git add commands must look like, and a deletion left in the wrong PR breaks the build in a way that is annoying to diagnose later.

Step 1: Group by concern, not by directory

This procedure produces exactly two layers in one pass: a lower commit and an upper PR that is everything else. A third concern is not a third output of this run - split it by running this same procedure again on the resulting upper branch, treating it as the new whole. "Groups" and "order" below describe how the whole diff's concerns relate to each other and to picking which one goes into the lower layer first, not a promise that Step 6 and Step 8 produce more than two.

A concern is a decision someone made, and it usually shows up as: one layer's API changing shape, one user-visible behaviour changing, or one file pair being deleted and replaced. Directory is a weak proxy - a single handlers/ directory routinely holds files belonging to two different concerns.

For each candidate group, write one sentence naming what a reviewer would be asked to approve. If you cannot write that sentence, it is not a concern, it is a pile of files.

Step 2: Establish dependency direction - the only test that matters

A lower PR must build and pass without anything above it. That is a claim about which symbols cross the cut, and it is cheap to check statically before touching git.

Find what the pre-change version of the upper half uses from the lower half's packages, then confirm those symbols survive the lower half's changes:

# what the old upper-half code reaches for
for f in "<upper-half file>" "<upper-half file>"; do
  echo "=== OLD $f ==="
  git show "$base:$f" | grep -oE "<pkg>\.[A-Za-z]+|<MethodName>|<TypeName>" | sort -u
done

# do those still exist after the lower half's changes?
grep -rn -- "func .*<Symbol>\|type <Symbol>" "<lower-half package dir>"

This is a heuristic, not a proof: it sees identifiers the regex happens to match and misses fields, constants, methods, aliases and files the upper half adds. Step 7 is the proof.

Two shortcuts that resolve most cases fast:

• Unexported symbols cross files, never packages. A lowercase symbol the lower half renamed or deleted can only be referenced from files in the same package, so the check narrows to upper-half files that live in the lower half's own packages and rules the rest out. It does not vanish: a lower commit that drops an unexported helper a sibling file in the upper half still calls does not compile.
• An unused exported method is harmless. The lower PR may introduce an accessor that nothing calls until the upper PR arrives. That compiles fine and is often the cleanest way to keep the cut tidy (see Step 4).

If the direction runs the other way - the lower half needs something the upper half introduces - the stack order is wrong. Reorder rather than forcing it.

Step 3: Refuse to split what is atomic

Some refactors have no compiling intermediate, and saying so is the correct answer.

The signature: removing a field, method, or interface entry requires simultaneously adding its replacement, flipping every reader, and updating every writer. Any partial state either does not compile, or holds the same fact in two places at once - and "limits live in two places for one commit" is harder to review, and easier to get wrong, than the whole change at once.

Test it by asking: what would the intermediate commit look like, and would it build? If the honest answer is "it would have both the old and new storage live simultaneously", do not split. Recommend keeping it whole and explain why in one or two sentences.

Do not split a group purely to hit a line-count target. A mechanical rename can usually be lifted out (it is skippable in review), but weigh the gain: twenty call sites moved into their own PR is a modest win for an extra branch to manage.

Step 4: Keep files whole - relocate small seams instead

The fiddliest case is one file whose diff serves both concerns, because splitting it means staging individual hunks and every later rebase becomes manual.

Before accepting a hunk-level split, look for a small seam you can move down-stack to make the file whole again. A seam is typically an interface entry plus its implementation plus any mocks that satisfy the interface - a handful of lines. Moving it into the lower PR (where it may sit unused) is almost always better than splitting a test file across two branches.

Watch specifically for mocks: any test double implementing an interface must gain a method the moment that interface does. That single fact often decides which PR an interface entry belongs to.

Step 5: Confirm the partition with the user before executing

Present the groups, the totals, which one is atomic and why, and the proposed order. This is a judgment call about history shape - use AskUserQuestion if more than one partition is defensible. In most repos branch and PR creation belongs to the user; check whether you are expected to run the commands or hand them over.

This confirmation is about the shape of the split, not the go-ahead to run it. Even when only one partition is defensible, wait for the user's explicit go-ahead before Step 6: everything from there on mutates the branch (git reset, git commit), and Step 8 ends by publishing (gt submit --stack) - an outward-facing action that is always the user's to run, obvious split or not.

Step 6: Commit the lower PR

Start from a clean tree. A mixed reset keeps whatever is already modified or untracked, and the staging below would fold it into one of the two PRs.

git status --porcelain=v1                    # must print nothing; if it does, sort the tree out first
git status --porcelain=v1 --ignored=matching  # must print nothing either

A plain git status --porcelain=v1 never lists ignored files, and git reset --hard in Step 7 does not touch them - so a stray ignored file already sitting in the tree (a stale .env, a leftover build artifact) rides through undetected and is present during Step 7's "lower half genuinely alone" validation for reasons that have nothing to do with either half. Sort it out - move it aside or delete it - before continuing; this procedure does not decide that for you.

git reset "$base"                     # mixed: HEAD and index move to the base, tree keeps the new content

# --- lower PR: stage its paths explicitly ---
git add -- "<path>" "<path>"          # one quoted argument per path
git commit -m "<type>(<scope>): <what this decision is>"

Stage the lower PR explicitly and leave the upper PR as "everything remaining". The reverse - explicitly staging the upper half - silently drops any file you forgot into the wrong commit. Quote every path and keep the --: a generated path list is data, and an unquoted substitution splits on whitespace and honours shell metacharacters.

Name files, not directories, for any path Step 1 flagged as serving both concerns: git add stages a directory recursively, and Step 1's own premise is that a directory routinely holds files from both. Staging one such directory here would sweep upper-half files into the lower commit right where the split was supposed to keep them apart. A directory is fine to name only once you have confirmed every file under it belongs to the lower half.

The upper half now sits uncommitted in the working tree. Do not commit it yet: Step 7 needs it absent, and it is far easier to set aside while it is uncommitted than once it has become a commit that HEAD points at.

Step 7: Prove the lower PR stands alone

Do not skip this, and do not infer it from the static check in Step 2. Build the lower half with the upper half genuinely absent - which is why it is still uncommitted: HEAD is the lower commit, and everything else in the tree is the upper half.

Set the upper half aside as one patch rather than file by file. A patch carries added, deleted, renamed and binary files uniformly, so there is no per-path cp to get wrong and no deletion to reconstruct by hand. Some repos forbid git stash because toolchain or lockfile churn makes pops conflict - look for a CLAUDE.md note or a sibling skill that says so; the patch route avoids it either way.

lower="$(git rev-parse HEAD)"                                 # the lower PR's commit
git status --short                                            # read the ?? lines first (below)
git add -A                                                    # stage the whole upper half, new and deleted files included
git diff --cached --binary "$lower" > "<scratchpad>/upper.patch"
git apply --check --cached -R "<scratchpad>/upper.patch"      # the patch must reverse cleanly before anything is reset
git reset --hard "$lower"                                     # tree = lower PR alone

git add -A sweeps untracked files too, including ones that have nothing to do with either half. Read the ?? lines before it: anything there that is not part of the split needs moving aside first, or it becomes part of the upper PR.

Then build and test every module the split touches, and stop if any of them fails:

failed=0
for m in "<module dir>" "<module dir>"; do
  (cd -- "$m" && go build ./... && go vet ./... && go test -count=1 ./...) || { echo "$m FAIL" >&2; failed=1; }
done
[ "$failed" -eq 0 ]                                           # a lower PR that does not stand alone is not submitted

Use -count=1. Go's test cache will happily report ok for a package whose result predates the change you are verifying, which makes a broken split look clean - and will do it while printing a plausible duration, so the output gives you no hint.

go build ./... writes a binary into the working directory for every main package it builds, and this repo's .gitignore does not name every one of them. Left in place, an untracked binary from validating the lower half rides along into the upper PR the moment Step 8 runs git add -A. Clean each module directory before bringing the upper half back:

for m in "<module dir>" "<module dir>"; do
  git clean -fd -- "$m"                                       # remove anything the build loop left behind
done

Bring the upper half back and confirm the tree is exactly what it was:

git apply --index "<scratchpad>/upper.patch"
git status --short                                            # the same upper-half paths as before the reset

A failure in this step means the cut is in the wrong place (see When to stop and ask), not that the lower PR needs patching up.

Step 8: Create the upper PR

Only once Step 7 passed and the upper half is back in the tree.

The branch you started on becomes the lower PR. If it already has an open PR, keep its name: GitHub ties a PR to its branch name, so renaming strands the PR and a resubmit opens a second one. Name only the new upper branch (gt create derives it from the message), and reach for gt rename only when no PR exists yet and the name now describes part of what the branch holds.

gh pr list --head "$(git branch --show-current)" --state open --json number,url   # non-empty: keep this name

git add -A                            # everything remaining is the upper PR
gt create -m "<type>(<scope>): <the other decision>"

gt submit --stack publishes both PRs. Stop here and confirm with the user before running it - the confirmation in Step 5 covered the split's shape, not this. Never run it without that go-ahead, whatever the split looked like.

gt submit --stack

Graphite notes

• gt split --by-hunk is the built-in way to divide one commit, but it walks every hunk interactively. For a many-file split, rewriting the tip with plain git (Step 6) and then gt create is faster and easier to verify.
• gt split --by-commit is the right tool when the branch already contains several commits that each belong in their own PR - Step 0's git log shows whether it does; if the commits already match the concerns, this is the whole job.
• Rewriting a tracked branch's history with plain git (Step 6 resets to the base) is fine when it has no children. If it does, run gt restack afterwards and expect to resolve conflicts upstack.
• gt log long shows the resulting stack shape with messages - read it once before submitting to confirm the order and the messages match the decisions.
• To check for unresolved conflicts, ask git rather than grepping for markers: git ls-files -u is empty when nothing is unmerged. A regex like ^======= also matches the =====... separator lines people put in comment banners, so it reports conflicts in files that were never touched.

When to stop and ask

• Two partitions are both defensible, and they imply different review stories.
• The split would require hunk-level surgery on more than one file after Step 4.
• The lower PR fails Step 7 and the fix would mean moving a substantial amount of the upper PR down - that is a sign the cut is in the wrong place, not that it needs patching.