monochromatic-git-policy-cli

A Git wrapper that enforces repository-owned policies. It is installed as git ahead of the real Git on PATH, forwards every command to the real Git, and checks the commands and the commits that pass through it against policies the repository configures as data.

The executable is cli-git-native. The library inside the crate is named git_policy_cli.

How the owner wants this package designed and judged is in DESIGN.md.

What it does

Every Git command goes through the wrapper first. Read-only commands, and commands the wrapper has no rule for, run the real Git unchanged.

• Rules before forwarding. require-root demands the repository top level as the working directory. linked-worktree-only keeps every command form that can discard uncommitted work out of the main worktree, and out of a linked worktree while it holds uncommitted work (see "Discarding commands"). branch-worktree-only sends branch creation through git worktree add. add-explicit rejects bulk staging patterns such as git add ..
• Commits landed by a service. A git commit is prepared in a private index, the content policies check the exact bytes, and only then does the commit service of the worktree land it on the real index and branch by compare-and-swap. The service is started by the operating system on the first commit (socket activation); the wrapper never starts it, and on a machine where no service can be reached or started a commit stops with a plain message and changes nothing. When another commit lands first, the wrapper prepares its own commit again on top of the new state. Concurrent commits from several worktrees or sessions therefore never overwrite each other. The wrapper takes no lock of its own: a commit meets only Git's index.lock, which the service holds for the branch move and the index build only (a median of 5 ms in the measurement in SPEC.md). A commit the service was landing when it died is settled by the service from its journal when it starts again.
• Merges that do not fail (the owner's rules of 2026-11-03: "Auto merging should never fail." and "Manual merge should never fail too."). Native Git's merge stashes first and dies when another process holds index.lock, changes the checkout before it moves the branch, and refuses every git merge that creates a commit while the index holds any staged change, even when the merge touches none of the staged files. The wrapper waits for the lock, lands the merge on a branch that moved meanwhile, and computes the merge from Git objects, commits it with the parents, message and signing native Git would use, and lets Git's own fast-forward update move the checkout to it. Every staged entry, unstaged edit and untracked file that the merge does not change stays byte for byte as it was. A merge that changes a path someone else has staged or modified, or that conflicts, is refused with the paths named and the ways forward, and changes nothing. Shapes the wrapper cannot make exact (--squash, --no-commit, -s, -X, several branches, an editor) run in native Git unchanged; see "Merging" for the full list.
• Package-manager state in a new worktree. After git worktree add has filled a worktree and copied its ignored files, the wrapper starts each tool's own offline repair there: pnpm install --offline --frozen-lockfile for the copied node_modules, a clone of the source's external Cargo build directory followed by cargo clean --package <member> for each workspace member, and mpm-shadow install --frozen-lockfile for toolchain launchers that still name another checkout's lock. A repair never fails the add: when a tool is missing, offline data is incomplete or a lock disagrees, one line names the command to run (see "Repairing package-manager state in a new worktree").
• Hooks. cli-git supports no external hook program: not a traditional hook (an executable script in .git/hooks or in a core.hooksPath directory) and not a hook configured with hook.<name>.event and hook.<name>.command. Every real Git the wrapper starts, forwarded commands and its own children alike, runs with -c core.hooksPath=/dev/null/cli-git-no-hooks, a directory that cannot exist, and -c hook.<event>.enabled=false for every hook event Git knows, after your global options and before the subcommand, so no -c of yours brings a hook back. The wrapper injects no hook of its own, so pre-push is switched off like every other event. A script that starts /usr/bin/git directly, outside any Git the wrapper started, still runs hooks. The wrapper runs no program that any configuration names: what a hook checked is a cli-git policy under policies in cli-git.config.jsonc, and there is no setting that brings a hook back. A postCommit key in cli-git.config.jsonc is a configuration error that says so. A commit in a repository that configures hooks prints one non-blocking configuration-warning event (code external-hooks-unsupported) naming the keys, the events and the ways forward, and git cli-git doctor lists the same hooks. A commit in a repository that sets core.hooksPath prints one such event too (code traditional-hooks-unsupported). The shims git lfs install writes are the expected case: see "Git LFS" for what replaces them.
• Git LFS. Every push the wrapper starts uploads the LFS objects it publishes pointers for, before any ref moves, in a repository whose own Git configuration (local, worktree or -c; not the global file) has a filter.lfs.* or lfs.* key. When the pushed commits add pointers the step is git-lfs push, which the wrapper runs itself as a child before it starts the push; the push carries no hook, the wrapper never starts itself and writes no shell text. The remote and refs come from the push's arguments (everything for --all, --tags, --mirror and glob refspecs; the branch Git would send for a bare push), and a failed upload stops the push before it starts. The wrapper, as the parent that ran the upload, prints the explanation. The wrapper first reads the pushed commits itself, with its in-process readers, and finds the pointer files they add (compared with each parent, so a pointer added and removed again inside the push still counts). A push that adds no pointer starts no git-lfs and sends no request to the LFS server; a push that adds some runs git-lfs push --object-id <remote> <oid>... for exactly those objects. The remote's tracking refs say what the destination already has only for a configured remote that is fetched from and pushed to one place; a push URL, a pushInsteadOf rewrite, a URL or path destination, a mirror push or an unconfigured remote makes the wrapper read the whole history of the pushed refs instead. A repository the readers decline, or an object missing from .git/lfs, falls back to git-lfs push <remote> <refs> (with --all in the untrusted cases), which scans itself. When git-lfs push fails, the wrapper does not run it again, whatever the cause. A server error (an HTTP 5xx answer other than 501, 507 and 509, which git-lfs reports on a line starting Fatal error: ) is outside our control, like a network that is down. That is pick 12 of doc/planning/cli-git-shipped-limits-audit.md, made by the coordinator under the owner's delegation and open to the owner's veto. The commit service's push retry plan (CLI_GIT_PUSH_RETRY_SECONDS, doubling to 600 seconds) tries the auto-push again by itself, and a typed git push ends at the first server error, with git-lfs's own words, so that you push again. The wrapper reads the standard error of git-lfs through a private temporary file and relays it with the user and password of every URL replaced by <credentials>. It covers a typed git push (including the merge push of a lost race), the auto-push after a commit and its reconcile pushes. --dry-run pushes and the wrapper's own probes never upload. --no-verify, GIT_LFS_SKIP_PUSH and -c hook.cli-git-lfs-pre-push.enabled=false skip it. Without git-lfs on PATH the push stops before anything starts and the message says to install it, remove the LFS configuration or use git push --no-verify. The other three hooks git lfs install writes (post-checkout, post-commit, post-merge) only keep lockable files read-only, do nothing without the lockable attribute, and are not run. See the SPEC section "Git LFS upload before a push" for the evidence and the limits.
• Content policies. final-newline, markdown/autofix, mono/forbidden-root-context, mono/dependent-version-bump, mono/catalog-specifiers and security/forbidden-strings check or rewrite what a commit contains. The same content policies scan what a manual git push newly publishes.
• After a landed commit. The wrapper checks the committed bytes again, then pushes the branch. A branch without an upstream is pushed to origin with --set-upstream when that remote exists. A branch with an upstream is pushed to it, including a branch whose upstream has another name, such as a linked worktree branch created from origin/main: Git itself refuses that push under its default push.default=simple, so the wrapper names the tracked branch. The resident service pushes the branch its record names and never reads HEAD. It sends the branch only where a plain git push of that branch would send it: a remote.<name>.push refspec is applied to the local refs the way Git applies it (an explicit refspec before a pattern, the first matching pattern, a negative refspec last), and push.default decides otherwise. Under push.default=matching the service runs the plain git push <remote> and lets Git match, so every local branch whose name exists on the remote is pushed, not only the recorded one, and a branch deleted on the remote is not recreated. When Git refuses the setting or would push the branch nowhere (several branch.<name>.merge values under simple or upstream, push.default=nothing, a triangular upstream, two refspecs that send different branches to one destination, a refspec that names a ref that does not exist, a refspec that does not cover the branch), nothing is pushed and the note names the branch, the setting and the way forward. The same push is handed over after a git merge, a git pull that merges, a git cherry-pick, a git revert and a git am that succeeded, when the branch tip moved forward and the branch is now ahead of its upstream and not diverged from it. The check reads the branch, the upstream tracking ref and the commit parents from the files and starts no Git. A rebase is not handed over (a rebased branch would need a forced push), git pull --rebase and pull.rebase included, and neither is a branch with no upstream, a diverged branch, a conflict, or a command that moved nothing. --continue of a stopped cherry-pick, revert, am or merge is the command that adds the commit, so it is handed over. Every decision, handed over or not, leaves a line with its reason in cli-git-push-trigger.jsonl in the common Git directory ("reason":"not-ahead" and the like), so a push that did not happen can be explained from the file. The file is shared by every worktree, so each line names its worktree, gitDir and instance, and the branch when it is known; a failed command's line carries the branch and the tip it was at before the command. level is warn for a step that could not decide or a hand-over that failed and info otherwise, and the userinfo of a URL in detail is masked. When the remote refuses that push because its branch moved after the commit landed, the wrapper merges the remote tip into the branch tip from Git objects alone and pushes the merge without force, at most ten times. While the merge is computed and pushed, the local branch, the index, the worktree and untracked files are not read or changed, so another session's uncommitted files cannot block the push. Afterwards the local branch follows the published merge by a fast-forward when nothing can be lost: a worktree with uncommitted work follows too, as long as the move writes none of the files that work touches, and the report says the work was left as it is. Otherwise the branch stays behind and the report names the overlapping files. A conflict that the word-level second merge pass cannot resolve in every file is reported with the conflicting paths and publishes nothing; when it resolves every conflicted file, the merge is published and its commit message and the report list each part merged by word (see "Word-level second merge pass"). An amended or rebased branch is not merged, and its rejection is shown as before. A failed push never fails the commit, and no configuration key turns the push off. Automatic maintenance, as native Git runs it after a commit, runs after each landing as work of the commit service, so a commit returns before it.
• A manual push that lost a race. When another session pushes to the same branch between your commit and your git push, Git rejects the push as non-fast-forward. The wrapper runs a plain push of one branch as a child process with your terminal, so Git's own output is unchanged. When that push fails and the remote branch has commits yours lacks, the wrapper merges them in from Git objects alone and pushes the merge by running itself, so every push policy reads the integrated result, at most ten times. A conflict is reported with the conflicting paths and publishes nothing, after the same second pass. The checked-out branch follows the published merge under the same conditions as after a commit; a detached HEAD stays where it is, and the report says so. Forced, deleting, mirroring, tag, --all, --set-upstream, --porcelain and --dry-run pushes, and pushes of a +refspec, are never integrated, and a failure that is not a divergence (authentication, offline, a refusing hook) keeps Git's own report.
• Linked worktrees. When a Git command creates a linked worktree, the wrapper copies the ignored state of the source worktree into it without any lock: the destination comes from the command line, the copy is a transaction of its own whose journal and stage claim name the process that owns them, and a copy a dead process left is recovered by the next command. Concurrent adds in one repository never wait for one another, and each worktree receives exactly its own copy. The wrapper reads the destination of every form Git's option parser accepts (abbreviated long options resolved as Git resolves them, bundled short options, --git-dir and --work-tree, a destination below directories Git creates); a form Git rejects keeps Git's own error, and a form the wrapper cannot read is made by Git alone with one line on standard error saying the worktree received no copy. git worktree add in the plain forms (-b <branch>, --detach, --quiet, --lock, --reason, one path and at most one commit-ish) is served by the first rung of a ladder that applies and otherwise falls to the next; when the add succeeds without being a snapshot, one line on standard error (unless --quiet) and one journal entry say why, in plain words, and what makes the next add quick (each step leaves a line in the diagnostic log, MONOCHROMATIC_VERBOSE=true):
  1. A btrfs snapshot of the worktree the command is run from (the main checkout included) when it is a clean subvolume on the destination's filesystem. Real Git runs with --no-checkout, the snapshot is made beside the destination, gets the new worktree's .git pointer and the source's index with refreshed stat fields, has the difference between the source's tree and the new HEAD applied by git read-tree -m -u, and is swapped into place with RENAME_EXCHANGE. The snapshot shares every data block with its source and carries the source's whole ignored state (node_modules included), so the ignored-state copy is skipped for it. A main checkout with uncommitted changes (staged, edited, deleted, renamed, mode-changed or untracked files) is snapshotted too: in the new worktree only, git read-tree --reset -u and git clean --force -d make it equal to what git worktree add makes from a clean checkout, the result is proven before the worktree is handed over, and the changes stay in the main checkout, which is never written. A failed proof falls to Git's own checkout and says why. A merge or other operation in progress, a held index.lock, a skip-worktree or assume-unchanged entry, a submodule and a changed .gitattributes file still send the add the slow way, and so does a dirty linked worktree.
  2. Clones (FICLONE) of every tracked file of a clean worktree, with its index, on btrfs, XFS with reflink and overlayfs. It shares disk but takes about as long as Git's checkout.
  3. Git's own checkout, then the ignored-state copy. A source is clean when no operation is in progress, its index equals the tree of its HEAD, every tracked file has the size, times and inode its index entry recorded, and Git lists no untracked file that it does not ignore. Rungs 1 and 2 are declined for a sparse checkout, a worktree-specific configuration, a submodule, .gitattributes files that differ between the source and the new HEAD, and --force, --no-checkout, -B, --orphan, --track, any -c option, and --no-worktree-copy; rung 1 from a linked worktree is also declined while a nested linked worktree is an ordinary directory (git cli-git migrate-to-subvolume or git cli-git convert-worktree-to-subvolume makes them subvolumes). A snapshot of the main checkout does not decline for one: it sets those worktrees, other checkouts, caches and run records aside in a trash directory under .claude/ and the commit service empties it afterwards, deleting only the directories the snapshot itself moved there (recorded with their identities) and never following a link. Those directories are emptied side by side, within the threads and memory the host allows. It does decline when a registered nested worktree holds a file the main checkout tracks. A trash that cannot be emptied is reported with its path and the reason, and its work item stays for the service. Another process that replaces or relinks directories of the new worktree while this runs is a risk the person accepts by using the command: only the cheap checks (no followed links, identities compared on opening) guard against it. The line reads like cli-git: this worktree was not made as a snapshot, so it took longer, because the main checkout has uncommitted changes in 2 paths (for example notes.md). Commit or stash them, and the next add is a snapshot that takes seconds. It tells uncommitted changes, a running Git command, an unfinished operation, an index that cannot yet vouch for its files (git status once fixes it), a source that is no subvolume, and a snapshot that started and failed; journalctl --identifier=cli-git finds the same entry afterwards, with the fields CLI_GIT_WORKTREE, CLI_GIT_SLOW_CAUSE and CLI_GIT_SLOW_DETAIL. On a filesystem that is not btrfs it says nothing. Every worktree the wrapper creates on btrfs is a subvolume, so rung 1 needs no migration after the first one. Git's post-checkout hook runs on no rung, because every real Git start of the wrapper switches hooks off. git worktree remove <directory> (no option) destroys a clean, unlocked subvolume worktree with one ioctl and deletes its administrative directory when the mount has user_subvol_rm_allowed (the doctor prints the commands that set it); in every other case Git removes it. It does so only for a worktree that the repository you run it in registered and that Git would pick for the name you typed (Git matches the name against the last components of every registered path before it resolves the path); an empty name, another repository's worktree and a name that Git resolves differently go to Git, which words its answer. A subvolume destroy cannot be undone, and the command prints a line that says so each time it destroys one: a write, a lock or a rename that another process makes in the moments between the last check and the destroy is destroyed with the worktree, as it is when Git deletes the files one by one. Do not run it on a worktree something else is still using. If the files are destroyed but the registration cannot be removed, the command fails with status 2 and a report that names the leftover .git/worktrees/<name> directory and the command that finishes the removal (git -C <repository> worktree prune); it does not run Git afterwards. From the main checkout the copy carries the dependencies and builds a new worktree needs to run at once (node_modules, target/, dist/, local configuration files and the repository's .cache/), and leaves out caches, the records of earlier runs and any directory that is another checkout.
• Signals. The wrapper catches SIGINT, SIGTERM, SIGHUP and SIGQUIT while it waits for Git. A signal sent by another process is relayed to the Git child, and the wrapper then ends without leaving state of its own behind.

Other processes changing files

The wrapper checks files, copies them and, in a few commands, removes them, and no check or lock it takes can stop another process from changing them in the same moment. A rename or a replacement of a path, a write through an open file or a shared mapping, and an edit between a check and the step that relies on it are outside what the wrapper guards against, in every command and in the commit service. Running the program is accepting that risk. The cheap guards stay (a lock where one is free, a second look before a removal, an identity check on what is removed), but they narrow the window and do not close it.

Requirements

• Linux with glibc 2.39 or newer. Release archives exist for x86_64-unknown-linux-gnu and aarch64-unknown-linux-gnu only. They are built on the ubuntu-latest runner and link pidfd_spawnp and pidfd_getpid, which glibc added in 2.39 (measured on the 0.2.2 x86_64 archive with objdump -T), so they do not start on Debian 12 or Ubuntu 22.04. The crate does not compile for Windows: cargo check --target x86_64-pc-windows-gnu --lib fails with 16 errors in eight files (the tempfile and ignore dependencies are unix-only and several modules use std::os::unix), so the Windows-specific branches in the sources have never run; each is marked UNTESTED ON WINDOWS in the source. A Windows port is deferred to a later version (owner decision, 2026-10-15).
• Real Git 2.55.0 or newer on PATH after the wrapper. The wrapper checks the version before work that depends on Git's behavior, such as a commit, and refuses older Git. The wrapper finds the real Git by walking PATH, skipping itself and anything that leads back to it: a link or a copy of the wrapper, its toolchain launcher (recognised by the .launch record beside it), and a script that starts the TypeScript wrapper. When PATH offers none, it uses the Git recorded for its commit service, with a notice (see "Install").
• git-lfs on PATH, only in a repository whose own Git configuration has filter.lfs.* or lfs.* keys.
• The scanner and the Markdown linter are linked into the executable and need no separate installation.

Install

The crate is published on crates.io. Install it as a cargo: tool of mise, the way the repository installs monochromatic-lint:

# mise.toml
[tools]
"cargo:monochromatic-git-policy-cli" = { version = "latest", os = ["linux"] }

mise install downloads the release archive of the crate's GitHub release when one exists for the host. Otherwise it compiles the crate from source with the Rust toolchain. cargo install monochromatic-git-policy-cli --locked compiles the same executable into ~/.cargo/bin.

The installed file is cli-git-native, not git. Link it as git in a directory that comes before the real Git on PATH:

# Run in the repository once per checkout, after the installation.
mkdir --parents .cache/cli-git-bin
ln --symbolic --force "$(mise which cli-git-native)" .cache/cli-git-bin/git

Never name a file git inside the tool's own bin/ directory. mise would then create a git shim for the whole account, and every Git command would run through the wrapper.

The wrapper never starts itself as Git. If every git after the wrapper on PATH leads back to cli-git (the link directory comes alone, or the real Git's directory is missing), it falls back to the real Git that its commit service was set up with, when the unit file of the service (cli-git@.service below $XDG_RUNTIME_DIR/systemd/user) names one in its --git option and that file is an executable that does not lead back to cli-git. It then runs your command through that Git and writes one line on standard error that says so and which directory to add to PATH to silence it. This fallback is the coordinator's pick under the owner's delegation, open to the owner's veto (doc/planning/cli-git-no-self-invocation.md). Without a usable record (no unit yet in this login, or a unit that names the wrapper, a launcher of it, or a missing file), it starts nothing and stops with exit status 2. The message lists the PATH it read and every candidate with the reason it was skipped, names the record when one was refused, and says how to fix it: install Git, or put the real Git's directory (for example /usr/bin) on the PATH of the program that ran the command. Every start of the real Git sets CLI_GIT_STARTED_AS_GIT, so a cli-git executable that a wrapper started as Git (another release, or a launcher without a record) stops on arrival with a message saying so; a hook, alias or editor that the real Git runs may still run an installed wrapper. The rule and the incident behind it are in doc/planning/cli-git-no-self-invocation.md.

PATH shadowing covers only processes that find git through PATH. An absolute path to the real Git, a PATH without the link directory, a GUI with an embedded Git library and direct libgit2 consumers bypass the wrapper completely, including its policies.

Commands

The wrapper adds one namespace to Git.

# Check or fix the whole repository, or the named pathspecs.
git cli-git check --all
git cli-git fix -- docs/ README.md
git cli-git check --all --policy final-newline
# List the hook files and configured hooks the repository holds; none of them runs.
git cli-git doctor
# Convert the main checkout into a btrfs subvolume (preview first); see the section below.
git cli-git migrate-to-subvolume --dry-run
git cli-git --help

The wrapper never starts its own executable, never writes shell text or scripts, and explains a failure from the parent process that owns the child (see doc/planning/cli-git-no-self-invocation.md). A Clippy rule bans std::env::current_exe and src/native/source_guard/no_self_invocation_tests.rs scans the source for the rest; nothing is on its exception list that starts the wrapper.

Exactly one scope is required: --all, or pathspecs after --. check and fix read the worktree files the scope selects, unstaged changes included; a commit or a push checks the staged or committed content instead, so a violation that exists only in the worktree fails check but not a commit that leaves it unstaged. --policy <id> limits the run to one policy and may be repeated. A selected policy that is off stops the run with exit 2, because it would check nothing and read as clean: a config-invalid event on standard output names the policy and the cli-git.config.jsonc key that turns it on ("policies": { "<id>": "error" }, or the value the file now sets to "off"). This holds for the four older optional policies as well, which used to exit 0 with no output in that case. A selected policy that is on but does not run for the command is refused the same way, with the command that does run it: fix --policy mono/catalog-specifiers stops with exit 2 and points to check. Without --policy an off policy is skipped without a word, and so it is when a commit, an add or a push runs. Events are JSON lines on standard output.

git cli-git doctor takes no options. It scans <git-common-dir>/hooks and the directory a configured core.hooksPath names, and prints the path of every traditional hook it finds, one per line, on standard output: each executable regular file and each symbolic link, except names ending in .sample. A hook file that is byte for byte the shim git-lfs writes for its own name is not listed there: it is explained on standard error, with its path, as normal and harmless (git-lfs writes these files again whenever it cleans a file in a Git the wrapper did not start; the wrapper does the LFS upload itself and never runs them; delete them or leave them). It exits 0 when it finds none or only such shims, 1 when it finds any other hook, and 2 when it cannot locate or read the repository. Traditional hooks are not supported, so a script that still checks something should become a cli-git policy under policies. It also lists the hooks configured with hook.<name>.event (one hook.<name> (events: ...) line each, exit 1) and notes a hooks key in cli-git.config.jsonc, which is accepted for older files and ignored. A commit does not scan for traditional hooks: it only warns when the configuration sets core.hooksPath or defines hooks, which it already knows without any extra file system access.

The doctor also says whether Git's own fsmonitor is off, on standard error and before the filesystem section (see "Git's own fsmonitor"): one line when core.fsmonitor is false or unset, and a finding with the file, the scope and the command that turns it off when it is on. It never changes the exit status and never writes configuration.

The doctor ends its standard error with a filesystem section for the main checkout of the repository Git selects (so -C, --git-dir and a linked worktree report the main checkout; a bare repository has none). It is informational: it never changes the exit status and starts no program. It reads statfs, the checkout's inode number, /proc/self/mountinfo, /run/ostree-booted and /proc/cmdline. Every line of the section after the first is indented, and the first begins with git cli-git doctor: filesystem:. Four cases are told apart: a checkout on a filesystem that is neither btrfs nor overlayfs gets one line saying that git worktree add uses Git's own checkout and no faster method applies; a checkout on overlayfs gets the detection, that cloning file data inside the overlay works and a snapshot does not apply, and no command; a checkout on btrfs that is an ordinary directory is told that making it a subvolume can help with performance (worktrees can then be made from btrfs snapshots, which took a median of 0.35 s to 1.09 s in the design measurements against 2.1 s to 2.7 s for Git's checkout, see doc/planning/cli-git-btrfs-worktree.md, and until then are made by clones of the tracked files) and is given git cli-git migrate-to-subvolume --dry-run and git cli-git migrate-to-subvolume; a checkout that already is a subvolume gets what a worktree add and a worktree remove do there, and the commands that set user_subvol_rm_allowed when the mount lacks it. On btrfs it also checks the mount option user_subvol_rm_allowed, which lets a regular user delete a subvolume in one call. When it is absent, the doctor prints the command to set it, for the host it finds: sudo mount --options remount,user_subvol_rm_allowed <mount point> until the next boot; on an ostree or bootc host, for the filesystem the kernel command line configures (the one mounted at /sysroot) with exactly one rootflags= word on the running command line, sudo rpm-ostree kargs --delete=rootflags=<old> --append=rootflags=<old>,user_subvol_rm_allowed (delete and append, because the value contains = and libostree splits each argument at the first =, which makes --replace read the old value wrongly), then a reboot; for any other filesystem, or on a conventional host, an /etc/fstab entry (or a systemd mount unit) for it. Nothing is run for the person.

The former trust, untrust and status commands are retired. Configuration is data, so nothing needs approval. They print an explanation and change nothing.

Converting the checkout to a btrfs subvolume

git cli-git migrate-to-subvolume turns the repository's main checkout into a btrfs subvolume at the same path. A subvolume is the precondition for creating worktrees from btrfs snapshots, and git cli-git doctor says when it would help. It runs inside the wrapper: it calls the btrfs ioctls, FICLONE, renameat2 with RENAME_EXCHANGE and its own tree walk, and it starts no program (no cp, rsync, btrfs or mv), so it needs no tool besides the wrapper. Run it inside the repository, from the main checkout or from any linked worktree: it finds the main checkout by reading .git itself. It takes minutes on a large checkout and there is no time limit on it. One command at a time works on a checkout: it owns the checkout through a claim file beside the journal (.cli-git-migrate-<name>.owner.<generation>), a second command is refused with the owner's process id, and a claim a killed command left needs no cleanup. The journal is replaced by writing a sibling file that is created exclusively under a name of its own (the journal's name, .tmp, the process number and a clock reading), synchronizing it and renaming it over the journal, so a link or a file planted or left at a fixed name is never followed, truncated or written through. Releases up to 0.5.20 wrote one fixed sibling name and followed a link planted there. The plain files a killed run left under the writer's own names (that fixed name included) are removed once a later write lands; a link, a directory or any other name beside the journal is left alone.

git cli-git migrate-to-subvolume --dry-run   # preflight and plan; nothing is created
git cli-git migrate-to-subvolume             # migrate; run it again after any stop and it resumes
git cli-git migrate-to-subvolume --rollback [--discard-unreadable]  # undo, finished or not
git cli-git migrate-to-subvolume --remove-old [--discard-unreadable]  # delete the original tree

The steps, each recorded in a journal beside the checkout (.cli-git-migrate-<name>.journal.json): create <root>.subvol-new as a subvolume; create each linked worktree inside the checkout as a nested subvolume (--no-nested-worktrees keeps them ordinary directories), so that a later snapshot of the root holds an empty stub for each; copy the tree, cloning file data so that the new tree shares every data block with the old one (only metadata is copied); repeat the copy as a delta that also deletes what the checkout lost meanwhile; hold a marked index.lock (Git's own lock for writing an index, the one lock that cannot be dropped) for the main index and for every linked worktree's index, in the original before the last delta and in the new tree before the swap, run the last delta and swap the two directories in one atomic call; rename the original to <root>.before-subvolume; carry over files that are newer in the original (writes that landed there after the swap), while the new checkout holds the locks; rewrite the stat fields of every Git index of the checkout and its linked worktrees from the copied files (--no-refresh-indexes leaves them; the first git status then hashes every file, and runs the Git LFS clean filter on LFS files); an entry is rewritten only when Git could trust it (its time is earlier than the index file's own) and the original file still has the stat fields the entry records, so a file edited behind Git's back stays dirty; release the locks and verify that the root is a subvolume with a readable .git/HEAD. The original tree is never deleted by the migration. One measured run, on a synthetic checkout of 121,260 entries (forty thousand tracked files in the checkout and in one linked worktree, forty thousand ignored files) under a scope of CPUQuota=200% and IOWeight=20: the copy took 8.34 s and 9.67 s in two runs, each delta pass 0.81 s to 1.02 s, the index rewrite 0.35 s and 0.40 s, and the whole migration 11.55 s and 14.74 s. The copy cost about 70 to 80 microseconds per entry and a delta pass about 7 to 8 microseconds per entry; the pass under the locks is the delta pass, so they are held for that long, for the exchange, and now also for the catch-up and the index rewrite. Each lock file holds a token that names the migration. A lock that another Git command holds when the migration needs it stops the migration with the file named, and is never taken over, exempted or removed; the migration's own locks are recognised by their token when a run resumes. Every entry is written under a temporary name and renamed into place, so a stop never leaves a half-written file under its real name, and a file that changes while it is copied is copied again.

A process whose working directory is inside the checkout keeps that directory: after the swap its relative writes land in the original tree. The preview and the run list such processes by number; stop agent sessions and editors first, and run cd "$PWD" in shells afterwards.

The command refuses, with the cause and every way forward, and changes nothing, when: the checkout is not on btrfs (overlayfs is named separately), is a mount point, is on another device than its parent or is a symbolic link; it already is a subvolume with no migration recorded; the parent directory has fewer free bytes than the metadata of the checkout's entries needs (2048 bytes per entry: one kibibyte measured per copied entry on btrfs, written twice by the DUP profile; the command lists every directory once to count the entries, and the refusal names the count and both byte figures); <root>.subvol-new or <root>.before-subvolume exists without a journal that explains it; Git is in the middle of an operation (any of index.lock, MERGE_HEAD, CHERRY_PICK_HEAD, REVERT_HEAD, BISECT_LOG, rebase-merge, rebase-apply, sequencer, HEAD.lock, config.lock, packed-refs.lock or shallow.lock in .git or in a linked worktree's administrative directory); the commit service has unfinished work in .git/cli-git/work/pending or claimed. Exit status 0 means done, already done or previewed; 1 means a refusal; 2 means a step failed after the journal began, and the message says to run the command again to resume or to add --rollback.

Files this user may not read (for example root-owned mode 600 logs written by a privileged run) stay in the original tree and are listed in .cli-git-migrate-<name>.unreadable.txt, and the run says so loudly. A directory this user cannot list stops the run before the swap, because copying around it with deletion on would erase its subtree from the new tree. Because the original tree is then the only copy of the unreadable files, --remove-old refuses while the list is not empty, unless --discard-unreadable is given with it; copy the files over with the needed privilege first (for example sudo cp --archive --reflink=always --parents).

--rollback carries back files that are newer in the new root, swaps the trees again (recorded in the journal before the call, so a rerun after a stop never swaps twice), rewrites the restored indexes, removes the migration's own index locks and deletes the new subvolume, nested subvolumes first. Files that only the new subvolume holds and this user may not read cannot be carried back, so --rollback refuses while there are any, unless --discard-unreadable is given with it. --remove-old first carries into the checkout what was written to the original after the catch-up (files newer than their copy, and files the checkout lacks that are no older than the catch-up; older ones were deleted on purpose), then deletes the original. Both deletions make the directories this user owns writable on the way down, so a mode-555 directory does not stop them. Both refuse, before they delete anything, when a directory below the tree is a nested subvolume or a mount point, and name it: its content is not part of the tree. Move it out, delete it with btrfs subvolume delete, or unmount it, then run the deletion again. Deleting a subvolume uses the kernel's destroy call where the mount has user_subvol_rm_allowed and otherwise empties it and calls rmdir, which the kernel allows for an empty subvolume root.

Converting nested worktrees to subvolumes

git cli-git convert-worktree-to-subvolume [--dry-run | --abandon] [--jobs <count>] (--all | <worktree>...) turns a linked worktree that is an ordinary directory into a btrfs subvolume at the same path, several worktrees side by side. A snapshot of the main checkout shows a nested worktree that is a subvolume as an empty directory, so converting the ones under .claude/worktrees/ once removes the work of setting each of them aside at every git worktree add.

• The worktrees share nothing but the filesystem (each has its own administrative directory, index, claim file and journal), so they are converted side by side, as many at once as the host allows: the threads this process may run less an eighth of them (at least one, none when only two), no more than the free memory holds at the measured peak of one conversion, never more than sixteen (the measured point past which the filesystem gains nothing), and no more than --jobs <count> when it is given. The report is the one a run of one worktree after another gives: each worktree's name and lines together, in the order named (with --all, sorted by path), whichever finishes first, with the lines of the first unfinished worktree shown as they happen. A worktree named twice is converted once and then found converted.
• A worktree that fails after its journal began does not stop the others. It keeps its journal and says so, every other worktree is converted, the closing line counts the failures, and the status is 2; running the same command again finishes the failed ones. Releases before this one stopped at the first failure.
• --dry-run lists each worktree and either would convert or why it is skipped: locked, an index.lock, an operation in progress, a modified tracked file, a process with its working directory or an open file inside it, a worktree registered inside it, or a converting directory that no journal owns. Run it first; it changes nothing.
• Without --dry-run the command takes no lock. It owns the worktree through a numbered claim file in the administrative directory (cli-git-convert.owner.<generation>) that names its process; a second command that finds a live owner skips the worktree and names that process, and a claim a killed command left needs no cleanup. A conversion is also refused while a migrate-to-subvolume of the main checkout runs, and that command is refused while a conversion runs. Then it creates <path>.cli-git-converting as a subvolume and records its identity, copies the worktree into it with cloned data, takes the administrative directory's index.lock with its own token only for the rest (the long copy runs without it), proves the copy equal with a second pass, swaps the two directories with one exchange, checks that the path is the subvolume it created and has the same .git bytes, rewrites the index stat fields so the first git status stays fast, gives the index.lock back, and removes the old tree only when it is the directory it recorded and the installed worktree holds all of it. A cli-git-convert.lock file that release 0.5.20 or older left is removed when no command holds it, and a conversion that release left unfinished is completed by running the command again. The registration never changes: the path and the bytes of the .git file are the same.
• The old tree is kept, and the run says why, when a process still holds it, when anything in it differs from the installed worktree (size, modification time, kind, link target) or is missing there, or when it contains a subvolume or a mount point, which this command does not delete. Nothing is ever copied from the old tree into the installed one, so an edit made to the installed worktree is never overwritten; whoever knows which version is wanted settles it by hand and runs the command again.
• A quoted line in any attributes file, an attribute source chosen with GIT_ATTR_SOURCE, attr.tree or --attr-source, or a filter, ident or encoding attribute that may apply to a file whose recorded size and times cannot vouch for it, makes the command refuse that worktree and name the reason, instead of deciding by the file's content.
• Where Git reads its system configuration file and its system attributes file from is a fact only the real Git the wrapper forwards to can give (git var GIT_CONFIG_SYSTEM, git var GIT_ATTR_SYSTEM, learned once per executable and cached). The command resolves that Git itself, as every command that forwards does, because the wrapper puts the variable that names it only into the environment of the Git processes it starts, never into its own. Release 0.5.20 and older read their own environment, found no Git named, and skipped every worktree whose index entries the stat data could not vouch for with "the Git configuration cannot be read without running Git"; this release judges those worktrees by content. Where no real Git is found, such a worktree is still skipped with that reason, which is the safe side.
• Risk you accept by running it: it deletes the old directory, and it cannot stop another process from changing the worktree while it works (see "Other processes changing files"). It says so once on standard error before it starts, and in --help.
• Run the same command again after any stop: it completes the conversion in place. --abandon undoes one that has not reached its swap, destroying only the subvolume its journal says it created.
• Measured on a fixture of four ordinary worktrees of 22.3 thousand entries each: 4.2 s per worktree under a scope of CPUQuota=400%, load average 6 to 10.
• Stop sessions that have a working directory in the worktree first; the command refuses while one does.

Repairing package-manager state in a new worktree

A copy keeps symbolic links and build state exactly as they were, which is wrong for tools that record where they were installed or built. pnpm's links climb out of the tree by a fixed number of levels to reach the shared store, so in a worktree at another depth they are dead, and Node then resolves the name through a parent node_modules to the source's packages without any error. Cargo bakes the source's paths into what it built (CARGO_MANIFEST_DIR, build script output). The toolchain launchers copied into .cache/toolchain-launchers name the source's toolchain.lock.json. After the ladder (or Git's checkout) and the ignored-file copy, the wrapper starts each tool's own repair in the new worktree:

• pnpm: pnpm install --offline --frozen-lockfile, when pnpm-lock.yaml and node_modules are both there. The pnpm is the first one on your PATH (in this repository the pnpm-shadow, whose warnings are shown only if the run fails).
• Cargo: for each Cargo workspace, the source's external build directory is cloned to the worktree's when the worktree has none, then cargo clean --offline --package <member> runs for every member (a Cargo.lock the clean would rewrite is put back as checked out, so the worktree is not dirty). Registry and git dependencies stay reused; the worktree's own crates rebuild with their own paths. A file or directory in the source's build directory that you cannot read is left out of the clone and counted in the success line; the source is never touched. The workspaces are repaired side by side, on as many threads as your machine's available parallelism.
• Toolchain: mpm-shadow install --frozen-lockfile. It also downloads the tools the worktree's lock names that are not installed yet, and can retire tool copies another checkout still links (both accepted by the owner), and its summary of what was installed is part of the success line.

Nothing runs for --no-worktree-copy or --no-checkout. --quiet silences the one line that reports success; a failure is always printed. When a tool is missing, its offline data is incomplete, a lock does not match or the run exceeds its limit, the add still succeeds and prints one line that names the reason and the command to run, such as pnpm --dir <worktree> install --offline --frozen-lockfile. Another process changing files while a worktree is copied or repaired is your risk and is not defended against. Times measured on this repository are in doc/planning/cli-git-main-worktree-ignored-copy.md, section "Package-manager state of a new worktree".

Version and help

The wrapper runs the real Git for the top-level --version, -v, --help and -h, so Git's own output and exit status come first and are never changed. Only when Git exits with status 0 does the wrapper add plain text to standard output, so a person can tell that a policy wrapper sits in front of Git. When Git exits with another status, or a signal ends it, nothing is added and the wrapper returns Git's status. Global options before the query, such as git -C <directory> --version, behave the same. A subcommand's own help, such as git commit --help, is untouched and so is every other query (--exec-path, --html-path, --man-path, --info-path, --list-cmds=<group>): the wrapper replaces itself with Git and their output is byte for byte Git's.

--version and -v add exactly one line, where <version> is this crate's version and <path> is the real Git the wrapper resolved:

monochromatic-git-policy-cli <version>, shadowing git at <path>

--help and -h add a blank line and a short section:

monochromatic-git-policy-cli <version> sits in front of this git
It adds repository policies on commit, push and worktree commands,
the `git cli-git` management command (`git cli-git --help`),
and `git cli-git doctor`, which lists the traditional hooks a repository holds.
real git: <path>

A script that compares the output of git --version with a fixed string should read the first line only, or call the real Git by its path. The wrapper's own findings stay JSON lines as described under "Output and exit status"; this text appears only for these four spellings.

Wrapper-only options

A wrapper-only option is removed before Git sees the arguments and applies to that one invocation.

• --no-enforce-<policy id> skips one policy, for example --no-enforce-final-newline or --no-enforce-security/forbidden-strings.
• --no-enforce-worktree skips linked-worktree-only for the forms listed in "Discarding commands". Write it directly after the subcommand, for example git restore --no-enforce-worktree -- <path>.
• --no-enforce-worktree-branch skips branch-worktree-only.
• --no-enforce-bulk-add skips add-explicit.
• --no-worktree-copy skips the ignored-state copy into a new linked worktree and the whole worktree ladder: Git alone makes an ordinary directory.
• --cli-git-keep-going reports every finding of the command in one run instead of stopping at the first blocking one.

Hints

Every block and every divergence from real Git's behavior says what you were probably trying to do and how to proceed. A JSONL event (a finding, an engine failure or a warn-unsafe notice) carries the hint in its own hint field, beside a message that is unchanged from the retired wrapper's text. A plain notice (an internal error, a push stopped by its deadline or its output cap, a lock wait that gave up, a failed worktree copy, a real Git that cannot start, a commit that normalization emptied) ends with the hint. The rejections that already carry their own alternatives (require-root, linked-worktree-only, branch-worktree-only and the commit-only rules) keep them in the message. src/native/policy/event_hints.rs holds the texts, and its tests fail when a finding code or an engine failure code has no hint.

Discarding commands

linked-worktree-only rejects a command form that can overwrite or delete uncommitted work by itself. The rule is:

• In the main worktree such a form is always rejected, clean or not.
• In a linked worktree it is rejected while the worktree holds uncommitted tracked changes (for git clean, also untracked files), and runs when the worktree is clean.
• A form that only moves HEAD (git checkout <branch>, git switch <branch>) is rejected in any worktree only while it holds uncommitted tracked changes.
• A status Git cannot read stops the command instead of reading as clean.
• Repositories under an allowlisted tool directory are exempt: uv's cache (UV_CACHE_DIR, else XDG_CACHE_HOME/uv, else ~/.cache/uv) and mise's data directory (MISE_DATA_DIR, else XDG_DATA_HOME/mise, else ~/.local/share/mise), where mise keeps its plugin clones and updates them with a forced checkout.

Every rejection names the form, what it would destroy, the alternative that fits the most likely intent, and the deliberate way to discard anyway. The deliberate way is always the same: save a copy (git diff > backup.patch) and pass --no-enforce-worktree directly after the subcommand. git apply backup.patch brings the edits back.

During a merge, a rebase, a cherry-pick or a revert there is one exception. git checkout --ours|--theirs -- <path>, git checkout <rev> -- <path>, git checkout --merge -- <path> and the git restore forms with --ours, --theirs, --source or --merge run for conflicted paths only (the paths with an unmerged index entry), in the main worktree and in a linked one, after the wrapper saves your current file. The copy goes to cli-git/discarded/<UTC time>/<path> in the common Git directory, beside cli-git-audit.jsonl, and the wrapper prints where before Git runs:

cli-git: saved your copy of a.rs to
  /path/to/repo/.git/cli-git/discarded/2026-11-02T14-01-07/a.rs
  before this command replaces it (a.rs was conflicted).

A path that is not conflicted keeps the rejection, and a line that names conflicted and clean paths together is rejected whole with the clean paths named. Write each conflicted file by its own name after --; a directory, a wildcard, pathspec magic, a missing --, --force, --patch, --pathspec-from-file and restore --staged are not covered and stay rejected. A token that is the value of --source, -s or --conflict is never read as --worktree or --staged, so git restore --source -W --staged -- f (a tag can be named -W) is an index-only restore and stays rejected, and the rejection names git restore --staged, not a restore of file content. A conflicted path that is deleted in your worktree has nothing to copy, and the command runs. When a copy cannot be written (no space, no permission), the command is rejected and says why. A symbolic link or a file where a directory in front of the name should be (dir for dir/file) is never followed: Git would replace it with a real directory and it would be lost unsaved, so the command is rejected, names dir and says to move it out of the way or save it first. A file that changes while it is being copied (an editor's atomic save, an append, a rewrite in place) is not reported saved: the copy is removed, the command is rejected, and running it again once nothing writes to the file goes through. A file that another program changes while the save runs is your risk. A pathspec mode (--icase-pathspecs, GIT_ICASE_PATHSPECS and the glob, noglob and literal forms) keeps the rejection, and when every path you named is a conflicted file the message names the mode and says the same command without it is saved first and then allowed. The same holds for pathspec magic on a conflicted file's name (:(icase)a.rs, :!a.rs, :/a.rs): the message names the magic and the file, and naming the file itself is saved first and then allowed. The folder is bounded, because it would otherwise grow for as long as the repository lives, and the copies are the only remaining content of replaced files. When a new copy has been written and synced, the wrapper removes the saves older than 30 days, then, while the folder is over 1 GiB, the oldest saves first. A save younger than 24 hours is never removed by either limit, and neither is the save just written. Only directories named like the wrapper's own (2026-11-02T14-01-07) are touched. No commit waits for this, and no lock is taken: passes that overlap share a save's files, and each copy is audited once. The folder is never followed through a link: if .git/cli-git or .git/cli-git/discarded is a symbolic link, nothing is removed, and the command's notice and git cli-git doctor say so; make it a real directory to apply the limits again. A save that cannot be read, a directory replaced while the pass ran, a removal that fails part way and an audit file that cannot be written are all kept and named the same way, never counted as removed. Each removed save leaves one line in cli-git-audit.jsonl ("outcome":"pruned", with the time, path, bytes, reason age or size, the limit it was judged against and the names of the copies in it). The limits are the discardedCopies section of cli-git.config.jsonc. The defaults are not yet based on a measurement of growth: on 2026-11-03 no repository on the development host had a saved copy, and a copy can be the only copy of your work, so the numbers are open to change (doc/planning/cli-git-shipped-limits-audit.md, "Defaults for saved conflict copies"). An example that keeps saves for 90 days and the folder under 256 MiB:

// cli-git.config.jsonc: keep saves for 90 days and the folder under 256 MiB
{ "discardedCopies": { "maxAgeDays": 90, "maxSize": "256m" } }

A value that is no limit (zero, a fraction, text that is not a size) is a configuration failure that names the key; to keep copies longer, set a large number. git cli-git doctor reports the folder's size, the number of copies and saves, and the oldest save, once it holds any.

Each decision about a discarding form (rejected, run in a clean worktree, run in an exempt cache, or run with the escape) is appended as one JSON line to cli-git-audit.jsonl in the common Git directory, with the time (time and timeMillis), a level (warn for a rejection), the release, the process and its parent, the worktree kind and the worktree's Git directory (gitDir), the form, the working directory and the arguments (a decision that saved conflicted files also lists where each copy went). When uncommitted work disappears, git rev-parse --git-common-dir locates the file; a line that says allowed or hatch near the time of the loss names the command that was let through. At one mebibyte the file moves aside to the next free number (cli-git-audit.jsonl.1, then .2, and on; the highest is the newest) and a new file starts. Rotated files that name no saved copy still on disk are kept up to four mebibytes, oldest removed first; a rotated file with a saved-conflict line whose copy still exists is never removed. The four mebibytes are derived from the measured rate of this file (about 23 KB a day) and the 30 days copies are kept; the timing logs and the push trigger log keep one rotated file, .1. A failed write never changes a command's result.

The forms and their alternatives follow. Each alternative is run through the built executable by src/native/binary/binary_discard_tests.rs.

• git stash (every form except list, show and create): commit work in progress with explicit paths on a topic branch in its own linked worktree (git worktree add -b <branch> <path>).
• git clean without --dry-run: git clean -n lists what would go, delete specific files with rm, or use the project's own clean task.
• git reset --hard, --merge, --keep: git reset --soft <commit> undoes a commit and keeps its changes, git diff <commit> compares, git worktree add <path> <commit> gives a clean copy.
• git reset in the default or --mixed mode: git commit --message <message> -- <paths> commits only the named paths and ignores other staged entries, git diff --cached --name-only lists what is staged.
• git restore --staged (unstaging): the same two commands as git reset.
• git restore and git checkout with paths, --force, --patch, --ours or --theirs (reverting files): git diff -- <path> and git show HEAD:<path> compare, git worktree add <path> <branch> opens another branch, or save a patch and use the hatch.
• git checkout <branch> and git switch <branch> while the worktree is dirty, and git switch --force: commit with explicit paths first, or open the branch with git worktree add <path> <branch>.
• git rm --force: git rm <path> refuses a modified file, and git rm --cached <path> stops tracking and keeps the file.
• git mv --force: pick a destination that does not exist, or git rm <destination> first.
• git checkout-index --force: git checkout-index --prefix=<directory>/ --all writes copies elsewhere.
• git read-tree --reset: git read-tree -m merges, or reset in a clean linked worktree.
• git merge --abort, git rebase --abort, git cherry-pick --abort, git revert --abort, git am --abort: the same command with --quit stops the operation and keeps the index and files, --continue finishes it after resolving conflicts, and --skip drops the current commit or patch.

The forms are read with Git's own option grammar, so abbreviations, clustered short flags (-rf) and a path after -- are handled as Git handles them. Commands without a table of their own (rm, mv, checkout-index, read-tree, restore and the --abort commands) take the hatch directly after the subcommand.

Merging

The owner's rules (2026-11-03): "Auto merging should never fail." and "Manual merge should never fail too." Native Git's merge that creates a commit does two things that break them in a checkout many sessions share: it first runs git stash create on the index (save_state() in Git's builtin/merge.c; no stash ref is ever stored), which dies with could not write index and stash failed when another process holds index.lock, and it changes the index and the worktree before it moves the branch, so a branch that moved meanwhile (another session's commit, or the auto-push's own reconcile merge) leaves the checkout half merged with fatal: update_ref failed for ref 'HEAD'. The wrapper therefore completes the merge itself whenever the command is one it understands, whether the index differs from HEAD (other sessions' staged work) or not. Native Git also refuses, when the index differs from HEAD, a git merge <branch> that needs a merge commit with error: Your local changes to the following files would be overwritten by merge, whichever files are staged. The wrapper completes that merge instead, when all of these hold:

• the command is git merge with one commit-ish and only the options -m, --no-edit, --ff, --no-ff, -q, -v, --stat, --no-stat, -n, --log, --no-log, -S, --gpg-sign, --no-gpg-sign, --no-verify, --allow-unrelated-histories and --no-progress (full spellings only);
• HEAD is not already ahead of or equal to the merged commit (native Git fast-forwards and answers "Already up to date" by itself);
• no merge, rebase, cherry-pick or revert is in progress, and the settings merge.autostash, merge.verifySignatures and branch.<name>.mergeOptions are unset;
• the merge would not open an editor: pass --no-edit, or set GIT_MERGE_AUTOEDIT=no (a merge started from a terminal opens one by default, and then native Git runs it as before).

The merge tree comes from git merge-tree --write-tree, the commit from git commit-tree, signed when -S is given or commit.gpgSign is true. Its message is Git's own (git fmt-merge-msg, then the whitespace cleanup of a commit that opens no editor), and the checkout then moves with git merge --ff-only <merge commit>. That is Git's two-way update: it writes exactly the paths the merge changes and keeps every other index entry and file. The merge commit skips pre-merge-commit, prepare-commit-msg and commit-msg, like the merge the auto-push makes from objects; post-merge runs once. The branch reflog reads merge <branch>: Fast-forward.

Two refusals change nothing:

• The merge changes a path that is staged, or modified in the worktree, by someone else. The finding merge-isolation/paths-in-use names the paths. The ways forward are to wait until the owner has committed or restored them, to merge in a clean linked worktree and bring the result back with git merge --ff-only, or, for your own files, to commit or stash them first.
• The merge conflicts textually, and the word-level second merge pass cannot resolve every conflicted file. The finding merge-isolation/merge-conflict names the conflicting paths. Resolve in a clean linked worktree or checkout, or rebase the branch. A conflict that does resolve by word completes the merge, and the lines of the merged parts come first on standard error.

An untracked file where the merge adds one is refused by Git at the last step; the finding merge-isolation/not-applied follows Git's message, and nothing is changed. A merge killed while it runs is finished or discarded by the next invocation (see "State on disk"). A git pull that merges is completed the same way when the command line or the settings say that it merges (--no-rebase, pull.rebase or branch.<name>.rebase set to false, or --ff or --no-ff) and every option and operand is one the wrapper reads exactly: the wrapper runs the pull's fetch, then the merge of FETCH_HEAD as a typed git merge --no-edit FETCH_HEAD would run. Every other pull is native Git's whole (a rebase, --ff-only, --autostash, pull.ff, submodules, a refspec with a destination, an abbreviated option, a pull with nothing stated, which Git refuses with its advice when the branches diverged), and it keeps native Git's failures (the held lock, the staged index, the late ref update); run git fetch, then git merge --no-edit origin/<branch>.

Waiting for the index lock

Before Git's fast-forward changes anything, the merge waits for any program's hold on index.lock (the commit service's, a foreign git's, or a lock nothing proves held): until LOCK_CEILING (10 minutes), whether its owner is proven alive or not (the line for an unproven owner says what is known and how to remove the lock if no program uses the repository); at the bound it names the holder and the merge ends with Git's refusal, non-zero, with nothing changed. The notice names the merge, says it has changed nothing yet, names the lock and the holder with its proof (an open descriptor or Git's PID file), and repeats every CLI_GIT_WAIT_NOTICE_SECONDS. No lock is ever removed. Native Git, measured in a fixture, never waits: a held lock fails the merge in a few milliseconds. A merge that native Git runs (a conflict, an editor) waits for the lock first in the same way.

A branch that moved while the merge was prepared

The merge commit is built on the head the merge was computed on, and Git's fast-forward moves the branch only from that head. When the branch moved first, Git's fast-forward refuses without changing anything, or, if it lost the race in the instant between its checkout and its ref update, the paths it installed are put back to the moved head's content (only paths that hold the old, the merged or the moved head's content; any other content stops it with the paths named and the record kept). The merge is then computed again on the moved head and lands as a merge commit whose first parent is that head, up to twenty times in a row; each round says so on standard error. No MERGE_HEAD is ever left behind by a merge that did not finish.

A conflict

With another session's staged work, the word-level second pass (see "Word-level second merge pass") runs on the conflicts of the merge computed from objects, and the merge completes when every conflicted file resolves; otherwise it is refused (merge-isolation/merge-conflict) and nothing changes. With a clean index the merge is native Git's, which writes the markers for the person to resolve in the worktree; the second pass then runs on the stopped merge.

What can still end a merge with a nonzero status

Each with its cause:

• A conflict that no resolution settles (exit 1 from native Git, or merge-isolation/merge-conflict, exit 2).
• Unsaved work in the worktree that the merge would overwrite: merge-isolation/paths-in-use (exit 2), or Git's own refusal with a clean index; an untracked file in the way (merge-isolation/not-applied).
• A lock that stays held for the whole wait limit: the merge stops, changes nothing and leaves the lock.
• A branch that moved in twenty rounds in a row: the merge stops and changes nothing; run it again.
• A form the wrapper does not complete and native Git runs unchanged: --squash, --no-commit, -s, -X, more than one branch, --abort, --continue, --ff-only, abbreviated options, an editor, merge.autostash, merge.verifySignatures, branch.<name>.mergeOptions. These keep native Git's failures (a held lock is waited for first, but the stash of a native merge commit and the late ref update are native Git's).
• A merge or sequencer operation already in progress (native Git reports it).
• A signer that fails when the merge is signed (nothing is changed).
• A repository or index the wrapper cannot read (the failure names the file).
• A git pull that needs a merge and is not read exactly: it runs its merge inside Git, out of the wrapper's reach.

Word-level second merge pass

Git's line merge calls two changes that touch a conflict: a line you edited and lines the other side inserted right after it, for example, in a paragraph that is hard-wrapped one sentence to a line. When a git merge, git rebase, git cherry-pick, git revert or git pull stops on such a conflict, the wrapper re-merges only the conflicted parts word by word (the way Jujutsu does with merge.hunk-level = "word"), and accepts a file only when every part of it resolves. A file that resolves is written to the worktree and staged as resolved. When every conflicted file resolves, the operation is finished the way Git finishes it after a clean run (through the wrapper's own command path, so the commit policies apply): a merge commits with its prepared message, a cherry-pick or revert commits and continues, a rebase continues, and the exit status is the finishing command's. --no-commit and --squash stop as they would after a clean run, with Git's line "Automatic merge went well; stopped before committing as requested" and status 0, and --edit opens the editor. git pull is covered, both its merge and its rebase; --autostash, merge.autoStash, rebase.autoStash and branch.<name>.mergeOptions are applied as Git applies them; a command with an option the wrapper does not read exactly is left stopped with the files staged and Git's status. What Git printed when it stopped (CONFLICT lines) is already out and stays. An autostash (--autostash, merge.autoStash, or a pull's) is applied by the commit that ends the finished merge, as in Git. When the stash conflicts with the merged files, the merge commit is still made, the stash is kept in your repository as stash@{0}: autostash, the files it conflicted in are unmerged in your index with the markers in the worktree, and the exit status is the commit's, as after Git's own apply. The same holds when you conclude such a merge yourself with git commit, and when you have edited a file the stash touches since the merge stopped (Git's apply then refuses and keeps the whole stash). Standard error lists every part that was merged, with the file and the lines in the base, your side and the other side, and the text it became:

cli-git: word-level merge: notes.md: part 1 of 1: base line 3, ours line 3, theirs lines 3 to 4 became "Flags are optional, see below.\nExit code is zero on success.\n"

The journal gets one entry per merged file (journalctl --identifier=cli-git), with the fields CLI_GIT_WORD_MERGE_FILE, CLI_GIT_WORD_MERGE_OPERATION and CLI_GIT_WORD_MERGE_PARTS. Only conflicts the command itself created are touched; a file you are resolving by hand is never replaced. A file that does not resolve stays exactly as Git left it. That includes two sides that append different text at the same place, two different edits of the same word, a file with a NUL byte in its first 8000 bytes (Git's own test), a file over 8 MiB, a path added on both sides or deleted on one, a link, a path whose mode differs between the sides, and a path whose merge attribute names a driver, binary or -merge.

The result can be text neither side wrote: two edits of different arguments of one call become one call with both edits. Nothing either side added or kept is dropped, and the parts are listed so you can check them. The wrapper's own merges from objects (the merge of a staged-index checkout above, the merge of an auto-push that lost a race, and the replay of a commit another commit beat to the branch) use the same pass and publish or commit a merge only when every conflicted file resolves. They read no worktree, so they read the .gitattributes of the branch tip (GIT_ATTR_SOURCE) unless you named an attribute tree yourself (--attr-source, GIT_ATTR_SOURCE or attr.tree), and the union line for doc/handover/*.md applies in them too. Design and measurements: doc/planning/cli-git-merge-auto-resolution.md; the rules: SPEC.md, "Word-level second merge pass".

What can still end a commit with a nonzero status

The owner's rule (2026-11-03): "Commits should never fail." A named-path commit beside another session's unfinished merge or stopped cherry-pick lands as an ordinary commit and the state is carried on (see SPEC.md, "Merges and commits never fail"). Each of these still ends a commit with a nonzero status, with its cause; none is waited out or retried on the moved head yet:

• A Git step the client starts shows no sign of work for its quiet window (child_class.rs: 60 seconds for a quick query step, 300 seconds for a bulk step; both unmeasured), and is quiet again when run a second time. A step that works is never ended for being slow, and a step that may not be run twice (a reference update) is waited for.
• The commit service does not start: a start the manager reports as under way is waited for until START_CEILING (60 seconds), a start it reports failed ends the wait at once ("commit service did not start"), and a unit it reports inactive is asked to start after START_WAIT.
• The service does not answer within ANSWER_CEILING (20 minutes, heartbeats or not): the client withdraws the request, so it never lands later by itself, and ends with "This commit waited N minutes for the commit service and was withdrawn"; if the commit landed at that moment the command reports success. Five seconds without a word ask the service again with the same commit, and a service that does not answer then ends the command with "The commit service stopped answering".
• The branch keeps changing: LANDING_ATTEMPTS (3, service_land.rs) landings lost to other commits end with BranchKeepsChanging, and REBUILD_ATTEMPTS (service_commit.rs) builds by Git on a moved branch end the commit.
• A commit that concludes a merge, an amend or a revert cannot replay onto a branch that moved (concurrent-commit/head-moved).
• index.lock is held by a program that does not let go within LOCK_CEILING (10 minutes), or was left by a dead owner (Git's lock PID file names a process that is gone; a diagnostic names it at once). The commit is withdrawn, kept on a side ref named in the message (refs/cli-git/kept/<batch>), your staged changes stay staged, and nothing lands later by itself; run the commit again.
• The content a commit captured was pruned from the object store before it landed (ClientFailure::ContentPruned).
• A policy finding at error severity, an empty message or a failing signer. A named path in conflict is no longer one of these: it is committed with the markers it holds and a warning that names it (owner, 2026-11-03: "Committing a conflicted path is fine. Warn and commit anyway, even if conflict markers will be committed."). A plain git commit with no paths while files are still unmerged finishes the merge the same way, with the markers in those files and the same warning for each (owner, 2026-11-03: "Warn and finish the merge (Recommended)"); Git's "Committing is not possible because you have unmerged files" no longer ends it. git commit -a over unmerged files commits their markers and warns for each file in the same words.

Output and exit status

Policy events of a wrapped command are JSON lines on standard error, one object per line, with schemaVersion 1. Real Git keeps the normal standard streams.

• When real Git does not run: 0 means clean or warnings only, 1 means final error findings, and 2 means a usage, configuration, commit or engine failure.
• When real Git runs, its exit status is preserved. A landed commit that a post-commit policy then blocks exits 2 and stays local.

Configuration

The wrapper reads cli-git.config.jsonc at the repository top level. The file is JSON with comments, at most 1 MiB, and it is data: it never names code to run. Without the file, the defaults apply. The former TypeScript wrapper read cli-git.config.mjs and cli-git.config.ts; those files are never executed or read. With no JSONC file beside them, they are a configuration failure that asks for the translation. Beside a JSONC file, they are reported and ignored. The keys plugins and trust are refused.

// cli-git.config.jsonc
{
  "policies": {
    // A severity alone: "off", "warn" or "error".
    "final-newline": "error",
    // A severity with options.
    "markdown/autofix": ["warn", { "rules": ["lfs-image-url"], "exclude": ["package/ssg/"] }],
    "mono/forbidden-root-context": "error",
    "mono/dependent-version-bump": "error",
    "mono/catalog-specifiers": "error",
    "security/forbidden-strings": ["error", { "builtinRules": true, "rulesFile": ["forbidden-strings.append.txt", { "path": "forbidden-strings.append.local.txt", "optional": true }] }]
  }
}

An unknown key, a wrong type or an unknown policy id is a configuration failure that exits 2 before any policy runs.

Policies

The wrapper ships these policies and no others. The first five run at their default severity without being listed. The last five run only where the file lists them. Listing one with a severity is the only way to turn it on, and a configuration that does not name a policy behaves exactly as it did before the policy shipped.

• require-root: default error.
• linked-worktree-only: default error.
• branch-worktree-only: default error.
• add-explicit: default error.
• final-newline: default warn. Text files end with exactly one line feed. A file whose Git text attribute is unset (-text, which binary includes) keeps its exact bytes, so a package declares its exact-byte inputs, such as fuzz seeds, with seed/** -text in its own .gitattributes. The attribute is read from the commit or index being checked, never from the worktree. Files below dist/final/node/ or bundle/node/ are also exempt, as generated output.
• markdown/autofix: default warn. Runs the linked monochromatic-lint fix engine over staged Markdown inside the commit, in-process on a guarded thread. Options: rules (only lfs-image-url exists) and exclude (gitignore-style patterns).
• mono/forbidden-root-context: default error. No CONTEXT.md at the repository top level enters a commit.
• mono/dependent-version-bump: default error. Packages that depend on a package with a bumped version follow the bump.
• mono/catalog-specifiers: default error, no options. Every dependency in a tracked package.json is catalog:, catalog:<name> or exactly workspace:* (peerDependencies take *, catalog: or catalog:<name>). The rule belongs to the monochromatic-catalog-specifier crate, whose README is its contract; the policy hands it each candidate package.json's bytes, never the worktree's, and reports its findings. A tracked package.json5, package.yaml or package.yml is a finding too, reported by name without reading it. The policy checks syntax only: it looks up no catalog entry and does not read pnpm-workspace.yaml. It runs when a commit is made, when git add stages, after a landed commit, before a push, and for check; it proposes no correction, so fix skips it. The finding codes are unsupported-manifest-format, invalid-manifest, malformed-section, duplicate-key, non-string-specifier and disallowed-specifier. A warn lets a literal version into the commit, so it earns the unsafe-warn notice. This repository's cli-git.config.jsonc does not list it yet, so the policy is off here.
• security/forbidden-strings: default error. Scans candidate bytes and path names with the linked scanner. Options: builtinRules (boolean, load the scanner's built-in baseline) and rulesFile (a rules file path relative to the repository top level, or an array of them). The rules files are rulesFile, else the environment variable FORBIDDEN_STRINGS_RULES, else forbidden-strings.local.txt in the repository top level, which may be absent. A file named by rulesFile or by the variable must exist, unless it is marked optional. Several files are read in order and scanned as one rule set, and every one of them is left out of the scan, as the single rules file always was.

The rulesFile option takes either form:

// One required file, the form that existed before lists.
{ "rulesFile": "forbidden-strings.append.txt" }
// Several files; a string is a required file, an object may mark one optional.
{ "rulesFile": [
  "forbidden-strings.append.txt",
  { "path": "forbidden-strings.append.local.txt", "optional": true }
] }

An optional file that is absent is skipped and recorded in the debug log. A configured string is one path exactly as written, so a comma or a ? in it needs no care; an array never needs the variable's list syntax. An empty array, an element that is neither a string nor an object with path, an object key other than path and optional, and a path named twice are configuration failures. Every path is checked the same way: relative to the repository top level, no .., no backslash. FORBIDDEN_STRINGS_RULES uses the scanner's comma-separated list (see package/forbidden-strings/cli/README.md, "Rules list"), where a trailing ? marks an optional entry; a value that is not valid UTF-8 is one required path. Because the variable's entries are resolved against the repository top level too, the same {{config_root}}-anchored value serves the scanner and cli-git. An invalid variable fails the policy only when a candidate is scanned, like a missing file. The new-worktree copy takes every configured file that Git ignores and the main checkout has.

Content already in history is not scanned again

security/forbidden-strings reads the bytes of a candidate only when no parent commit's tree already holds that exact blob. A pure rename, a copy of an existing file, and a merge that brings a file only one parent has therefore pass, because a blob's object name is the hash of its bytes and the bytes are already in history. What counts as a parent depends on the command:

• git add, and a commit being made: the commit the new commit is based on (the recorded base, else HEAD).
• The post-commit pass over a landed commit: every parent of that commit.
• The pre-push pass: every parent of a pushed commit that is not itself being pushed, so a parent that is part of the push vouches for nothing. A push that includes a tip with no known published history skips nothing.
• git cli-git check, a root commit and an unborn branch have no parent tree, so everything is scanned.

The comparison is by exact blob object name per path, never by path name or by similarity, so one changed byte makes the file new and it is scanned in full. The pathname of a skipped file is still scanned, because a rename gives the content a new name. A file that a fix corrected during the same invocation is never skipped, because its bytes are no longer its object's. If the parent trees cannot be listed, every candidate is scanned in full and a warning says why.

Each skipped file is recorded on standard error as a debug line when MONOCHROMATIC_VERBOSE=true: the masked pathname, the blob's object name, and the commit whose tree holds it.

Tradeoffs. Content that reached a parent through an earlier bypass, or that sits at a path the scanner never reads (the rules file and the scanner's own rule sources), stays unscanned wherever it is copied or moved to afterwards. That is also what lets the scanner's own package move to a new directory. A rule added to the scanner after a file entered history is applied to that file again only when its bytes change.

A warn setting reports findings without blocking. A policy whose protection warn would remove produces a configuration warning event when the file sets it to warn.

Discarded copies

• discardedCopies.maxAgeDays: positive integer, default 30. Saves of conflicted files older than this many days are removed when a new save is written.
• discardedCopies.maxSize: a number of bytes above zero or a string such as "256m" (k, m and g, powers of 1024, as Git's integer settings), default "1g". While the folder holds more than this, the oldest saves are removed.

Neither limit reaches a save younger than 24 hours. Zero, a negative or fractional number, and text that is not a size are configuration failures that name the key. See "Discarding commands" for the pass.

Concurrency

• hooks.concurrentCommits: boolean. Inert: it tuned the hook lock, and the wrapper runs no external hook, so there is no lock. Older files that set it still load; the value is checked and ignored, and git cli-git doctor says so.
• indexLock.unprovenOwnerTimeoutMs: non-negative integer, default 1000. Inert: it was how long the wrapper backed off for an index.lock whose owner is dead or unproven, and nothing reads it now (the commit service waits for a foreign lock until LOCK_CEILING, proven alive or not). Older files that set it still load; the value is checked and ignored, and git cli-git doctor names the key and asks for it to be removed.
• landing.reserveAfterLostRaces: positive integer, default 1. Inert: it was how many lost landing races passed before a commit asked for the landing reservation, and nothing reads it now (the commit service lands in arrival order). Older files that set it still load; the value is checked and ignored, and git cli-git doctor names the key and asks for it to be removed.

Concurrent commits cannot be turned off.

Waiting and time limits

Nothing the wrapper waits for waits for ever (owner, 2026-11-03: "nothing should ever wait forever or indefinitely"), and the wrapper never waits without saying so. Every bound is in src/native/service/wait_ceilings.rs with its measurement, and the list of every wait is doc/planning/cli-git-no-unbounded-waits.md. A process that works is waited for however long it takes: a limit ends a wait for quiet, never for slowness. Every Git step the wrapper starts is ended at a ceiling of its class (reads 120 seconds, bulk steps 10 minutes, reference updates 60 seconds) whatever it shows; a reference update ended after it renamed the branch reports the commit as landed, and one ended before it withdraws the commit as at the lock ceiling. A lock holder is waited for until LOCK_CEILING (10 minutes), the commit service's answer until ANSWER_CEILING (20 minutes), a start until START_CEILING (60 seconds), the service's hold for a forwarded git add until SERVICE_HOLD_CEILING (10 minutes), and a push whose connection is open and silent until CONNECTED_PUSH_QUIET_WINDOW (150 seconds, then the service's retry plan keeps it). At a bound on the commit path the request is withdrawn, the commit is kept on a side ref, and the command exits non-zero naming the blocker. A service that goes silent for five seconds is asked again. Where no inotify watch can be set, the waits poll every 50 ms. These were chosen by the owner's answers of 2026-11-03 and the coordinator's reading under the delegation, and are open to the owner's veto. The environment variables below set the limits and the pace of the notices, each a positive whole number of seconds; a zero, negative, fractional or non-numeric value is ignored.

• CLI_GIT_CHILD_WAIT_SECONDS, default 60 for a quick query step and 300 for a bulk step. How long one Git process that the wrapper starts for itself may show no sign of work before it is ended. A step that is slow and works (bytes moved, processor time, a task running or waiting for the disk) is never ended for being slow, however long it runs. The two defaults are not measured: they were wall-clock limits, no log held a quiet gap, and the timing records now keep the longest quiet gap of every step (longestQuietMs and quietMs in cli-git-commit-timing.jsonl) so a measured number can replace them. A quick query step only reads refs, commits and trees (rev-parse, rev-list, ls-tree, cat-file, diff-tree, merge-base, write-tree, update-ref and the like). A bulk step is every other step: it moves bytes (pack-objects, index-pack, add, maintenance run), or compares the index with the worktree (reset, update-index, ls-files, diff), or is not known to be a query. The object reader (git cat-file --batch) is a bulk step too, because one request can name a large blob. Setting the variable gives every class the same window. A step that is safe to run twice and shows no sign of work for the window is ended and run once more; if it is quiet again, the command reports which step, its class, the window and how to go on. A step that may not be run twice (a reference update, a commit Git builds) is announced and waited for as long as it lives. No hook of yours runs below a step the wrapper starts, and the stop never reaches a process beside the step.
• The tools and checks that are not Git end on the same rule, with windows that are not measured: a package-manager tool of the worktree repair after 600 seconds with no sign of work (with its whole process group), the Markdown linter after 60 seconds in which its thread shows none, and the standards-review call after 10 seconds of silence on its connection (30 seconds for the script's default). A tool that works, such as an install that downloads for a long time, is never reported as having run past a limit. The 20 second budget around the whole review is gone.
• CLI_GIT_WAIT_NOTICE_SECONDS, default 10. How often a waiting command repeats its notice (see "What a waiting command prints"), and how often the commit service repeats the line that names the holder of a foreign index.lock.
• CLI_GIT_PUSH_DEADLINE_SECONDS, default 30. How long the push after a landed commit may show no sign of work and hold no open connection (the stall window) before it is stopped. The 30 seconds has no measurement or owner decision behind it (the audit keeps it as an open question). The push and every process it started are watched: one that moves bytes (on its pipes and files, and on its connected TCP sockets), uses processor time, or has a task able to run or waiting for the disk (an fsync on a slow volume) is working and is not stopped here, however slow; one that does none of these is stopped after this many seconds. A run whose connection to the remote is open (the push, the fetch that follows a refused push, and the Git LFS upload) is waiting for the remote and is ended only when it shows no sign of work for CONNECTED_PUSH_QUIET_WINDOW (150 seconds, unmeasured); the service's retry plan keeps the push and tries it again. The owner's answer of 2026-11-03, "Bound it, keep retrying (Recommended)", replaced "Wait while connected (Recommended)". The wrapper says so once, in a line that starts with cli-git: git push is waiting for the server (git fetch or git-lfs push for those runs), and waits until the remote answers or the connection closes. "Connected" means that a process of the run holds a TCP socket in state ESTABLISHED, read from the kernel. A connection that dies without a reset is noticed only when TCP gives up; the wrapper sets no time for it. The wrapper adds no low-speed abort of its own to any run (it used to set GIT_HTTP_LOW_SPEED_LIMIT=1000 and GIT_HTTP_LOW_SPEED_TIME=60, which ended a silent check after 60 seconds): Git sets neither, and a limit and time you set in the environment or in http.lowSpeedLimit and http.lowSpeedTime still reach Git and still end the wait. A push to a local path has no connection to wait on and is stopped after this many seconds, as before. The name is kept from the time when this was a fixed deadline, so an existing setting now sets the stall window. Where the system cannot count bytes (no /proc/<pid>/io, as on macOS), it is a fixed deadline for the whole push, as before. The commit stays saved, the failure is reported, and git push can be run by hand. Git's own read-only queries before the push are ended by the same stall window when they show no sign of work; a query that works is never ended for being slow.
• CLI_GIT_PUSH_RETRY_SECONDS, default 10. The pause before the commit service tries a stopped or failed auto-push again on its own. A failure that can pass is tried again, with the pause doubled each time up to 600 seconds, until the push lands; no new commit is needed, and the service's journal says each time that it will try again (cause, number of the try, pause). Being offline, a stall, an HTTP 5xx answer of the LFS server and a held lock can pass. Three failures stop the retry until the person's next commit, because trying again would get the same answer (owner decision, 2026-11-03: "Offline is retried, but auth and server hook declines stop until your next commit."): a reported conflict, a login the service does not have (Git's "Authentication failed", "terminal prompts disabled", an HTTP 401 or 403, or SSH's "Permission denied ("), and a hook on the server that declined the push (a [remote rejected] line whose reason ends in "hook declined"). The failure record then starts with plain words that say which one it was, above Git's own words (for a decline, the message the server's hook printed). A rejection by the server whose reason is none of these is tried again, and the journal names its line. The failure record the next command shows says that the service keeps trying, and goes when the push lands. The plan survives a restart of the service: it is kept on disk, one small file per branch, in the service instance's own journal (push-retries), and the service that starts next rebuilds it and tries again with no wrapper command. The record the next command shows is not the plan, since showing it removes it. A pause that was running when the service stopped starts again from the restart, and the try number is kept.
• There is no ceiling on a push, and no variable sets one: no total time ends a push that shows progress or holds an open connection. The report of a stopped push names limit=stall, the bytes the push moved and the seconds it ran. A stopped Git process is first asked to end with SIGTERM, which lets Git remove its own .lock files, and is killed only if it is still running after five seconds. A process that was killed, or that could not answer in time, can leave a lock file behind; once it is gone the wrapper judges the reference locks of the repository it worked in and removes only those whose holder is proven gone (the rules are in the push deadline entry of SPEC.md), so a later git fetch or git commit does not fail with "Unable to create ... File exists". A lock whose pid file names a process of the stopped tree is removed on the wrapper's own knowledge that it stopped that process, also inside a container, where the general proof is not available. A lock it cannot prove stale, such as an empty lock without a pid file, is left, because a live Git of another program holds exactly such a lock for as long as its transaction lasts; the wrapper says what it found and prints the command that clears it. The same stop applies to a Git step that shows no sign of work for CLI_GIT_CHILD_WAIT_SECONDS.
• A lock that a Git killed in any other way left (a closed terminal, the out-of-memory killer, a reboot, a release before 0.5.23) is judged before the wrapper's own fetch and push and before your plain git fetch and git push. It is removed only when its holder is proven gone: the process its pid file names has ended, or its number now belongs to a program that is not Git, and no process of a nested PID namespace has that number as its own; and never while a process has it open or a running Git is named. The age of a lock by the clock is not a proof, because a correction of the clock moves it: a lock left by a reboot is removed when it has a pid file, and otherwise left with the command that clears it. Git writes no process number into a lock, and a live Git holds its reference locks with no file open, so nothing else can prove a lock stale. A lock that is not proven is left, and the wrapper prints what it found and the exact rm command that clears it (SPEC.md, "Reference locks left by a Git that was killed").

A command that passes through to Git unchanged, and the editor of a commit, run as long as they run and receive Ctrl-C like any Git command.

What a waiting command prints

Every wait of the wrapper writes to standard error, so a queued command can be told from a hung one. A wait that ends within one second prints nothing, because healthy commits queue behind each other briefly all the time. After that the first line comes, and it repeats every CLI_GIT_WAIT_NOTICE_SECONDS seconds until the wait ends. A wait whose holder is already known when it begins (a live index.lock, another process's push) prints its first line at once. Each line names what the command waits for, the process that holds it when the wrapper can tell, how long the wait has lasted and where it gives up.

• An index.lock held by a live process, which the commit service waits for before it lands a commit: the holder is proven by an open descriptor or Git's PID file and the proof is named, and an index.lock that nothing proves held is said to be unproven while the wait lasts. The first line of a wait adds a hint, and the repeats stay short.
• A git merge that waits for the index lock of any holder before it changes anything: it says it has changed nothing yet, names the lock and the holder with its proof, and repeats.
• The commit service's own hold on the index, which a forwarded git add waits for.
• Another process's push of the same branch.
• A long run of monochromatic-lint for a Markdown file, which names the file and stops waiting for the linter only after 60 seconds in which its thread shows no sign of work, leaving its thread to end with the command. The guard contains a hang and a panic, not a stack overflow or an abort in the linter.
• The object reader and the Git steps of a commit that run long, which name the Git command and the process, and end it only when it shows no sign of work for the window of its class (CLI_GIT_CHILD_WAIT_SECONDS). The first notice of a quick query step comes after 5 seconds and that of a bulk step after 10 seconds, so a step that only needed a moment prints nothing.

// standard error of a git add that waits for the commit service to finish a commit
cli-git: git add is waiting for the commit service to finish a commit in this repository, 12 seconds so far;
this waits for the service's hold for up to `SERVICE_HOLD_CEILING`, then runs and reports what Git says.

Reading the repository in process

The wrapper starts real Git as rarely as it can. For read-only questions it reads the repository itself and falls back to Git for anything it does not fully understand. The readers live in src/native/repo_read_*.rs; the design, the measurements and the limits are in doc/planning/cli-git-in-process-reads.md.

What is answered in process, with the standard output, standard error and exit code Git itself would give:

• the repository location every wrapped command asks for (rev-parse with the bare, Git directory, common directory, top level and prefix options), and the Git paths and object format the commit path asks for;
• ls-files --stage -z and ls-files --unmerged -z;
• diff-index --cached (name-only and raw, with --diff-filter and literal pathspecs), diff --cached --name-only -z, diff-tree --root -r -m for a commit with at most one parent, diff-tree -r --name-only -z between two commits or trees (with or without --no-renames) and diff-tree --no-commit-id --name-only -r -z for a commit with exactly one parent (the landing and the review ask for these; only the trees that differ are read), and show --no-patch --format=%s;
• config --list -z --show-scope, remote get-url, for-each-ref with the wrapper's format, rev-parse --verify --quiet <ref>^{commit}, rev-parse --abbrev-ref --symbolic-full-name @{upstream} and symbolic-ref --quiet HEAD;
• cat-file --batch fed full object names, in the commit preparation.

Every reader answers only after checking what Git would check, and otherwise leaves the question to Git: GIT_* variables that can move a path, global options other than -C, bare repositories, links, other file systems, reftable, conditional includes, unmerged or intent-to-add entries in a diff, a config.worktree that makes the repository bare, moves the worktree or holds a value the readers cannot prove Git accepts (with extensions.worktreeConfig on, the file is otherwise read and listed under the scope worktree), a file that became a directory, a gitlink, a possible rename, replace refs, shallow or grafted history, split or sparse indexes, and configuration keys that can change the output. Differential controls run real Git and the readers over the same command lines (src/native/repo_read_*_tests.rs, src/native/binary/binary_budget_tests.rs).

Set CLI_GIT_IN_PROCESS_READS=off in the wrapper's environment to start Git for every one of these questions, for a measurement that compares both paths or to rule the readers out when output differs.

Operating conditions

This package is one of the few that must keep working under the conditions its host lives in every day, not only on a quiet machine. The wrapper, the commit service and their tests are expected to work under these conditions:

• high CPU load from other sessions and builds;
• memory pressure and heavy swap use;
• slow SSD and other I/O trouble;
• a nearly full disk, so a write can fail with no space left in the middle of a commit;
• process and descriptor limits close to exhaustion;
• a process killed at any instruction, which the service's journal exists for;
• a closed terminal or a broken pipe;
• slow fsync and unusual filesystems;
• DNS or mirror trouble, where the wrapper recovers or falls back instead of wedging.

Latency requirements

Two requirements bound the latency the wrapper adds, and native:perf enforces both:

• A forwarded real Git command gains strictly less than 2,000 ms (SPEC.md).
• Every commit made through the wrapper must not take 5,000 ms or more longer than the same commit made with native Git (each commit, not a percentile of them).

No wrapped command is required to finish within a fixed time, and there is no 5-second bound on wrapped commands. The owner's words (2026-11-01):

" 5-second bound on wrapped commands" - there is never a 5s bound on wrapped commands. The measure is always just commits vs native git should not take 5s longer.

The 5-second timeout --signal=KILL of the binary tests is a hang guard of the test harness (TIME_BOUND_SECONDS in src/native/binary/binary_support.rs), short so that a mutant that loops fails every affected control quickly. It says nothing about how long a correct wrapped command may take.

What pins each condition

Each condition has permanent controls that start the built executable on real Git fixtures. The wrapper's side of a commit no longer writes state that outlives the command, so the call-by-call fault sweeps of the earlier landing (strace --inject failing, killing or delaying every write, rename, flush, descriptor and process start of a commit, and the records they left to recover) retired with that landing. The commit service's own faults are pinned by the kill points of its journal (binary_service_tests.rs: a kill at each point of a batch, requests sent again, late and older services), the pruned-content controls (binary_service_pruned_tests.rs) and the unit controls of its landing (service_land_tests.rs); the call-by-call sweeps are ported to the service below. What the faults found, and the change for each, is in the fault injection note.

• A closed terminal or a broken pipe: binary_fault_stream_tests.rs closes the standard streams, gives them a pipe with no reader or a full device, and requires that the status matches the one Git gives when started directly.
• A Git step that stalls: binary_child_bound_tests.rs delays the return of every start of a Git step past its class bound (strace --inject, a ptrace of the child, no privilege and no load) and requires that the step is stopped, named with its class, and that the repository recovers. The control reports a skip and checks nothing on a host without strace.
• A process killed at any instruction elsewhere: the worktree-copy and merge flows are killed at their phase markers (binary_merge_isolation_tests.rs, binary_worktree_copy_tests.rs) and, in binary_fault_trace_tests.rs, at a spread of the Nth write, fsync, rename, unlink, directory and process-start calls of each flow (strace --inject signal=SIGKILL, positions counted from an uninjected run, first and last included), followed by the next commands and git fsck. The commit service is killed the same way at any call of itself and of the Git it starts. A staging directory of a killed process stays on purpose and is ignored by every reader.
• A signal to the wrapper alone: binary_fault_signal_tests.rs signals the wrapper's own process ID (no timeout in front, which would forward the signal to its whole group) and checks that a stand-in Git received exactly that signal through the wrapper's relay, that the wrapper ends with 128 + signal, that nothing runs on and that the repository recovers; a hangup of the whole group is a separate control. The SIGQUIT core dumps of Git seen on the development host came from the earlier controls, which signalled timeout and so reached Git and the editor directly.
• DNS or mirror trouble: binary_fault_remote_tests.rs makes the push after a landed commit and an explicit push meet a refused connection, a name under .invalid, a server that hangs up, a server that never answers (stopped at CLI_GIT_PUSH_DEADLINE_SECONDS) and a git-lfs that never exits, and requires that the commit stays, the failure is worded, no lock is held and no process the push started is left running; binary_push_environment_tests.rs, binary_lfs_pre_push_tests.rs and binary_lfs_server_tests.rs cover the stalled push and the failing upload. The push is the service's work after the command returned, so each control waits for the work records to settle and reads the failure from the notice the next command shows.
• Process and descriptor limits, memory pressure and a starved processor: binary_fault_limits_tests.rs sets RLIMIT_NOFILE, RLIMIT_NPROC and RLIMIT_AS (setrlimit between fork and exec, no privilege) on the wrapper and on the commit service, one side at a time, sweeps the descriptor and address-space limits from the smallest at which the program starts to the smallest at which a commit lands (both found by running, so a new build moves them), lowers and raises the service's RLIMIT_NPROC in place with prlimit, and runs a commit with the wrapper, the service and Git on one processor. After each, the pressure ends, the next ordinary command succeeds, a further commit lands and nothing is left. A failed allocation ends a Rust program with an abort; the controls accept that end and require the recovery.
• Slow storage, slow fsync, a full or failing disk: binary_fault_trace_tests.rs runs the service under strace --inject with every flush returning late (100 ms), with ENOSPC and with EIO from the Nth storage call on (write, fsync, rename), and requires the recovery after each. The wrapper's own fsync and writes are not swept because a commit no longer writes state that outlives the command.
• High CPU load, memory pressure and slow I/O in the test harness: the 5-second hang guard, which loaded hosts have tripped, and the byte copy of the wrapper that keeps a cold start inside it (the troubleshooting note below).

What no control injects, and why:

• Swap and real memory pressure: the kernel's reclaim needs a cgroup the test does not own; RLIMIT_AS stands in for the allocation failures, and the gate container's --memory cap bounds the whole run.
• A real full filesystem, real slow storage, and unusual filesystems (FUSE, NFS, a loop mount): mounting needs a privilege the gate does not give. strace --inject stands in for the errors and delays, and binary_subvolume_tests.rs and fs_kind_tests.rs cover btrfs and the filesystem classification where the host offers them. The gate runs on an overlay filesystem.
• High CPU load: a synthetic load is banned. The container gate's --cpus cap and the single-processor control stand in.
• RLIMIT_NPROC for a user who may exceed it (a privileged run) is exempt from the limit; the two process-limit controls say so and check nothing there.

Known gaps: an editor that never exits is not judged by CLI_GIT_CHILD_WAIT_SECONDS (it runs as long as it runs, as documented, and is ended by the signals above); the removal of a name (unlink) is not swept with failing errors, because a lock file left by a failed removal needs a person with native Git as with the wrapper; a step stopped at its bound ends with the processes it started, found in /proc, so only Linux ends them.

Tests therefore never wait for a quiet host. The 5-second hang guard on one wrapped command in the binary tests is deliberate, and is not a requirement on the wrapper; its measurements and the evidence needed to change it are in the troubleshooting note.

State on disk

The wrapper keeps its state beside the repository's Git data under cli-git* paths in the common Git directory. Delete nothing there while a Git command is running. A commit keeps a scratch directory cli-git-service-<id> in the worktree's Git directory while it is made; the commit service removes it when the command ends, however it ends. The service's after-commit work records and the notices of a failed backup push are kept under cli-git/. The timing record of each wrapped commit is appended to cli-git-commit-timing.jsonl (see "Splitting a slow commit"). A directory cli-git-transactions in the worktree's Git directory belongs to an earlier release: this release no longer uses or reads it. A command that finds commit records in it prints a note on standard error, names the directory, gives the exact rm --recursive --force -- '<directory>' command that removes it (with the path quoted for the shell), and goes on. Look at git log -1 and git status to see whether that commit was made, then run the command.

A merge in progress keeps a one-line record in cli-git-merges in the worktree's Git directory, from before Git starts updating the checkout until the branch has moved. Do not remove it by hand unless a command says so.

The ignored-state copy into a new linked worktree stages its files in a .cli-git-worktree-copy-* directory beside the new worktree, which can hold gigabytes. The wrapper records that directory's name in the common Git directory before it creates it, so a copy that was killed while staging cannot leak it: the next invocation in the repository removes the directory and says so on stderr. Delete a .cli-git-worktree-copy-* directory by hand only when no git command is running in that repository.

A worktree added from the main checkout receives the main checkout's ignored state minus what a new worktree does not need. Left out by directory name, wherever the name sits below the main checkout: a directory that holds a .git file or directory (another checkout, including every registered worktree); the caches .pnpm-store, .gradle and .kotlin; a .cache directory other than the one in the repository root (the root .cache holds the tool launchers and generated inputs that tasks read, so it is copied); incremental and zig-global-cache below a target directory; and the records of earlier runs, .monochromatic below a node_modules directory, target/verification, mutants.out, mutants.out.old, playwright-report, test-results and .serena/cache. The run records follow the constant COPY_RUN_RECORDS in src/native/worktree_copying/worktree_copy_prune.rs, which is false. A linked or bare source copies everything it ignores except its nested registered worktrees. An ignored file or directory that the user cannot read, such as a log written by root, is left out of any copy instead of failing it, and one stderr line says how many and which first paths: cli-git: left out 2 unreadable ignored entries from "/path/to/source": "a/b", "c". The copy clones files with a copy-on-write request, in parallel on a large tree, and trusts the clone: it proves a cloned source file unchanged by its size, modification time, change time and inode instead of reading both files again. A source file written to during the copy still fails the copy, and a file the filesystem would not clone is compared byte for byte.

Development

The commands below run in the repository that holds the crate, Aquaticat/Monochromatic, from the repository top level. Every task runs through mise.

# Run from the repository top level.
mise run //package/git-policy/cli:native:check
mise run //package/git-policy/cli:native:clippy
mise run //package/git-policy/cli:native:build
mise run //package/git-policy/cli:native:test:host -- <test name filter>
mise run //package/git-policy/cli:native:test:binary:host -- <test name filter>
mise run //package/git-policy/cli:native:test:container

• native:check type-checks every target. native:clippy lints every target with warnings denied. native:build writes the debug executable to package/git-policy/cli/target/debug/cli-git-native.
• Every process a unit test starts dies with the test thread, however the test ends: start helpers with spawn_owned in src/native/testing/test_support.rs, which gives the helper a process group of its own (so what it starts dies with it) and kills the group when the thread's fixtures are removed, the moment a failing or panicking test unwinds (a_loop_started_by_a_panicking_test_is_gone_when_the_test_ends). A shell loop that a failing push-progress test never ended ran for 87 minutes on 2026-11-03 before this.
• native:test:host runs the unit tests and native:test:binary:host runs the tests that start the built executable, both against the host's Git while iterating.
• native:test:container is the authority. It builds the crate and runs the tests and Clippy in bounded containers with Git 2.56.0 and no network. Set GIT_POLICY_NATIVE_IMAGE_TAG=<name> to keep the images of parallel runs apart. The test run uses cargo nextest run --no-fail-fast, so every test runs even after a failure and one flaky test never hides the rest; the exit status is still non-zero, and nextest prints its summary and the list of failed tests last. The image build is layered so that a rebuild reuses everything the change did not reach, with no network and --locked throughout (bin/native-layered-build.ts, bin/native-test-containerfile.ts):
  • The dependency layer compiles every registry crate from the manifests, the lock files, the vendored sources and empty stand-in source files, so its cache key holds no source file. Only a change to a manifest or a lock file rebuilds it, which includes a version bump of the crate or of a path crate.
  • The path-crate layer copies the JSONC crate and the scanner with its engine and compiles them (the scanner's build script precompiles the embedded baseline), so it is rebuilt only when one of them changes.
  • The subject layer copies the wrapper's src/native and README.md (the tests include it) and compiles only the wrapper, so a test sidecar edit or a source edit recompiles one crate.
  • The linter stage follows the same split: its registry dependencies compile from stand-ins, then its own sources.
  • Each real-source layer touches the copied files in the same step, because Cargo compares file times with the stand-in build and would otherwise treat the stand-in output as current.
  • Build cache mounts (RUN --mount=type=cache) are not used: they keep target/ outside the image, but the test run, the planted controls and the mutation campaigns reuse the image's own target/, so the compiled output has to stay in the image.
  • The image build keeps the same CPU limit as the run. In paired builds at the same time, a cold dependency layer compiled about 1.7 to 1.9 times faster with --cpus 4 than with the default (47 s and 58 s against 23 s and 31 s) and the path-crate layer about 1.6 times faster, while the wrapper-only layer that a test or source edit rebuilds did not change (about 12 s either way). A cold build happens only after a manifest or lock file change, so the default is unchanged; pass --cpus 4 after the -- for that case (the test concurrency follows it too).
  • Every build step costs a fixed amount on a busy host (measured as tens of seconds, up to minutes, for a step that does no work), so the Containerfile keeps its step count low and puts the steps that rarely change first.
• native:perf builds the release executable and measures the latency it adds to forwarded Git commands and to commits (explicit-path and service, unsigned and SSH-signed), with paired direct-Git samples in disposable repositories. It reports the median, p95 and median absolute deviation of both sides and the added delta, and for each commit the median, p95 and maximum of the paired deltas (the wrapped time minus the direct time of the same round). It fails when a forwarded command's added p95 reaches the 2,000 ms bound, when any commit's paired delta reaches the 5,000 ms bound (the owner's measure, see "Latency requirements"), or when any sample fails. -- --output <file> stores the raw samples (baselines live in perf/; schemaVersion 2 adds the paired deltas and names each scenario's kind). The judging lives in bin/native-perf-judge.ts, and bin/native-perf-judge.unit.test.ts runs it on the recorded baseline, edited to break each bound. Options for measuring a commit at the size the person commits at (doc/planning/cli-git-commit-latency-under-disk-pressure.md): --shape <repository> makes every fixture a reflink copy of a prepared repository without remotes (a clone of the main checkout, for example), refreshes its index and gives it the tracked file README.md to edit; --hooks-from <git directory> copies that repository's policy hooks, --tool <path> puts an executable such as git-lfs on the fixtures' PATH, and --only <scenario,...> measures the named scenarios. --cold drops the pack, index and packed-refs pages of the fixtures from the page cache before every sample on both sides, and --trace-directory <dir> leaves an strace record of each sample and of each commit service for counting syncs, bytes and process starts. --pause <ms> waits after each sample, so the work a commit leaves running after its answer is over before the next one. With --output, the report also holds the wrapped commits' timing records and the service's answered lines (evidence), which split a slow sample into its phases.
• native:lock:members records a changed dependency list or version in Cargo.lock without moving any other locked version.

A release build of the executable:

# Run in package/git-policy/cli.
cargo build --release --locked --bin cli-git-native

The release profile keeps panic = "unwind" and overflow-checks = true. The linked scanner turns a matcher panic into a failing finding by catching the unwind, and the crate refuses to compile under panic = "abort".

Git's own fsmonitor

Git's file-system monitor is off, by the owner's decision of 2026-11-03:

• "If git's fsmonitor is fundamentally broken and cannot be easily or meaningfully worked around, disable git's fsmonitor globally entirely and make doctor say so."

Why: Git 2.56.0's Linux daemon registers an inotify watch for every directory before it starts answering (compat/fsmonitor/fsm-listen-linux.c, then ipc_server_start_async), and a Git process that finds no listening daemon starts one and waits for it for up to fsmonitor.startTimeout (60 s by default; builtin/fsmonitor--daemon.c). On the owner's main checkout, where linked worktrees live inside the checkout, one daemon watched 2,269,377 directories and spent 117.7 s of CPU and 734 MB to start, and 261 daemons held 4,591,335 watches in all. A commit of 7.7 s spent 6.5 s with the wrapper's Git child waiting for git fsmonitor--daemon start, while a running daemon saved about 7 ms per git status (16 ms off, 9 ms on, five alternating runs each). The measurements and the cause are in doc/troubleshooting/git-fsmonitor-daemon-start-wait.md.

What the wrapper does now:

• It still never passes -c core.fsmonitor=... to any Git child, never writes core.fsmonitor in any repository, and never starts, stops or kills git fsmonitor--daemon. The setting is the person's configuration, and it is off there: the person's global ~/.gitconfig holds core.fsmonitor = false.
• git cli-git doctor says so on standard error. When core.fsmonitor is false or unset it prints one line that Git's fsmonitor is off by the owner's decision of 2026-11-03, with the file and scope of the entry that decided (or "not set"), the reason in one sentence, and the path of the document above. When the last entry of core.fsmonitor in any scope turns it on (true, a true word, a number other than zero, a key without a value, or a hook path) it prints a finding that names the file and the scope and the command that turns it off for that scope, such as git config --global core.fsmonitor false. A value from -c or the environment has no file, so the finding says to remove it there. Doctor reports and never writes configuration, and its exit status does not depend on this line. Doctor reads one more configuration listing with --show-origin for this, because the snapshot the commit path reads names scopes but not files.

The owner's earlier words, kept as history; each was superseded on 2026-11-03 by the decision above:

• ""To stop the fsmonitor program" - we don't need to."
• (2026-11-02) "Why are we messing with git's fsmonitor? I already told you it's fine for git to run it."
• (2026-11-03) "Why are some of your agents still messing with fsmonitor? I've said it at least twice that we shouldn't mess with git's own fsmonitor s."
• (2026-11-03) ".git/config overriding it to false is wrong."

Before that decision the person's ~/.gitconfig set core.fsmonitor = true. A main checkout's .git/config had held core.fsmonitor=false until 2026-11-03, written by no code in this repository; it was removed at the owner's word. Tests that reproduce core.fsmonitor = true inside their own temporary repositories stay: they check that the wrapper behaves as Git does under that setting (repo_read_config_typed_tests.rs, reader_fsmonitor_tests.rs, binary_budget_tests.rs), and they never touch a real repository's monitor. The controls of the doctor line are in doctor_fsmonitor_tests.rs, with an injected global file. SPEC.md (search "core.fsmonitor") records the decision for each code path.

No --quiet on the wrapper's own Git runs

No Git run the wrapper or the service starts carries --quiet or -q where the option only hides messages or progress. The owner's words:

• (2026-11-03) "No the wrapper should never append --quiet. Letting native git emit whatever it wants to is expected."
• (2026-11-03) "Drop it everywhere (Recommended)", about the wrapper's own internal runs: maintenance, the auto-push fetch, fast-forward merges, reset, clean and init.

So no other output-hiding option or redirect stands in for it either. A --quiet stays on a run whose documented effect is the exit status or what is computed (rev-parse --verify --quiet, symbolic-ref --quiet, diff-index --quiet, diff-files --quiet; read in the Git 2.56.0 sources, listed in doc/planning/cli-git-slow-server-and-inner-quiet.md). The words of a run that said something are kept in the journal (src/native/worktree/inner_git_words.rs): the service writes one info entry in its log, and the wrapper writes one client entry (journalctl --identifier=cli-git), both with the field CLI_GIT_INNER_COMMAND. A test that starts such a run records its arguments with recording_git.rs and checks quiet_starts. git reset without -q lists the tracked paths whose files differ from the index, so the service's journal now names the paths of other uncommitted work for every landing.

Container base images

Every container image of the verification tasks starts from the newest CachyOS image the host CPU can run, on its latest tag (doc/decision/container-base-distro.md at the repository top level). The tag is never pinned to a version or a digest, so an upstream CachyOS update that breaks a task is a bug to fix at once. Two derived base images carry what the network-free image builds need:

• native-test-base.Containerfile builds localhost/git-policy-native-test-base: CachyOS, the distribution's Rust (Cargo and Clippy) and build tools, and the audited Git 2.56.0 built from its checksum-pinned kernel.org tarball at /usr/bin/git, where the tests expect the real Git. The gate image and the mutation images build on it.
• e2e/concurrent-commits.Containerfile builds localhost/cli-git-concurrent-e2e: CachyOS, Node 24.11.0, and Git 2.55.0.

Each run of native:test:container, test:git-image, test:e2e:concurrent:image and test:e2e:concurrent:native:image first pulls the latest tag with podman pull --policy=newer (one small request when nothing changed; a registry that cannot be reached is reported and the copy the host holds is used). The derived base is tagged <first digits of the pulled image ID>-<first digits of the Containerfile hash>, so a new CachyOS image or an edited Containerfile builds a new base (with network, once) and every run after that finds it. The gate image then builds from the base's ID with no network, exactly as before, and its layer cache is keyed on that ID. bin/native-test-base.ts holds the pull, the tag and the build. The resolver that picks cachyos-v4, cachyos-v3 or cachyos is package/test-harness/core-matrix; a host that is not x86-64 stops with an error that names the image, because these images have no Alpine variant.

Container limits

Every container the verification tasks start is bounded. The defaults are 2 GiB of memory, 2 CPUs and 512 processes, defined once in bin/container-limits.ts. Each task accepts overrides after the -- of mise run, with these exact option names:

• --memory <size> is a whole number with an optional unit b, k, m or g, such as 512m or 2g.
• --cpus <count> is a number from 0.01 with at most two decimals, so 1, 1.5 and 0.25 are valid. The build CPU quota is derived as cpus * 100000 microseconds per 100000 microsecond period.
• --pids-limit <count> is a whole number of processes from 1. It is accepted only where the container has a process cap, and rejected with a message otherwise.

Every container also starts at the lowest disk weight, --blkio-weight=10 and --cgroup-conf=io.weight=10 (lowestDiskWeight and diskWeightArguments in bin/container-limits.ts); no option raises it. Podman writes the first to io.bfq.weight, which the BFQ I/O scheduler reads, and the second to io.weight, which the iocost controller reads, and each leaves the other file at the default weight, so passing both puts whichever file the host's disk scheduler reads at the lowest value. podman build accepts neither option, so the containers of image builds keep the default weight.

A value is a separate token (--memory 1g); the --memory=1g form is not read and reaches the task's own option parser, which rejects it as an unknown argument. An invalid value stops the task before any container starts. The message names the option and the accepted forms. Options repeat with the last one winning, and the number of build jobs (CARGO_BUILD_JOBS, make --jobs) follows the whole CPUs of the limit, never below one.

# Run from the repository top level.
mise run //package/git-policy/cli:native:test:container -- --memory 4g --cpus 4
mise run //package/git-policy/cli:native:mutation:scoped -- --memory 1g --cpus 1 --file src/native/invocation/startup_gate.rs
mise run //package/git-policy/cli:test:e2e:concurrent:native --memory 1g --cpus 1
mise run //package/git-policy/cli:test:git-image -- --memory 4g

• native:test:container accepts all three options, and so do native:mutation:scoped and test:git-image.
• native:mutation runs the gate and then the mutation step, and mise passes the options to the mutation step only, so the gate runs at the defaults. Run native:test:container with its own options first when both need changing.
• test:e2e:concurrent:native takes --memory and --cpus as task flags and applies them to the run. The image build it depends on uses the defaults, and test:e2e:concurrent:native:image and test:e2e:concurrent:image take the same options after the --. The e2e containers have no process cap either.
• The mutation manifest records the limits it ran with, and the gate and image-build manifests record the container argument arrays.
• native:mutation:scoped reuses the image of the last gate, whose CARGO_BUILD_JOBS was fixed when it was built. Run the gate with the same --cpus for the job count to match.
• The package/git-policy/cli.fuzz campaign uses the defaults from the same module and has no override options yet.

Who runs a mutation campaign

The session's main agent starts every campaign of this package and waits on it, never a subagent; the owner's reason is in the launcher's Who runs a campaign.

Resuming a mutation campaign

native:mutation:scoped and native:mutation:parallel (and so native:mutation) are resumable. A kill, a reboot, an out-of-memory stop or systemctl --user stop <unit> loses no scored mutant: run the same task with the same options on the same gate image and it continues where the campaign stopped, scoring only mutants that no earlier run scored, whether they were caught, missed, timed out or unviable. For native:mutation:parallel, start the same launcher command again; each file resumes its own campaign.

The campaign's results are in target/verification/native-mutation-<key>/, where the key names the gate image, the exact options, the digest of the planted controls and the identity of the code that runs and judges them, so a changed control table or runner rescores. While a campaign runs, attempt-NNN/mutants.out/ there holds the outcomes scored so far in caught.txt, missed.txt, timeout.txt and unviable.txt. When it ends, mutants.out/ holds the same four files as the union of every attempt, with outcomes.json, log/ and diff/, and exit.json records the status of the whole campaign (attempt and skippedMutants say how many earlier mutants the last run left out). To score everything again, delete that directory; a rebuilt gate image gets a new one by itself. The mechanism, the choice of cargo-mutants --iterate, and the limits are in package/mutation-test/cli-cargo-mutants/README.md.

What a mutation campaign costs per mutant

Every mutant runs the package's tests, so the cost of a campaign is mutants times the time the tests need (doc/planning/mutation-campaign-per-mutant-cost.md has the measurements). Two measures cut it; both keep every kill and every result.

The baseline is skipped when the gate already passed on the same image. native:test:container writes manifest.json (the full ID of the image it tested) when it starts and passed.json (tests, clippy and the same image) only after the test run and Clippy both exited zero, in target/verification/native-<random>/. native:mutation:scoped resolves the ID of the image it layers cargo-mutants over and runs the unmutated baseline only when no such directory names that ID with a passed record; the log says which it did (Baseline: skipped, because <directory> proves ... or Baseline: run, because no gate evidence ...) and the campaign manifest records the directory as baselineProof. A rebuilt snapshot has another image ID, so it has no proof until its gate has passed. The proof is the gate's pass on that image, and only for a campaign whose --memory and --pids-limit are at least the ones in the gate manifest's limits: with fewer processes than the gate had, the fork failures of 2026-11-01 would come back as caught mutants, so the log names the gate directory and the limits and the baseline runs. A manifest that records no readable limits proves nothing.

The planted controls (a full suite run per guard removal, about 24 minutes per campaign) are skipped by the same idea. A campaign whose controls were all noticed leaves planted-controls-passed.json in its evidence directory (in its pass-1 directory for a two-pass campaign), written to a temporary file, flushed and renamed so a reader never sees half of it. The record holds plantedControls: true, the full image ID, limits (the container arguments the controls ran with), the control names in controls, and controlsSha256, a digest of every control's file, replaced text, new text and named test. native:mutation:scoped looks through target/verification/native-mutation-* before it plants anything, and skips the controls when a record names the same image ID, the same controls, and no more --memory or --pids-limit than this campaign has. The log says Planted controls skipped, because <record> proves ... or Planted controls: run, because no campaign evidence proves .... A rebuilt snapshot has another image ID, and a control added, removed or edited has another digest, so either one runs the controls again. A record made at larger limits is named in the log and not used. A control that was not noticed, or one whose planted source did not compile, leaves no record. --cpus is recorded but not compared, because fewer CPUs only slow the suite. The record also holds harnessId, the plantedHarnessId literal in bin/native-planted-controls-harness.ts: the digest of the sources that run and judge the controls (bin/native-planted-controls.ts, bin/native-planted-controls-outcome.ts and bin/native-verification-process.ts). It is part of the loaded code, not hashed from disk afterwards, so an edit between loading and writing cannot publish a record for code that never ran, and it is content-derived, so two branches that change the judge differently cannot share an identity the way a counter would. A unit test asserts that the literal equals the digest of the sources and prints the digest to write, so editing any of them without updating the literal fails the test, and a record of another identity is not proof. Limits are not parsed as Podman parses a command line. The runners write only --rm, --init, --network=none, the two disk-weight arguments (a record made before they existed lacks them and is still read), --cpus= and one --memory= and one --pids-limit=, so a recorded argument list with anything else (a repeated bound, a split or attached short option, a cluster, an option whose value looks like a bound, a trailing bare option) was not written by a runner and gives no proof, whatever it would mean to Podman. A zero or unlimited value (--pids-limit=-1) is no bound a campaign can be compared with either. The same reading applies to the gate's passed.json proof. Campaigns take no lock for this. Two campaigns started together over one image, with no record yet, both run the controls and each writes its own record; neither waits for the other, because waiting would serialize work that is independent. The cost is that both pay the 24 minutes, so start one first and the rest after its record exists. NATIVE_MUTATION_SKIP_PLANTED=1 still skips without any record (a person's claim that another campaign ran them). A resumed campaign is held to the same rule as a new one, and a changed control table or a changed runner or judge rescores: the digest of the controls and the runner and judge identity are part of the campaign key, so a campaign run after either changed is another directory that keeps no scored results.

native:mutation:parallel runs the controls beside the file units, never before them (bin/native-mutation-parallel.ts, bin/native-planted-parallel.ts). Before it starts anything it resolves the image ID of every unit's tag, and looks for a proving record with the reader above, at the memory and process limit the units will run with. With a record, no extra unit starts, and each file unit logs the record it found. Without one, it starts one more capped unit per distinct image, named <unit prefix>-<label>-planted-<first 12 of the ID>, through the same slot limiter. That unit runs only native:mutation:planted, and writes the record into a native-mutation-planted-* evidence directory. The file units start at the same moment and do not wait. Each skips the controls, because NATIVE_MUTATION_PLANTED_UNITS names the unit for its image, and logs Planted controls skipped here, without proof, with the unit and the image. That skip is not a pass: the results of those campaigns hold only once the controls unit ends with a planted-controls-passed.json for the image. The launcher prints the controls unit, its log and that rule, and puts the unit first in the systemctl wait hint. The controls unit's log ends with Planted controls passed; recorded in <record>, or with Planted controls FAILED over image <ID>, followed by the controls that were not noticed, and the unit then shows as failed. If systemd refuses the controls unit, or a tag does not resolve to an image, the launcher says so, and the file units of that image plant the controls themselves, as native:mutation:scoped does alone. --setenv NATIVE_MUTATION_SKIP_PLANTED=1 after the -- still skips without proof, and no controls unit starts. The controls unit runs the full image ID the launcher inspected (--image), never the tag, which may be moved while the unit waits for a slot. The launcher inspects with the environment a unit gets (the tag, the --setenv entries, and its own TMPDIR, PATH and HOME last) and nothing inherited from its shell but XDG_RUNTIME_DIR, so it looks in the container storage the units will use, and it refuses a --setenv of the image tag variable or of NATIVE_MUTATION_PLANTED_UNITS, which it sets itself, because an override would run another image than the one inspected. The controls unit gets the same extra environment as the file units, with the launcher's own TMPDIR, PATH and HOME last. It succeeds only when it published a matching record or found one, so a record that could not be written fails the unit. A campaign that deferred keeps its obligations in its manifest as the list plantedControlsDeferred. Each holds the image, the limits the campaign ran with, the digest and names of the controls, the runner and judge identity, and the controls unit with its systemd invocation ID and its log. The launcher names the invocation it started in the coordination entry, so a campaign defers only to that invocation and never to a later unit that took the name. Whether an obligation is met is asked on demand and never kept: a matching record exists, found with the same reader and by the saved keys, not by the asking process's own table or code, or it does not. Finalization and native:mutation:reconcile run that lookup and print the record path, or "no record yet" with the unit, its invocation and its log. With no record, an obligation is pending while that very invocation runs, and unvalidated once it ended, after one more look for a record published in between. Unvalidated fails the campaign's finalization, so a campaign whose controls failed never passes. So does a manifest that is unreadable, holds an incomplete obligation, is missing after the campaign finished (exit.json without manifest.json), or no longer holds an obligation the running process itself decided to keep. A campaign's key includes the controls digest, the runner and judge identity, and the memory and process limit (the memory normalized to bytes), and campaign.json records those key parts in full so that a clash of the shortened key is refused, so a resumed campaign has the same controls, runner and limits as the attempt that deferred, and a proof or a local run in a resume pays that obligation by construction. An attempt that proves or runs the controls writes a manifest without the field, and a person's skip does the same. The two passes of a two-pass campaign share the key and each directory is validated by its own manifest. A pass that runs the controls itself while the first pass is still waiting for a record must publish its own, and fails the campaign with the reason when it cannot, because a debt in another directory is paid only by a record the lookup finds. A directory in which a container ran (an attempt-* holds mutants.out) and that has no manifest is invalid, finished or interrupted; a controls-only directory, an attempt that never started, and the aggregate root of a two-pass campaign whose pass 2 never ran have none and are exempt. The controls unit's name carries a token unique to the launch, so no other launch's unit can be taken for it. A file unit that starts after the controls unit is gone and left no record runs the controls itself. mise run //package/git-policy/cli:native:mutation:reconcile looks up the obligation of every campaign under target/verification later, and exits nonzero unless all are validated.

--passes 2 runs each file's campaign in two passes. Pass 1 runs only the tests that can reach the mutated file: the tests of its module and of every module that imports it, directly or through other modules (the closure of importers over the crate's use paths under src/native), plus the test targets that refer to one of those modules, and never the binary suite (native_binary). Pass 2 runs the whole suite on the mutants pass 1 missed or timed out, and on those only. The report in mutants.out and the exit status read as one campaign: a mutant pass 2 scored has pass 2's outcome.

# Run from the repository top level; one file, two passes, measured on its own before a queue is started.
mise run //package/git-policy/cli:native:mutation:parallel -- --passes 2 --timeout 900 m1 repo_read_query.rs

--passes 2 refuses a glob --file, no --file and a file the library does not declare, instead of quietly running one pass. It also refuses to start without the gate proof above: pass 1 runs a filtered baseline at best, so run native:test:container first, at limits no larger than the campaign's. A campaign directory started with --passes 2 refuses a run without it, and the other way round; move the directory away to switch. The selection is written to the manifest (passOne) and to the log. How the two passes are seeded, composed and resumed is in package/mutation-test/cli-cargo-mutants/README.md under "Two passes".

Source layout

The crate is one library and two executables, and its sources live under src/native/. Each folder there is one subsystem, so a contributor opens a folder for a concern instead of searching a flat list of prefixed file names. The small-file rule stays: a folder holds many small modules, each with its sidecar test file.

How a module is reached

• lib.rs is the map. It declares each folder (pub mod command;) and re-exports the folder's modules with a glob (pub use command::*;).
• Each folder's mod.rs declares the modules inside it with one documented line each.
• The glob re-export keeps every module at its old path, so crate::command_options and git_policy_cli::command_options still resolve. The fuzz crate, the two executables and the integration tests name modules that way and needed no change.
• A glob re-export loses to an item written in lib.rs, so no folder may share its name with a module.
• use super:: in a module that a folder declares directly names the folder, not the crate root. Write use crate::module_name to reach a module of another folder.
• Library test names now start with the folder: command::command_options::tests::name. The integration test roots (native_binary, native_candidates) keep their names.

Subsystems

• invocation/: what one invocation does from arguments to Git: the decision, real Git resolution, forwarding, startup, management words and the wrapper's own Git version check.
• config/: reading and validating cli-git.config.jsonc and the Git configuration a command is judged by.
• command/: the option grammar of each Git command the wrapper understands and the rules that rewrite, refuse or hint.
• policy/: the policy engine: registry, triggers, events, passes over candidates and the direct fix.
• markdown/, dependent_version/, standards_review/: three policies large enough for a folder each.
• candidate/: what a commit or push would add, read from the index or objects, and the scanner run over it. Its candidate_consumer_tests.rs is the native_candidates test root.
• commit/, shadow/, transaction/: the wrapper's commit path, the shadow repository it prepares on and replays onto, and the durable transaction records with their recovery.
• landing/ and service/: what the resident commit service does to land a commit, and the service program around it: protocol, server, client, journal, provisioning and the unit files in service/service_units/.
• push/, lfs/, merge/, conflict/: pushes and their waits, Git LFS, merges with the corpus in merge/word_merge_corpus/, and conflict state around commits and checkouts.
• worktree/, worktree_copying/, subvolume/, repair/, health/: worktree commands, the ignored-state copy, btrfs subvolumes and the filesystem primitives under them, package-manager repair, and the doctor report.
• repo_read/, index/, lock/, process/, foundation/: reads without a Git process, the index file, locks and claims, facts about processes, and shared building blocks that depend on no subsystem.
• testing/: test-only fixtures shared by many unit tests, never part of the release executable.
• source_guard/: the controls that read the crate's own source text.
• binary/: the consumer-level controls that start the built executable; binary_tests.rs is the native_binary test root.
• At the top: lib.rs, main.rs (cli-git-native, installed as git) and service_main.rs (cli-git-service).

Where new files go

• A new module goes in the folder of the subsystem it belongs to. Declare it in that folder's mod.rs with a documented line, never in lib.rs.
• Its sidecar test file <module>_tests.rs sits beside it and is declared in the module with #[path]. Test support used by one subsystem stays in that folder under #[cfg(test)]; support used by many goes in testing/.
• A data file that include_str! reads sits in a subfolder beside the module that reads it, and its path is added to the include list in Cargo.toml when the published crate needs it.
• A new subsystem is a new folder with a mod.rs, a pub mod and pub use pair in lib.rs, and a line in the list above. Its name must differ from every module name, singular where it reads naturally.
• Do not put a module in lib.rs or beside it; the folders exist so the top level stays short.
• The source scans (source_guard/) follow mod declarations from the Cargo.toml targets, so a file that no declaration reaches fails them.
• Paths in tooling name the folder, for example src/native/command/command_add.rs. A mutation scope may name subfolders (--file src/native/command/command_add.rs); pass 1 of a two-pass campaign selects tests by top-level folder, so its first pass covers more tests than it did when every module was a top-level entry.
• Records under doc/ that predate this layout name the old flat paths. git log --follow -- <new path> finds a file's history across the move.

The native directory layer

The sources sit under src/native/ although the TypeScript implementation that the name told apart from is gone. This is a layout decision, picked by the coordinator's agent, veto open. Removing the layer needs no binary rename and no behaviour change, because the targets are named in Cargo.toml.

• Keep src/native/ for now and drop it in a follow-up.
  • Pros: no second mass rename while other branches still edit these files; nothing else changes paths. The shared mutation scope rules (nativeSourcePrefix) and their golden scenarios keep their prefix.
  • Cons: the redundant segment stays in every path until the follow-up.
• Move to src/ now (src/lib.rs, src/main.rs, src/service_main.rs and the folders).
  • Pros: the conventional Cargo layout and shorter paths, in one step with the folder move.
  • Cons: a second rename of every file under review in other branches, a change of the prefix in the mutation package and its golden scenarios, and hundreds of path mentions in other packages and documents.
• Move to src/ with the executables in src/bin/.
  • Pros: Cargo finds the executables without [[bin]] paths.
  • Cons: everything the previous option costs, plus the executable entry points move away from lib.rs.

Ranking: keep and drop later > move to src/ now > src/ with src/bin/, because the first avoids a second rename while branches are in flight and loses only a path segment, the second reaches the same layout as the third at a smaller cost, and the third only adds moves that nothing requires.

License

LGPL-3.0-or-later. The license texts are in LICENSES/.

The commit service

A scoped commit (git commit -m <message> -- <paths>, with the message, identity, sign-off, trailer, empty-commit, reused-message, fixup and -a options listed in doc/planning/cli-git-resident-service.md) is landed by a resident commit service instead of by the wrapper itself. The service is a separate program, cli-git-service, that the operating system starts on demand; the wrapper never starts it. Commits queue in arrival order, a batch of whatever is queued lands with one durable write, and a commit that waits says how many commits go first. There is nothing to install and no setting.

• The first commit that needs the service sets up its socket activation itself, and every later one checks it is still right: it compares the two installed units with what this release would write and, only when they differ (an upgrade or a downgrade moved the service program), rewrites them, reloads the manager, clears failed state and starts the socket. When they match it opens no connection to the bus. Setting up writes two unit files into $XDG_RUNTIME_DIR/systemd/user (nothing is left in the home directory; the next first use after a login does it again) and asks the user's service manager, through the session bus, to start the socket. The manager starts the service when the first connection arrives. Units that an earlier release wrote into ~/.config/systemd/user are recognized by their content and removed; a file of the same name that something else wrote is never touched, and the commit stops and names it.
• The single service of the releases before one service per worktree is stopped before the first instance starts, wherever its unit files are (including ~/.config/systemd/user); the commit waits for it to be inactive, and fails naming it when it will not stop. Limit: going back to one of those releases starts its single service beside the per-worktree instances until their idle hours pass; stop cli-git@*.service and cli-git@*.socket with systemctl --user stop before committing with the older release (doc/planning/cli-git-resident-service.md, "Limit: a downgrade").
• A failed backup push of a linked worktree, and a push the service had not run yet, survive git worktree remove and git worktree prune: the wrapper copies them first, and the next command in any worktree shows them once. The copy is kept while the command that took it may still be removing, and until the worktree's checkout is seen intact (Git refused) or its Git directory is gone (Git removed it); a record that cannot be copied or kept is printed on standard error, and what could be copied stays. Limit: a worktree that Git's own gc prunes is not seen by the wrapper, so its records are lost with it.
• The service keeps the records of its newest 1024 finished batches in its journal and removes older ones when it starts; a request is answered as already landed only while its own client retries it, which is minutes at most.
• The service exits after three hours with nothing to do (Bazel's server has the same default); the next commit starts it again. A service started from another release of the wrapper is asked to finish its batch and exit, and the next commit reaches the current one; if it does not exit, the commit says so and names systemctl --user restart cli-git.service.
• git cli-git service status says whether the service answers; it is a diagnostic only.
• When the service cannot be set up (no session bus, no user service manager), the commit stops with a plain message that says what to do, and changes nothing: the service is the only way a commit lands, so there is no other path. Every command line the service does not build from the command line itself (amend with an editor, an editor, signing, templates, a repository whose configuration sets a template, a clean-up mode, signing or trailers) is built by real Git in a scratch repository and landed by the service as Git made it.
• The backup push and the automatic upkeep of a commit the service landed run in the service after the command returned. A push that fails does not fail the commit; the next commit through the service prints a notice once. The service pushes with its own environment, so a variable only your terminal has (an ssh agent address) must also be in the user manager's environment.
• Where a worktree holds the tracked file doc/audit/standards-review-gate-shadow.md, the service also runs a standards review of each commit it landed, after the command returned, and appends one entry to the repository's review log, <common Git directory>/cli-git/standards-review-shadow.md. That log is untracked and inside the Git directory, so no checkout shows it as a change; there is one per repository, shared by every worktree's service, and the service creates it when missing. A linked worktree's entries go there too, never into that worktree's own Git directory (.git/worktrees/<name>), which Git removes with the worktree. Each entry is written in one append, so entries of concurrent instances never interleave, and no lock is taken. The tracked file only switches the review on and is never written by the service. The log has no rotation. SHADOW_REVIEW_LOG points it at another file; the runbook is doc/runbook/standards-review-gate.md.

Splitting a slow worktree add

The wrapper appends one JSON line per git worktree add it serves to <common Git directory>/cli-git-worktree-add-timing.jsonl, in one append with no flush to disk, beside the commit timing file. steps gives each step's milliseconds and calls (steps do not overlap, and otherMs is the rest of totalMs), work gives the time spent on worker threads by name (the Cargo workspaces, the tools' repairs and the source checks run side by side, so a name's sum can exceed the step that waited for it), and notes.route names the rung that served the add.

# sh: the newest add of this repository, one line
tail --lines=1 -- "$(git rev-parse --path-format=absolute --git-common-dir)/cli-git-worktree-add-timing.jsonl"

Splitting a slow commit

Two logs exist without anyone having switched anything on, and neither is ever shown at the terminal. Together they split the time of a wrapped commit by phase.

The wrapper appends one JSON line per commit to <common Git directory>/cli-git-commit-timing.jsonl, beside cli-git-audit.jsonl, in one append with no flush to disk. Every worktree of the repository writes to the one file, and a worktree's instance field tells the lines apart. The file moves aside to cli-git-commit-timing.jsonl.1 at one mebibyte, replacing the older one. At the rate measured on 2026-11-03 (760,605 bytes in about 30 hours) that keeps three to four days of timing. A failed write never changes the commit's result. A commit that stops before it reaches the service, because the machine has no way to start one, writes no line. Read the newest lines of the repository you committed in:

# sh: the five newest commits of this repository, one per line
tail --lines=5 -- "$(git rev-parse --path-format=absolute --git-common-dir)/cli-git-commit-timing.jsonl"

The fields, all durations in whole milliseconds taken from the monotonic clock:

• startMillis is the wall-clock start (milliseconds since the Unix epoch), and time is when the line was written, in the form of every other record of the wrapper. requestId is the identity of the commit's request, the same text as request="<id>" on the service's answered line and as the journal field CLI_GIT_REQUEST_ID, so the two records join on a key and not on a time; it is null for a commit that never made a request. level is warn for a command that stopped and info otherwise. instance, gitDir, branch, release, pid, outcome (landed, ended or stopped) and exitCode say which commit it was.
• totalMs is the whole command. It is the sum of prepareMs (policies, hooks, the private index, the tree and the objects; attempts counts the preparations), buildMs (the message, the identities, the tree check, a commit Git builds, the hand-over of the scratch directory), provisionMs (checking and, after an upgrade, rewriting the socket-activation units), submitMs (sending the request and waiting for the answer), afterLandingMs (the summary and the files a policy corrected), gateMs (the post-commit gate), handoffMs (leaving the push and the upkeep for the service) and otherMs, which is whatever no phase covers.
• submitMs contains reachMs, the time to reach a service of this release. connectTries, refusedConnects and startWaitMs show the manager starting the service, updateRounds counts the retire requests sent to a service of another release, releasesSeen lists every release that greeted, in order, and answeredByRelease is the one that took the request.
• submitMs also contains answerWaitMs, the time from the request being written to the answer: the queue, another program's hold on the index and the landing itself. positions lists the queue positions the service reported, waitingNotices counts its notices that it was waiting for another program, lockHolderKind says what the last of them said of that program (proven-alive for a process proven to hold the lock, unproven for an owner nobody could prove alive, null when there was none), reconnects counts connections that broke and sends counts requests written.
• handoffMs is one figure: leaving the work records and waking the service, which reaches the service a second time. It is not split into a reach and a wait.

The service writes to its standard error, which the service manager keeps in the journal of its unit. The unit is cli-git@<instance>.service, and <instance> is the instance field of the commit's line. Without a line, it is 32 hexadecimal digits, the 128-bit FNV-1a hash of the real path of the worktree's own Git directory (.git in the main worktree, .git/worktrees/<name> in a linked one):

# sh: the unit instance of the worktree you are in
node --input-type=module -e '
import { realpathSync } from "node:fs";
let hash = 0x6c62272e07bb014262b821756295c58dn;
for (const byte of Buffer.from(realpathSync(process.argv[1]))) { hash ^= BigInt(byte); hash = (hash * 0x0000000001000000000000000000013bn) & ((1n << 128n) - 1n); }
console.log(hash.toString(16).padStart(32, "0"));
' "$(git rev-parse --absolute-git-dir)"

systemctl --user list-units --all 'cli-git@*.service' lists the instances that exist. Read the unit around the commit's startMillis:

# sh: the service's lines for one worktree on 2026-11-01 between 22:52 and 22:54
journalctl --user --unit="cli-git@<instance>.service" --since="2026-11-01 22:52:00" --until="2026-11-01 22:54:00"

Or find the service's entry for one commit by the requestId of its timing line:

# sh: the service's entry for one request, with its fields
journalctl --user --output=json-pretty CLI_GIT_REQUEST_ID=30cbb78e-677f-41ff-b059-87d9f49d3409
# sh: only what failed or did not go as intended
journalctl --user --unit="cli-git@<instance>.service" --priority=warning

The service sends each line to the journal's native socket when its standard error is the journal (the service manager's JOURNAL_STREAM names the very stream the process has as standard error). The journal then dates the line by the moment it was sent, not by the moment it got round to reading it, which on a loaded machine could be seconds later and out of order with the manager's own lines. A line carries its priority (err for something the service needed and could not do, warning for something that did not go as intended or a request that failed, info for the rest) and fields to filter on: CLI_GIT_RELEASE, CLI_GIT_INSTANCE, CLI_GIT_AT_MS (the time of the event), and on some lines CLI_GIT_REQUEST_ID, CLI_GIT_BRANCH, CLI_GIT_OUTCOME, CLI_GIT_BATCH and CLI_GIT_OID. When the journal is not listening or the machine's socket queue is full, the line goes to standard error instead, prefixed with its priority as <n> for the stream to read, and a service run by hand writes plain lines to the terminal. A value with a newline in it reaches the journal as written (the protocol's length-prefixed form); the line on standard error is one line. The write to standard error never waits for room: a line the stream took only in part is finished before the next one starts, and a line that finds no room is dropped and counted, and the next entry that gets through says how many (CLI_GIT_LOG_DROPPED on a native entry, a warning line on the stream). Every line ends with at_ms=<milliseconds since the epoch>, the time of its event, on both paths. The userinfo of any URL in a line is replaced by <credentials>.

The lines, each prefixed cli-git-service: on standard error and identified as cli-git-service in the journal:

• started release="<release>" protocol=<n> pid=<pid> instance=<instance> comes first on every start, then ready. An instance=any is a development run.
• answered request="<id>" branch="<ref>" outcome=<word> ... comes once per request answered. outcome is landed, already (a retry of a request that had landed), retry-prepare, branch-moved or failed:<code>. batch, size and position say which requests landed together. The phases, in order of time: read_ms (the client sending its line), queued_ms (waiting for the batch to start), settle_ms (finishing a batch a killed service left open), landing_ms and answer_ms; total_ms is the whole stay from the connection being accepted. landing_ms is the sum of lock_wait_ms (taking the index lock: flushing the marker to storage, lock_stage_ms, with lock_stage_kind=pre when the marker was made while the service was idle and cold when the commit made it, and waiting for a hold on the index, lock_place_ms, with lock_holder="Process <pid>" when the evidence names one, lock_looks for how many times the lock was found held, and lock_found and lock_last for what the first and the last look found: service seq=.. pid=.. state=running|gone|unknown age_ms=.. for a lock with the service's marker, foreign no-marker size=.. age_ms=.. for any other, with open_by=[pid (comm)] when a process has it open: the program's name, at most fifteen bytes, never its arguments or environment; the scan for them is made once, with the evidence for the waiting notice, and stops after a quarter of a second; the marker's birth identity, boot ID and PID namespace are not in the line), reads_ms (the branch reads and comparisons), objects_ms, journal_ms, ref_ms (moving the branch) and index_update_ms (removing the conclusion state, then building and installing the new index); records written before 2026-11-04 carry the same duration as sync_ms, which was never a disk sync; lock_hold_ms is how long the index lock was held, from its placement to its release (the branch move and the index build), and is no phase of that sum; attempts is more than one when the branch moved under the batch. Each figure is cut to whole milliseconds on its own, so a sum can fall short of its total by less than one per term. The line never holds the commit's message, its author or its address.
• waited for the index lock batch=<n> result=taken|refused lock_wait_ms=.. lock_stage_ms=.. ... comes when the lock was found held or when flushing the marker or waiting took a second or more, at the moment the wait ends; it carries the same lock_* fields as the answered line, and stays when the landing fails afterwards. A slow commit with a long lock_stage_ms and no lock_looks lost its time to the disk and not to another program.
• pushed branch="<ref>" oid=<oid> items=<n> outcome=done|failed|skipped push_ms=<n> comes once for each push the service ran for a branch after commits (a failed push is a warning and carries Git's words, credentials masked, after a colon), automatic maintenance ran maintenance_ms=<n> and the review of <oid> is done review_ms=<n> say the other work ran, and settled the batch an earlier service left open batch=<n> decision=finish|drop (a warning) comes for each batch a restart found open.
• retire requested by a client; this service is release "<release>" and exits after its batch comes once for each retire request the service reads, and a client sends one in every round while the old service is still there, so the number of these lines is at most updateRounds of the commit's line. The request carries no release, so the client's side is in its line: releasesSeen, updateRounds.
• exiting reason=<idle|retire|error|signal-<n>> uptime_ms=<n> comes last. idle is three hours with nothing to do, retire is a retire request, error is the service's queue closing, and signal-15 is the manager stopping the unit. Uptime counts from the start of the process for a signal and from the start of serving for the others. A service killed by SIGKILL or the kernel's out-of-memory killer writes nothing, which is itself the evidence.

To split the slow commit of 2026-11-01 that showed Updating the commit service..., read its line in cli-git-commit-timing.jsonl: a large reachMs with updateRounds above zero and two entries in releasesSeen is a service of another release being replaced, and the journal of the instance shows that service's retire requested and exiting lines and the new started line; a large answerWaitMs is the queue or a held index, and the answered line of the request in the journal says which, and which phase of the landing.

The design, the measurements and the migration plan are in doc/planning/cli-git-resident-service.md.