[medium] fix(git): stop respawning git outside repos and across session cwds - #631

BLUF

• Priority: medium. Removes git subprocesses that run on every render and never get cached.
• Outside a repository: rev-parse --is-inside-work-tree and the origin/upstream remote lookups used to spawn git on every render, because with no git dir there was nothing to cache against. They now return null without spawning, but only when git is certain to fail. Spawns per render (heavy config): 4 → 0.
• Two sessions in one repo at different cwds (root and a subdirectory) wiped each other's persistent cache on every render. The cache file is now keyed by (git dir, cwd). Spawns per render with the two sessions alternating: 14 → 0 once warm.
• CPU median per render: -13 % to -19 % on git-heavy configs. stdout is byte-identical.

Details

(a) No repository. getGitRepoMetadata returns null when no .git is found. readPersistentCacheEntry and writePersistentCacheEntry then return early, so a statusline in a non-repo cwd spawns git on every render forever. runGitArgs now skips the spawn only when all of these hold:

• A cwd was resolved from the payload. With no cwd, git still runs in the process cwd as before.
• GIT_DIR, GIT_WORK_TREE and GIT_COMMON_DIR are all unset, since any of them can point git elsewhere.
• No .git entry of any kind (dir, gitfile, malformed gitfile, symlink) and no HEAD (a bare repo, or a cwd inside a git dir) exists anywhere on the physical path (realpath) from cwd up to the filesystem root. Git discovers from getcwd() after the chdir, so a logical path through a symlink could miss a repo that git would find.
• GIT_CEILING_DIRECTORIES and filesystem boundaries only make git stop earlier, so walking all the way to the root is the conservative direction.

Every command that runGit/runGitArgs issues (rev-parse, status, diff, rev-list, ls-files, remote, symbolic-ref) needs a repository. Where the skip applies, git would exit 128 and the function would return null anyway.

(b) Cache keyed per cwd. The cache file was named after the git dir but stored one file-level cwd. A read with a different cwd missed, and the next write replaced the whole file. Hashing (gitDir, cwd) into the file name gives each session its own file. This also removes the read-modify-write race between them. Sharing entries across cwds would be wrong anyway, because some outputs depend on the cwd: ls-files --unmerged only lists conflicts under it. The existing per-repo files become one orphan per repo and are no longer read.

Measurements

Spawn counts (strace, heavy-ttl0, i.e. TTL 0 so freshness is decided by HEAD/index mtimes only):

bench2: round-robin fork/exec, 20 interleaved passes. CPU is user+sys including reaped children. The host was heavily loaded, so compare ratios rather than absolute values.

Run 1: git17-ttl0 (19 git widgets), load1 42-53 on 6 cores.

Run 2: heavy-ttl0 (heavy config with a transcript, so the git share is smaller), load1 44-64.

In the non-repo case the CPU difference is within noise, because a git process that exits immediately with 128 costs little CPU. The gain there is the spawn count and latency.

Checks

• bun run lint: the full run flagged only two unnecessary casts in the new test, now removed. After the fix, tsc --noEmit is clean and eslint is clean on the changed files.
• bun run build: OK.
• bun test: the new and existing git tests pass (59 in git.test.ts). Red before the fix: the two new behaviour tests fail on main. The new tests hide any stray .git/HEAD above the temp fixture, so they do not depend on the host's directory layout.
• The full suite on this very loaded host shows the same timeout-driven failures on unmodified main as on this branch: fetchUsageData error handling, custom-command capture, and several TUI menu tests such as ColorMenu, which also fails in isolation on main. None of them touch git.
• Same file as my sibling perf PRs on git.ts (render-scoped cache I/O), which will need a trivial rebase whichever lands first.

Overlap

• [medium] perf(jj): detect repos without spawning and cache jj commands #632 (also from me) edits src/utils/git.ts as well; the two branches conflict there. Whichever merges first, I will rebase the other.

🤖 Generated with Claude Code