[TheGreatAxios]

Summary

• Floats isomorphic-git to 1.44.2 via the existing ^1.27.2 catalog range (single lock resolution change; 1.42.7 index-write-once-per-walk and 1.43.3 packfile-checksum-once land on the context store's hot paths for free).
• Batches the context store's per-file add/remove loop (src/session/optimized-context-store.ts:~1335) into array-form calls on 1.43.0+ — one GitIndexManager.acquire per call instead of one index write per path. On-error net state preserved: the same resetIndexPaths(dir, extraPaths) reset covers the add∪remove union, so a mid-batch failure still leaves nothing staged. Sole behavioral delta: multiple simultaneous add failures surface as MultipleGitError instead of the first error; the surrounding catch is type-agnostic.
• The findRoot() conversion for project-key was evaluated and declined with evidence: the API is async-only while all four consumers (session/index.ts, crash/report.ts, permission/path-restriction.ts, config.ts) are synchronous by design, and findRoot ignores GIT_DIR/GIT_WORK_TREE/core.worktree semantics that rev-parse --show-toplevel honors — different semantics under worktrees and env-var overrides (probes showed matching behavior under plain worktrees/submodules; the env-var cases are the decliners).

Verification

• bun run check exit 0: full one-process suite 8939 pass / 0 fail across 581 files.
• Focused: bun test src/session/optimized-context-store.test.ts 58 pass / 0 fail; bun test src/session 563 pass / 0 fail across 41 files.
• Lock diff is exactly one line (isomorphic-git 1.41.9 → 1.44.2); no pin change, so the known bun re-dedupe shake (tracked in CL-10681) does not apply.

Fixes CL-10654

[linear-code]

CL-10654

[corbits-triage]

Superseded by the comment below.

[corbits-triage]

Superseded by the comment below.

[corbits-triage]

Triage: Unsure

A maintainer will review this pull request.

[TheGreatAxios]

Independent perf verification for this PR (measure-only lane, no source touched).

Method: benchmark pitting the exact pre-CL loop (git show 10042685: per-path git.remove) against the landed code (optimized-context-store.ts:1927-1949: batch + omit-one-out retry) on synthetic repos (fresh repo per run, 1 warmup + 7 reps at N=10/100, 5 reps at N=1000, 3 full runs), isomorphic-git 1.44.2.

• fs proof: batch performs exactly 1 .git/index write; per-path performs exactly N. Every run, deterministic. (Lock writes 0 both sides — acquireLock is an in-process mutex.)
• Wall medians: N=10 → ~6–9x (0.19–0.31 ms vs 1.61–1.79 ms); N=100 → ~75–135x (0.23–0.36 ms vs 16.9–49.2 ms); N=1000 → ~1125–1160x (0.53–0.54 ms vs 602–621 ms). Batch time is flat; per-path is linear. Perf claim CONFIRMED, effect scales with N.
• On-error net-state equivalence CONFIRMED on all four fault tests: persistent path fault → byte-identical statusMatrix; transient EIO → identical state at write cost 13 (old) vs 3 (new); persistent EIO on batch → abort-before-write, index sha unchanged; cross-strategy restore via resetIndexPaths → identical snapshots.

No follow-up needed — recording the numbers here.