turbo

acd5ae5b - perf: Skip unused repository indexing for package listings (#13792)

perf: Skip unused repository indexing for package listings (#13792) ## Summary Avoid building the tracked Git index, scanning untracked files, and calculating unused cache SCM provenance for `turbo ls`. The listing path keeps its SCM handle so `--affected` and Git-based filters continue to work; all other `RunBuilder` callers retain the existing behavior. ## Models - Author: `openai/gpt-5.6-sol` - Opposite-model reviewer: `anthropic/claude-fable-5` (`fable_performance_reviewer`) ## Hypothesis `turbo ls` does not hash files, execute tasks, or use cache provenance, so its eager repository index and dirty-state work is unnecessary. Removing that work only from the listing path should reduce warm end-to-end latency and filesystem syscalls without changing output or affected filtering. ## Methodology - Built and preserved baseline and candidate with `cargo build --profile release-turborepo -p turbo`. - Baseline SHA-256: `5a29b912b6f4790eb95dd8f66426335f29b5500140d68ebd6327dc6c918a630e`. - Candidate SHA-256: `37812f33869d97c8fd51361b46110e5f88bd5883dce4d04b5a347e4f12701cdc`. - Both binaries: 50,462,240 bytes. - Timed warm `turbo ls --output=json` with telemetry disabled and output redirected. - Used 3 warmups and 20 measurements per binary per corpus, split into four balanced AB/BA blocks. - Bootstrapped the four block-level median speedups with 100,000 resamples; no samples were excluded. - Corpora: create-t3-turbo `8f945b7` (13 packages), Payload `24ac895` (52), and Turborepo `0cfccdd8` (86). - Environment: Linux x86_64, 2-vCPU Intel Xeon 2.90 GHz, 4.2 GiB RAM, Rust nightly 2026-05-22, hyperfine 1.19.0. ## Results The clean-checkout Payload command improved from **58.0 ± 8.1 ms** to **34.2 ± 1.5 ms** across 20 runs. Balanced block results: | Corpus | Mean block speedup | 95% bootstrap CI | | --- | ---: | ---: | | create-t3-turbo | 10.86% | [7.50%, 13.32%] | | Payload | 42.82% | [34.07%, 54.43%] | | Turborepo | 12.73% | [7.55%, 17.90%] | Every block improved in both command orders. The claim is limited to warm-state `turbo ls`. `strace -f -c` on Payload corroborated the mechanism: - total syscalls: 19,052 → 3,748 - `newfstatat`: 8,897 → 3 - `getdents64`: 2,261 → 4 - `openat`: 1,564 → 313 - `fstat`: 1,162 → 15 The strace timing itself is diagnostic and is not used for the wall-clock claim. ## Correctness - `cargo test -p turbo --test command_ls_test` — 5 passed - `cargo test -p turbo --test affected_test` — 32 passed, including affected listing/filter coverage - `cargo clippy -p turborepo-lib --all-targets -- -D warnings` — passed - Final release build — passed - `git diff --check` — passed - Baseline/candidate JSON was byte-identical across all three corpora, the dirty local checkout, and Payload with an injected untracked file. ## Review The opposite-model reviewer approved the exact final diff with no blocking findings. It found the timing evidence, syscall corroboration, output equivalence, and affected-filter coverage sufficient. Non-blocking limitations were the noisy 2-vCPU environment and the warm-state-only scope. Co-authored-by: vercel-gh-bot-4[bot] <312518292+vercel-gh-bot-4[bot]@users.noreply.github.com>