mirror of
https://github.com/yamadashy/repomix.git
synced 2026-05-30 11:18:53 +02:00
Kazuki Yamada
67e1d7d8a5
Reorder pack() startup so two pre-existing latencies stop sitting on the critical path between searchFiles and produceOutput: 1. Pre-init the metrics worker pool BEFORE searchFiles. Each metrics worker pays ~200-285ms to parse gpt-tokenizer's BPE table; we already ran this in parallel with the post-search pipeline, but the search phase itself (~100-150ms) was outside the overlap window. Pool sizing no longer depends on the file count, capped at METRICS_PREWARM_THREAD_CAP (=3) — enough to absorb the typical metrics workload while small repos stop over-spawning workers they will never use. 2. Lazy-load `produceOutput` via the dynamic-import deps pattern already used for `packSkill`. The output-style module chain (Handlebars compile + style modules) now resolves in parallel with searchFiles/collect instead of blocking packager.ts entry. 3. Drop the `await metricsWarmupPromise` and `await tokenCacheLoadPromise` barriers before `produceOutput`. Neither produceOutput nor file collection depend on those promises; the awaits only need to fire before `calculateMetrics` actually dispatches tasks, so they move inside the metrics branch of the trailing Promise.all. The metrics taskRunner construction moves OUTSIDE the try block so a synchronous throw never escapes without cleanup; the surrounding try expands upward to wrap searchFiles + sort/regroup so a search failure still triggers `metricsTaskRunner.cleanup()`. A new packager.test asserts this leak path. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>