Gabriel Vidal ebae75df94 perf(ai-agent): stop the backend re-walking, re-parsing and re-forking everything ...

Idle CPU sat at ~58-110% and the catalogs took seconds, because the hot paths
all did O(everything) work on a timer or per request:

- the indexer os.walk'd the whole workspace every 10s to find ~280 files, and
  the watcher rglob'd ~2600 transcripts every 2s (pathlib allocates a Path per
  entry) -> fsscan.walk_files (os.scandir) + a cached discovery walk, refreshed
  every WALK_INTERVAL_SECS and forced on save;
- a live session's growing transcript was re-parsed from byte 0 on every tick
  (O(n^2) over a session) -> conversations.ParserState is now resumable, and the
  indexer feeds it only the appended lines from a stored byte offset;
- the watcher already knew which transcripts changed but threw it away ->
  scan_transcripts(only=…) imports + re-parses just those, instead of re-walking
  both source dirs and the whole archive;
- /api/services forked two `git log`s per service per request (11s for 34) ->
  githist.BucketedHistory walks each repo's log once, buckets commits per dir and
  caches on the repo HEAD; standalone project repos get a TTL cache;
- every list request json.loads'd ~900 summary blobs and deep-copied the meta
  sidecar per card -> decoded summaries are cached in the store (read-only) with
  a version counter, and MetaStore.peek() skips the copy on hot paths.

Also: SQLite in WAL, file word/char stats stored on change, and the skills
rollup no longer SELECT *'s every summary blob to read one column.

Verified: parse output is byte-identical to the old parser across all 905
archived transcripts (full and incremental), and idle CPU drops ~58% -> ~7%
measured against the live container on the same workload.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2026-07-13 19:34:34 +02:00

5.3 KiB

Raw Permalink Blame History

View Raw