Skip to content

indexio 0.1.10

Choose a tag to compare

@github-actions github-actions released this 22 Oct 13:30

Hooks that follow how agents actually read files, a PowerShell hook, Claude Code memory notes in recall, fewer repeated answers, and compaction that no longer rewrites the whole index. 30 changes, every one A/B tested against 0.1.9 on a copy of a real 49-repo, 12k-file index by replaying real calls from two weeks of Claude Code sessions.

A/B results

Measured on Windows 11, 32 cores. "Before" is 0.1.9.

Case 0.1.9 0.1.10
Greps with 5+ alternatives (1,972 real calls) some files dropped 82 calls gain lines, 0 lose any
Hook that cannot decide in time waits for the 10.7 s timeout allows after 3 s
Time to deny a read 40.8 ms median 17.8 ms median
PowerShell reads of indexed files not hooked routed to indexio (108 denies over 2,065 calls, 394.6k chars)
Identical read_span answers repeated in a session sent in full one line; -10.3% read_span output
Repeated code_grep / code_search / file_outline answers sent in full -5.9% / -12.3% / -15.1% output
Reads of an agent worktree's identical copy not hooked routed to indexio (0.93M of 1.75M chars)
Failed calls from real sessions (55) 39 answered 45 answered, 0 broken
Grep across all repos (fn locate) 17.4 ms 9.7 ms
Shard compaction while shards are being written 28-30 s, ~2.5 GB 13-25 ms, 220 MB or less
indexio compact of 48 shards 32.0 s 0.9 s
Vector compaction of 13 segments 11.2 s, 2.8 GB 0.8 s, 101 MB
Hybrid answers after vector compaction (vs before it) 83% identical 100% identical

1,500 real hook calls and 1,520 real search calls return the same verdicts and answers as 0.1.9 wherever a change was not meant to alter them.

Hooks

  • New PreToolUse hook for PowerShell: Get-Content, Select-String, Get-ChildItem | Select-String, -Tail N, and $c = Get-Content f followed by line ranges or loops get the same treatment as the Bash hook. Run indexio hook install again to add it.
  • A hook gives a verdict within 3 s or allows the call.
  • The re-index after a deny runs in a detached process, and is skipped for the session's own repo (the server already watches it).
  • Reads of 60 lines or less go through, whether a Read range, sed -n, head, tail, Get-Content, or a whole file that short.
  • Sending a denied call again unchanged lets it through.
  • Byte views and counts (cat -A, head -c, tail -f, | wc, | uniq -c), grep -c/-l/-L/-q/-o and Python byte-analysis one-liners are no longer denied.
  • tail -n N and multi-range sed -n 'A,Bp;C,Dp' are answered with exact read_span ranges instead of a whole-file outline.
  • A read of a file in .claude/worktrees/<wt>/ that is byte-identical to the indexed file is redirected to it; an edited copy is left alone.

Search and lookups

  • Regex prefilters keep every alternation branch (5+ branches lost files).
  • (?i) regexes use the trigram index instead of a brute-force scan.
  • Lookaround regexes return a superset with a note instead of an error.
  • code_grep and lexical code_search leave session transcripts out of the search itself unless asked for (repo:sessions).
  • File tools without repo use the session repo when it has the path, else the one repo that does; repo plus a path that starts with the repo folder name resolves.
  • code_search accepts code_grep's pattern, repo and path.
  • An answer identical to one the session got in the last 20 minutes from read_span, code_grep, code_search or file_outline (400+ chars) comes back as one line. Sending it again returns the text; a compaction or /clear resets this.
  • read_span, code_grep, file_outline, code_search, recall and list_files are marked always-load, so Claude Code skips a ToolSearch round trip before using them.

Sources

  • Claude Code memory notes (~/.claude/projects/*/memory/) are indexed with the transcripts; recall searches them, this project's notes first.
  • A session whose unregistered checkout sits beside three or more registered repos adds it (git or a plain project folder up to 20,000 files, never a folder of repos). Opt out with INDEXIO_NO_AUTO_ADD or <data_dir>/auto-add-skip.
  • A repo with no commit yet is indexed from its working tree.
  • Config files opened by name are indexed: Caddyfile, Dockerfile.*, .gitignore, .dockerignore, *.example, *.inf, *.patch and similar. Secret files are still skipped.

Compaction

  • Shard compaction merges the cheapest run of adjacent shards instead of always the oldest, so the big base shard is not rewritten for every few small deltas. Once tombstones pass 30% the oldest run is merged, which drops dead docs.
  • Vector compaction merges only the small segments when that is enough and carries the large one over untouched.

Full Changelog: v0.1.9...v0.1.10