Skip to content

Tags: lambiengcode/reify

Tags

v0.2.3

Toggle v0.2.3's commit message
v0.2.3

For Windows users this is a correctness fix rather than a feature.

The index lock in 0.2.2 did nothing on Windows. `process_is_alive` was
implemented for unix and stubbed to `false` everywhere else, so every lock
looked stale and reclaimed itself, and two `reify index` runs could proceed
against the same store. No configuration avoids it; upgrading is the fix.

It was found by testing on the platforms releases actually ship for. CI ran on
Linux only, so the Windows binary added in 0.2.2 went out on the strength of a
type-check - which is exactly what a type-check cannot catch. Windows and
macOS now run the full suite and an end-to-end CLI smoke on every change; the
job found the lock bug in its first run, by failing to recognise its own
process as alive.

Also here:

  retrieval, SWE-bench Verified   84.6% -> 87.0%  (three rounds, n=500)
  offered every touched file      77.0% -> 81.4%
  end to end vs BM25              73.3% against 67.3%, p = 0.24

Retrieval improved by capping how much of the window a single file may claim:
one loosely-matched file used to take 13 of 20 slots. The end-to-end arm is
ahead of BM25 for the first time, but not significantly so at this sample
size, and the README says that before it says the percentages.

`reify index` no longer fails outright when git cannot read history on a
partial clone - the code still indexes, and the missing history is reported.
Three more MCP tools, within the same per-turn token ceiling as the original
three.

No store schema change: upgrading from 0.2.2 needs no reindex.

v0.2.2

Toggle v0.2.2's commit message
v0.2.2

Indexing roughly twice as fast and `reify why` seventeen times faster, measured
against the previous build on the same machine:

  full index            6.75 s -> 4.25 s
  reindex, no change     256 ms -> 101 ms
  reindex, one edit      974 ms -> 486 ms
  reify why             1517 ms -> 87 ms median

Discovery stats past files whose size and mtime are unchanged instead of reading the
whole tree, and hashes the rest across all cores. Reference resolution re-resolves
only references the edit could have affected rather than all 144,309. Query-time git
runs with GIT_NO_LAZY_FETCH=1, which stops a blobless clone silently fetching blobs
over the network - 29.5 s of network for 0.07 s of work - and makes the offline
guarantee hold for the whole process tree, not just this binary.

Also: a hand-drawn mascot and derived logo set, a colour demo recorded with vhs, a
documentation site at lambiengcode.github.io/reify, and a link card leading with the
SWE-bench result.

The store schema is v7. Run `reify index --force` once to rebuild an older store.
Includes everything from the untagged 0.2.1.

v0.2.0

Toggle v0.2.0's commit message
reify v0.2.0 — SWE-bench Verified results, self-management commands, …

…feature-tour demo

v0.1.0

Toggle v0.1.0's commit message
reify v0.1.0 — first release