Skip to content

release: FixMap v0.8.2 - #432

Merged
aryamthecodebreaker merged 1 commit into
mainfrom
codex/v0.8.2-docs
Aug 2, 2026
Merged

release: FixMap v0.8.2#432
aryamthecodebreaker merged 1 commit into
mainfrom
codex/v0.8.2-docs

Conversation

@aryamthecodebreaker

Copy link
Copy Markdown
Owner

B12 of the sweep, plus the version bump. 13 issues. npm run ci green; server manifest and Action metadata checks pass.

Suites re-recorded — and confidence moved

held-out regression
Top-1 / 3 / 5 7/12 · 8/12 · 9/12 11/16 · 16/16 · 16/16

Hit rates unchanged. Calibration changed, for the better:

held-out confidence bands     before        after
  high                        3/6 = 0.50    2/3 = 0.667
  medium                      3/4 = 0.75    4/7 = 0.571
  low                         1/2 = 0.50    1/2 = 0.50

The high band halved in size and became more accurate. That follows from scanComplete now meaning "every candidate was actually read" rather than "the file limit was not reached" — a repository with files past the sample ceiling no longer earns a label it cannot support. high is now more accurate than medium on held-out, which is the ordering the label is supposed to have.

CHANGELOG (#280, #308)

0.8.1 shipped without an Evidence section; 0.8.2 has one. It states plainly that the held-out composition changed during the 0.8.1 follow-up — got rotated into regression once it had informed a fix — so 8/12 is not a regression against the 9/12 published before it. It also records the two ranking proposals that were measured and rejected (#282, #347). Two bullets beginning "Added" were sitting under 0.8.1's Fixed heading.

A note at the top says accuracy figures inside released entries are historical by design, with the live numbers on the generated evidence page.

The gate named the wrong suite (#316)

evaluate-external.mjs --suite heldout reported every failure as "External evaluation" — poor behavior for the check standing between a regression and a release. It says "Held-out evaluation" now.

Documentation drift

#405 the npm README documented a plan+verify surface three releases old — no doctor, --explain, --compare, --limit, --exclude, or --working-tree. #417 the docs page counted four explain statuses; there are five (outside-limit was missing). #404 the get-started Action snippet had drifted from the README on both actions/checkout version and permissions — aligned to the README's known-working set rather than guessing a narrower one. #336 / #399 the growth log, launch kit and launch checklist now carry a dated banner pointing at the generated evidence page. #386 / #324 / #413 the stop-word tradeoff, the Windows ENOTEMPTY workaround, and the fact that using: node24 is GitHub's runner runtime rather than a requirement on your Node are all written down now.

Version bump

Root, core, cli, action, server.json and the lockfile move to 0.8.2. apps/web keeps 0.0.0 deliberately — it is unpublished, and publish.yml:93 validates only its @aryam/fixmap-core dependency, never its version. The lockfile was patched by hand: regenerating it on Windows drops the @emnapi entries Linux CI requires (9 present, verified).

Closes #280, closes #285, closes #308, closes #316, closes #324, closes #336, closes #377, closes #386, closes #399, closes #404, closes #405, closes #413, closes #417

🤖 Generated with Claude Code

Re-records both accuracy suites from this tree. Hit rates are unchanged —
held-out 7/12, 8/12, 9/12 and regression 11/16, 16/16, 16/16 — but confidence
moved: the high band on held-out went from 6 cases to 3, and from 3/6 correct to
2/3. That follows from scanComplete now meaning "every candidate was actually
read" rather than "the file limit was not reached", so a repository with files
past the sample ceiling no longer earns a label it cannot support. high is now
more accurate than medium on held-out, which is the ordering the label is for.

The CHANGELOG gains the Evidence section 0.8.1 shipped without, and states that
the held-out composition changed in the 0.8.1 follow-up — so 8/12 is not a
regression against the 9/12 published before it, just different repositories. It
also records the two ranking proposals that were measured and rejected.

The held-out gate named the wrong suite in every failure it reported, which is a
poor property for the check standing between a regression and a release.

Documentation: the npm README described a plan+verify surface three releases
old; the docs page counted four explain statuses when there are five; the
get-started Action snippet had drifted from the README on both checkout version
and permissions; and the growth, launch-kit and checklist documents now carry a
dated banner pointing at the generated evidence page rather than reading as
current truth. The stop-word tradeoff and the Windows ENOTEMPTY workaround are
written down where they will be found.

Versions: root, core, cli, action, server.json and the lockfile move to 0.8.2.
apps/web keeps 0.0.0 deliberately — it is unpublished and publish.yml validates
only its core dependency. The lockfile was patched by hand; regenerating it on
Windows drops the @emnapi entries Linux CI needs (9 present, verified).

Closes #280, closes #285, closes #308, closes #316, closes #324, closes #336
Closes #377, closes #386, closes #399, closes #404, closes #405, closes #413
Closes #417

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@aryamthecodebreaker
aryamthecodebreaker merged commit 943b178 into main Aug 2, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment