v0.5.2

Summary

QAMap 0.5.2 improves the local pull-request review brief with removed-code evidence,
heuristic change signals and an explicit map of evidence it cannot fully show.
It makes no model calls or source uploads; the calling agent still uses tokens.

Changes

• Preserve removed lines before reference excerpts, label heuristic signals and list incompletely shown files with recovery commands.
• Highlight rewritten guards, dropped waits, error handling, unassigned fields and callers that may need to handle a new throw, without presenting static evidence as executed QA.
• Ask the reviewer to keep the brief in the foreground, recover omitted diffs and check complete changed functions and base behavior before making definite claims.
• Avoid client QA scenarios from server-only changes, resolve dotted module names and keep joined nouns in one flow stage.
• Reuse repository walks, import graphs and large-monorepo index caches while preserving changed files and nearby configuration under index limits.
• Add the repository plugin marketplace, isolated Claude directory package, listing links and paired installation guides. Both READMEs show recorded results alongside their limitations.
• Let fork PRs use their own branch names while retaining title, template and confirmation checks.
• Limit coverage to two concurrent test files without lowering thresholds or skipping tests. Distinguish benchmark command timeouts from output-buffer failures and include compact-call diagnostics.
• Set the post-release smoke default to the previously published 0.5.1 package; publication checks explicitly select 0.5.2.

Validation

The earlier clean release gate passed 837/837 tests, 44 static QA contracts,
11/11 repository-index checks, 10/10 context checks and three execution contracts
with all three seeded regressions caught. Scan found zero findings. Plugin
metadata and the packed 276-file local installation smoke passed. Coverage
execution passed 837/837: 92.11% lines, 88.80% branches and 95.73% functions.

After the documentation corrections, the final PR head passed 837/837 tests with
no skips, plugin metadata, the policy scan, all 44 static contracts, the local
plugin installation smoke, package dry-run, whitespace and README line checks.
All four required CI checks passed on that exact head. Coverage figures remain
from the earlier clean run, not a new coverage run after the corrections.
The initial warm-arm coverage failure was not reproduced; its cause remains
unknown, and both failed and passing receipts are retained.

Historical host measurements used Claude Code CLI 2.1.292 and one fixed model on
42 public PRs from 12 projects. The build that became 0.5.2 used 48.6M tokens
against 86.1M for standalone, a 44% reduction in the sum of per-PR medians.
Blind preference favored standalone 18 to 14 with four ties; valid findings
were 19 versus 16, incorrect findings 7 versus 7 and unverifiable findings
5 versus 12. The design followed inspection of these cases, so this is not
independent evidence or a new measurement of the released package. Codex was
not measured.

On the separate internal 19-case comparison with CLI 2.1.292, 0.5.1 used 65.4%
fewer tokens and a separate candidate used 54.6% fewer. That candidate is not the
build measured on the 42 public PRs. These results do not establish universal
savings or superiority in bug finding.

See the 0.5.2 validation record.

Install

npm install -g @ivorycanvas/qamap@0.5.2

GitHub Action users can pin IvoryCanvas/QAMap@v0.5.2. Plugin users need the
matching 0.5.2 CLI; directory updates are separate from npm publication.
Claude Code users can install the repository plugin with:

/plugin marketplace add IvoryCanvas/QAMap
/plugin install qamap@ivorycanvas

QAMap itself makes no model calls, while agent invocation and interpretation
still use model tokens. Static reports do not mean tests were executed.

Full changelog | Release record