Skip to content

Refresh macOS OCR smoke and qualify resumable scans - #225

Merged
silverstein merged 4 commits into
masterfrom
codex/ocr-current-smoke-20261013
Oct 13, 2026
Merged

silverstein merged 4 commits into
masterfrom
codex/ocr-current-smoke-20261013

Conversation

@silverstein

Copy link
Copy Markdown
Member

Outcome

Refresh the real macOS OCR protocol smoke so it qualifies the current source
instead of stopping at September's obsolete 57/58-tool counts. Optional OCR
remains disabled by default, separately installed, and explicitly confirmed.
No product runtime, bundled engine, native PDF text, or source/share bytes change.

The optional two-page check exercises actual Vision recognition through stdio,
retains page one, closes only its own server process, reads the checkpoint in a
fresh server, completes page two, and requires a repeated completed resume to
do no new page work. It reopens retained proposal/commit/image/review bytes and
checks source preservation. Catalog checks bind current default names and full
optional input/output schemas and annotations. Negative controls require the
correct consent, source-identity and encryption refusal reasons.

Qualification

  • Independent adversarial source review accepted exact 7eb5a42c after fixing
    a valid-shape optional output-schema substitution and the encrypted adapter's
    exact exit-code contract.
  • 22 new dependency-free cases pass on Linux and macOS. macOS native OCR plus
    these cases passes 49/49; focused job, discovery, documentation and test-runner
    suites pass 183/183 with a retained process exit of 0.
  • Actual macOS 26.6 arm64 / Node 22.23.2 / Vision 1.0.1 protocol run passes:
    61 default tools, 65 configured tools, one standalone page plus a two-page
    resumable raster-scan job. Source remains unchanged. Completed pages are not
    rerun. All proposals remain explicitly unverified.
  • Actual isolated browser checks pass for both retained pages at 1440px and
    390px: six frozen synthetic lines per page, matching source-box selection,
    warnings, private artifact hashes, no horizontal overflow, no scripts or
    HTTP(S) resource timing entries. This is not an OS network audit.

Initial focused partition expectations failed and are corrected here. One
non-detached SSH test invocation ended without a final process receipt and is
not claimed green; the isolated rerun retains actual exit 0. The narrow-browser
harness initially attempted an offscreen click; the corrected check requires a
visible target and uses normal scrolling/clicking, with no product code change.

These are current-source protocol/engine/browser checks, not installed Claude
or ChatGPT acceptance, general OCR accuracy, a new benchmark or release. No
personal desktop restart/configuration change, provider/model call or upload.

Bead: pdf-toolkit-mcp-9gew, discovered from pdf-toolkit-mcp-478i. The parent
retains actual installed-desktop acceptance as a separate gate.

@silverstein
silverstein merged commit 01823d3 into master Oct 13, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant