Final Edition

Keep the passage. Let the agent catch what the reader lost.

Comment

What it does

Final Edition turns deliberately selected saved pages and bookmark metadata into a portable, searchable reading edition. It keeps supported code, tables and figures alongside source digests and visible omissions. Missing, ambiguous or unsupported captures stay visible rather than becoming invented content.

The local Python builder lets the owner review source pairings, choose a complete or reading-only delivery profile, and export a static HTML edition. Complete delivery includes original source bytes; reading-only excludes those originals and auxiliary bookmark metadata. Reading-only is an inclusion boundary, not redaction or permission to share.

The problem

Keeping a saved page is not the same as preserving the passage a reader needs. Conversions can lose code, table structure, figures or source context. A passing build can also miss a broken reading interaction. Final Edition makes these losses and associations inspectable, then uses a bounded GitLab-assisted workflow to investigate failures in the reading contract.

GitLab-assisted post-code workflow

The reading-contract reviewer uses GitLab Duo Agent Platform. It checks the merge request, exact failed pipeline and source identity, job receipts and reviewed reference before proposing a bounded repair or abstaining. Its five read tools do not write to the repository, merge changes or release artifacts. Human application and publication remain separate decisions.

In a disclosed controlled evaluation, a one-line mutation deliberately removed the reading-tree reset. The initial flow abstained because the reader failure occurred before its supported jobs. A narrow routing update let the same read-only agent inspect that early failure. It diagnosed the missing reset and proposed a small patch plus a regression test. Independent local application passed checks. A negative control with a deliberately wrong SHA then produced an evidence-based abstention.

These are observed runs, not a general success-rate claim:

The preserved live outputs also contain two reporting caveats: the positive output incorrectly said real Duo execution was still pending, and the negative output overstated a distinct-reference requirement. Those statements are not evidence of current status.

How it is built

Python's standard library parses the selected HTML as data into a deliberately limited representation. The JavaScript reader constructs supported content with DOM element and text APIs. HTML and CSS provide the local review interface and static reader. GitLab CI runs checks; the tested Duo flow diagnoses and proposes changes. Unsupported or omitted content is reported. This is not a general sanitization or security certification.

The local builder listens on loopback. Selected files are processed by the local Python process, not wholly inside the browser. The project is not a hosted service.

What testing taught us

A later real-browser test found that refresh lost the open item and query. A separate, independently developed improvement now preserves edition-scoped browser history state, including search and caveat filter, and supports Back and Forward without trapping the reader. This improvement is not attributed to the GitLab agent.

The assembled local candidate passed 157 Python tests and 36 Node tests, including 16 builder and 20 reader tests. Cloud Chrome HTTP checks observed refresh and navigation behavior in both delivery profiles. After the improved candidate was published to a review branch, all 67 downloaded source files matched the tested clean candidate. The actual downloaded CI editions then passed a fresh cloud Chrome HTTP smoke check for article retention on refresh, query retention on Back, and replacement of prior content when opening a second item.

These results do not establish file:// cold-open, cross-browser behavior, security qualification or no-network behavior. The release gate intentionally remains closed.

Current draft status and next steps

Intended category: Path A, Start Fresh: Assisted / Best Assisted Agent. This is a draft, not a final entry.

The improved candidate is available in draft GitLab merge request !2 at source commit cf47ffc529007a6b171bd1d6cab802d3c24b7531. Its four automatic CI jobs passed. The pipeline remains blocked at the deliberately untouched manual release_readiness gate: it is not an overall green or release-approved pipeline. No merge, tag or release has been made.

The project still needs file cold-open and security qualification of the intended distributed bytes, an evidence-bound release review, final judge testing instructions, selected screenshots, and a publicly visible YouTube demonstration under three minutes showing the actual automation.

No autonomous merge, deployment, release, general success rate, quantified sustainability benefit or hosted application is claimed.

Source repository: https://gitlab.com/gitlab-ai-hackathon/transcend-october-2026/27415396/showcase

Built With

Share this project:

Updates