Zero Defect 1.8.0

Added

• The review page renders an HTML deliverable as the page itself, with its own styles, inside an isolated view. Scripts, frames, forms, links, and external resources are removed and nothing is fetched. Findings are placed by finding their visible text in the rendered page, with the anchor line choosing among repeats. A Rendered / Source toggle shows the markup line by line.

• Local images in Markdown and HTML deliverables are embedded by the fill script and shown; web images show as labelled placeholders instead of broken icons.

Changed

• Review page redesign after screenshot review and critiques from GPT-6 Astra and a UX-laws review: a Review sidebar of cards with Accept, Keep original, and See options (each option with its own Use this) replaces the popover; quiet severity underlines in the document with only the selected finding shown as a redline; a sticky footer with Copy decisions, the undecided count, and Undo; a live readiness line; Needs information for accepted changes that still contain [NEED: ...]; bulk acceptance limited to Style rule and Should fix with a confirmation; larger touch targets and a phone bottom sheet that keeps the decision buttons in view.
• The page no longer carries its own Content-Security-Policy tag, which published hosts moved outside <head> and ignored with a console warning; the fill script's sanitizing and the host's own policy apply.
• fill-review-form.mjs refuses binary deliverables (PDF, Word, PowerPoint, images) with a message to supply a Markdown, HTML, or text rendition, instead of embedding unreadable bytes.

Install: Claude Team organizations upload zero-defect-1.8.0.zip (Organization settings > Plugins). Marketplace installs pick up v1.8.0 from the tag. Codex: codex plugin marketplace upgrade, then reinstall zero-defect.

Zero Defect 1.7.0

Added

• Supporting material: name the notes, fact sheet, pricing sheet, or prior agreement a deliverable was written from, and every lens checks the deliverable against it. The Claude skill passes those paths to each reviewer. The Codex collector takes an optional supportingPaths array (or --supporting-path), shows it to every reviewer and the adjudicator as evidence, and keeps it out of the style census.
• A precision section in the review contract: four checks before reporting, and boundary examples for necessary contrasts, ordinary follow-ups, email greetings and sign-offs, closing courtesies by document type, correct arithmetic, accurate gap reports, unsupported assurances, and first-party facts.
• A new review page: your document with every finding in place as a tracked change, and a to-do list on the right sorted by severity. Clicking an item jumps to it and opens its decision panel; each option shows its redline, what it gains, and what it Might lose. Decided findings go back to plain text with a check. Changes, Original, and Final views; Copy decisions for agent.
• The page is a fixed template (review-form.html) filled by scripts/fill-review-form.mjs from review data that follows review-form.schema.json. The model writes only the data; the script validates it, reads the deliverable, and builds the page. No model writes HTML. Every host publishes the page from a file path.

Changed

• A claim that could not be checked is no longer a defect by itself. The author's own prices, terms, results, product capabilities, and metrics are not flagged only because no evidence was supplied. Unchecked external claims the audience would rely on are grouped into one Should fix finding. Must fix requires a demonstrated problem.
• When supporting material exists, any difference that changes a claim counts: wording drift is Should fix; a changed number or an unauthorized promise is Must fix.
• Ordinary conditional follow-ups are no longer commitment findings. Greetings, sign-offs, and signatures are never slop findings; a closing courtesy is fine in an email and Should fix in a formal document.
• Repairs target the smallest span that fixes the defect, so repairs for several findings on one line stay compatible.
• Codex reviewers and the adjudicator run on GPT-6 Astra by default. ZERO_DEFECT_CODEX_MODEL overrides it; an empty value uses the Codex default.

Install: Claude Team organizations upload zero-defect-1.7.0.zip (Organization settings > Plugins). Marketplace installs pick up v1.7.0 from the tag. Codex: codex plugin marketplace upgrade then reinstall zero-defect.

Zero Defect 1.5.0

MECE is now the eighth lens within the existing Zero Defect review. Proposed responsibilities receive an observed responsibility-to-role matrix showing accountable owners, contributors, overlaps, gaps, and unclear handoffs. Findings use the normal ZD identifiers and merge with other lenses. Documents without role content omit the matrix.

Upload zero-defect-1.5.0.zip to your Claude organization plugin settings, or install from the repository's Claude marketplace. The archive contains the plugin files at its root.

Validation: package layout and manifest verified; the shared prompts were exercised by seven successful live Codex acceptance cases. A new native Cowork live run was not performed for this release.

See the changelog and evals/codex/verification.md for details. The SHA-256 checksum is attached.

Zero Defect for Codex 0.2.0

MECE is now the eighth lens within the existing Zero Defect skill. The normal invocation remains $zero-defect:zero-defect.

• Applicable documents receive an observed responsibility-to-role matrix. Shared contributors are allowed; duplicate accountability, missing owners, and unclear handoffs become normal ZD findings.
• Unrelated documents omit the MECE section. Unknown scope remains explicit.
• Report validation rejects contradictory verdicts, incorrect style counts, duplicated IDs, and missing observed matrix rows while preserving diagnostic evidence.
• Canonical style records remain supported, and early reviewer exits no longer crash the collector.
• Inputs must be readable UTF-8 text, such as Markdown, HTML, CSV, or TSV. Native PDF and Office documents require an accessible text export; unsupported or empty inputs are rejected before dispatch.

Validation: 94 automated tests passed on Node 22.22.0 and Node 26.7.0; six live corpus reviews and a fresh installed-skill invocation passed. The installed-skill test completed all eight lenses and correctly reviewed mixed Markdown and CSV inputs. See evals/codex/verification.md for the evidence summary and boundaries.

Install from the GitHub marketplace:

codex plugin marketplace add mattlgroff/zero-defect --ref master
codex plugin add zero-defect@zero-defect-codex

Start a new Codex task after installing or updating. The full Codex package archive and its SHA-256 checksum are attached. Claude is independently released as 1.5.0.

Zero Defect for Codex 0.1.9

First production-ready Codex distribution of Zero Defect.\n\n- Adds a repository marketplace and Codex plugin manifest.\n- Reuses the canonical Claude 1.4.0 workflow, contract, and seven reviewer prompts without modifying or duplicating them.\n- Runs seven concurrent, ephemeral, read-only Codex reviewer processes and a separate adjudicator.\n- Fails closed when any reviewer cannot complete.\n- Requires Node.js 22 or newer and an installed, authenticated Codex CLI.\n\nClaude remains independently published as v1.4.0 and is byte-for-byte unchanged.

Zero Defect 1.4.0

Adds unique ZD-### identifiers to every published issue after adjudication and cross-lens merging. Identifiers remain stable during follow-up on the current report, and the README documents precise fix, leave, and explain replies.

Zero Defect 1.3.0

Adds contextual detection for mannered prose that replaces direct meaning with metaphor or polish, while preserving literal and useful figurative language. Includes focused positive and false-positive acceptance cases.

Zero Defect 1.2.0

Zero Defect 1.2.0 expands the AI-writing slop lens with contextual detection for Claude-associated rhetorical patterns.

Highlights:

• detects cataphoric teasers such as “Here’s the part that nobody tells you,” “Here’s what most people get wrong,” and “Here’s where it gets interesting”
• detects dramatic fragments, importance signaling, canned validation and pushback, structural metaphor clusters, and aphoristic endings
• preserves contextual controls so phrase matches are candidates, not automatic findings or claims of AI authorship
• adds focused acceptance cases for detection, grouping, repair direction, and false-positive controls

The attached ZIP is rooted at the individual plugin directory for direct Cowork organization upload.

Licensed under Apache-2.0 with the adapted anti-slop taxonomy under CC-BY-SA-4.0.

Zero Defect 1.0.0

Zero Defect is a prompt-only Claude Cowork and Claude Code plugin for adversarial review of English-language business deliverables.

It runs seven focused reviewers in parallel across commitments, evidence, numbers, consistency, decision completeness, language precision, and AI-writing slop.

The attached ZIP is rooted at the individual plugin directory for direct Cowork organization upload.

Licensed under Apache-2.0 with the adapted anti-slop taxonomy under CC-BY-SA-4.0.