Writing Diagnostic 1.2.0

Writing Diagnostic reviews whether prose has earned what it says. The packaged skill reasons about 18 writing primitives; the MCP server validates those supplied judgments and presents an interactive diagnostic. It does not independently analyze writing, retrieve documents, verify facts, or publish edits.

Findings distinguish violation, pressure test, and pass. The repair order is think first → keep test → wording options. Selecting a marked phrase or whole-passage check opens its reasoning and repair bank. Filters work by verdict, primitive group, and individual primitive. Keep it and repair-direction choices remain local until the writer chooses Revise with these decisions, which sends one follow-up message to a compatible host conversation. The widget still does not edit the passage itself.

Status

This ZIP is a corrected local development package. Public submission remains incomplete: it needs a deployed HTTPS endpoint and widget origin, verified publisher and country choices, actual support/privacy/terms pages, a real installed-host recording, and host/portal verification. See submission/REMAINING.md and submission/verification.md for the exact status. Missing values have not been replaced with fictional URLs or attestations.

Run and check locally

Requires Node.js 22 or later. ZIP packaging and its integration tests also require Python 3.9 or later. There are no npm or Python dependencies to install.

In the Writing_Assistant repository, this product lives at products/writing-diagnostic/. From the repository root, npm run diagnostic:dev serves its sample on port 8790, and npm run diagnostic:check checks the product. The standalone commands below use port 8787 by default.

npm test
npm run check
npm start

To create an independent plugin ZIP, run npm run package from this directory, or npm run diagnostic:package from the repository root. The result is dist/writing-diagnostic-plugin-v1.2.0.zip; it contains only this product under one writing-diagnostic/ directory, including the hidden compatibility manifests. Generated ZIPs, secrets, dependencies, and build caches are excluded.

The HTTP server binds to 127.0.0.1:8787 by default. Open http://127.0.0.1:8787/preview.html to inspect the prepared sample through the real local renderer and a simulated Apps host. This sample is not an independent analysis service or a substitute for testing the installed plugin in ChatGPT or Codex.

npm run check validates the package and reports remaining public requirements separately. A passing package check does not mean the plugin is ready to submit.

The root plugin.json and mcp.json describe the portable plugin. .codex-plugin/plugin.json and .mcp.json provide the compatibility layout. Both MCP configurations use the bundled stdio server for local installation. To run stdio directly:

npm run stdio

Send newline-delimited JSON-RPC requests. The server supports legacy MCP initialization and current stateless discovery. All tool output containing structured data has a declared output schema.

Tools and UI

findings is required, even when it is an empty array. Every finding needs a unique ID across both finding arrays. Span quotes must match the original passage exactly; repeated quotes use the occurrence field. Spans cannot overlap. Whole-passage findings belong in whole_passage_checks and omit quote, occurrence, and scope. The renderer derives scope and character offsets without rewriting the passage. See public/sample-diagnostic.json for a complete input example.

The UI resource is ui://writing-diagnostic/v1.1.0.html. It includes a narrow CSP, the standard Apps bridge, and the direct window.openai.toolOutput compatibility bridge. Text is rendered through DOM text nodes rather than interpreted as HTML. Validation errors produce recovery guidance instead of a false clean diagnostic.

Deploy when authorized

The package includes a Node HTTP server, a Dockerfile, and Vercel API handlers/configuration. No deployment was completed for this archive. Provider builds, a public endpoint, and real host behavior still need verification.

Routes include POST /mcp, GET /health, and GET /.well-known/openai-apps-challenge. The local server also serves the product page, preview, sample input, and widget. Vercel serves those files from public/ and routes the API paths through api/.

After deployment, copy submission/public-settings.example.json to a settings file and supply real values for every field. Use the helper below to check URL availability and the deployed tool listing, then update the manifest and MCP configurations:

node scripts/prepare-public.js /absolute/path/to/completed-settings.json

The helper cannot verify legal coverage, recording playback, publisher identity, or installed-host behavior. Complete those steps before rebuilding the public submission ZIP. submission/review-cases.json contains five positive and three negative cases; submission/demo-walkthrough.md describes the recording to make. submission/policy-and-support-inputs.md separates implementation facts from decisions the publisher still needs to make.

Source layout