GraphPaper 0.3.1: external Graphify with Codex and a public bundled runtime

[AronAxe]

Completed implementation

• Run the actual, unmodified public Graphify 0.9.80 CLI in a separate process, with GraphPaper routing its bounded text inference through the selected extraction provider, model and reasoning setting.
• Support official Codex / ChatGPT OAuth without another writing API key or API-billing fallback; retain the existing OpenRouter, compatible-API and Anthropic adapters.
• Share request accounting and budgets; remove the duplicate Native extraction pass; preserve file provenance into the graph inspector and bounded angle retrieval without promoting it to proof.
• Validate completed graphs before committing them and cover missing runtimes, malformed/empty output, dangling edges, cancellation, timeout and previous-project compatibility.
• Include pinned public Graphify/SDK dependencies, metadata, tokenizer cache and licences in the Windows distribution. Verify installed files against distribution RECORD hashes: 240 Graphify files, 1,938 OpenAI SDK files and 14 tiktoken files. No private fork or unpublished upstream patch.
• Supply the upstream hash-seed launch contract so a frozen worker does not restart GraphPaper.exe as a general Python interpreter. This packaging-only issue is fixed and regression-tested.
• Keep the native bridge restricted and the application CSP unchanged. Retain voice, project folders, humanizer/deslop and scientific research features. The user's README header image is unchanged.

Actual verification

• 217 local Windows automated tests passed, one Windows symlink-permission skip.
• 44 real-HTTP browser workflow checks passed.
• 10 native-source and 15 actual compiled-executable interaction checks passed, including external Graphify extraction, correct model/effort routing, source inspection, real Save As cancellation and pending-edit persistence on normal close. The compiled close completed in 8.329 seconds.
• Real Codex-backed extraction using the packaged Graphify worker passed: two synthetic evidence documents, explicitly selected gpt-6-luna / low, one model request, five nodes and four edges. Voice input excluded; no extra Native pass or API fallback. The live-provider test uses the GraphPaper API with its packaged worker; native GUI extraction is independently verified with synthetic model transport.
• All source matrix jobs on Windows and Ubuntu, Python 3.12 and 3.13, and the complete hosted Linux browser workflow passed for application commit e1b64107fb7f8db11e9182a6218acd09f86165b4 (run 37672455238). macOS was not directly tested.
• The older UI tests contained async-save/project-refresh races; they now wait for the actual persisted/refreshed state rather than asserting too early. Deadline unit checks no longer depend on sub-clock-resolution sleeps.

Verified Windows release artifact

GraphPaper-v0.3.1-Windows-x64.zip — 187,511,280 bytes.

SHA-256: 1a0690506aaa6a45d0e2d0ebc6967f715d8da96dfdad5fd02443bf6d2a5bbed3.

Versioned evidence and platform-specific reports are under docs/validation/v0.3.1; actual compiled-UI screenshots are included in docs/images. The archive is CRC/checksum verified, its UI files match the tested source, and the root licence is present. No user manuscripts, credentials, databases, build environments or binaries are committed. Existing installed applications were not updated as part of this delivery.

The public release is published only after source merge and a final GitHub-upload checksum verification. Later hosted rebuilds cannot silently overwrite the verified asset.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-07T19:29:32.627069Z	4952218	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.