Proofweave

Proofweave is a local-first workspace for shaping deep-research questions, reading cited reports, and keeping a durable personal library. It uses only Google Deep Research; it does not fall back to a general model or grounded search request.

Run locally

Requirements: Node.js 20.9 or newer and npm.

git clone https://github.com/ASHR12/proofweave.git
cd proofweave
cp .env.example .env.local
npm install
npm run dev

Add your GEMINI_API_KEY to .env.local, then open http://localhost:3000. The key is read only by Node route handlers. .env.local and all other real environment files are ignored; no API keys are committed.

Run the provider-free verification suite with:

npm run lint
npm run typecheck
npm test
npm run build

These commands do not contact Google. A paid Deep Research call occurs only after explicit confirmation in the New Research screen.

Custom report exports

Open a saved report and choose Custom export to shape a reusable text or Markdown template. The editor previews the current report as you type, keeps the latest template in the browser, and can copy the rendered text or download it as a Markdown file.

Template values use double braces. Available values include title, question, provider, agent, citationCount, citationLabel, duration, completionDate, and reportMarkdown. The complete report is included in the default template.

Architecture

src/providers/google-research.ts is the sole provider boundary. It starts Deep Research agent deep-research-preview-04-2026 through Google's Interactions API, persists the interaction before acknowledging submission, and then resumes, streams, or polls that same interaction. Provider credentials never reach the browser.

Proofweave keeps the provider's cited Markdown and URL annotations as the authoritative report. A deterministic parser derives optional summaries, finding cards, sections, and comparison rows without fabricating missing structure.

Persistence and recovery

• Saved reports live in data/reports/*.json.
• Active and historical runs live in data/runs/*.json.
• Writes use a temporary file followed by an atomic rename.
• On restart, persisted active runs are reconciled with Google and resumed from their last event when possible. Recovery never creates a replacement paid interaction.
• Runtime report and run records remain ignored by Git, apart from directory placeholders and the public sample below.

This single-user workspace supports one Node process against its data directories. Recovery also depends on Google's interaction-retention window.

Included sample

data/reports/sample-glm-5-3-vs-kimi-k3.json is a genuine prior report comparing GLM 5.3 and Kimi K3. It was produced by Deep Research agent deep-research-preview-04-2026, not relabeled or regenerated.

The public sample preserves the report text, structured sections, and public citation metadata. Its Google interaction ID, resume event ID, local run ID, provider runtime status, and progress log were removed or replaced with a stable sample ID. The original private runtime record remains outside version control. A fresh clone loads the sanitized sample automatically; future runtime reports remain ignored.

Local API

• GET /api/reports and POST /api/reports
• GET /api/reports/:id and DELETE /api/reports/:id
• GET /api/research and POST /api/research
• GET /api/research/:id
• POST /api/research/:id/cancel

For optional browser screenshots, start the app and run npm run capture:screenshots. Generated artifacts and downloaded exports are ignored.