All notable changes to this project are documented here. The format is based loosely on Keep a Changelog and the project follows SemVer once it ships to the Chrome Web Store.

• IMPLEMENTATION.md — step-by-step handbook for Phases 2-4. Locked decisions, code patterns, file paths, acceptance criteria, anti-patterns, and commit cadence. Designed so a less-capable model can execute the work without making bad architectural calls.

• Phase 2 default image model: Organika/sdxl-detector at q8. Benchmark deferred to Phase 2.5.
• Phase 2.5: new explicit phase for the local-image-model benchmark, run after Phase 2 ships and after the maintainer curates a 60-image test set.
• Phase 3 default text model: Hello-SimpleAI/chatgpt-detector-roberta at q8. Default text confidence threshold raised to 0.70 (vs 0.55 for images) — biases toward "Uncertain" given known false-positive risk.
• Phase 4 boost-failure behavior: fall back to local model with a visible toast "Boost failed (rate limited) — using local model". Toggleable in Options, default on.
• Phase 4 rate limiting: client-side, 10 RPM per provider by default, configurable in Options.

• ARCHITECTURE.md — added explicit default-models table, quantization choice (q8), backend strategy (device: 'auto'), and boost-failure / rate-limit behavior.
• ROADMAP.md — Phase 2 cross-references the handbook; new Phase 2.5 section; Phase 3 + 4 reflect the locked defaults.
• AGENTS.md — points cheaper agents at IMPLEMENTATION.md; added the new locked decisions; updated repo layout for the post-Phase-1 reality.

• New Netlify-ready static site under site/ with:
  • site/index.html (product marketing page),
  • site/docs/index.html (documentation hub),
  • site/docs/architecture.html (architecture overview),
  • custom SVG graphics under site/assets/.
• New root netlify.toml configured to publish site/ with security headers.

• Updated README.md with website and docs URLs plus current extension build command (npm run build).

• ARCHITECTURE.md — source of truth for v1 system design.
• ROADMAP.md — phase-by-phase plan to Chrome Web Store launch.
• AGENTS.md — persistent guidance for AI agents working on this repo.
• This CHANGELOG.md.

• Rewrote README.md to honestly describe project state; dropped the unsubstantiated "CLIP zero-shot via Vision Transformers" claim.
• Rewrote CONTRIBUTING.md for the upcoming TypeScript/Vite stack.
• Moved icons from repo root to assets/icons/.

• All v0.1 Python "detectors" (simple_ai_detector.py, huggingface_detector.py, google_ai_detector.py, ml_model_example.py, train_model.py, deploy_model.py, test_api.py, setup_api_key.sh, requirements*.txt, .python-version) — none of these were functional AI detectors; the wired-in one was a hash-of-bytes heuristic. See archive/legacy-python/README.md.
• All v0.1 documentation (docs/, sphinx-docs/, DOCS.md, SETUP_GUIDE.md, netlify.toml) — described a CLIP system that did not exist in code. See archive/legacy-docs/README.md.
• v0.1 extension files (manifest.json, background.js, content-script.js, content-style.css, popup.{html,js}, options.{html,js}) — kept as a reference for the v1 rewrite. See archive/legacy-extension-v0.1/README.md.

• Vite 8 + CRXJS 2.4 + React 19 + TypeScript 6 + Tailwind 4 build system.
• Typed manifest.config.ts — dropped host_permissions: <all_urls>, kept activeTab + scripting + storage + offscreen + contextMenus.
• src/background/index.ts — service worker with typed message routing, offscreen lifecycle, context menu setup, keyboard shortcut handler.
• src/offscreen/ — offscreen document stub (Phase 2 will add ONNX session).
• src/content/index.ts + badges.ts — visible-image discovery, scroll/resize-aware badges, image-to-dataURL via canvas, text-selection toast, timeout handling.
• src/popup/App.tsx — "Scan visible images" + "Scan selected text" buttons, mock-mode indicator, settings link.
• src/options/App.tsx — provider radio selector (local / Gemini / OpenAI / HF), API key input, confidence threshold slider, save button.
• src/shared/types.ts — discriminated-union message bus, ClassificationResult, ExtensionSettings.
• src/shared/storage.ts — typed chrome.storage.sync wrapper.
• src/shared/classify-mock.ts — deterministic mock classifier for development.
• src/detectors/ — router stub + Phase 2/3 placeholder files.
• tests/unit/classify-mock.test.ts — 5 unit tests, all passing.
• .github/workflows/ci.yml — GitHub Actions: typecheck + lint + test + build on push/PR.
• vitest.config.ts, eslint.config.js, tsconfig.json — toolchain configuration.

Extension builds to dist/, loads in Chrome via developer mode, and shows "MOCK" badges on visible images when "Scan visible images" is clicked. All CI checks pass (0 errors, 0 warnings). Ready for Phase 2 (real local image inference).

Original public release. Plain-JS MV3 extension + localhost Flask server claiming CLIP-based AI detection. The on-page UX was solid; the "ML" was a hash-based placeholder. Preserved verbatim under archive/legacy-extension-v0.1/ and archive/legacy-python/.