AI software quality workbench for agent-generated code — desktop-first, local-first, and focused on finding bugs that normal AI review misses.
CodeVetter should end as a personal verification layer for AI-built software. The durable scope is:
- code review
- bug finding
- agent-written code verification
- debugging and replay
- synthetic user QA for software quality
- AI step-through debugging
- codebase history explanation
The near-term wedge is not beating Claude, Codex, or hosted PR bots at generic review. It is a self-first workflow that makes agent output trustworthy: inspect the diff, understand the repo and prior intent, exercise the changed behavior, preserve evidence, fix one finding at a time, and re-check that the issue is gone.
| Capability | Current state | Main gap |
|---|---|---|
| Code review | Review tab runs local diffs through CLI agents and persists findings. | Needs multi-pass specialist review, better AGENTS.md/project-context ingestion, and benchmarked catch-rate evidence. |
| Bug finding | Findings, severity, code viewer, and re-review loop exist. | Needs runtime evidence from tests/browser sessions/logs, not only static diff judgment. |
| Agent-written code verification | Product is aimed at agent output and can fix/re-review selected findings. | Needs agent provenance: which agent changed what, prompt/task context, and whether the fix actually resolved the original user goal. |
| Debugging/replay | History indexes Claude/Codex sessions and can replay conversations. | Replay is not connected to files, diffs, failures, screenshots, tests, or review findings. |
| Synthetic user QA | Not implemented as a first-class workflow. | Needs browser/app automation that performs user tasks, captures screenshots/traces, and converts failures into review findings. |
| AI step-through debugger | Not implemented. | Needs an execution timeline across agent actions, file edits, commands, test failures, and UI observations. |
| Codebase history explainer | Repo Unpacked generates repo briefs; History indexes agent sessions. | Needs commit/decision mining tied to touched files so reviews can catch intent regressions. |
The product should prefer narrow, evidence-backed loops over broad "code intelligence" surfaces. A feature is on-strategy when it helps answer: "What changed, why did the agent change it, what could break, can we reproduce it, and did the fix actually work?"
| Concern | Service |
|---|---|
| Desktop app | GitHub Releases — Tauri 2 macOS build, with @tauri-apps/plugin-updater auto-updater (latest.json manifest) |
| Landing page | Cloudflare Pages (codevetter, codevetter.com) — static Next.js export |
| Database | Local SQLite via @tauri-apps/plugin-sql (desktop only, no server) |
| Auth | None — LLM provider API keys stored in user settings |
| AI | User-supplied keys (Anthropic / OpenAI / OpenRouter) |
| CI/CD | GitHub Actions — release.yml builds Tauri binaries on GitHub release; deploy-landing.yml deploys the landing page to Cloudflare Pages on push to main |
Give your coding agent this prompt:
Install CodeVetter from the latest GitHub release:
https://github.com/sarthak-fleet/CodeVetter/releases/latest
Detect this machine's OS and CPU architecture, download the matching CodeVetter app archive, verify the release asset hash when available, extract it, install CodeVetter.app into /Applications on macOS, remove the quarantine attribute if needed, and launch the app once to verify it starts.
Prefer the app archive over the DMG until the macOS bundle is Developer ID signed and notarized.
# Clone and install dependencies (uses npm workspaces)
git clone https://github.com/sarthak-fleet/CodeVetter.git
cd CodeVetter
npm installRequires Rust + Tauri prerequisites for the desktop app.
- Install dependencies (see above)
- Launch the desktop app in development mode:
cd apps/desktop && npm run tauri:dev
- Open the Review tab, pick a local repository, and run your first review through an installed CLI agent.
Run the desktop app (dev mode)
cd apps/desktop
npm run tauri:devRun Playwright end-to-end tests for the desktop app
cd apps/desktop
npm testBuild the landing page
cd apps/landing-page
npm run buildapps/
desktop/ Tauri 2 + React 19 + Vite desktop app — the core product
landing-page/ Next.js marketing site (static export, deployed to Cloudflare Pages — codevetter.com)
| Layer | Technologies |
|---|---|
| Desktop frontend | React 19, Vite, Tailwind CSS, shadcn/ui |
| Desktop backend | Rust (Tauri 2), SQLite |
| Review engine | TypeScript — runs in the webview, no server required |
| Landing page | Next.js 15 (static export → Cloudflare Pages) |
| Testing | Playwright (e2e) |
| Package manager | npm workspaces |
ISC (root package); MIT (landing-page template — Copyright 2022 Themesberg)
This section is maintained by the SaaS Maker Active-AI product/design loop so future agents do not reopen duplicate UI tasks.
- Business lane: Core/status context
- Rule: do not create another broad "improve the UI" task unless the acceptance criteria differ materially from the tasks listed here.
- Source of truth for task status: SaaS Maker task board. README entries are durable context only.
| Task ID | Title | Status |
|---|---|---|
| d6d19901 | CodeVetter: add verification summary handoff proof | done — compact verification summary panel added to QuickReview sidebar with fixed/reproduced/unchecked counts and copy-proof button |
| a59acaa7 | CodeVetter: add unchecked finding risk summary | done — QuickReview sidebar now lists unchecked findings grouped by severity with per-bucket risk copy explaining why each unchecked item still matters (above the verification handoff proof) |
| 79eff0b9 | CodeVetter: add revalidation checklist after fixes | done — when a finding's re-check status is "fixed", QuickReview renders a checklist derived from the finding's evidence fields (file/line, artifact, level, notes) so the user can tick off concrete revalidation steps; checklist state persists per finding alongside other evidence |
| 2b9ac8d9 | CodeVetter: add copyable reviewer handoff template | done — QuickReview's "Copy proof" button now emits a full markdown reviewer handoff (heading, score/agent/finding tallies, per-finding evidence with status icons, and a ### Next actions checkbox list derived from unchecked findings, reproduced findings, and unticked revalidation items for fixed findings) so reviewers can paste proof directly into PRs/Slack |