Skip to content

PRODUCT.md

Latest commit

 

History

History
90 lines (67 loc) · 3.67 KB

PRODUCT.md

File metadata and controls

90 lines (67 loc) · 3.67 KB
Raw

Product

Platform

macOS desktop application (Tauri 2 with a React webview), with CLI/MCP and machine-readable local artifacts as primary verification surfaces.

Users

CodeVetter serves developers and technical product owners who use coding agents against real repositories and need to understand whether the resulting change is correct. They work locally, often in unfamiliar or large codebases, and need both a human-readable view of the system and compact evidence that another agent can query.

Product Purpose

CodeVetter determines whether a coding agent completed a software task correctly using reproducible execution evidence. Success means the user can move from a task and an agent-authored change to executable verification, inspectable evidence, and a measurable verdict without trusting another model's opinion alone.

Positioning

CodeVetter is an execution-backed verification and evaluation system rather than a generic code reviewer or agent control center. Its durable mechanism is the connection between task intent, source-backed repository context, executable checks, captured evidence, and measurable outcomes.

Operating Context

The primary workflow is local and repository-scoped:

task → agent change → executable verification → evidence → measurable verdict

The CLI/MCP and machine-readable verification bundle are primary product surfaces. The Tauri desktop application is a local viewer over the same evidence, graph, history, review, and testing systems. Users bring their own Codex or Claude CLI/provider access and keep project data in a local SQLite database.

Capabilities and Constraints

  • The active core is TypeScript/Node web-task evaluation with browser and API behavior, deterministic graders, benchmark cases, failure taxonomy, and reliability/cost/latency measurement.
  • The canonical structural graph is local, syntax-aware, source-backed, trust-qualified, and queryable by both humans and agents.
  • Structural topology is navigation evidence, not runtime proof and not an independent source of findings.
  • CodeVetter has no hosted review server, multi-tenant collaboration layer, or automatic authority to publish local evidence.
  • Desktop, CLI, MCP, JSON, Markdown, and offline HTML outputs must preserve the same evidence identities and limitations.

Brand Commitments

The product name is CodeVetter. Product language is direct, technically honest, and explicit about what is verified, inferred, stale, partial, blocked, or unsupported. It avoids theatrical agent metaphors on verification surfaces and does not present correlations, graph paths, model judgments, or fixture results as executable proof.

Evidence on Hand

  • The committed benchmark cases, graph qualifications, warm-verification receipts, differential-verification receipts, and release artifacts are the available evidence.
  • Synthetic or sample fixtures demonstrate contracts only; they must not be presented as proof of real agent improvement.
  • No real paired evidence currently establishes that structural context improves coding-agent task outcomes.

Product Principles

  1. Executable outcomes outrank model opinion.
  2. Every useful claim carries source identity, scope, and limitations.
  3. Human views and agent tools share one canonical evidence model.
  4. Local, bounded, reversible workflows are the default.
  5. New capability earns investment through measured reliability, not feature accumulation.

Accessibility & Inclusion

Human-facing evidence must remain understandable without color alone, usable by keyboard, responsive at narrow and wide widths, and available through a non-visual machine-readable representation.