RepoAssay/repo-test-architect

Audit-first test strategy tooling for codebases.
Repo Test Architect builds a deterministic audit graph before asking any model or agent to reason about tests. The goal is to identify repo-native, high-value test work from facts the tool can inspect locally: project roots, framework signals, existing tests, source classifications, blockers, and remaining risk.
The current implementation can:
- audit JavaScript/TypeScript, PHP, Python, Ruby, and Swift projects through supported adapters
- audit supported, bounded Kotlin/JVM fixtures through the same shared artifact model
- audit conventional Go modules and literal repository-contained go.work members through a supported bounded adapter
- audit conventional Cargo packages and literal repository-contained workspace members through a supported bounded Rust adapter
- audit one conventional SDK-style C# test project or one unique literal production/test project edge, including bounded native xUnit MTP-v2 and MSTest.Sdk v4 ownership, literal and one-hop root-aliased target frameworks, one repository-contained direct *.cs compile include, finite target-conditioned package shapes, bounded nearest-file build and central-package props, exact immutable test-field receivers, stable direct-call results, inline out results, framework exception and collection/string assertions, one-hop test helpers, and guarded well-known System type collisions, through a supported bounded .NET adapter
- audit one conventional Composer/PSR-4 project with PHPUnit through a supported bounded PHP adapter
- audit one conventional Mix application with exact MixProject/Mixfile, bounded source and Mix-task ownership, literal local ExUnit wrapper discovery, and test-body evidence through a supported bounded Elixir adapter
- detect polyglot project roots and report unsupported ecosystems without hiding them
- produce a complete repository analysis, findings, ranking, plan, execution hints, and verification commands in one audit pass
- classify source files by likely test value and defer low-value direct tests
- rank candidates and generate test plans from the audit graph
- derive provider-neutral execution, context, parallel-safety, and repository-reasoning hints without selecting models or spawning subagents
- analyze conservative test placement findings across project boundaries
- collect project-level stats for coverage, candidate counts, frameworks, commands, and adapter usage
- provide disabled-by-default local MCP diagnostics, safe internal-error report IDs, runtime checks, and inspectable sanitized bundles without external reporting
- expose the same deterministic behavior through CLI commands, a local invoke harness, and a stdio MCP SDK server
- lock behavior with golden snapshots, model-consistency scenarios, package checks, and cross-OS CI
Native test generation is intentionally deferred. generate_selected_test returns a structured deferred artifact until adapter-specific generation policy and repair-loop fixtures exist.
Repo Test Architect 1.0.0 is available on npm as latest and in the Official MCP Registry, with stable support for the documented CLI, MCP tools, configuration and versioned artifact contracts. Treat its findings as evidence-backed review input rather than an automatic instruction to change a repository. See the 1.0.0 release notes for changes, compatibility and publication status.
Ten adapters are supported within their documented boundaries. Dart/Flutter remains experimental, and native test generation remains deferred.
Maintenance is provided as time permits, with no guaranteed response time or feature schedule. Focused community PRs are welcome; see Contributing and Support.
Install
Node.js 20 or newer is required.
Run the CLI without a global install:
Pin the stable version:
Or install the CLI and MCP server binaries:
Add the local stdio MCP server to an MCP-capable client:
To keep an MCP client on this exact release, pin its version:
The client launches the server locally. Repository source stays on the machine unless the client or another configured tool sends it elsewhere. Connected models receive instructions to start with analyze_repository for a general repository review. See MCP client config and agent install paths for global-install, local-checkout, and host-specific guidance.
Quick Start
For a human-readable review of the current repository:
analyze detects every project root, runs each supported adapter once, and derives the audit summary, top findings, candidate ranking, test plan, execution hints, project stats, and verification commands. Markdown stays compact; JSON preserves the complete evidence bundle: