Agent Dx

agent-dx — Agent Developer Experience Audit

Analyze a repo and generate a prioritized report of verification tools coding agents need to self-check their work. Report-only — recommends tools and commands, does not auto-install.

How It Works

1. Detect project language, framework, and runtime
2. Scan for existing verification tools
3. Generate a Markdown report with prioritized recommendations
4. Include copy-paste install commands and agent-friendly run commands

Step 1: Detect Project Context

Read these files to understand the project:

package.json                    # JS/TS: deps, devDeps, scripts, workspaces
tsconfig.json                   # TypeScript config
bun.lockb / bun.lock            # Bun runtime
pnpm-lock.yaml                  # pnpm
yarn.lock                       # Yarn
package-lock.json               # npm
pyproject.toml                  # Python: deps, tools
requirements.txt                # Python: deps
setup.py / setup.cfg            # Python: legacy
Cargo.toml                      # Rust
go.mod                          # Go
.github/workflows/*.yml         # CI configuration
CLAUDE.md                       # Project-level agent context
~/.claude/CLAUDE.md             # Global agent context

Determine language:

• package.json exists → JS/TS
• pyproject.toml or requirements.txt or setup.py → Python
• Cargo.toml → Rust
• go.mod → Go
• No manifest → scan file extensions as fallback

Determine package manager / runtime (JS/TS only):

• bun.lockb or bun.lock exists → already using Bun
• pnpm-lock.yaml exists → already using pnpm
• yarn.lock exists → already using Yarn
• package-lock.json exists → already using npm
• None of the above → recommend Bun (see references/tool-matrix.md for rationale)

Determine framework (from dependencies):

• electron → Electron (dual runtime: Node + Chromium)
• react → React
• next → Next.js
• vue → Vue
• svelte → Svelte
• express / fastify / koa → Node API server
• django / flask / fastapi → Python web framework

Determine monorepo:

• workspaces field in package.json
• pnpm-workspace.yaml exists
• lerna.json exists
• nx.json exists
• turbo.json exists

If monorepo detected, analyze root config + each workspace independently.

Step 2: Scan Existing Tools

Use references/tool-matrix.md for the full lookup table of tools per language.

For each category in the tool matrix:

1. Check devDependencies in package.json (JS/TS) or pyproject.toml [tool.*] sections (Python)
2. Check for config files (e.g., vitest.config.*, tsconfig.json, biome.json)
3. Check package.json scripts for: test, lint, typecheck, e2e, format, check
4. Scan .github/workflows/*.yml for CI steps — if a tool already runs in CI, mark it as "CI-covered" not "Missing". Only recommend tools not already covered by CI.

Step 2b: Audit CLAUDE.md Files

Check for agent context files at two levels:

Project-level: CLAUDE.md in the repo root (or .claude/CLAUDE.md) Global-level: ~/.claude/CLAUDE.md

Project CLAUDE.md

Check if it exists. If it does, scan for these sections (doesn't need to match exactly — just check if the topics are covered):

If project CLAUDE.md is missing, recommend creating one. If it exists but is thin (missing 3+ of the 8 topics above, or any topic covered in fewer than 50 words), note which topics are missing.

Recommendation template for project CLAUDE.md:

# [Project Name]

[One-line description of what this project does]

## Tech Stack
- Runtime: [Bun / Node / Python / etc.]
- Framework: [React / Next / Django / etc.]
- Database: [Postgres / SQLite / etc.]

## Development
- `bun install` — install dependencies
- `bun dev` — start dev server
- `bun test` — run tests
- `bun run lint` — lint code
- `bun run typecheck` — type check

## Code Conventions
- [Key patterns, naming conventions, directory layout rules]

## Things to Avoid
- [Known pitfalls, deprecated patterns, etc.]

Global CLAUDE.md

Check if ~/.claude/CLAUDE.md exists. This gives agents personal context about the developer that applies across all projects.

If missing, recommend creating one with:

# About Me
- Name: [your name]
- Role: [what you do — e.g., "Full-stack engineer", "Founder at X"]

## Preferences
- Preferred runtime: [Bun / Node / etc.]
- Preferred package manager: [bun / pnpm / npm / etc.]
- Code style: [terse vs verbose, functional vs OOP, etc.]
- Communication style: [brief vs detailed responses]

## Current Projects
- [Project 1]: [one-line context]
- [Project 2]: [one-line context]

## General Instructions
- [Anything agents should always do or never do across all your projects]

The global file is especially useful for things like: "I use Bun everywhere", "I prefer terse code", "Always use TypeScript strict mode" — context that agents otherwise have to ask about every session.

Step 2c: Verify Detection

Spot-check findings before reporting. For each tool marked "Installed", confirm it actually runs:

• JS/TS: bunx vitest --version, bunx biome --version, tsc --version
• Python: pytest --version, ruff --version, pyright --version
• Rust: cargo clippy --version
• Go: golangci-lint --version

If a command fails, downgrade that tool's status from "Installed" to "Not found (listed in deps but not runnable)".

Step 3: Generate Report

Use the tool-matrix reference (references/tool-matrix.md) for recommendations per stack. Use the agent-commands reference (references/agent-commands.md) for agent-optimized commands.

Report Template

# agent-dx Report

## Agent Summary
```yaml
missing_critical: [list of missing critical-tier tools]
missing_important: [list of missing important-tier tools]
existing_tools: {category: tool, ...}
readiness_score: X/12

Project Context

• Language: [detected language]
• Framework: [detected framework(s)]
• Runtime: [Node version / Python version if detectable]
• Monorepo: [Yes (tool) / No]
• Package Manager: [npm / yarn / pnpm / pip / poetry]

Agent Context Files

Existing Tools

Agent Readiness Score

[X/12 categories covered — includes CLAUDE.md files and runtime]

Recommendations

Critical (agents can't self-verify without these)

[Only list categories marked Missing from the table above]

[N]. [Category] — [Recommended tool] Install:

[install command]

Add to scripts:

[script entries if applicable]

Agent verification command:

[command with JSON output flag]

Important (significantly improves agent effectiveness)

[Same format, for Important-tier missing tools]

Nice-to-Have (for thorough verification)

[Same format, for Nice-to-Have-tier missing tools]

Agent Cheat Sheet

Quick reference of all verification commands for this project:

[list all run commands, both existing and newly recommended]

### Priority Tiers

**Critical** (agents can't operate effectively without these):
1. Project CLAUDE.md — agent needs to know how to run/test/lint the project
2. Unit/integration testing
3. Linting
4. Type checking

**Important** (significantly improves agent effectiveness):
5. Runtime / package manager (recommend Bun for new JS/TS projects)
6. E2E / browser testing (only if frontend/Electron detected)
7. Structured logging
8. GitHub CLI for CI checks
9. Global ~/.claude/CLAUDE.md — agent context about the developer

**Nice-to-Have** (for thorough verification):
10. Security scanning
11. Static analysis (only for larger/security-sensitive codebases)
12. Test coverage
13. Visual regression (only if UI components/Storybook detected)

## Tool Conflict Strategy

When an existing tool overlaps with what you'd recommend:
- **Never suggest replacing** an existing tool — respect the team's choices
- **Note faster alternatives** as an FYI footnote only (e.g., "FYI: Vitest is 10-20x faster than Jest for future consideration")
- **Fill gaps only** — if ESLint exists but no type checker, recommend tsc, don't push Biome
- **Complement, don't compete** — recommend tools from different categories, not same-category alternatives

## Edge Cases

| Scenario | What to do |
|----------|-----------|
| No manifest file (no package.json, no pyproject.toml) | Scan file extensions, give generic recs, note "add a package manager first" |
| Empty repo | Report: "No source code detected. Set up your project first." |
| Monorepo | Report per workspace + root-level summary. Note monorepo orchestrator if missing (Turborepo/Nx). |
| Electron app | Flag dual-runtime. Recommend agent-browser via CDP for Electron window + standard testing for Node. |
| Python project | Switch entirely to Python tool matrix (pytest, Ruff, Pyright, structlog). |
| Rust/Go project | Give basic recs (cargo test/clippy or go test/golangci-lint). Note: full support is JS/TS + Python focused. |
| All tools already installed | Congratulatory report. Check for missing agent-friendly flags (JSON output). Suggest agent cheat sheet. |

## Output

Print the full Markdown report directly to the user. Do not write it to a file unless asked.