detect-stack

Detect Stack

Analyze the current codebase and generate .agents/stack-context.md. Run this on first use in a project, or when the stack context is missing or stale.

What to Detect

Scan the project root for these signals:

Language — Count file extensions (.rs, .py, .ts, .go, .swift, .java, .rb, .cs, .ex, etc.). Primary = most files. Note secondary languages.

Framework — Read the package manifest:

• Cargo.toml → check [dependencies] for axum, actix, rocket, sqlx, diesel, tokio
• package.json → check for react, next, svelte, express, fastify, prisma
• go.mod → check for gin, echo, fiber, chi, sqlc
• pyproject.toml / requirements.txt → check for django, flask, fastapi, sqlalchemy
• Gemfile → check for rails, sinatra
• Package.swift → check for Vapor; scan for SwiftUI imports
• mix.exs → check for phoenix, ecto

Tests — Look for test directories, test file naming patterns, test config files.

CI gates — Scan .github/workflows/, .gitlab-ci.yml, etc. for blocking checks (format, lint, type check, test).

Conventions — Check for CLAUDE.md, AGENTS.md, CONTRIBUTING.md. Note directory structure and naming patterns.

Output

Write .agents/stack-context.md — keep it short (under 40 lines):

# Stack Context

Generated: YYYY-MM-DD

## Stack
- **Language**: [primary] [version if pinned]
- **Framework**: [web framework], [db layer]
- **Build**: [build tool]
- **Test**: [test command + framework]
- **Lint**: [linter] [CI gate: yes/no]
- **Format**: [formatter] [CI gate: yes/no]

## Secondary Languages
- [language] ([what it's used for])

## Conventions
- Error handling: [pattern]
- Module structure: [pattern]
- Naming: [pattern]
- Tests: [where they live, how they're organized]

## CI Gates
- [list of blocking checks]

Stack-Specific Principle Application

After writing the stack context, do not generate static reference files. Instead:

1. Use your built-in knowledge of the detected language/framework to apply each coding principle idiomatically
2. If you need specific framework API details, use context7 MCP (resolve-library-id then query-docs) to pull live documentation
3. If context7 doesn't cover it, use web search as fallback