Detect Stack

detect-stack — detect the repo's stack and install the matching skills

Read the repository, decide which technology stacks it uses, then install the per-stack skills + agents tagged for exactly those stacks. Core SDLC skills (anything with no stack: tag) install regardless. This is a one-time (re-runnable) bootstrap.

Inputs

Run from the root of the target repo (or a codebase/-of-repos umbrella). No artifact inputs — you discover everything from the files on disk. If your instruction names harness targets (claude-code, cursor, codex), use them; otherwise infer them (below).

Stack vocabulary

Emit only these tokens (they match the stack: tags on the skills/agents):

go · java · kotlin · python · react · vue · angular · node · rust · flutter · android · db

Detection method

1. Scan for manifests and sources (breadth-first; ignore node_modules, vendor, .git, build output). Map signals to tokens — a repo can match several:

   Signal (files / dependencies)	Stack token
   go.mod, *.go	go
   pom.xml, build.gradle(.kts), *.java (Spring Boot / Quarkus deps still ⇒ java)	java
   *.kt, *.kts, Kotlin Gradle plugin	kotlin
   requirements.txt, pyproject.toml, setup.py, Pipfile, *.py (Django/FastAPI deps ⇒ still python)	python
   package.json with react/next	react
   package.json with vue/nuxt	vue
   package.json with @angular/core	angular
   package.json with @nestjs/*, or Node/TS backend without a UI framework	node
   Cargo.toml	rust
   pubspec.yaml	flutter
   AndroidManifest.xml, com.android.* Gradle plugin	android
   SQL migrations dir, or postgres/mysql/redis/prisma/clickhouse deps	db

2. Prefer dependency evidence over file extensions when they disagree (a .ts file in a Go repo's tooling dir is not a node stack). Only emit a token you can point to a concrete signal for; list that evidence.

3. De-duplicate and sort the tokens.

Install step

1. Pick harness targets. Include every already-installed Maestro host you can evidence: .claude/skills/maestro/SKILL.md ⇒ claude-code, .cursor/skills/maestro/SKILL.md ⇒ cursor, and .agents/skills/maestro/SKILL.md ⇒ codex. If none is present, default to claude-code. Do not treat a generic .agents/ folder as proof that Maestro was installed for Codex.
2. Run the installer with the detected filter. Prefer the repo-local wrapper; fall back to the installer script:

   ./maestro install <harness-targets> --stack <detected-tokens>
   # or, if there is no ./maestro wrapper:
   bash install.sh <harness-targets> --stack <detected-tokens>
   

   e.g. ./maestro install codex --stack go,react,db. The installer always adds the core SDLC skills; --stack only gates the per-stack ones.
3. Report what installed (the installer prints N installed, M skipped by --stack).

Standards

• Detect what the repo actually uses today — never install a stack you can't evidence.
• Re-runnable: re-running re-detects and re-installs; it overwrites the installed copies, never the skills/ source of truth.
• If you detect a stack the vocabulary doesn't cover, say so plainly and skip it (don't invent a token — the installer would match nothing).

Safety

• Read-only against application code. The only thing you run is the installer, which writes ONLY into harness config dirs (.claude/, .cursor/, .agents/) — never application code, never .maestro/ run state.
• If no stack is detected, install just the core pack (--stack with no tokens installs everything; to install core-only, pass a token that matches nothing is wrong — instead report "no stack detected" and let the human choose).

Output contract

Return stacks (the detected stack tokens, comma-separated), installed_count (skills + agents installed), and summary (one line naming the stacks and harness targets).