Onboard

Onboard Skill

Automated data collection + LLM synthesis + interactive guidance.

Architecture

/onboard
  │
  ├─ Phase 1: collector.js (pure JS, zero LLM)
  │   ├─ scanManifest()    → package.json/Cargo.toml/go.mod
  │   ├─ scanStructure()   → directory tree with file counts
  │   ├─ readFileIfExists()→ README.md, CLAUDE.md
  │   ├─ scanCI()          → .github/workflows, Dockerfile
  │   ├─ getGitInfo()      → branch, commit count, remote
  │   ├─ getRepoIntel()    → onboard + hotspots + conventions +
  │   │                       projectInfo + entry-points + slop
  │   └─ getRepoMap()      → symbols, imports, exports (if available)
  │
  ├─ Phase 2: onboard-agent (Sonnet)
  │   ├─ Synthesize collected data into readable summary
  │   ├─ Read key source files for architecture understanding
  │   └─ Present summary to user
  │
  └─ Phase 3: Interactive (conversational)
      ├─ "What do you want to do?"
      ├─ Guide to files using coupling + ownership + symbols
      └─ Answer follow-up questions with full context

Depth Levels

Data Flow

The collector produces a single JSON object:

{
  "manifest": { "type": "npm", "name": "...", "language": "typescript", "scripts": [...], "dependencies": {...} },
  "readme": "# Project\n...",
  "claudeMd": "# Rules\n...",
  "structure": [{ "path": "src/", "depth": 1, "files": 23, "dirs": 5 }, ...],
  "ci": { "github": true, "workflows": ["ci.yml", "release.yml"], "dockerfile": false },
  "gitInfo": { "branch": "main", "commitCount": 232, "lastCommit": "2026-03-15", "remoteUrl": "..." },
  "repoIntel": {
    "onboard": { "language": "typescript", "structure": "single package", "health": "active", ... },
    "hotspots": [...],
    "conventions": { "commit": "conventional", ... },
    "projectInfo": { "readme": {...}, "manifest": {...} },
    "entryPoints": [{ "path": "src/main.rs", "name": "main", "kind": "main" }, ...],
    "slop": {
      "orphanExports": [...], "passthroughWrappers": [...],
      "alwaysTrueConditions": [...], "commentedOutCode": [...],
      "counts": { "orphanExports": 3, "passthroughWrappers": 0, "alwaysTrueConditions": 0, "commentedOutCode": 1 }
    }
  },
  "repoMap": { "totalFiles": 45, "totalSymbols": 312, "keyExports": { "src/index.ts": ["Queue", "Worker", ...] } }
}

This is passed as the agent's prompt context. The agent reads key files to fill gaps, then presents a synthesized summary. No data collection happens in the LLM - only synthesis and guidance.

Repo-Intel Data

Expected: the orchestrator (the command that spawned this agent) has already checked <stateDir>/repo-intel.json and either pre-fetched the data into your context or skipped (user declined to generate). Do not call AskUserQuestion here - subagents cannot interact with the user.

If the pre-fetched data is empty, proceed with the available context. The orchestrator has already made the decision on the user's behalf.

Binary: agent-analyzer auto-downloads to ~/.agent-sh/bin/ from agent-sh/agent-analyzer GitHub releases (~10 MB) on first use. The lib/agentsys resolver locates the agentsys install (CC marketplace clone, npm global, or sibling repo).