claudehut-init

name	claudehut-init
description	Use once per project before starting work (or when ClaudeHut reports no codebase index) to bootstrap ClaudeHut for a Java/Spring repository - detects the stack, generates the project memory + index + path-scoped rules, and wires the always-load @import slice. Invoked as /claudehut:init. Idempotent.
allowed-tools	Read Write Grep Glob Bash

ClaudeHut Init (Bootstrap prerequisite)

Bootstrap is a deterministic script, not a hand-generation task. Run it; it writes the canonical project plane + stack-gated rules + the @import slice with zero guesswork. Then optionally enrich the seeded stubs. Do NOT hand-write these files or emit a JSON analysis instead — the script is the source of the writes.

Flow

flowchart TB
  start(["/claudehut:init"]) --> gen["run claudehut-init generator + ls plane<br/>(deterministic script — never hand-write)"]
  gen --> verify{"all 5 present? MEMORY · PROJECT ·<br/>LANGUAGE · architecture · reuse-index"}
  verify -- "no (and attempts ≤ 1)" --> fix["fix the reported error → re-run with --refresh"]
  fix --> gen
  verify -- "no / cap hit" --> halt(["BLOCKED: init incomplete —<br/>surface the missing file + error"])
  verify -- "yes" --> enrich["enrich stubs below provenance line<br/>(optional — raises quality)"]
  enrich --> mcp{"interactive (AskUserQuestion available)?"}
  mcp -- "yes" --> ask["AskUserQuestion multi-select →<br/>emit claude mcp add per pick (suggest, never run)"]
  mcp -- "no (-p)" --> block["print recommended lines as copy-paste block"]
  ask --> fin(["Bootstrapped — commit .claude/ except state/"])
  block --> fin

1. Generate the project plane + verify (REQUIRED)

Call the Bash tool to run the generator and list the result in one command (a tracked tool call, not shell auto-exec at skill-load):

"${CLAUDE_PLUGIN_ROOT}/bin/claudehut-init" "${CLAUDE_PROJECT_DIR}" && ls "${CLAUDE_PROJECT_DIR}/.claude/claudehut/"

The SessionStart hook already auto-runs the generator when the plane is absent, so this skill is the explicit / --refresh path — run it here so the listing confirms the plane.

It detects the stack from the build files and writes, under ${CLAUDE_PROJECT_DIR}/.claude/claudehut/: MEMORY.md, PROJECT.md, LANGUAGE.md, architecture.md, reuse-index.json, learnings.jsonl, state/ — plus the stack-gated rule tree under .claude/rules/, and appends the always-load @import slice to CLAUDE.md. Idempotent: it skips existing plugin-owned files (pass --refresh to regenerate) and never clobbers learnings.jsonl.

Verify-and-retry per the Flow. Init is not complete until all five judgment files exist (P3: the binding prerequisite for project-adaptive memory and cross-session learning).

2. Enrich the seeded stubs (best-effort — raises quality, not required for correctness)

The script seeds judgment fields as TBD — refine. Improve them by reading the code (keep edits under the provenance line — re-init treats them as authoritative and won't overwrite them):

• architecture.md / PROJECT.md: fill dependency direction, transaction strategy, error mapping, messaging topology.
• reuse-index.json components[]: catalog existing @Service/@RestController/@Repository/@Component classes (id, kind, path, purpose, tags) so the Brainstorm reuse-scan can find them.
• LANGUAGE.md: refine the canonical term meanings to this project's real usage.

3. Suggest MCP servers (optional, opt-in — never auto-install)

ClaudeHut ships no active MCP config and connects nothing automatically. Read the catalog at ${CLAUDE_PLUGIN_ROOT}/templates/mcp-recommendations.md and match it against the detected stack to build the candidate list:

• tech-stack bucket — each server whose detect-when matches a detected dependency (gives the Review auditors live data; without them they review statically).
• memory bucket — the knowledge-graph memory MCP.
• research bucket — the docs MCP (context7) for current library best-practice.

Interactive vs -p selection follows the Flow: emit a claude mcp add --scope project … line only for each selected server.

The developer substitutes their own connection string / token — never print or store real secrets, and do not run these commands yourself (suggest, don't force).

Finish: "Bootstrapped. Commit .claude/ (except state/) to share with the team."