mtk-setup

---

name: mtk-setup description: One-stop setup entry point that bootstraps a repo or re-runs architecture audit type: skill allowed-tools: Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion argument-hint: [--audit|--audit-only] [--merge] [--preview] [--non-interactive] [--no-verify-commands] [--update-guidelines] [--refresh [--dry-run]] [--check] [--converge] user-invocable: true

MTK Setup — Unified Entry Point for Bootstrap and Audit

You are the single setup entry point for MTK. You dispatch to the right workflow skill based on the engineer's argument and the current state of the repo.

Overview

MTK distinguishes between two setup tasks:

• Bootstrap (first-time setup): Detect tech stack, pull coding guidelines, audit the codebase once, generate CLAUDE.md, .claude/tech-stack, and .claude/references/architecture-principles.md. This is the one-time preparation.
• Re-audit: Re-run the architectural audit only. Regenerates .claude/references/architecture-principles.md (and conventions.md) to reflect current reality. Use after significant architectural change.
• Merge: Unify architecture audits from multiple repos (e.g., a team hub repo that aggregates payfac, collection-system, etc.) into a single team-wide document.
• Refresh: Drift-scoped refresh of ALL generated rules and findings (architecture-principles, conventions, detected-tools, reference pruning, AGENTS.md/tool configs, indexes) — not just the audit doc. --check is the CI gate that reports staleness without writing anything.
• Converge: Inverse of refresh — code is judged against the agreed architecture-principles.md/conventions.md, and violations are reported as graded work items (blocking/flag/note), never auto-fixed.

When To Use

• First time onboarding a repo → run /mtk-setup with no flags.
• Architecture has drifted and you want architecture-principles.md refreshed → /mtk-setup --audit.
• You have audit files from multiple repos in .claude/references/audits/ and want one unified doc → /mtk-setup --merge.
• Generated docs (architecture-principles, conventions, detected-tools, AGENTS.md, indexes) have drifted, or you just merged significant changes → /mtk-setup --refresh (add --dry-run to preview first).
• CI or a quick staleness question ("are the generated docs still accurate?") → /mtk-setup --check.
• You want to know where the code has drifted from the agreed architecture-principles.md/conventions.md, graded as reviewable work items → /mtk-setup --converge.

Workflow

STEP 0: Parse Arguments

Parse the argument string into a mode and flags:

Flag	Meaning
--audit or --audit-only	Run audit workflow only (skip stack detection, CLAUDE.md generation)
--merge	Multi-repo unification mode (implies audit)
--preview	Show proposed changes, ask before writing (bootstrap only)
--non-interactive	Skip interview questions (bootstrap only)
--no-verify-commands	Skip STEP 3.5a's build/test/format command execution during CLAUDE.md generation (bootstrap only)
--update-guidelines	Bump the pinned moberghr/coding-guidelines SHA in .claude/manifest.json to current HEAD and refresh recorded sha256 hashes. Does not re-run bootstrap; engineer must invoke /mtk-setup afterward if they want the new guidelines applied. Cannot be combined with other flags.
--refresh	Drift-scoped refresh of ALL generated rules and findings (not just the audit doc). Delegates to setup-refresh.
--dry-run	Only with --refresh: print the invalidation plan, write nothing.
--check	CI staleness gate — run scripts/setup-refresh-plan.sh --check, print its output, propagate its exit code. Writes nothing.
--converge	Inverse of --refresh: judge the code against architecture-principles.md/conventions.md and report drift as graded work items (blocking/flag/note). Delegates to setup-converge. Read-only outside .claude/.mtk-cache/.

Default mode (no flags): bootstrap.

Flag combination rules:

• --update-guidelines is mutually exclusive with every other flag. Reject with a clear message if combined.
• --merge implies audit and is mutually exclusive with --non-interactive.
• --refresh and --check are each mutually exclusive with --audit, --merge, --update-guidelines, and each other. Reject with a clear message if combined.
• --converge is mutually exclusive with --audit, --merge, --update-guidelines, --refresh, and --check. Reject with a clear message if combined.
• --dry-run requires --refresh. Reject with a clear message if passed without it.
• --no-verify-commands is valid only in bootstrap mode. Reject with a clear message if combined with --audit, --merge, --refresh, --check, --converge, or --update-guidelines.
• --refresh --non-interactive is allowed: the refresh runs headless, and every engineer-edited file defers to NEEDS REVIEW per the regen-diff-contract §6 — proposals are never force-applied without the interactive gate.

STEP 1: Decide the Target Skill

Argument pattern	Invoke
--update-guidelines present	Inline workflow (see below). Does not invoke a target skill.
--check present	Inline workflow (see below). Does not invoke a target skill.
--refresh present	.claude/skills/setup-refresh/SKILL.md (pass --dry-run through)
--converge present	.claude/skills/setup-converge/SKILL.md (no flags)
--merge present	.claude/skills/setup-audit/SKILL.md (pass --merge)
--audit or --audit-only present	.claude/skills/setup-audit/SKILL.md (no flags)
None of the above	.claude/skills/setup-bootstrap/SKILL.md (pass --preview / --non-interactive / --no-verify-commands through)

Inline workflow for --update-guidelines:

1. Resolve current HEAD SHA and validate it before using it for anything:

   CURRENT_SHA=$(git ls-remote https://github.com/moberghr/coding-guidelines HEAD | awk '{print $1}')
   if [[ ! "$CURRENT_SHA" =~ ^[0-9a-f]{40}$ ]]; then
     echo "ERROR: git ls-remote did not return a valid 40-char SHA (got: '$CURRENT_SHA') — network issue or unreachable repo; stopping." >&2
     exit 1
   fi
   

2. Read pinned SHA from the resolved pin file: PM="${CLAUDE_PLUGIN_ROOT:-.}/.claude/manifest.json"; [ -f "$PM" ] || PM=".claude/mtk-version.json". Read coding-guidelines.sha from $PM (both files carry this key at the same path). If identical to $CURRENT_SHA, report "Already at HEAD ($SHA). Nothing to do." and exit.
3. For each file listed in coding-guidelines.files (read from $PM), fetch at the new SHA and compute sha256 — compute all files' hashes before writing anything (no partial pin updates):

   curl -fsSL "https://raw.githubusercontent.com/moberghr/coding-guidelines/$CURRENT_SHA/$FILE_PATH" | sha256sum | awk '{print $1}'
   

   Check curl's exit status for every file (e.g. via ${PIPESTATUS[0]}, since it's piped into sha256sum) — -f makes curl itself fail on HTTP error responses instead of emitting an error page as if it were file content. If any fetch fails, STOP immediately: report which file(s) failed and leave $PM untouched — do not write a pin update for only some of the files. Only proceed to step 5's write once every file's hash has been computed successfully.
4. Print a diff summary: old SHA → new SHA, plus any sha256 changes per file.
5. Write rule — never write into the plugin cache. Determine the write target from $PM, not just from CLAUDE_PLUGIN_ROOT being set (the fallback below can also resolve to a plugin path):
   • If $PM is plugin-prefixed — under $CLAUDE_PLUGIN_ROOT or under ~/.claude/plugins (a marketplace install reading the plugin's own manifest) — do not write there. Instead write the updated pin to the repo-local .claude/mtk-version.json, under its coding-guidelines block (repo, sha, files.<path>) — create the block if absent, leaving every other top-level key (version, installed, source) untouched.
   • Otherwise — $PM is the local .claude/manifest.json (the toolkit repo itself, no plugin prefix), or $PM already fell back to .claude/mtk-version.json (no manifest reachable at all) — update that same file in place: coding-guidelines.sha and each coding-guidelines.files.<path> value.
6. Tell the engineer: "Guidelines pin updated in <file>. Run /mtk-setup --audit or /mtk-setup to apply the new guidelines to this repo." The report line MUST name the file that now carries the updated pin (.claude/manifest.json or .claude/mtk-version.json). Do not auto-invoke bootstrap — the engineer decides when to apply.

Inline workflow for --check:

1. Run bash scripts/setup-refresh-plan.sh --check (resolve the script path per the MTK File Resolution section).
2. Print its output verbatim — the per-artifact plan table plus the summary line.
3. Propagate its exit code as the outcome: 0 → report "✅ Generated docs are fresh."; 1 → report "⚠️ Staleness detected — run /mtk-setup --refresh to reconcile."; 2 → report the usage/setup error.
4. Write nothing. --check is a read-only gate, suitable for CI.

Ambiguity check: if the repo has no .claude/tech-stack file and the engineer passed --audit, warn:

"No tech stack detected — this looks like a first-time setup. Audit alone won't generate CLAUDE.md. Run /mtk-setup with no flags to do a full bootstrap. Proceed with audit only? [y/N]"

Use AskUserQuestion for this prompt.

For --refresh on a repo with no .claude/tech-stack, do not offer to proceed — setup-refresh hard-stops on an un-bootstrapped repo (its STEP 0 preconditions). Tell the engineer directly: "This repo has not been bootstrapped — run /mtk-setup with no flags first; --refresh is for re-runs."

For --converge on a repo with no .claude/tech-stack (or no stamped architecture-principles.md/conventions.md), do not offer to proceed — setup-converge hard-stops on an un-bootstrapped/unaudited repo (its STEP 0 preconditions). Tell the engineer directly: "This repo has not been bootstrapped — run /mtk-setup with no flags first; --converge is for judging code against an already-agreed principles doc."

STEP 2: Read and Follow the Target Skill

Read the target SKILL.md (paths resolved per the MTK File Resolution section below). Follow every step of that skill inline — do NOT summarize or skip steps. The target skill owns its own verification section; run those checks before reporting back.

STEP 3: Report

When the target skill completes, print a one-line summary:

✅ MTK Setup: [mode] complete in [duration]. See [output file(s)].

Where [mode] is one of: bootstrap, audit, merge, refresh, check, converge.

MTK File Resolution

MTK skills and shared references live either in the project (local install) or the plugin cache (marketplace install). Resolve once:

1. If $CLAUDE_PLUGIN_ROOT is set, prefix .claude/skills/ and .claude/references/ reads with it.
2. Otherwise, if .claude/skills/context-engineering/SKILL.md exists locally → project-relative paths work as-is.
3. Otherwise, fall back to find ~/.claude/plugins -maxdepth 8 -name "SKILL.md" -path "*/mtk/*/context-engineering/*" -type f 2>/dev/null | sort -V | tail -1 | sed 's|/.claude/skills/context-engineering/SKILL.md||'. If empty, MTK skills are unavailable — warn the engineer and proceed with CLAUDE.md only.

Always project-relative (never prefixed): CLAUDE.md, .claude/tech-stack, .claude/rules/, tasks/, docs/, .claude/references/architecture-principles.md, .claude/references/pre-commit-review-list.md, .mtk/ (workflow state). Resolve skills and scripts from the same root: a split (skills from a local dev checkout, scripts from the plugin cache) risks version drift — anchor both the same way.

Verification

• Correct target skill selected based on flags
• Target skill's verification section was executed
• All files the target skill was supposed to create exist
• Bootstrap mode: CLAUDE.md, .claude/tech-stack, and .claude/references/architecture-principles.md exist
• Audit mode: .claude/references/architecture-principles.md updated
• Merge mode: unified .claude/references/architecture-principles.md written; source audits in .claude/references/audits/ untouched
• Refresh mode: setup-refresh's own Verification section executed; refresh report printed with Updated / Created / Preserved / Needs review counts
• Check mode: plan table printed, exit code propagated, no files written
• Converge mode: read-only outside .claude/.mtk-cache/; report printed with blocking/flag/note counts