Init

/mck-scaffold:init

Status: Phase 7 — fully functional. Runs the interview (Steps 1–6), then executes scaffolding (Step 7), triggers upstream installs (Step 8), and prints the final summary (Step 9). Acceptance: same answers always produce the same target-project files; re-runs are idempotent.

Pipeline overview

Step 0: research-refresh staleness check (NEW in v0.11.0)
Step 1: detect project context
Step 2: load interview definition (questions.md + routing.md)
Step 3: run interview (Q1–Q8 with pruning)
Step 4: apply routing (deterministic lookup)
Step 5: produce decision summary
Step 6: confirm or iterate
─────────── confirmation point ───────────
Step 7: execute scaffolding (compose + write files)
Step 8: trigger upstream installs (auto, not optional — Phase 11 contract)
Step 9: final summary + manual-steps callout

Steps 0–6 do not write to the target project. Step 0 may write to mck-scaffold itself if user opts in to a research refresh. Steps 7–8 write to the target project.

---

Step 0 — research-refresh staleness check

New in v0.11.0. Surface upstream-ecosystem drift before routing decisions get made against a stale catalog. The dogfood on consultant-helper (2026-05-15) immediately surfaced Spec-Kit drift less than 30 days after the catalog was written — staleness is real and routing against stale data produces broken installs.

Read research/last-run.json from the plugin root. Extract timestamp_iso and compute days_since against today.

Render an appropriate prompt based on staleness:

Staleness	Prompt tone
< 30 days	Brief, skip-by-default: "Catalog refreshed {days_since} days ago — fresh. Skipping refresh by default. Pass --refresh to override."
30–90 days	Neutral: "Catalog refreshed {days_since} days ago. Refresh before this scaffold? It catches upstream drift (~5-10 min, 4 parallel research streams)."
> 90 days	Strong: "Catalog refreshed {days_since} days ago — drift is likely. Strongly recommended to refresh before this scaffold. Skip only if you've verified the framework you expect to route to hasn't shifted."

Use AskUserQuestion with options sized to the staleness tone:

• < 30 days → ["Skip refresh, proceed to interview" (default), "Force refresh anyway"]
• 30-90 days → ["Refresh first (recommended)", "Skip refresh — proceed to interview"]
• > 90 days → ["Refresh first (strongly recommended)" (default), "Skip — proceed without refresh"]

If user opts to refresh:

Invoke the research-refresh skill conversationally. It will run its own pipeline (4 parallel research streams → diff → per-section confirm → apply). When it completes, return here and proceed to Step 1.

If user opts to skip:

Proceed directly to Step 1. The current catalog state in research/streams/*.md and interview/routing.md will drive the routing.

Argument handling at Step 0:

• --refresh — force the refresh prompt regardless of staleness (useful for "I know there was a release yesterday, re-check anyway")
• --no-refresh — skip Step 0 entirely without prompting (useful for scripted / CI invocations)

If neither flag is passed, follow the staleness-tone logic above.

---

Step 1 — detect project context

Run these checks in the current working directory and report what you found:

• Does .git/ exist?
• Does CLAUDE.md exist?
• Does .claude/ exist?
• Do any SDD framework markers exist? Check for .specify/, docs/constitution.md with MUSUBI banner, openspec/, docs/superpowers/, ~/.claude/plugins/.../superpowers/.
• Does docs/decisions/INTERVIEW.md exist? This is the re-run indicator.

Print a one-line context summary, e.g.:

Detected: .git present, no existing CLAUDE.md, no prior SDD framework. Treating this as a fresh init in an existing repo.

Re-run handling

If docs/decisions/INTERVIEW.md is present, route to the re-run subroutine (see "Idempotency" near the end). Do NOT proceed with a fresh interview when a prior scaffold record exists.

---

Step 2 — load the interview definition

Read these files from the plugin root (the directory containing .claude-plugin/plugin.json):

• interview/questions.md — Q1–Q8 question tree, options, pruning rules
• interview/routing.md — answer-combination → preset routing table + install command list

These are the source of truth. If their content conflicts with anything in this SKILL.md body, the markdown files win.

---

Step 3 — run the interview

For each question Q1–Q8 (respecting pruning):

1. Read the question wording and options from interview/questions.md.
2. Build an AskUserQuestion call using the verbatim wording.
3. Q5 is multiSelect: true; all others single-select.
4. For Q6: set the auto-default per Q3 (solo → CLI, team10+ → MCP) as the first option, but always ask.
5. Apply pruning before each next question:
   • After Q2: if brownfield, skip Q3 and Q4.
   • After Q1: if ≠ d, skip Q8.
6. One question per AskUserQuestion call. Never batch.

Record answers in a running dictionary.

---

Step 4 — apply routing

Read interview/routing.md. Match answers against the primary routing table — this is a deterministic lookup, not a judgment call. Resolve:

• methodology_preset: one of L1-lean, L1-spec-kit, L1-superpowers, L2-musubi, L2-openspec-brownfield
• methodology_overlays: zero or more of L2-csdd-security, L3-watch
• stack_adapters: from Q5 (each selected stack name, e.g., dotnet, angular, unity, python)
• mcp_orientation: yes/no from Q6
• multi_agent_appetite: none/limited from Q7
• install_commands: from routing.md's "Meta-installer triggers per preset" table — collect commands for preset + each overlay

If the routing table has no exact match for the answer combination, that's a routing bug. Report it to the user and abort — do not silently pick a default.

---

Step 5 — produce the decision summary

Print the markdown block defined in Phase 2 (interview answers table, resolved methodology, rationale paragraphs, scaffolded-files preview, honest caveats from methodology-picker/SKILL.md). Do not write to disk yet.

---

Step 6 — confirm or iterate

After the decision summary, ask via AskUserQuestion:

"Does this routing look right?"

• Yes — proceed to scaffold and install. → continue to Step 7
• No — try a different preset. → surface 2 next-closest routing rows; user picks one; re-render summary
• Tweak an answer. → ask which question to revisit; rerun forward from there
• Dry-run only. → skip Steps 7-8; print "what would be written / installed" without writing

If --dry-run was passed as an argument, force the dry-run path.

---

Step 7 — execute scaffolding

Order matters. Compose files first (Steps 7a–7g), THEN run installs (Step 8). Some overlay operations (e.g., CSDD constitution.append.md) must happen AFTER the methodology's install creates the base constitution — those are handled at the end of Step 8.

Define the variable dictionary first:

{{project_name}}            = basename of cwd
{{project_path}}            = absolute cwd
{{scaffold_date}}           = today (ISO format YYYY-MM-DD)
{{mck_scaffold_version}}    = from .claude-plugin/plugin.json "version"
{{re_eval_date}}            = scaffold_date + 6 months
{{git_present}}             = yes/no
{{prior_claude_md}}         = yes/no
{{prior_claude_dir}}        = yes/no
{{prior_sdd_markers_or_none}} = comma-separated detected markers or "(none)"
{{methodology_preset}}      = from Step 4
{{methodology_overlays_or_none}} = csv or "(none)"
{{methodology_overlays_suffix}}  = " with overlays X, Y" or ""
{{preset_one_liner}}        = first sentence of methodologies/{preset}/README.md
{{stack_adapters_csv}}      = csv from Q5
{{mcp_orientation}}         = yes/no
{{multi_agent_appetite}}    = none/limited
{{q1_answer}} ... {{q8_answer}} = verbatim option labels from questions.md
{{q5_answer_csv}}           = csv of selected stacks
{{pruning_notes}}           = e.g. "Q2 was 'brownfield', so Q3 and Q4 were skipped." or "No pruning applied."
{{user_notes_or_none}}      = free-text "Other" notes from interview or "(none)"
{{rationale_paragraphs}}    = generate from preset + answers (1-3 short paragraphs)
{{alternatives_list}}       = next-closest routing rows as bulleted "not chosen because..." list
{{caveats_block}}           = caveats from methodology-picker that apply to this preset
{{scaffolded_files_list}}   = filled in as Step 7 progresses
{{installed_plugins_list}}  = filled in as Step 8 progresses

7a — read all template fragments

From the plugin root:

• templates/shared/CLAUDE.md.tmpl
• templates/shared/AGENTS.md.tmpl
• templates/shared/settings.json.tmpl
• templates/shared/docs/decisions/INTERVIEW.md.tmpl
• templates/shared/docs/decisions/METHODOLOGY.md.tmpl
• templates/methodologies/{methodology_preset}/CLAUDE.append.md
• templates/methodologies/{methodology_preset}/settings.merge.json
• For each overlay: templates/methodologies/{overlay}/CLAUDE.append.md, settings.merge.json, and constitution.append.md if present
• For each stack in stack_adapters: templates/stack/{stack}/CLAUDE.append.md, settings.merge.json

Cache content in memory before any writes.

7b — compose CLAUDE.md

1. Start with shared CLAUDE.md.tmpl, substituted.
2. Build {{methodology_specific_content}} = methodology preset's CLAUDE.append.md, substituted. Append any overlay's CLAUDE.append.md, substituted, in routing order.
3. Build {{stack_specific_content}} = concatenation of each stack adapter's CLAUDE.append.md, in Q5 selection order, each substituted.
4. Substitute the two content blocks into the shared template.
5. Verify no {{...}} placeholders remain. If any do, report a substitution gap and abort the write.

Write to {{project_path}}/CLAUDE.md. See "Step 7h — write contract" for conflict handling.

7c — compose AGENTS.md

Substitute variables on AGENTS.md.tmpl. Write to {{project_path}}/AGENTS.md.

7d — compose .claude/settings.json

Start with settings.json.tmpl. Deep-merge each layer's settings.merge.json in order: methodology → overlays → each stack (in Q5 selection order).

Merge rules:

• Arrays: append. For permissions.allow and permissions.deny, deduplicate.
• Objects: deep-merge recursively.
• Scalars: later wins.

Apply MCP overlay if mcp_orientation = yes (Phase 7 placeholder: leave hooks arrays as merged; MCP-specific allowlist fragment lands when MCP overlay template is built — currently a no-op).

Write the merged JSON to {{project_path}}/.claude/settings.json.

7e — write docs/decisions/INTERVIEW.md

Substitute the shared INTERVIEW.md.tmpl with answers and project context. Write to {{project_path}}/docs/decisions/INTERVIEW.md.

7f — write docs/decisions/METHODOLOGY.md

Build {{rationale_paragraphs}}, {{alternatives_list}}, {{caveats_block}} per the variable dictionary. {{scaffolded_files_list}} is the running list of files written so far. {{installed_plugins_list}} will be filled after Step 8.

Substitute and write to {{project_path}}/docs/decisions/METHODOLOGY.md. This file is updated again after Step 8 to fill in {{installed_plugins_list}}.

7g — L3-watch overlay

If L3-watch is in methodology_overlays, write templates/methodologies/L3-watch/horizon/L3-WATCH.md (substituted) to {{project_path}}/docs/horizon/L3-WATCH.md. (The L3-watch template ships in Phase 8.)

7h — write contract

For each file about to be written:

Situation	Action
Target file does not exist	Write. Add to scaffolded_files_list.
Target file exists, content matches what we would write	Skip silently. No-op.
Target file exists, content differs	Show a unified diff. AskUserQuestion: overwrite / skip / abort the whole scaffold. Default = skip.

Never silently overwrite a file the user has changed. Mck-scaffold is opinionated about respecting user edits.

---

Step 8 — trigger upstream installs

Read the install_commands resolved in Step 4. There are two command types:

8a — bash-runnable commands

These are auto-executed. Step 8a is not optional. After Step 6 confirmation, every bash-runnable install in the resolved preset's row of interview/routing.md runs via the Bash tool. Do not defer to "user runs this later" — the whole point of the scaffolder is that it scaffolds.

Reference the canonical install commands in interview/routing.md § "Meta-installer triggers per preset". For each:

1. Platform detection. Check $IsWindows (PowerShell) or uname -s (bash). Some preset install commands branch on this (e.g., Spec-Kit needs --script ps on Windows + UTF-8 env vars).
2. Environment setup. For Windows + Python-based upstream tools (specify-cli, musubi-sdd if Python-based, etc.), set $env:PYTHONIOENCODING = "utf-8" and $env:PYTHONUTF8 = "1" BEFORE invocation. Without this, Spec-Kit's banner crashes with UnicodeEncodeError on the default cp1252 codepage.
3. Conflict pre-check. Detect existing markers:
   • Spec-Kit: .specify/ already exists → conflict
   • MUSUBI: docs/constitution.md with MUSUBI banner + .claude/skills/@orchestrator → conflict
   • OpenSpec: openspec/ already exists → conflict
4. If conflict found, AskUserQuestion:
   • Install alongside — proceed; the upstream tool's own --force handles the merge (most modern installers support this)
   • Skip this install — note in summary, continue
   • Abort entire scaffold — stop everything (files already written remain; document this in summary)
   • Default: install alongside with --force (auto-install is the explicit Phase 11 contract; only escalate to user prompt if conflict detection finds something genuinely ambiguous)
5. No conflict → run the command via Bash tool with output captured. Apply long timeouts (180–600 seconds) — first-time uvx invocations download from GitHub which can be slow.
6. For Shotgun: prompt the user for the "<short description>" argument BEFORE the install runs — it's free-text and required.
7. Verify success — check expected post-install markers exist (.specify/, openspec/, docs/constitution.md). If markers missing, capture stderr and surface to user. Continue with the next install — don't abort the whole run for a partial failure.
8. Surface Spec-Kit's .claude/ security advisory — if Spec-Kit ran, its post-install output recommends adding parts of .claude/ to .gitignore to prevent agent credential leakage. Append this note to the final summary (Step 9) so the user sees it.

8b — slash-command installers

/plugin install … is a Claude Code harness command, not a Bash command. A skill body cannot programmatically invoke another slash command. Surface them to the user explicitly:

⚠ Manual step required:
The following slash commands must be run in this Claude Code session to complete the install.
Please run each one and confirm before continuing:

  /plugin install superpowers@claude-plugins-official
  /plugin reload

Capture in installed_plugins_list regardless (with a (manual) marker). Trust that the user ran them.

8c — version pinning (policy)

Per DESIGN.md §10, the locked default is "pin to known-good; quarterly re-eval." In Phase 7, we do NOT yet hard-pin upstream versions — install commands use @latest or the marketplace's chosen version. Pinning lands as a Phase 7.1 iteration once known-good versions for each upstream are vetted.

8d — CSDD constitution append (post-install)

If L2-csdd-security is in methodology_overlays:

1. Wait until docs/constitution.md exists in the target project (MUSUBI's npx musubi-sdd init creates it).
2. Read templates/methodologies/L2-csdd-security/constitution.append.md (substituted).
3. Append it to the end of {{project_path}}/docs/constitution.md.
4. Add the file to scaffolded_files_list as "modified".

If MUSUBI's install didn't create docs/constitution.md (failure, skipped, conflict), skip the append and surface a warning.

---

Step 9 — final summary

Re-substitute METHODOLOGY.md to fill in {{installed_plugins_list}}.

Print a closing markdown block:

✅ mck-scaffold complete.

## Files written

{scaffolded_files_list — relative paths}

## Plugins / commands installed

{installed_plugins_list — including `(manual)` slash-command items}

## Manual steps still required

{slash commands the user must run, if any — usually `/plugin install superpowers@...` + `/plugin reload`}

## Re-evaluation cadence

This methodology is recommended for review on **{{re_eval_date}}** (six months from today). At that point, re-run `/mck-scaffold:init` — it will detect this scaffold and offer reconfigure or diff.

## Next

1. Read `CLAUDE.md` — it's your project's authoritative conventions.
2. Read `docs/decisions/METHODOLOGY.md` — it's why this setup was chosen.
3. {preset-specific next step, e.g.:
   - L1-superpowers: "Try `/superpowers:brainstorm` on your first feature."
   - L1-spec-kit: "Try `/speckit.specify` on your first feature."
   - L2-musubi: "Try `@requirements-analyst` to draft your first EARS spec."
   - L2-openspec-brownfield: "Run `uvx shotgun plan` again if your codebase changes significantly."
}

---

Idempotency — re-run handling

When docs/decisions/INTERVIEW.md exists at Step 1:

1. Read the prior answers and the prior methodology_preset from the file.
2. AskUserQuestion:
   • Reconfigure from scratch — re-run Steps 2–9 with fresh answers (this overwrites INTERVIEW.md and METHODOLOGY.md)
   • Refresh templates with same answers — re-run Steps 7–9 only; useful when mck-scaffold has been updated and you want the new template content
   • Show diff vs current state — iterate through what would be written, output a unified diff per file, no writes happen
   • Abort
3. Default: abort. Re-runs are destructive; require explicit user opt-in.

---

Argument handling

• --dry-run — force the dry-run path at Step 6 (skip Steps 7–8 entirely)
• --overlay=<name> — Phase 10 feature; Phase 7 echoes "(overlay support arrives in Phase 10)" and continues normally

---

Acceptance criteria (Phase 7)

• Phase 2 criteria still pass (interview is deterministic, no files in Steps 1–6)
• Step 7 composes CLAUDE.md / AGENTS.md / settings.json correctly per the shared → methodology → overlays → stacks order
• No {{...}} placeholders remain in any target-project file after substitution
• Existing target files are never overwritten without explicit user confirmation
• Step 8 runs bash-runnable installs and surfaces slash-command installs as manual steps
• Conflict detection fires before each install
• CSDD constitution append happens AFTER MUSUBI's docs/constitution.md is created
• Re-run on a folder with prior INTERVIEW.md never silently overwrites
• --dry-run skips Steps 7–8 cleanly
• Final summary lists every file written and every install triggered (including (manual) items)

---

Implementation notes for Claude (not user-facing)

• Use Write and Edit for file ops. Write creates parent dirs automatically.
• For variable substitution, do a straightforward string replace per variable in the variable dictionary. Don't try to be clever with regex — templates are authored by humans.
• For deep-merge of JSON: implement as nested object walk. Arrays append; dedup permissions.allow and permissions.deny; objects merge recursively; scalars overwrite.
• For Bash installs:
  • Run with reasonable timeouts (60–180s typically; Shotgun on large repos may need 600s).
  • Capture both stdout and stderr; surface stderr if non-empty.
  • Don't use --no-confirm flags speculatively — only when documented by the upstream tool.
• If a Bash install fails, continue with the next install. Log the failure for the final summary. Don't abort the whole scaffold on one upstream failure.
• Cache the plugin root path (containing .claude-plugin/plugin.json) at the start of Step 1. All template loads are relative to it.
• For the diff-display path (re-run option c, and the per-file conflict resolution): use a clear unified-diff format (--- old, +++ new, hunks). Don't print giant files; truncate at ~200 lines and link to the source paths.