fwd-setup

---

name: fwd:setup description: Setup wizard for HeadingFWD's optional Claude Code conventions. Runs unattended — zero prompts, always project-local scope: installs every convention in one batch — smartlint Stop-hook, lessons memory file, gitignore entries for fwd runtime artefacts, clear-context prompt on plan accept, and no-attribution (disables the Claude commit/PR byline). Copies bundled payload files into .claude/hooks/ or .claude/lessons/, merges JSON snippets into .claude/settings.local.json, injects an instructions section into CLAUDE.md (lessons), or appends a marker-bracketed block to .gitignore. Idempotent and modular — each feature lives in scripts//. Prints a fixed-format Dutch summary from the bundled OUTPUT.md template, never free-form prose. Runs on Haiku. Use only when the user invokes /fwd:setup explicitly. context: fork model: haiku allowed-tools: Bash disable-model-invocation: true

fwd:setup

Install HeadingFWD's optional conventions with zero prompts, always project-local. One batch, one fixed-format summary. Each feature is independent: its installer lives in scripts/<feature>/install.sh, accepts --scope user|project, and is idempotent. Re-running the skill is safe.

Apply + report

!bash "${CLAUDE_SKILL_DIR}/scripts/apply-all.sh" && echo "---OUTPUT-TEMPLATE---" && cat "${CLAUDE_SKILL_DIR}/OUTPUT.md"

Instructions

The block above already ran all five installers at --scope project — that already happened, do not run them again. It contains two parts:

1. Above ---OUTPUT-TEMPLATE---: one ### <feature> report per installer (exit=, stdout:, stderr:), always in this order: smartlint, lessons, gitignore, clear-context-on-plan, no-attribution.
2. Below ---OUTPUT-TEMPLATE---: the fixed report template (OUTPUT.md) — read its own instructions and fill in every placeholder from part 1.

Print the filled-in template verbatim as your only response. Do not add commentary before or after it, do not call any tool, do not ask a question.

Per-feature exit code semantics (uniform across all installers):

• 0 — installed (fresh) or already present / refreshed in place. Idempotent.
• 2 — collision: the installer found a conflicting existing value and refused to overwrite. It printed the conflicting entry + remediation hint to stderr. Relay that verbatim per OUTPUT.md's "Aandachtspunten" rule; do not retry.
• other non-zero — argument or I/O error. Same handling as a collision.

Feature-specific collision causes (for context when relaying stderr):

• smartlint — existing Stop-hook with a different command.
• lessons — corrupt sentinel markers in CLAUDE.md (start without end, or out of order).
• gitignore — corrupt sentinel markers in .gitignore.
• clear-context-on-plan — showClearContextOnPlanAccept explicitly set to false.
• no-attribution — attribution.commit or attribution.pr holds a non-empty custom string.

Invariants

• Skill is invoked explicitly only — disable-model-invocation: true keeps the wizard out of automatic-trigger flows.
• Zero prompts. The wizard never calls AskUserQuestion — every feature is always installed, always at project scope. allowed-tools: Bash enforces this structurally: the skill has no tool to ask with.
• Always project scope. scripts/apply-all.sh hardcodes --scope project — this skill never touches ~/.claude/. (The installers themselves still accept --scope user if invoked manually outside this skill.)
• Runs on Haiku (context: fork + model: haiku) — the apply step is a single deterministic bash pre-flight block and the report is a fill-in-the-blanks template, so no larger model is needed.
• The final message is always the filled-in OUTPUT.md template — never a free-form summary. See OUTPUT.md for the exact fill rules and worked examples.
• Each feature installer lives at scripts/<feature>/install.sh, accepts --scope user|project, and is idempotent. Re-runs detect existing installs and either refresh in place or exit 2 on a corrupt/conflicting state.
• scripts/apply-all.sh always exits 0 itself (it reports per-installer failures, it doesn't propagate them) so the && chain in "Apply + report" always reaches cat OUTPUT.md.
• JSON merging (smartlint, clear-context-on-plan, no-attribution): the shared scripts/lib/merge-json.sh deep-merges objects, concatenates arrays, and lets new scalars win on conflict. Each installer dedupe-checks its own entries before calling the merger so re-runs do not double the array. jq is required; when missing the merger backs up the target to <file>.bak and replaces it with the new snippet.
• Markdown injection (lessons): the lessons installer brackets its section with sentinel HTML comments (<!-- fwd:lessons:start --> / <!-- fwd:lessons:end -->) and uses inline head / tail to replace the region between them on refresh. No shared lib yet — extract to scripts/lib/merge-markdown.sh if a second feature also needs it.
• Gitignore injection: marker-bracketed block (# fwd:setup:gitignore:start / # fwd:setup:gitignore:end) in .gitignore.
• Hook commands use literal $CLAUDE_PROJECT_DIR — Claude Code's shell expands it at hook runtime.
• One feature's collision (exit 2) never aborts the batch — apply-all.sh always runs all five.

Adding a new feature later

1. Create scripts/<feature>/install.sh (accept --scope user|project, idempotent, exit 2 on collision).
2. Drop static files in scripts/<feature>/payload/. Use placeholders (__HOOK_DIR__, __LESSONS_PATH__, etc.) for path substitution at install time.
3. Pick a merge style appropriate to what you're injecting:
   • JSON settings → call ../lib/merge-json.sh
   • Markdown / CLAUDE.md section → use sentinel markers + inline head/tail (see scripts/lessons/install.sh)
   • Plain file copy → just cp into <scope>/.claude/<feature-dir>/
4. Add it to the FEATURES array in scripts/apply-all.sh (order = display order).
5. Add the feature-specific collision cause to the bullet list above.
6. Add a row for it to the slug→label table in OUTPUT.md.