aidlc-init

Bring any repo up to the team's 18-item "AI-native repo" template. Diagnose what's present, ask only what can't be inferred, and for each branch pick the LOWEST-dependency option that fits — raising a dependency is always an explicit opt-in. Create only the MISSING pieces (never overwrite). Finish by re-scoring with the chosen profile.

<Use_When>

• /codepresso:aidlc-init <target-path> invoked
• User says "set up AIDLC structure", "make this repo AI-native", "scaffold AGENTS.md/ADR/policy" </Use_When>

- **Minimal external dependencies.** Default every branch to the zero/low-dependency option (raw git hook over husky, bundled node context index over an external tool, skip CI when there's no host). Only add a dependency when the user opts in, and state the tradeoff when you ask. - **Situational, not universal.** The 18 items are tool/host-specific. Unused tools and absent hosts are scored `na`, never penalised. - **Non-destructive.** Only ever create missing files; never overwrite. 1. Resolve `` (default = current project). Confirm it's a git repo; if not, warn and ask whether to continue.

2. Analyze (CLI): run node "${CLAUDE_PLUGIN_ROOT}/scripts/aidlc-cli.mjs" detect <path> then ... scan <path>. Show a compact 18-item table (status + evidence) and the overall %. If secrets[] is non-empty, surface a 🔴 warning at the top and tell the user to rotate them — never copy a secret value anywhere.

3. Interview (AskUserQuestion — 7 branches; only the non-inferable). Build a profile from the answers. Pre-fill each branch from detect:

   • Structure/stacks/submodules — confirm detected values; let the user correct.
   • Agent tools (multi-select, prechecked from detect.tools): which tools the team actually uses (claude, cursor, opencode, cline, copilot, gemini, amazonq). Only selected tools are scaffolded/scored; the rest become na.
   • CI host — default to detect.host. If host is none, ask "create GitHub Actions anyway, or skip?" (default skip → ci-pr na).
   • pre-push install — default raw (a .git/hooks/pre-push wired by hand, zero deps). Offer husky/lefthook only if detect.hookFramework indicates it or the user asks. skip → na.
   • Context index — default regen-node (bundled zero-dep scanner, stays fresh). Opt-in: external tool (richer, adds a dependency) or static (one-time, will go stale) or off.
   • Local dev — if detect.localDev shows a one-command bring-up, set localDev: 'detected' and inject that command into AGENTS.md. Otherwise offer localDev: 'scaffold' (apply-static writes an executable scripts/local-up.sh) or localDev: 'skip'.
   • Integrations (Notion/Figma/Google/AWS): default ALL OFF; opt-in per integration.
   • Ticket convention: confirm prefix (e.g. TSK) or "none".

   Write the answers to .codepresso/state/aidlc-profile.json: { tools[], ciHost, prePush, contextMode, localDev, integrations[], ticketPrefix }.

4. Preview (CLI, profile-aware): run ... scan <path> --profile .codepresso/state/aidlc-profile.json (honest na-scoped score) then ... plan <path> --profile <profile>. Show the file tree to be created, marking each static (templated) vs authored (skill-written). List any still-MISSING item NOT in the plan and not authored below (e.g. hooks, unit-tests, runbook) as 'manual setup required — not auto-created'. Ask for explicit confirmation. Write nothing before confirmation.

5. Apply: a. Static: ... apply-static <path> --profile <profile> (templated, non-destructive; host-correct CI variant picked from ciHost). b. Wire, don't just create:

   • pre-push: after the script exists, actually WIRE it — raw: copy/symlink to .git/hooks/pre-push + chmod +x; husky/lefthook: add the pre-push entry. A created-but-unwired script only scores partial.
   • contextMode=regen-node: merge the SessionStart hook from templates/aidlc/.claude/settings.codesight-hook.json into the repo's .claude/settings.json, then run node "${CLAUDE_PLUGIN_ROOT}/scripts/codesight-scan.mjs" <path> once to seed .codesight/CODESIGHT.md. c. Authored (you write, only if scan marked missing, existence-checked first):
   • AGENTS.md — the single authoritative entry point: real build/test/run commands (from the detected stack + the local-dev branch), conventions, short architecture overview.
   • CLAUDE.md — thin pointer to AGENTS.md (no duplicated content).
   • Pointer files for each SELECTED tool only (.cursor/rules, GEMINI.md, .github/copilot-instructions.md, .clinerules, .opencode/…, .amazonq/…) — each a thin pointer to AGENTS.md.
   • If structure=mono: author <submodule>/CLAUDE.md per submodule (stack-specific).
   • .codesight/CODESIGHT.md only if contextMode=static (otherwise the regen hook owns it).

6. Re-score (CLI): run ... score <path> --profile <profile> and report the new % + remaining gaps. Note that functional scoring counts only WIRED hooks / FRESH indexes / host-matched CI — explain any item that stayed partial and why.

7. End with: 💡 /codepresso:aidlc-doctor <path> — re-check compliance anytime.

<Tool_Usage>

• Bash for the aidlc-cli.mjs subcommands (always pass --profile after step 3) and for wiring hooks.
• AskUserQuestion for the 7-branch interview + preview confirmation.
• Read/Write for the profile JSON and authored files (Write only after existence check — non-destructive). </Tool_Usage>