groundtruth-setup

You are running the Groundtruth setup check. Some things a Claude Code plugin cannot set itself — the main statusLine and any env var — so this command DETECTS the current state and gives the user the exact, copy-pasteable action for anything missing. Be concise: a 5-line checklist (✓ / ✗ + the one-line fix), then a one-line summary. This command is the single point for rules — the things it DOES execute, each only after the user says yes, are (a) arming clean literal rules in warn (item 1), (b) running the AI prose pass /groundtruth-rules-ai and arming its clean proposals (item 1, only when rule-docs exist), and (c) the settings.local.json badge/block wiring (the closing offer — it does NOT write an integrity key, item 4). Everything else you DETECT and hand back the fix; for settings.local.json edits SHOW the user the block to paste (it's their file). Never enable block mode or arm a review-status rule on your own.

First, resolve two paths you'll need:

• the project root = the cwd.
• the plugin hooks dir: find the newest */hooks/groundtruth-statusline.mjs under ~/.claude/plugins/cache/groundtruth/groundtruth/ (use the highest version). Call it <HOOKS>. If you can't find it, fall back to telling the user to point at their groundtruth checkout's hooks/.

Then check each item and report ✓/✗:

1. Rules — arm the clean ones inline (this command is the installer; don't punt to /groundtruth-rules). Read .claude/groundtruth/proposed-rules.json and compiled-rules.json. Count clean-proposed (status:"armable", not already armed) vs any that status:"review" (already match committed code — over-broad risk).
   • If clean rules are unarmed → show the count and ask before arming: "Found N clean rules extracted from your docs (they match zero existing code, so they can only fire on new code). Arm them in warn now? [yes / review each / skip]." This is safe because arming in warn only adds a line to the verdict card — it never blocks — and each firing prints its [id] so you can /groundtruth-rules unarm <id> in one step. Warn ≠ enforce; block stays a separate, deliberate step (item 3).
     • yes → arm exactly the clean candidates: write .claude/groundtruth/compiled-rules.json as the union of already-approved + newly-approved, keeping only runtime fields (id, source, kind, file_re, line_re, severity, message, unless_re if present), severity defaulting to "warn", preserving any existing rule's severity, deduped by id. (This write is legitimate — /groundtruth-setup is a ratified writer of compiled-rules.json, same as /groundtruth-rules.) Confirm: "Armed N rules at warn — active next turn."
     • review each → hand off to /groundtruth-rules (the ledger) for a rule-by-rule look; arm nothing here.
     • skip → arm nothing; note it stays available via /groundtruth-rules approve-all later.
   • Never arm a status:"review" candidate here — those match existing code and need a human eye in /groundtruth-rules. Surface their count as a heads-up.
   • If proposed-rules.json is missing, init hasn't run: say "start a fresh session, or run node \"<HOOKS>/compile-rules.mjs\" .".
   • Richer rules from prose (/groundtruth-rules-ai) — this command is the single point, so OFFER TO RUN it here; gate the offer ONLY on whether rule-docs exist. The deterministic extractor only catches rules written in the literal use `X` not `Y` / never `X` forms; rules written as prose sentences need the model pass. Decide by doc PRESENCE, not by candidate count (finding N literal rules does NOT mean every prose rule was caught — a doc-heavy repo with 26 literal hits can still have many prose rules the literal pass can't see):
     • Check for ANY rule-docs: CLAUDE.md, AGENTS.md, **/SCHEMA.md, **/ARCHITECTURE.md, docs/*.md, .claude/skills/**/SKILL.md, .claude/agents/*.md, .cursorrules, .windsurfrules (parity with the engine's RULE_SRC_RE).
     • If there are NONE → do NOT offer the AI pass (there is nothing for it to read; say so in one line and move on). This is the only skip condition.
     • If any exist → OFFER to run it, regardless of how many literal rules were found. Say plainly what it does before running: "Your repo has prose rule-docs (CLAUDE.md, …) — the literal pass can't see rules written as sentences. Run the AI pass? It fans out an agent ONCE, only to draft rule candidates from those docs — the one and only place Groundtruth uses an LLM. It just proposes: nothing arms until you approve, and after that no model is ever called again — every per-turn audit stays fully deterministic + offline. [yes / skip]." (This transparency is required — the user must know it's a one-time, propose-only agent pass, not ongoing LLM use.)
       • yes → invoke /groundtruth-rules-ai inline; when it returns proposals, arm the clean (status:"armable") ones the SAME way as the literal candidates above (warn, review-status excluded). One flow — literal-arm + AI-propose + approve, all here; do not send the user to separate commands.
       • skip → note it stays available as /groundtruth-rules-ai later.