Grilling

Interview user until shared understanding. Walk every branch of decision tree, dependencies one by one. ONE question at a time — wait for answer before next. Facts findable in environment: look up, never ask. Decisions are user's — put each to user, wait. Do not act until user confirms.

Naming: question S«s»Q«q», answer S«s»Q«q»A«n».

Steps

1. resolve-store.sh path — locates store harness already cloned. Resolve, never clone. Exits non-zero naming its own fix when store is unnamed, unresolved or absent; unnamed → say so out loud and record nothing (silent skip is indistinguishable from successful record). Store URL lives in environment only — never hardcoded, never echoed into artifacts.

2. resolve-store.sh unmerged — any records → refuse to open session until principal merges previous one, per recording.md.

3. resolve-store.sh preferences — emits active set as JSON array for session's preferences: copied, never retyped. Inject that set ONLY, never decision history.

4. Author every decision point as session JSON — whole session so far, answers included. Contract: render/decision-context.ts beside this file; semantics: format.md. NEVER hand-format a question; renderer appends free-text slot itself.

5. Render both user-facing forms — validation failure names offending field; fix JSON, re-run:

   node --experimental-strip-types <skill-dir>/render/render.ts <session.json> --out <dir>
   

6. Publish session.html as artifact, redeploying same URL as session grows; unavailable → print session.md into chat verbatim — never as file attachment, never as timed dialog (dialogs close while user still typing).

7. Answers arrive as page's copied JSON pasted into chat, or as chat replies: answer id ("S1Q2A1" or "1"), correction via "N, but actually because …" — shorthand "N, BAB …". No live page-to-agent channel yet (pandoscope/skills#139). What answers mean: reading.md. Follow-ups: append new questions plus answers received, re-render; user may revisit earlier answers.

8. Session end: record every ruling to store, embed rejection reasons in session's target artifact, per recording.md.

9. check.sh <session.json> — re-validates session, verifies citations are verbatim store lines, fails while records stay unmerged, prints residue to verify by hand.

Question rules

• Slot 1 carries prediction: preference-driven when active rules match (named in matches), else cold pick. Diverging recommendation sits at slot 2, if-clause arguing when it beats slot 1.
• Wildcard ONLY when genuinely plausible unexplored branch exists.
• If-clause = when option beats recommendation. Add "why not recommended" only when it differs from negated if-clause.
• Near-ties MUST be marked; never fabricate weaknesses for close calls, and never score them as misses.
• Correction ("N, but actually because …", shorthand "N, BAB …") accepts option, overrides its stated reason — highest-signal event; flag in record.
• Drill down with ONE follow-up (2-3 ranked guesses + free text) only when free-text answer leaves rejection reasons unstated. Guesses count as predictions.

Non-goals

No automatic preference-rule acceptance — human in loop always. No embedding/RAG tooling. No upstream sync.