Verification record — generate-skill-offer

Evidence from S2-D3 live runs (2026-07-03; CLI built from 744539ec0; scratch project cloned from the example-project fixture with @prisma/client/@prisma/config symlinked; curated env -i environment; per-run isolated XDG_CONFIG_HOME with prisma-nodejs/ pre-created and commands.json seeded 30 days back per the two suppression traps).

• First interactive run (pty via script -qec): offer shown ("Install Prisma's agent skills so AI coding tools work better with this project? (y/N) — This prompt closes in 30s and can be suppressed with --no-hints"); answer n → exit 0, generate normal, skills-offer.json = {"offeredAt":"2026-07-03T21:19:37.040Z","outcome":"declined"}.
• Second run, same config dir: no prompt of any kind; exit 0.
• Accept path (fresh dir): answer y → S1 runner streams installs; 24 SKILL.md (8 skills × .agents/.claude/.windsurf), skills-lock.json present; generate succeeds; acknowledgement outcome: "accepted".
• Non-TTY (plain pipes): no prompt; generate normal; no acknowledgement written — a gated-out run leaves the offer available for a future interactive session.
• Mutual exclusion: in both prompted runs the skills offer was the only prompt; no NPS question appeared in any run (no active NPS timeframe remotely; the mechanism itself is pinned by S2-D2 unit tests).
• Container/CI gates verified against their real implementations for this sandbox (pid 1 is systemd; no /.dockerenv etc.) — nothing was faked.

Cosmetic observation (expected): on accept, the installer output streams before the "Generated Prisma Client" line because Generate.parse returns the summary for the caller to print after the offer resolves.