projects/agent-native/slices/init-skill-install/briefs/d3-r1.md
Pre-staged during D2's run; confirm against D2's report before issuing. Deviation from the slice plan's original sketch: evidence returns in the implementer's report (the implementer is read-only on
projects/); the orchestrator appends it toplan.md § Verification.
Prove the slice end-to-end and file the externalities: run the freshly built CLI's
prisma init for real in a scratch directory (live network) and capture the skill
installation; simulate an install failure and capture the non-fatal path; file the
per-ORM-minor tagging ask on prisma/skills.
In: building the CLI in the worktree (turbo build --filter=prisma or equivalent);
scratch-directory runs OUTSIDE the repo (e.g. under /tmp); gh issue create on
prisma/skills; code changes ONLY if the real runs prove something broken (separate commit +
report — do not silently redesign).
Out: everything else. No new features; no test changes unless a real-run fix requires them.
init (default path, no flags) exits 0 and
produces .claude/skills/prisma-*/SKILL.md, .agents/skills/prisma-*/SKILL.md, and
skills-lock.json; the summary lists them. Captured: exact command, trimmed output,
find listing of the produced skill paths.SKILLS_CLI_VERSION
unreachable via a temporary local override, or network cut for the child — choose the
least invasive; document the mechanism) exits 0, prints the warning with the manual
command. Captured: command + output.--no-skills real run: exits 0, produces no .claude/ / .agents/ entries.gh issue create --repo prisma/skills requesting per-ORM-minor
release tags (context: the ORM CLI now installs from the default branch and wants
#v<minor> pinning; reference prisma/prisma TML-2968 work). Captured: issue URL.## Verification evidence heading,
ready for verbatim inclusion in the PR description.Stay focused on the goal; control scope. If a real run fails due to a bug in D1/D2 work, fix it (separate commit), re-run the affected capture, and report; if it fails due to something structural (spec assumption false), halt and surface.
skills CLI's real behavior diverges from the spec's assumptions
(e.g. writes different paths, prompts despite -y); network unavailable; gh not
authenticated for prisma/skills issue creation (capture the prepared issue body in the
report instead and say so).