Slice plan: init-skill-install

Spec: projects/agent-native/slices/init-skill-install/spec.md Branch: tml-2968-s1-install-prisma-skills-during-prisma-init (cut from origin/main; dispatches execute in isolated worktrees; slice PR targets main and excludes projects/ and drive/ files).

Dispatch plan

Dispatch 1: skill-install runner module

• Outcome: packages/cli/src/init/skill-install.ts exists and exports detectRunner(cwd) and installSkills({ cwd }) per the spec's Chosen design (pinned SKILLS_CLI_VERSION, agent list, --yes on the npx path, never-throws failure shape), with a new packages/cli/src/__tests__/skill-install.vitest.ts passing locally: command assembly for npm/pnpm/yarn/bun, the no-package.json fallback to npm, and the { ok: false, manualCommand } failure shape.
• Builds on: The spec's chosen design.
• Hands to: A unit-tested runner interface (installSkills, detectRunner, exported constants) that Init wiring consumes without touching the runner's internals.
• Focus: The module and its unit tests only. No Init.ts changes; no real network execution in tests (execa mocked or injected).

Dispatch 2: Init wiring and integration tests

• Outcome: prisma init calls installSkills after scaffolding unless --no-skills (flag in arg spec + help text); success appends skill locations to the summary; failure prints the manual-command warning and init exits 0. Init.vitest.ts gains integration tests with the runner mocked: default path invokes it, --no-skills skips it, failure is non-fatal.
• Builds on: Dispatch 1's runner interface.
• Hands to: Feature-complete init behavior, fully covered by mocked tests — the state Dispatch 3 verifies end-to-end.
• Focus: Init.ts and its tests only. Do not modify the runner except to fix an interface mismatch surfaced by consumption (report it if so).

Dispatch 3: end-to-end evidence and externalities

• Outcome: The three slice-DoD conditions hold with evidence: (a) a real prisma init run in a scratch directory (built CLI, live network) produces .claude/skills/prisma-*/, .windsurf/skills/prisma-*/, .agents/skills/prisma-*/, and skills-lock.json — command + output captured; (b) a simulated failure (unreachable skills version) leaves exit code 0 with the hint printed — output captured; (c) the per-ORM-minor tagging ask is filed on prisma/skills via gh and its URL recorded (as executed: filing was permission-gated → accepted deferral D14, body preserved for the operator). Evidence lands in a ## Verification block appended to this plan file.
• Builds on: Dispatch 2's feature-complete behavior (and Dispatch 1's manualCommand shape for the failure capture).
• Hands to: Slice at DoD — PR-open evidence for the build workflow to include in the PR description.
• Focus: Verification and filing only; code changes limited to what the real runs prove broken (report, don't silently redesign).

Sizing

Dispatch	Size	dispatch-INVEST note
1	M	One outcome: tested runner. Mechanical scope, no design judgment left (spec fixed it).
2	M	One outcome: wired init. Consumer-side only; substrate untouched.
3	S	Verification + two external filings; binary evidence checklist.

Handoff completeness

Slice-DoD condition (a) ← D3; (b) ← D2 tests + D3 live capture; (c) ← D3. CI-green + reviewer-accept inherited via the slice PR.