v0.12.0 — AI cleanup prompt handoff (Improvements + Unused + CLI)

[jonathan-pap]

Summary

Closes the gap between "Lineage tells you what's dead" and "Lineage helps you delete it" — without breaking the read-only / zero-dep promise. Lineage never mutates the model itself; instead it generates a markdown prompt the user pastes into Claude Code (with Kurt Buhler's pbi-desktop plugin) or any AI agent that can drive TOM / Tabular Editor CLI.

Three surfaces:

• Dashboard: Unused tab gets a top toolbar with a single smart "Generate AI cleanup prompt" button — auto-picks category from what's flagged. Clicking opens a modal preview with Copy + Download .md.
• Markdown: the improvements.md doc's "unused measures" and "dead-chain" findings each embed a <details> block carrying the same prompt — survives ADO Wiki + GitHub render, so wiki readers get it too.
• CLI: powerbi-lineage prompt --category <c> --report <path> [--output <file>] for batch / CI flows.

The prompt itself carries Stage 1 / Stage 2 ordering, an EXTERNALMEASURE guard, a _-prefix helper-measure flag, a git-commit pre-flight (PBI Desktop's Ctrl+Z does NOT undo TOM changes), and a soft pin to pbi-desktop as the recommended tool. No DAX bodies embedded — the AI is already connected to the live model and can read DAX itself before deleting.

Zero new runtime deps. Tests: 299/299 (was 276) — 14 new for the builder, 4 new for the MD embed, 6 new for the CLI smoke.

Why this design

A read-only doc generator is what users picked us for. Vendoring pbi-desktop (GPL-3.0) would license-contaminate the repo; shelling out to TOM via PowerShell would require a 200 MB NuGet install and break the browser-mode + cross-platform promises; TMDL-on-disk pruning would reimplement what TOM already validates atomically, with a huge test surface for edge cases.

Generating an AI prompt keeps every existing constraint intact and adds the verb users actually wanted. The user re-points Lineage at the project after running the prompt; the audit shrinks; that's the verification — no diff view needed.

Full decision table and rejected alternatives in claudedocs/intake-specs/design-ai-cleanup-handoff.md (gitignored — internal-only).

Commits in this PR

1. cf2e2e0 — Stops 1–4: pure prompt builder + MD embed + modal + Unused-tab toolbar + inline-script export-stripper hardening.
2. 951ca69 — Stops 5–6: CLI subcommand + version bump + CHANGELOG + WHATS-NEW + readme badge.

Test plan

Automated (already green):

• npm test — 299/299 pass
• npm run build — clean

Manual (defer to merge / pre-tag):

• Open a real .pbip in the dashboard. Confirm the Unused-tab toolbar appears when there are flagged measures and doesn't when there aren't.
• Click Generate AI cleanup prompt — modal opens with the correct category and a count matching what'll end up in the prompt (EXTERNALMEASURE proxies excluded even if they're in the cards below).
• Modal Copy + Download both produce identical content.
• Esc, backdrop click, and the × button all close the modal.
• Documentation → Improvements: <details> block expands cleanly; prompt is select-all-able. Paste into a test ADO Wiki page + a GitHub gist — both render correctly.
• CLI: node dist/app.js prompt --category measures-all --report <real-pbip> prints a valid prompt; --output cleanup.md writes the same bytes to a file.
• End-to-end (the load-bearing one): paste the generated prompt into Claude Code with the pbi-desktop plugin installed against a sacrificial .pbip. Delete one disposable test measure via the AI. Reopen the project in Lineage. Confirm the audit shrinks by exactly that measure.

Notes for review

• The inline-script export-stripper change in src/html-generator.ts is the load-bearing infrastructure fix — ai-prompts.ts has export function declarations (it's a dual-use module for both server and inline-script), and the old stripper only handled trailing empty export {};. Caught mid-implementation when the dashboard rendered blank; covered now by every inline script in the bundle parsing via new Function() at smoke-test time.
• findDeadChainMeasures is duplicated in src/ai-prompts.ts to keep that module self-contained (so it can concat into the dashboard's classic-script bundle without dragging improvements.ts along). Both copies are exercised by tests on each side.

🤖 Generated with Claude Code