Roadmap

What's shipped, what's in flight.

Today — MCP-first, four surfaces

Hover is an open-source Vibe Testing suite organised around four surfaces and one artifact:

Surface	Role	Status
MCP — `@hover-dev/mcp`	author — add it to your own agent; `/mcp__hover__test_app` explores + crystallizes a suite	✅ Shipped (on npm)
VS Code — `hover-dev`	review — an optional Business Map + Dashboard cockpit; drives no agent	✅ Shipped (Marketplace)
CI	run — the crystallized specs run on every PR as plain Playwright	✅ Shipped
Cloud	watch — hosted parallel runs, monitoring, flakiness dashboard, on-failure self-heal	Planned

The through-line is the artifact: owned, portable @playwright/test in your repo and your CI. The AI authors it once; nothing AI runs after.

Recently shipped

API testing. Hover passively captures the app's browser traffic off the same CDP connection (no proxy) and crystallizes the checks worth pinning into a plain *.api-test.spec.ts — verified with replay_request first, so no status code is confabulated.
Self-heal — mode A (local, on-demand). /mcp__hover__heal replays a saved spec against the live app, finds the first drifted step, and re-grounds only that step. Record == replay preserved.
Page Object extraction. NON-login flows repeated across specs are lifted into pages/ + fixtures.ts and the specs folded onto them — deterministically.
Auth as a fixture. A detected login is lifted into an auth.setup.ts setup project that logs in once and reuses storageState; the password is redacted to process.env.HOVER_PASSWORD.
The .hover/ test wiki, self-maintaining. Your app's knowledge compounds across runs and stays honest on its own: progressive recall (recall_business_knowledge returns an index once rules pile up; recall_fact pulls one on demand), /mcp__hover__lint (cross-checks the map against your specs + run ledger for drift), relationships (a ## Relationships block records depends-on / shares-state / navigates-to edges between flows), a run log (.hover/log.md, an append-only history), and /mcp__hover__ask (answer a question from the wiki, with citations).

What's next

The wiki, rendered. Surface the new wiki data in the cockpit: relationship edges in the Business Map graph, drift badges from lint, and a timeline from the run log. The data ships today; the views are next. No source uploaded.
CI-driven self-heal (mode B). A red CI run offering to heal the drifted spec automatically. Local, on-demand heal ships today; the CI-triggered path is planned.
Hover Cloud (planned, not shipped). A hosted layer over the specs you already own: parallel runs, scheduled monitoring, and a flakiness dashboard. Authoring always stays local and free; the cloud only ever runs and monitors specs you already own — never authoring lock-in.

Version history

How Hover got here. The engine evolved through the version line below; today it ships as the @hover-dev/mcp server (and the optional VS Code cockpit).

Version	Theme	Status
v0.0.1-poc	Phase 0 — end-to-end feasibility (`claude -p` drives Chrome via CDP)	✅ Shipped
v0.1.x	Phase 1 — Vite plugin + chat UI + persistent service + Save as Spec	✅ Shipped
v0.2.x	Phase 2 — multi-agent (claude + codex), dark widget v2, Result + Findings cards, custom tooltip	✅ Shipped
v0.3.x	`@hover-dev/next` — Next.js 16+ Turbopack-native integration	✅ Shipped
v0.4.x	Click → Suggest fix prompt	✅ Shipped
v0.5.x	Merged Record + Assert workflow (Record / Exists / Says / Equals sub-toolbar)	✅ Shipped
v0.6.x	Voice mode — push-to-talk STT + spoken progress narration, browser-native, zh/en autodetect, Chrome 139+ on-device	✅ Shipped
v0.7.x	Security testing + plugin API — `@hover-dev/api-test` (HTTPS MITM, captured-flow inspector, IDOR / authz probing MCP) + `defineHoverPlugin` manifest API behind it	✅ Shipped
v0.8.x	Multi-framework source attribution + integration overhaul — JSX / Vue / Svelte / Astro `data-hover-source` stamps via the private `@hover-dev/transform-source`; `@hover-dev/next` gains plugin support via `register()`'s second arg	✅ Shipped
v0.9.x	Widget plugin-UI protocol + cursor-agent — `window.__HOVER_WIDGET__` host API (namespaced CSS / DOM mutations / toolbar buttons / overlays / WS message handlers / lifecycle); `@hover-dev/api-test` migrates onto it; `cursor-agent` joins the registry	✅ Shipped
v0.10.x	Multi-tab agent reliability + 3 more agents — system-prompt addendum for popup checkouts / OAuth chains / post-popup state, `pnpm bench-multi-tab` for A/B'ing prompt changes, two-step card+OTP `examples/payment-provider`, `aider` + `gemini-cli` + `qwen-code` in the registry	✅ Shipped
v0.11.x	Spec resilience: ⟳ Re-record + Saved-sessions overlay + FAQ — when a saved spec breaks (UI changed), Re-record button (widget) or `pnpm hover re-record <spec>` (CLI) replays the JSDoc `Original prompt:` against the current UI and overwrites the file. Widget gains the Skills+Specs tabbed overlay (Save-as-Skill later retired — spec + Re-record covers replay; the overlay is now Specs + Seeds). README + docs site gain a top-level FAQ explaining why we don't self-heal at CI time	✅ Shipped
v0.12.x	Security spec recording semantics — `replay_flow` MCP tool gains `intent` + `expectStatus` parameters that record the replay as a security check. Save-as menu sprouts a "Security spec" entry that writes `__vibe_tests__/<slug>.api-test.spec.ts` — plain Playwright with the `request` fixture, one `test()` per recorded check. Also: server `HoverPluginManifest.saveHandlers` + widget `WidgetPluginSpec.saveEntries` plugin extension points	✅ Shipped
v0.13.x	Record/replay parity — (1) visibility prelude in `writeSpec` emit (`{ const el = …; await expect(el).toBeVisible(); await el.<action>; }`) makes UI drift fail in ~3 s with `Locator expected to be visible` instead of 30 s actionability timeout; (2) Record mode emits `page.goto` as the first step (was: spec started on `about:blank` and failed mysteriously); (3) opt-in "Reload before recording" setting aligns record-time and replay-time starting state. Plus a `?drift=on` visibility-drift lab and 9 new vitest cases	✅ Shipped
v0.14.x	Single-Chrome security + landing site — security mode drops its second browser (one debug Chrome; a `hover:service:start` hook sets launch flags), gethover.dev lands, the widget gains `--mode-accent` plugin theming + CJK output mirroring	✅ Shipped
v0.15.x	Structured spec output + CLI mode — named `test.step` Given/When/Then stages, `.hover/<slug>.json` sidecars, Page Objects + fixtures (`hover extract`), `Promise.all` popup / upload / download pairing, the off-by-default AI optimization pass (`hover optimize` — diff-reviewed candidate, original always kept, `// KNOWN BUG` flagging), the `.hover/rules/` seed library (full catalogue ships built-in), terminal-native `hover run`, an optional model API key passed to the spawned CLI's env, the `add`→`setup` rename, and one Clack-style frame across every command. Save-as-Skill retired (spec + Re-record + seeds cover it)	✅ Shipped
v0.16.x–v0.22.x	MCP-first maturity — API testing (passive CDP capture → `replay_request` verify → deterministic `*.api-test.spec.ts`), self-heal mode A (`replay_spec` drift detection + the `/mcp__hover__heal` prompt, plus a "🏥 Heal in agent" CodeLens in the cockpit), Page Object extraction (`detect_shared_flows` + `extract_page_objects`), and auth as a fixture (`auth.setup.ts` setup project + `HOVER_PASSWORD` redaction + scaffolded `playwright.config.ts`)	✅ Shipped (you are here)
VS Code extension (`hover-dev`)	The review cockpit — a Business Map graph of your flows + coverage and a Dashboard (pass / fail / flaky + CI results), optional AI optimize pass, PR-CI workflow generation. Optional — it reviews, it doesn't drive the agent.	✅ Shipped (on the VS Code Marketplace)
`@hover-dev/mcp`	The authoring engine — an MCP server you add to your own agent (BYO-CLI). Grounded actuation tools + `crystallize_spec` + the `test_app` prompt; record == replay, deterministic crystallize.	✅ Shipped (on npm)

For real-time tracking, see github.com/Hyperyond/Hover/issues.