[nova] feat: attach existing Chromium apps through explicit local CDP endpoints

[bigduu]

User outcome

Part of bigduu/Zenith#185. The user wants Chromium-based applications to participate in Nova interaction UX, including discovery when debugging has not been opened. This issue delivers the bounded connection half; independent application discovery and supported enablement remain separately reviewable slices.

Connect Nova's existing DevTools MCP integration to an explicitly selected, already-running local Chromium application endpoint. Preserve the application and Bodhi windows when Nova connects, disconnects, or fails.

Verified current behavior

At Nova master 8f029d22dbc9ab317a68560788d1e6360b71120e, src/chrome_devtools.rs only exposes isolated Chrome and existing stable Chrome auto-connect. The pinned chrome-devtools-mcp@1.8.0 already supports explicit browser HTTP and browser WebSocket endpoints, pageId targeting, and disconnecting without closing a caller-owned browser.

Acceptance slice (4–6 hours, Nova only)

• Add explicit endpoint attachment to the existing launcher; preserve isolated mode as the default. HTTP browser URL and browser WebSocket URL inputs are mutually exclusive with each other and with launch/auto-connect options.
• Validate endpoint syntax before spawning the sidecar: explicit loopback IP literal and port, appropriate HTTP/WS schemes, no userinfo or fragment, and a documented browser-level path contract. Do not echo rejected credentials in errors. Preserve the fixed upstream version and current telemetry/network-header privacy defaults.
• Clearly state the actual boundary: these are trusted local endpoint inputs, not network confinement. Pinned Puppeteer follows HTTP discovery results and WebSocket redirects. Do not claim input validation prevents all remote traffic; a hardened transport is a separate acceptance slice.
• Attachment failure must not launch another browser. Detaching or replacing this MCP runtime must leave the target app and Bodhi running. No headless or application launch flags in endpoint mode.
• Document list_pages followed by explicit pageId targeting for reads/actions, including two-window examples. Selecting an application for discovery does not itself grant full CDP control.
• Explain browser-level CDP versus renderer-only sockets and Node main-process inspector endpoints. Electron/CEF compatibility must be stated per tested runtime; upstream officially supports Chrome/Chrome for Testing, so successful tools/list alone is not an Electron compatibility claim.
• Connection/setup errors give actionable instructions without requiring Bodhi to quit. Existing native AX and extension capabilities remain independently available.

In-scope files/systems

src/chrome_devtools.rs, launcher CLI/e2e tests, README and plugin documentation; a URL parser dependency only if needed. No new CDP engine, proxy, daemon, or provider router.

Test plan

• Fake-npx real-process tests for literal IPv4/IPv6 arguments, mutual exclusions, rejection before spawn, private defaults, and stdio transparency.
• A test-owned local Chromium fixture and pinned upstream handshake/list_pages; disconnect/SIGTERM preserves the fixture process, failed connection does not launch Chrome, and no Browser.close is sent to an attached browser.
• Any advertised Electron version requires its own test-owned two-window fixture with exact pageId snapshot/fill/click and observed UI effect. Otherwise label Electron support experimental/unverified.
• Do not use the user's daily browser, profiles, or installed apps as automated fixtures.

Non-goals

Automatic app discovery, enabling disabled debugging, restarting or altering installed apps, consent persistence, a menu-bar picker, Chrome extension provider unification, arbitrary raw CDP commands, and remote endpoint support.

Evidence

• Nova launcher
• Pinned upstream options
• Pinned upstream ownership cleanup
• Pinned Puppeteer discovery and redirect behavior

Coordination

Unclaimed proposal. GitHub Projects GraphQL is rate-limited on 2026-09-05; Ready/capacity and board placement are unverified. Check those before implementation rather than treating this issue as claimed.

Public API clarification (2026-09-05)

The user requires these mechanisms to remain internal, with simple application-level public methods. Treat endpoint flags in this slice as internal/advanced transport plumbing, not the normal user workflow. Do not add a separate public vocabulary of scan-port/select-protocol/connect-CDP steps that the caller must learn. The application-level consumer should use #42 discovery internally and accept an app selector; normal callers should not copy a discovered endpoint or configure a second MCP server manually. Preserve existing advanced launcher compatibility. Full internal attachment/provider routing still needs its own bounded acceptance slice before claiming the finished app UX.

Current execution is on #42 only. This issue remains unclaimed; do not infer implementation from the tracker update.

[bigduu]

🔒 claimed by codex-root at 2026-10-02T05:03:47+08:00

Continuing the explicitly requested Nova issue batch in an isolated cumulative worktree after #38 code/review handoff. The user has authorized this named issue; current Roadmap status is Triaged, with no existing claim, branch or competing Nova implementation.

Acceptance slice (4–6 hours): advanced/internal explicit local endpoint attachment in the existing pinned DevTools launcher, pre-spawn validation and literal argv/privacy tests, pageId targeting documentation, and a test-owned Chrome for Testing fixture proving reads/actions, failure without fallback launch, and detach/SIGTERM without closing the target app or Bodhi. Preserve isolated default; normal app-level discovery/selection remains #42 and is not replaced by manual port setup.

In scope: src/chrome_devtools.rs, launcher tests, README/plugin docs only. No new CDP engine, proxy, provider routing, automatic discovery, application restart, permission or production signing changes. Electron/CEF stay experimental/unverified unless exact runtime acceptance is performed. Ordered merges remain gated on #34 live acceptance.