Compile the browser-automation core for wasm32: lib/bin split, rt abstraction, injected CDP transport, embedding seams

[Glavin001]

Summary

Makes the browser-automation core of agent-browser compile and run as a library on wasm32-unknown-unknown so non-native hosts (e.g. Cloudflare Workers) can embed the command core (dispatch, snapshot/refs, interaction, network, cookies, storage) and drive a remote browser over an injected CDP WebSocket, without forking the automation semantics. Native CLI/daemon behavior is unchanged, and no host-specific code is added.

Building blocks:

1. lib/bin split — adds src/lib.rs so the crate exposes a library (agent_browser) alongside the existing binary; main.rs imports from the lib. Native-only top-level modules (chat, doctor, mcp, skills, upgrade) are gated off the wasm build.

2. rt module (src/rt.rs) — a small runtime abstraction over task spawning and timers:

   crate::rt::{spawn, spawn_blocking, sleep, sleep_until, timeout, timeout_at, interval, Instant, JoinHandle}

   Native delegates to tokio (rt::Instant is std::time::Instant). The wasm implementation drives the same API from the JS event loop (setTimeout, wasm_bindgen_futures::spawn_local, an abortable JoinHandle that tracks completion). The gating rule is deliberate and narrow: only code that is compiled on wasm switches to crate::rt; test modules and cfg-gated native-only functions keep calling tokio directly, exactly as on main.

3. Injected CDP transport (cdp/client.rs) — the WebSocket is abstracted behind:

   pub trait CdpTransportSink { async fn send_text(&self, text: String) -> Result<(), String>; async fn send_ping(&self) -> Result<(), String>; }
   pub enum TransportEvent { Text(String), Close(Option<String>), Error(String) }
   pub fn CdpClient::from_transport(sink: Arc<dyn CdpTransportSink>, events: TransportEventStream) -> Self

   connect/connect_with_headers become thin native wrappers that build the tungstenite transport and call from_transport; command multiplexing, the pending-response map, and event broadcast are unchanged and now portable. BrowserManager::from_client(client, ws_url, direct_page) builds a manager around such a client (native connect_cdp now goes through it too).

4. Library-host seams — output.rs gains begin_capture()/end_capture() -> CapturedOutput: internal outln!/out!/errln! macros buffer the exact CLI output text when a capture is active and print normally otherwise. src/artifacts.rs routes file-producing commands (screenshot, PDF, HAR, diff image) through artifacts::write, which falls back to std::fs::write unless a host installs a writer via set_artifact_writer. Native behavior is identical when no capture/writer is active.

5. Target gating — native-only subsystems (daemon socket server, local Chrome launch/discovery, inspect server, stream/dashboard, WebDriver/Appium transports, ffmpeg recording, plugin subprocesses, install, the read command) are #[cfg(not(target_arch = "wasm32"))]-gated, with explicit "not supported on this platform" errors where a runtime path could be reached. read stays native-only because its client-level timeout and per-hop redirect allowlist cannot be enforced by reqwest's wasm backend. Cargo.toml splits target-specific dependencies: native keeps full tokio, tokio-tungstenite, reqwest-rustls, socket2; wasm adds wasm-bindgen, js-sys, wasm-bindgen-futures with tokio limited to macros+sync. (The wasm reqwest block intentionally omits rustls-tls-webpki-roots: on wasm32 reqwest delegates TLS to the host fetch implementation and the rustls stack does not build there.)

6. wasm-safe gen_id — commands::gen_id derives its microsecond timestamp from js_sys::Date::now() on wasm32; std::time::SystemTime::now() aborts on wasm32-unknown-unknown (found by running the core in workerd, where parse_command trapped).

The Default impls (DaemonState, RefMap, EventTracker, RecordingState, TracingState) and the #[allow(clippy::should_implement_trait)] on WaitUntil::from_str are needed because clippy lints these pub items now that they are on a lib target (CI runs clippy -- -D warnings).

Verification

All from cli/:

• cargo check --all-targets: clean
• cargo check --lib --target wasm32-unknown-unknown: clean (0 errors, 0 warnings)
• cargo clippy --all-targets -- -D warnings: clean
• cargo fmt -- --check: clean
• cargo test: 1065 passed, 0 failed
• Downstream validation: the wasm library was exercised end-to-end in a Cloudflare Worker (workerd) via the injected transport — open/snapshot/screenshot against a real Chrome over CDP, with byte-identical CLI stdout via the output-capture seam.
• Native e2e (cargo test e2e -- --ignored --test-threads=1) was verified earlier to behave identically to main in the same environment; native code paths keep using tokio directly wherever they are not compiled for wasm, so native behavior is unchanged by construction (on native targets crate::rt re-exports the tokio primitives).

Link to Devin session: https://app.devin.ai/sessions/5c759851c2274cd7a02ace63935862ef
Requested by: @Glavin001

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

Tested native parity of this branch end-to-end against real headless Chrome.

Result: no behavioral difference from upstream main observed.

• Full CLI flow on the branch (open → snapshot → click @ref → snapshot → screenshot → close) works; AX ref tree renders as expected (button "Click me" [ref=e2]), and clicking via ref actually mutated the page (heading "Clicked!" in the next snapshot).
• Ran the identical sequence with a main build in a separate worktree: transcripts identical, and the two screenshots are byte-identical PNGs (cmp clean).
• cargo check --lib --target wasm32-unknown-unknown: clean, exit 0.
• cd cli && cargo test: 1065 passed, 0 failed (97 ignored e2e).

Branch CLI transcript (all exit codes 0)

✓ AB Test Page
  http://localhost:8000/test.html
- heading "Hello" [level=1, ref=e1]
- button "Click me" [ref=e2]
✓ Done                      (click @e2)
- heading "Clicked!" [level=1, ref=e1]
- button "Click me" [ref=e2]
✓ Screenshot saved to /tmp/abtest/branch.png
✓ Browser closed

Environmental note on the earlier DevToolsActivePort failure

The "Chrome exited early without writing DevToolsActivePort" failure in this environment is caused by google-chrome being a browser shim script, not a real Chrome. Pointing AGENT_BROWSER_EXECUTABLE_PATH at a real Chrome binary with --args "--no-sandbox --disable-dev-shm-usage" resolves it on both this branch and main.

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 1 additional finding in Devin Review.

[devin-ai-integration]

🟨 wasm reqwest client drops rustls-webpki-roots feature used on native

The new wasm dependency block configures reqwest without the rustls-tls-webpki-roots feature that the native block keeps (cli/Cargo.toml:38 vs cli/Cargo.toml:46). On wasm32 reqwest delegates TLS to the host fetch implementation, so this is not directly exploitable, but the asymmetry means any future non-browser wasm host would get a reqwest build with no configured trust anchors rather than the pinned webpki root set the native build relies on.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

Intentional: on wasm32 reqwest uses the browser/host fetch backend and delegates TLS entirely to the host, so rustls-tls-webpki-roots has no effect there — and the rustls/ring stack it pulls in does not build for wasm32-unknown-unknown. Enabling it in the wasm block would break the wasm build without adding any trust-anchor behavior. Noted the rationale in the PR description.

[devin-ai-integration]

End-to-end test results (Devin)

Verified native CLI behavior is unchanged on this branch by building the release binary (Rust 1.97.1) and exercising it against real Chrome 143 (--no-sandbox --disable-dev-shm-usage).

Golden path (all passed): open → snapshot (refs) → type → click (form submitted) → screenshot → pdf → network har stop <path> → close. Screenshot/PDF/HAR files were actually written through the new artifacts::write route (valid PNG 1280x581, %PDF-1.4 2 pages, HAR JSON with 10 entries).

Output seam: --help output is byte-identical to a freshly built main binary (diff clean), confirming the outln!/out! refactor didn't change stdout formatting.

Full e2e suite against real Chrome

cargo test --release e2e -- --ignored --test-threads=1
test result: ok. 96 passed; 0 failed; 0 ignored; 1044 filtered out; finished in 70.96s