Add named Browser REPL executors for concurrent REPL calls

[rgarcia]

Summary

POST /repl runs every call in one Node runtime per browser, so callers can't run independent stateful tasks at the same time, and a timeout or reset wipes everyone's state. This adds named executors to the Browser REPL, matching the model POST /playwright/execute already uses.

• POST /repl takes an optional executor name (^[A-Za-z0-9_-]{1,64}$). Each executor is its own runtime process with its own bindings, modules, and browser connection. Calls on different executors run concurrently; calls on one executor run one at a time.
• Calls without executor run on an always-present default executor that behaves as before.
• A timeout, crash, OOM, or reset only ends that executor's state. A timeout still ends the executor's process: unlike Playwright calls, REPL cells share state with later cells, so code abandoned by a timeout can't be left running.
• Tabs: each named executor owns a background tab the API opens on its first call. Its helpers attach to that tab, and newTab() creates tabs, without making them the active tab, so a named executor doesn't take focus from the live view or the default executor. ensureRealTab() returns to that tab even while it's blank. A named executor never falls back to another page: if its tab is closed mid-cell, helpers fail with a clear error, and the next call opens a new tab without losing state.
• Limits and endpoints: at most 8 named executors (409 with the executor list). GET /repl/executors lists them, default first. DELETE /repl/executors/{name}?close_tab=true stops a named executor; the running call and queued calls fail with "executor was deleted", and its tab is closed. Deleting default restarts it: the running call fails with "executor was restarted", and queued calls run on a new process without waiting. No idle reaping; graceful API shutdown removes named executors and closes their tabs.
• Response fields: new_state (call started from empty state), terminated, and tab: {target_id, created}. repl_id and repl_terminated are kept but marked deprecated.
• Custom WebMCP tools stay owned by default. Named executors don't start the custom-tool registry, because its page-side cleanup would remove default's registrations. webmcp.addCustomTools, listCustomTools, and removeCustomTool throw a clear error there. Every executor can still discover and invoke tools. Moving the registry to a shared tools host so every executor can register tools is a follow-up.
• Heap logging: each runtime keeps the 512 MiB heap limit and logs a warning when it passes 90% (re-armed below 80%). The API watches runtime stderr and logs an error with the executor name when a runtime hits the limit.

Changes

• cmd/api/api/browser_repl.go: browserReplManager is now one executor, with its own socket, an optional tab owner, a recorded tab ID, a restart path that doesn't wait for queued calls, and a retire path for delete and shutdown.
• cmd/api/api/browser_repl_executors.go: registry of named executors, the cap, and the execute/list/delete handlers.
• cmd/api/api/browser_repl_heap.go: stderr watcher for V8 heap-limit aborts.
• cmd/api/api/executor_tabs.go: the tab owner and name pattern, moved out of the Playwright executor code so both executor kinds share them.
• runtime/: executor name and home tab; non-activating attach for named executors; tab_missing answered without running code; attached tab reported in each response; repl.executor; heap monitor.
• openapi.yaml (+ regenerated lib/oapi/oapi.go and lib/events/category_gen.go) and docs/repl.md.

Testing

• New unit tests (browser_repl_executors_test.go) run against the real REPL runtime bundle: concurrency across executors, state isolation, a timeout ending only one executor, tab ownership across a reset, the cap and 409, invalid first calls not holding slots, a failed first call not tearing down a queued call, delete during tab open (tab still closed), delete with in-flight and queued calls, close_tab=false, default restart without waiting for queued calls, idle restart, custom-tool rejection in named executors, a start failure not holding a slot, shutdown closing tabs, and the heap watcher. They passed 3 runs with -race. The full cmd/api unit suite and the runtime tests pass.
• New e2e subtest TestBrowserReplAPI/<image>/named_executors, against real Chromium:
  • a named executor navigates its own tab without changing the active tab (checked through Playwright's default executor);
  • ensureRealTab() stays on a blank home tab;
  • two executors' 2s waits overlap;
  • closing an executor's tab reopens it without losing state;
  • closing it mid-cell fails helpers instead of attaching to another tab;
  • newTab() in a named executor leaves the active tab unchanged;
  • delete closes the tabs.
• The full TestBrowserReplAPI (headless and headful), TestPlaywrightExecuteAPI (including the custom WebMCP subtests), and TestPlaywrightDaemonRecovery pass locally against chromium-headless/chromium-headful images with this branch's API binary and runtime bundles layered on.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit fdf9e8d. Configure here.}