Fixes for issues found during the hackathon

[david-hummingbot]

Summary

A collection of fixes for issues found during the hackathon. More fixes will be added to this branch before it's ready to merge.

Notifications: "Clear all" replaces "Mark all read"

The notification bell's header action now deletes the user's notification history instead of only marking it read. Opening the bell still marks everything read, so the unread badge behaves as before.

• Backend: new clear(user_id) in condor/notifications.py, scoped to the caller's own notifications and run under the store's write lock. Exposed as DELETE /api/v1/notifications.
• Frontend: api.clearNotifications(). The bell shows Clear all with an inline two-step confirm (Clear all? · Clear · Cancel) that resets when the dropdown closes.
• Test: clearing empties only the caller's notifications and leaves other users' untouched.

Agent code can no longer freeze or exit the main process

run_code snippets and routines run on Condor's main event loop, and the existing timeout could not stop synchronous code. During the hackathon an agent's snippet ran glob.glob("/**/_theta_derive.py", recursive=True), walked /sys on the main thread for hours, and froze the dashboard, every agent and the shutdown handler. Separately, exit() in a snippet or routine stopped Condor outright (exit code 0, nothing logged).

New condor/runtime/code_guard.py, scoped to the run and any tasks it spawns:

• Timeout interrupts synchronous code via a sys.monitoring line/jump callback and an audit hook, so a stuck snippet is cut at its budget on the offending line.
• Refused: listing /, /proc, /sys, /dev; input(); signalling Condor's own process; time.sleep over 1s; os._exit / os.abort / signal.raise_signal (checked before running).
• exit() / sys.exit() / KeyboardInterrupt become an ordinary error.
• Tasks a snippet left running are stopped when its run ends, but runs Condor starts on its behalf (start_routine) are detached and outlive it. help() no longer opens an interactive pager.
• After a timeout, finally / except / with / async with cleanup gets up to 2s to run, whether the snippet was busy or awaiting, so locks and files are released.

New condor/runtime/loop_watchdog.py logs the loop thread's stack when the event loop is blocked for 10s or more.

Not covered in-process: a single long call into C, and memory use. Both need snippets in a separate worker process, which is a follow-up.

Test plan

• uv run pytest tests/test_notifications.py (29 passed)
• Frontend typecheck and npm run build
• Local dashboard: Clear all button renders in the bell
• uv run pytest (6468 passed), including 39 tests in tests/test_code_guard.py
• Live dashboard via /api/v1/code/run: the recursive glob from / is refused in 2ms, while True: pass is cut at its 3s budget, exit() returns an error, and the dashboard keeps serving
• Manual: Clear all → Cancel keeps the list; Clear all → Clear empties it and the list stays empty after a reload

[greptile-apps]

Confidence Score: 5/5

[High impact] The PR appears safe to merge; the cleanup finding is addressed and no new blocking issue was found.

Summary

This PR adds notification deletion, guards against agent code freezing or stopping Condor, and a watchdog that logs blocked-loop stacks.

Since the last review:

• Agent code gets checks that stop selected calls from taking Condor down.
• A watchdog logs where Condor’s event loop gets stuck.
• The notification bell lets users clear their history with confirmation.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Snippet reaches its deadline] --> B{Handling an exception?}
  B -->|No| C[Interrupt at a safe instruction or audited operation]
  B -->|Yes| D{Cleanup grace ended?}
  D -->|No| E[Allow cleanup]
  D -->|Yes| C
  C --> F[Record timeout]

Loading

_{Reviews (5) · Last reviewed commit: "fix(runtime): let cleanup run when a sni..." · Reviewed by Greptile}