🚅 Deployed to the rivet-pr-5065 environment in rivet-frontend

Service	Status	Web	Updated (UTC)
kitchen-sink	❌ Build Failed (View Logs)	Web	Jun 5, 2026 at 3:22 pm
frontend-inspector	❌ Build Failed (View Logs)	Web	Jun 3, 2026 at 1:14 am
frontend-cloud	❌ Build Failed (View Logs)	Web	Jun 3, 2026 at 1:14 am
website	😴 Sleeping (View Logs)	Web	May 30, 2026 at 9:03 pm
ladle	✅ Success (View Logs)	Web	May 19, 2026 at 8:30 pm
mcp-hub	✅ Success (View Logs)	Web	May 19, 2026 at 8:20 pm

• donotmerge: dummy inspector api #5085
• donotmerge: insert and clear actor queue #5065 👈 (View in Graphite)
• feat(workflow-engine): add retryOnTimeout opt-in for step timeout retries #5089 : 1 other dependent PR (#5125 )
• fix(rivetkit): wire CancellationToken through waitForNamesAvailable to eliminate KV busy-polling #5063 : 1 other dependent PR (#5050 )
• fix(rivetkit): fix BigInt serialization in inspector workflow history and state deserialization #5058
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

Code Review: PR #5065 — Insert and Clear Actor Queue

Two new inspector HTTP endpoints: DELETE /inspector/queue (clear all messages) and POST /inspector/queue (enqueue a message). The cross-layer wiring (Rust core → NAPI → TypeScript native + wasm runtimes) is consistent with how existing inspector routes are structured.

Bugs / Correctness

reset() doesn't wake waiting consumers (queue.rs:569)

After deleting all messages and resetting metadata, queue_notify.notify_waiters() is never called. Any actor code blocking in queue.next() / queue.next_batch() will keep waiting indefinitely rather than waking and observing an empty queue. Compare with every other write path (e.g. send() at line 346): they all fire self.0.queue_notify.notify_waiters() after the mutation. The reset should do the same, or this choice should at least be documented.

Non-atomic delete → metadata write

The reset deletes message KV entries, then writes the new metadata in a separate kv.put. A process crash between the two leaves persisted metadata with size > 0 pointing to keys that no longer exist. On restart, ensure_initialized() will reload the stale size from KV. The existing send() path uses kv.batch_put to commit message + metadata atomically — consider doing the same here (write metadata first and delete keys second so the count is always conservative, or batch the metadata into the last deletion call).

next_id resets to 0 → IDs repeat from 1

Resetting next_id: 0 means the next send() will issue ID 1 again. If any caller holds an old message ID (e.g. from a POST /inspector/queue response), it will silently alias a newly enqueued message. For an inspector-only tool this is low risk, but it deserves a comment.

Protocol / Architecture

Missing WebSocket inspector parity

Per docs-internal/engine/inspector-protocol.md:

When updating the WebSocket inspector, also update the HTTP endpoints.

The new DELETE and POST operations have no counterpart in the WebSocket inspector (src/inspector/client.browser.ts only handles QueueUpdated / QueueResponse read messages). If the browser-based inspector panel needs these actions, they must be wired into the WS transport as well.

Missing documentation updates

Also per inspector-protocol.md, new endpoints must update:

• Tests in rivetkit-typescript/packages/rivetkit/tests/
• website/src/metadata/skill-base-rivetkit.md
• website/src/content/docs/actors/debugging.mdx (the "Get Queue Status" section at line 435 only documents the GET endpoint)

Minor / Style

InspectorEnqueueBody missing rename_all (mod.rs:244)

Other inspector request bodies (e.g. InspectorWorkflowReplayBody) use #[serde(rename_all = "camelCase")]. The two fields here are single-word so it doesn't change behavior, but the annotation should be added for consistency.

Number(message.id()) precision risk (native.ts:3567)

Message IDs are u64 in Rust. Number() loses precision above 2^53 - 1. The existing GET inspector handler returns m.id directly — the POST response should be consistent. For an inspector-only tool the practical risk is low, but the conversion is wrong in principle.

Error response inconsistency (inspector.rs:133)

The DELETE handler propagates errors with Err(error).context(...), which produces an unstructured 500. The GET handler wraps errors in inspector_anyhow_response() for a structured JSON error body. Consider using the same pattern for write endpoints.

name defaults to "" without validation

InspectorEnqueueBody::name defaults to an empty string via #[serde(default)]. An empty queue message name may be unexpected. A quick validation check or a comment noting this is intentional would help.

Summary

The core idea is clean and the cross-language layering is correct. The main items to address before merging: fire queue_notify on reset, fix the non-atomic delete/metadata ordering, add test coverage, and update the inspector docs per the protocol convention.