Greptile Summary

This PR adds a new before_dispatch plugin hook that fires synchronously after message_received and before LLM invocation in dispatchReplyFromConfig, enabling plugins to abort dispatch and send a direct reply without consuming any LLM tokens. The implementation is well-structured and follows existing hook patterns.

Key observations:

• Security bypass via absent sessionKey — The hook is guarded by && sessionKey (line 214 of dispatch-from-config.ts), so it silently skips for any message where ctx.SessionKey is falsy. Auth gate plugins that expect universal coverage cannot intercept those messages. The sessionKey field in the event type is declared as required (string), creating an inconsistency between the type contract and the actual call site behaviour.
• queuedFinal inferred rather than captured — On the blocked path, queuedFinal is set to Boolean(beforeDispatchResult.replyText) rather than the actual return value of dispatcher.sendFinalReply(...). If sendFinalReply returns false for the given payload (e.g. silent-reply token), the returned queuedFinal would be inaccurate.
• Fail-open error handling — Hook errors are swallowed and the message is permitted through. This is consistent with other hooks in the codebase but is particularly risky for security hooks; a note in the JSDoc would help plugin authors know to handle errors defensively.
• No new unit tests — The PR reports existing tests continuing to pass but does not add dedicated unit tests for the new before_dispatch path (no hooks.before-dispatch.test.ts, no new dispatch-from-config.test.ts cases covering the blocked/unblocked/no-reply variants).

Confidence Score: 3/5

• The feature works for the common case but has a documented bypass condition and no new unit tests, making it risky to rely on for security-critical gates.
• The core implementation is sound and consistent with the existing hook infrastructure. However, the && sessionKey guard means auth gate plugins cannot universally block all inbound messages — messages without a session key silently bypass the hook entirely. This is a meaningful correctness gap for the primary use-case (authentication gates) described in the PR. Additionally, there are no new unit tests covering the blocked, unblocked, or reply-less variants of the new code path.
• Pay close attention to src/auto-reply/reply/dispatch-from-config.ts — specifically the sessionKey guard on line 214 and the queuedFinal calculation on line 233.

_{Last reviewed commit: e84a114}

All bot review conversations have been addressed — replied inline and pushed a follow-up commit (fafeff9) to capture sendFinalReply return value directly and add fail-open JSDoc documentation.

@vincentkoc @joshavant — would appreciate a review when you have a moment. This adds a before_dispatch plugin hook to enable pre-LLM message interception for security plugins (auth gates, rate limiters, etc.). The motivation and design are detailed in #43418.

The change is small (+104 lines across 3 files), follows existing hook patterns, and all 111 existing tests pass.

This looks like a solid PR to me -- it extends the hooks consistent with such strategies as #20067 and discussed at #36671! Very supportive! More hook power!

This is a super useful feature indeed! LGTM

I think this is a very useful patch. The existing hooks does not have similar capabilities, so that we can hook on the message processing link and do more secure authentication and identification.

Thanks for the contribution @loveyana — the gap you identified was real and the use case is legit.

However, the before_dispatch hook already landed on main via #50444 (commit a10d587). The current implementation uses runClaimingHook (shared hook infrastructure), has a richer event shape (content, body, channel, timestamp), and integrates with routeReplyRuntime for cross-channel delivery.

Closing as superseded — the feature you need is already available. Let us know if the landed version doesn't cover your auth gate use case and we can iterate from there.