Agent studio and askAI: `sidepanel-js: data-guardrail-violation events dropped silently; guardrail-triggered responses render as one token`

[JudyHFisher]

Description

When an Agent Studio input guardrail fires (e.g. off_topic_questions), the backend emits a data-guardrail-violation event containing a fallbackResponse string. @docsearch/sidepanel-js has no handler for this event type, so the fallback string is dropped and the UI renders only the aborted text-delta chunks — typically a single letter (e.g. "I").

Steps to reproduce

1. In Agent Studio, enable a guardrail, such as off_topic_questions input guardrail on your assistant
2. Mount @docsearch/sidepanel-js with agentStudio: true and that assistant's ID
3. Ask an off-topic question, e.g. "What's the capital of France?"

Raw EventStream (verbatim from the streaming API response):

data: {"type": "start", "messageId": "alg_msg_tVT46Wu8tMHAe2jK", "messageMetadata": {}}
data: {"type": "start-step"}
data: {"type": "text-start", "id": "alg_msg_tVT46Wu8tMHAe2jK"}
data: {"type": "text-delta", "id": "alg_msg_tVT46Wu8tMHAe2jK", "delta": "I"}
data: {"type": "text-end", "id": "alg_msg_tVT46Wu8tMHAe2jK"}
data: {"type": "finish-step"}
data: {"type": "data-guardrail-violation", "data": {"category": "off_topic_questions", "guardrailType": "input", "fallbackResponse": "I'm sorry I couldn't respond to that, I can only answer questions about Workato's documentation."}}
data: {"type": "finish"}
data: [DONE]

The fallbackResponse field contains the full intended message; nothing in the compiled library consumes it.

Suspected root cause

grep of the compiled @docsearch/sidepanel-js@4.7.0 bundle for guardrail or fallbackResponse returns zero matches. The library's stream reducer switches on:

start, start-step, text-start, text-delta, text-end, finish-step, finish, tool-call, tool-calls, tool-input-start, tool-input-delta, tool-input-available, tool-input-error, tool-output-available, tool-output-error, tool-result, data-code

data-guardrail-violation is not in that set, so the event has no effect on the message-parts state.

Expected behavior

The fallbackResponse string from the data-guardrail-violation event renders as the assistant's message.

Actual

Only the first token from the aborted text-delta stream renders. On repeated near-duplicate off-topic asks, the delta content shrinks further (single letter → nothing).

[Attach: Screenshot 2026-07-29 at 2.43.11 PM.png — shows the empty response bar with just a cursor and the feedback buttons, after asking "what's the capital of sweden"]

Environment

• OS: macOS
• Browser: Chrome, Chrome Incognito
• @docsearch/sidepanel-js: 4.7.0 (reproduced on 4.6.2 as well)
• @docsearch/css: 4.7.0
• Panel variant: inline
• agentStudio: true
• Guardrail: off_topic_questions, guardrailType: input

Contrast case

docs.algolia.com uses the same library version (@docsearch/sidepanel-js@4, currently resolving to 4.7.0) and does not exhibit this bug — because its assistant does not have any input guardrails enabled, so data-guardrail-violation events are never emitted. Off-topic behavior there is handled by the model via its system prompt, producing normal text-delta chunks that the library renders correctly.

[Attach: Screenshot 2026-07-29 at 3.15.53 PM.png — EventStream from docs.algolia.com showing only start/start-step/tool-input-* events, no data-guardrail-violation]

Suggested fix

In the stream reducer, add a case for data-guardrail-violation: append a new message part rendering data.fallbackResponse as the assistant's text, replacing (or appending after) the truncated text-* sequence. Any consumer of an Agent Studio input guardrail is currently unusable without this.

[8bittitan]

Thank you for the find! We will get this planned and implemented 👍

[JudyHFisher]

Thank you!