fix(chat): stop conversations dropping silently or on over-eager moderation#38
Merged
Conversation
…empty completions The AI SDK reports a mid-stream provider failure as an in-band fullStream 'error' part (it calls onError but does not throw), and result.response only rejects when zero steps completed. After a completed tool step the error was captured but never surfaced, and an empty model completion rendered as a blank turn — both ended the conversation with no response and no error. - client loop (use-agent-chat.ts): handle the 'error' part and re-throw the captured streamError after the loop; add a fallback assistant message for empty completions / empty sub-agents / step-limit-on-tool; mark tool-error invocations done so chips stop spinning; wrap the sub-agent generator so a failing sub-agent shows a visible error instead of an unhandled tool-error - server gateway (router.ts): re-throw on an 'error' part so the catch emits a proper error chunk instead of a clean [DONE] that the client reads as empty - mock model: add "empty" and "stream error" seams - add chat-silent-drop e2e covering both paths Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…or prompt Writing a description/summary for an application (or dataset, processing, catalog) is legitimate metadata management, but the moderator classified it as "writing an essay" and blocked it. Because the gateway moderates every model call by caller role, this most often hit sub-agent task delegations (e.g. "write a description for application X"), interrupting the conversation with a content_filter refusal. - buildModerationSystemPrompt: add an explicit in-scope rule that authoring/ editing resource metadata (title, description, summary, topics, tags, keywords of an application/dataset/processing/catalog) is always allowed; narrow the off-scope "essay" criterion to general-purpose prose unrelated to a platform resource - add a unit test guarding the metadata whitelist language Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…gent is moderation-blocked A delegated sub-agent that gets content_filter'd previously handed the main assistant the same vague user-facing refusal, so it could only dead-end. Now: - sub-agent moderation blocks return a distinct, assistant-directed notice (SUBAGENT_MODERATION_NOTICE) via a non-rendered `moderationBlocked` flag read by toModelOutput: it states this is a content-policy decision (not a tool error), allows re-formulating a legitimate task or explaining to the user, and discourages resubmitting identical wording. No auto-retry loop — the model decides within stepCountIs(10) and each re-delegation is itself re-moderated. - the user-facing refusal (AgentChat.vue en/fr + DEFAULT_REFUSAL fallback) now names the decision as content moderation generically and invites rephrasing, without leaking a category/reason. - update the moderation e2e refusal constant to match. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes conversations that ended with no useful outcome, on two fronts: silent drops and moderation-driven interruptions.
Silent drops — the AI SDK surfaces a mid-stream provider failure as an in-band
errorpart (it does not throw), andresult.responseonly rejects when zero steps completed. After a tool step the error was captured but never surfaced, and an empty model completion rendered as a blank turn.errorpart (the gateway re-throws so it emits a real error chunk instead of a clean[DONE])tool-errorinvocations settle (no perpetual spinner); a failing sub-agent shows a visible error rather than an unhandled tool-errorModeration — legitimate work was being blocked, especially sub-agent delegations:
Why: users reported conversations ending with no response and no error, and legitimate requests (notably sub-agent tasks like writing a resource description) being interrupted by moderation.
Heads-up:
[DONE]