Repository navigation
fix: scope analyst evidence and align agent phase prompts - #174
Merged
Merged
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
devhims
changed the title
The frame analyst drops sentences now covered by the shared visual scope guidance. The transcript analyst states its scope and warning rules once each. Shared visual guidance no longer claims relevanceContext always carries the research question. The comparison routing rule names visualEvidence and drops repeated branches. Research evidence guidance drops a closing line the completion instruction already provides, and topic research step 6 no longer duplicates step 7.
This branch was successfully deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Problem / Motivation
A multi-video request can produce misleading warnings when an isolated analyst treats its input as the entire research result. For example, a call assigned the 10-second frame can report that the 60-second frame is missing even when another batch handles it. Transcript analysts can similarly warn about other videos or infer incomplete captions from a short transcript.
What changed
Transcript, frame and storyboard analysts now receive explicit instructions to extract evidence from their assigned source or images. Broader questions provide relevance context. Findings and warnings must stay within that local scope, while the finalizer assesses the combined evidence and overall request coverage.
Research prompts now use evidence-collection guidance when a separate finalizer owns the answer. The legacy answer-producing path retains its writing instructions. Routing explicitly requires visual research when saved transcripts cannot establish the requested visual facts. Final-answer instructions also clarify that stored-context gathering has already finished.
Tests
TypeScript checking and 335 automated tests passed. Eleven live Fireworks cases passed across transcript scope, visual scope, genuine readability limitations and routing with saved transcripts. Synthetic image fixtures make the visual checks repeatable.
These are prompt changes, so live checks cannot guarantee every model response. Production validation remains after deployment. Previously saved analyses retain their existing warnings.