Repository navigation
Prompt reduction - #270
Open
jgindin wants to merge 6 commits into
Open
Prompt reduction#270jgindin wants to merge 6 commits into
jgindin wants to merge 6 commits into
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request optimizes LLM chat history management and token usage by introducing failed/cancelled turn context exclusion, compacting JSON serialization for schemas and history, and dynamically injecting multimodal guidelines into system prompts when image attachments are present. It also adds robust unit tests to verify these behaviors. The review feedback highlights a potential runtime TypeError in ChatCleaner.isUserLayoutSnapshot if msg.content is not a string, and suggests adding a defensive type guard before calling .trim().
Contributor
⚡ A2UI Composer PR PreviewYour automated preview is successfully live (commit
|
Contributor
Author
Browser Verification Evidence:
|
| Scenario | main Request Bytes |
prompt-reduction Request Bytes |
main contents Turns |
prompt-reduction contents Turns |
prompt-reduction Multimodal Guidelines Included? |
|---|---|---|---|---|---|
| Prompt 1 (Text-Only Form & Action) | 144,726 B | 52,590 B (-63.7%) | 1 (formatted snapshot) | 1 (compact snapshot) | false |
| Prompt 2 (Follow-Up Incremental Edit) | 150,478 B | 54,319 B (-63.9%) | 3 (formatted model turn + duplicate snapshot) | 3 (compact model turn, no duplicate snapshot) | false |
| Prompt 3 (Text-Only Dynamic List) | 158,876 B | 56,802 B (-64.2%) | 5 | 5 | false |
Prompt 4 (Include screenshot ON) |
167,123 B | 70,048 B (-58.1%) | 7 (with 3,436 B PNG) |
7 (with 3,436 B PNG) |
true |
Prompt 5 (mockup.png Uploaded) |
150,197 B | 66,339 B (-55.8%) | 1 (with 5,464 B PNG) |
1 (with 5,464 B PNG) |
true |
| Prompt 6a (503 Error + Text-Only Retry) | 150,180 B | 52,449 B (-65.1%) | 1 (leaks failed user text + 5,464 B PNG) |
1 (clean retry turn, 0 B PNG) |
false |
| Prompt 6b (Stopped Stream + Text-Only Retry) | 150,273 B | 52,449 B (-65.1%) | 3 (leaks stopped turn + 5,464 B PNG) |
1 (clean retry turn, 0 B PNG) |
false |
jgindin
force-pushed
the
prompt-reduction
branch
2 times, most recently
from
October 19, 2026 16:18
ba89dbf to
aebfe38
Compare
suyangw-g
reviewed
Oct 19, 2026
suyangw-g
left a comment
suyangw-g
left a comment
Contributor
There was a problem hiding this comment.
Reviewed commit aebfe382 across all 14 modified files. The prompt compaction, catalog $defs pruning, compact snapshot serialization, and failed/cancelled turn context exclusion are well-structured and backed by thorough unit tests. Left 3 inline comments on user snapshot classification in ChatPanelBase, previewUrl MIME header construction in FileIngestionService, and promptId fallback/test coverage in ChatCoordinator, plus 1 minor test assertion nit in chat-panel.spec.ts.
Add memoized text-only and multimodal system prompt signals behind
buildSystemPrompt({includeMultimodalGuidelines}) in
ChatPromptFactoryService. Include multimodal guidelines in
ChatCoordinator and ChatPanelBase only when active context or staged
inputs contain an image or screenshot, and infer missing file MIME
types in FileIngestionService via a2a-attachments helpers.
Serialize COMMON_TYPES_SCHEMA and the active catalog with compact JSON.stringify instead of pretty-printed formatJson. Prune unreferenced $defs entries from the catalog before prompt injection while preserving transitively referenced definitions and without mutating input objects.
Replace the Simple Example and verbose vacation_booking Complex Form Example in ChatPromptFactoryService with a compact form and action example while retaining the dynamic list template example. Document optional updateDataModel.path updates and clarify root vs. relative template paths in Data Binding.
Keep activeDraft() and Monaco editor text formatted via formatJson() while compacting sanitized layout snapshots and model responses written to chatHistory via JSON.stringify(). Add ChatCleaner isUserLayoutSnapshot(), stamp isSnapshot on synced layout messages, track LLM-committed layouts in lastSynchronizedLayout, and compute componentCount for user snapshots in ChatPanelBase.
Filter out USER prompt messages and attachments paired with failed (ERROR or parseError) or cancelled (isCancelled) MODEL outcomes in ChatCoordinator.getFullMessageContext() and systemPrompt while keeping them in chatHistory for UI rendering. Preserve earlier succeeded turns when retries reuse promptId, deduplicate adjacent layout snapshots, and correlate MODEL placeholders by promptId.
Stamp isSnapshot: false on submitted USER prompt messages in ChatCoordinator.submitPrompt and gate USER snapshot inference on ChatCleaner.isUserLayoutSnapshot in ChatPanelBase.visibleChatHistory and getBubbleClass. Construct FileIngestionService previewUrl from the inferred MIME type and base64 payload, restrict the findTargetModelMessageIndex fallback to untagged MODEL turns, and expand unit test coverage.
jgindin
force-pushed
the
prompt-reduction
branch
from
October 20, 2026 15:28
79b56b3 to
59d3a0f
Compare
This branch has not been 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
2 participants
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.
Description
Reduce LLM prompt payload size and prune failed turn history
Reduce prefill token volume across system instructions, schemas, layout snapshots, and conversation history:
Pre-launch Checklist