feat(aiagent): allow giving messages history #7395

centdix · 2025-12-16T18:10:23Z

Enregistrement.de.l.ecran.2025-12-16.a.21.35.03.mov

cloudflare-workers-and-pages · 2025-12-16T18:16:56Z

Deploying windmill with Cloudflare Pages

Latest commit:	`4b338f6`
Status:	✅ Deploy successful!
Preview URL:	https://9969e789.windmill.pages.dev
Branch Preview URL:	https://fg-arraymessages.windmill.pages.dev

View logs

claude · 2025-12-16T20:42:58Z

Claude finished @centdix's task in 3m 38s —— View job

Code Review Completed ✅

I've completed a comprehensive review of PR #7395 which adds support for providing message history to AI agents. This is a valuable feature that allows users to bypass the built-in memory system and provide custom conversation histories.

Summary

This PR adds a new messages parameter to the AI agent that allows users to provide a custom array of conversation messages. When provided, this bypasses the memory system and uses the explicit messages as the conversation history.

Key Changes:

Backend: Added messages: Option<Vec<OpenAIMessage>> field to AIAgentArgs
Backend: Made user_message optional (was previously required)
Backend: Logic to use explicit messages when provided, bypassing memory read/write
Frontend: Added messages field to AI_AGENT_SCHEMA with proper OpenAI message structure
Frontend: Updated field descriptions and made user_message optional in schema
Frontend: Minor improvements to nullable handling and JSON parsing

Code Quality Assessment

✅ Strengths

Well-structured feature: The implementation cleanly separates the explicit messages path from the memory-based path
Backward compatible: Existing AI agents continue to work as user_message remains functional
Clear validation: Proper validation ensures at least one of messages or user_message is provided
Good documentation: Field descriptions clearly explain when memory is bypassed
Type safety: Proper use of Option types in Rust reflects the optional nature of fields

⚠️ Issues & Recommendations

I've left 5 inline comments on specific code sections with detailed recommendations:

Performance optimization (ai_executor.rs:425) - Unnecessary clone of messages array
Edge case handling (ai_executor.rs:413) - Validation logic for image output type
Documentation clarity (flowInfers.ts:47) - Clarify interaction with system_prompt
Input validation (types.rs:114) - Consider max message array size limit
Code improvement (JsonEditor.svelte:43) - Positive feedback on cleaner error handling

Critical Issue (Not in diff, but important)

The tool_calls and tool_call_id fields in OpenAIMessage (types.rs:46-48) are marked as nullable: true in the frontend schema but use Option<T> without #[serde(default)] in Rust. This could cause deserialization failures when workers are behind the API version.

Recommended fix:

#[serde(skip_serializing_if = "Option::is_none", default)]
pub tool_calls: Option<Vec<OpenAIToolCall>>,
#[serde(skip_serializing_if = "Option::is_none", default)]
pub tool_call_id: Option<String>,

Security Considerations

✅ Good: No SQL injection risks - uses parameterized queries
✅ Good: No XSS risks - data is properly serialized/deserialized
✅ Good: Input validation ensures required fields are present
⚠️ Minor: Large message arrays could consume significant memory - consider adding size limits (see inline comment)

Performance Considerations

✅ Good: No N+1 query patterns
✅ Good: Efficient use of async/await
⚠️ Memory: The .clone() on line 425 could be avoided (see inline comment)
⚠️ Scalability: Large message arrays are cloned and extended multiple times

Testing Instructions

To test this feature through the Windmill app interface, navigate to the Flow Builder and create or edit a flow with an AI Agent step. In the AI Agent configuration, you should now see a new "messages" field (appearing after "messages_context_length" in the form) that accepts an array of message objects. Each message object requires a "role" field (user/assistant/system) and optional "content", "tool_calls", and "tool_call_id" fields. When you provide messages through this array field, the agent will use these as the conversation history instead of loading from memory. You can test the bypass behavior by providing some messages, observing that memory is not updated after execution. Also verify that user_message is now optional - you can provide either messages alone, user_message alone, or both together (where user_message gets appended as the final message). The output_type field should default to "text" and both user_message and the new messages field should only be visible when output_type is set to "text".

Overall Assessment

Rating: 7/10 - Good feature implementation with some opportunities for improvement

Strengths:

Clean separation of concerns
Backward compatible
Well-documented in code
Type-safe implementation

Areas for improvement:

Critical: Add #[serde(default)] to nullable fields in OpenAIMessage
Important: Memory optimization opportunity
Minor: Error handling for edge cases

This is a solid feature that will be useful for users who want more control over conversation history. The implementation is clean and maintainable. Addressing the critical serde issue is recommended before merging to prevent potential deserialization failures in distributed deployments.

claude · 2025-12-16T20:45:47Z

backend/windmill-worker/src/ai_executor.rs

+    if let Some(ref explicit_messages) = args.messages {
+        if !explicit_messages.is_empty() {
+            // Use explicitly provided messages (bypass memory)
+            messages.extend(explicit_messages.clone());


Performance: Unnecessary clone

This .clone() allocates memory for the entire messages array. Since args.messages is never used after this point, consider taking ownership instead.

If you can make args mutable, you could use:

if let Some(explicit_messages) = args.messages.take() { if !explicit_messages.is_empty() { messages.extend(explicit_messages); } }

This avoids the allocation, which could be significant for large message histories.

claude · 2025-12-16T20:45:49Z

backend/windmill-worker/src/ai_executor.rs

+
+    // Validate: at least one of messages or user_message must be provided
+    if !use_explicit_messages && !has_user_message {
+        return Err(Error::internal_err(


Edge case: Consider image output type

When output_type is Image, the user might not provide user_message (since the description states tools and messages are ignored for image generation). However, this validation would still trigger an error.

Consider either:

Exempting image output from this check

Making the error message more specific

Ensuring the frontend prevents this invalid state

claude · 2025-12-16T20:45:50Z

frontend/src/lib/components/flows/flowInfers.ts

+		messages: {
+			type: 'array',
+			description:
+				'An array of conversation messages to use as history. When provided, these messages replace any messages that would be loaded from memory, and memory will not be updated at the end of the run. If user_message is also provided, it will be appended as the final message.',


Documentation clarity

Consider clarifying the interaction with system_prompt:

description: 'An array of conversation messages to use as history. When provided, these messages replace any messages that would be loaded from memory, and memory will not be updated at the end of the run. Note: system_prompt, if provided, will be prepended as the first message. If user_message is also provided, it will be appended as the final message.',

This makes the complete behavior clear to users.

claude · 2025-12-16T20:45:51Z

frontend/src/lib/components/JsonEditor.svelte

 	const dispatchIfMounted = createDispatcherIfMounted(dispatch)

 	function parseJson() {
 		try {


Good fix!

This change properly handles the empty string case by setting value = undefined and clearing the error, while the else branch handles the parsing. This is cleaner than the previous early return pattern.

backend/windmill-worker/src/ai/types.rs

centdix · 2025-12-16T22:32:42Z

/ai change the input schema of the ai agent, instead of adding messages array, we want to have a history field, which is a oneOf.
It is either 'auto', with a single "number" value, or "manual" with an array of messages.
We can then also remove the messages_context_length input.
We should make sure this change is backward compatible with the schema that only had messages_context_length (no need for messages as this was added in this PR).
That back comp should be both on the back and the front.
Add a comprehensive description to the history field.
Do one commit for the backend, one commit for the frontend.
Do one additional commit for correctly rendering a oneOf of that type on InputSchemaTransform, we do not handle that correctly

claude · 2025-12-16T22:36:35Z

Claude encountered an error —— View job

Command failed: git checkout fg/arraymessages --

I'll analyze this and get back to you.

centdix · 2025-12-16T23:01:16Z

/ai change the input schema of the ai agent, instead of adding messages array, we want to have a history field, which is a oneOf.
It is either 'auto', with a single "number" value, or "manual" with an array of messages.
We can then also remove the messages_context_length input.
We should make sure this change is backward compatible with the schema that only had messages_context_length (no need for messages as this was added in this PR).
That back comp should be both on the back and the front.
Add a comprehensive description to the history field.
Do one commit for the backend, one commit for the frontend.
Do one additional commit for correctly rendering a oneOf of that type on InputSchemaTransform, we do not handle that correctly

claude · 2025-12-16T23:04:34Z

Claude finished @centdix's task in 11m 37s —— View job

AI Agent Schema Changes Complete ✅

Successfully replaced the messages array and messages_context_length fields with a unified history field using a oneOf discriminator pattern.

Summary of Changes

Backend (Rust):

Added History enum with two variants:
- Auto { context_length: usize } - automatically manages conversation history with memory
- Manual { messages: Vec<OpenAIMessage> } - bypasses memory and uses explicit messages
Implemented backward compatibility: old messages_context_length parameter automatically converts to Auto mode
Updated message loading and persistence logic to use the new history field
Fixed borrow checker issues with proper referencing

Frontend (TypeScript/Svelte):

Updated AI_AGENT_SCHEMA in flowInfers.ts with new history field schema
Added comprehensive descriptions explaining the behavior of each mode
Updated field order to replace messages_context_length and messages with history
Enhanced ArgInput.svelte to support mode as a discriminator (in addition to existing kind and label)
Added proper type casting to resolve TypeScript strict type checking

Validation:

✅ Backend: cargo check --features $(./all_features_oss.sh) passed
✅ Frontend: npm run check passed (0 errors, 2 pre-existing warnings unrelated to changes)

Commits

feat(backend): Replace messages/messages_context_length with history oneOf field (0a49373)
feat(frontend): Replace messages/messages_context_length with history oneOf field (714b54e)
fix(frontend): Add support for 'mode' discriminator in oneOf rendering (4b338f6)

Backward Compatibility

The changes maintain full backward compatibility:

Existing flows using messages_context_length will automatically work with the new code
The backend converts old-style parameters to the new History::Auto format transparently
No migration needed for existing data

--- |

…oneOf field Replace the separate 'messages' array and 'messages_context_length' fields with a single 'history' field that uses a oneOf discriminator. The 'history' field can be either: - 'auto' mode: automatically manages conversation history with memory, takes a 'context_length' number parameter - 'manual' mode: bypasses memory and uses explicitly provided messages array Backward compatibility is maintained: if 'messages_context_length' is provided in the old schema format, it is automatically converted to 'auto' mode with the specified context_length. Co-authored-by: centdix <[email protected]>

… oneOf field Replace the separate 'messages' array and 'messages_context_length' fields with a single 'history' field in the AI agent schema. The 'history' field uses a oneOf discriminator with two modes: - 'auto': { mode: 'auto', context_length: number } - automatically manages conversation history with memory - 'manual': { mode: 'manual', messages: array } - bypasses memory and uses explicitly provided messages The schema includes comprehensive descriptions for each mode explaining the behavior. The order array has been updated to include 'history' in place of the old 'messages_context_length' and 'messages' fields. Co-authored-by: centdix <[email protected]>

Update ArgInput.svelte to properly handle oneOf schemas that use 'mode' as the discriminator field, in addition to the existing 'kind' and 'label' support. Changes: - Updated tagKey derivation to check for 'mode' first, then 'kind', then 'label' - Added 'mode' to the onOneOfChange function to track mode changes - Added 'mode' to the list of keys excluded from enum validation - Added 'mode' to hiddenArgs to prevent it from being shown in the form - Added title fields to the history oneOf variants in flowInfers.ts This allows the AI agent's history field to properly render with toggle buttons for 'auto' and 'manual' modes. Co-authored-by: centdix <[email protected]>

centdix changed the title ~~feat(aiagent): allow setting messages history for ai agent~~ feat(aiagent): allow giving messages history Dec 16, 2025

centdix marked this pull request as ready for review December 16, 2025 20:42

centdix requested review from alpetric, hugocasa and rubenfiszel as code owners December 16, 2025 20:42

claude bot reviewed Dec 16, 2025

View reviewed changes

backend/windmill-worker/src/ai/types.rs Outdated Show resolved Hide resolved

centdix added 7 commits December 16, 2025 23:00

handle messages array for ai agent

ce2c64c

better

82dcbe4

nit

4bf070e

make tool_calls and tool_call_id nullable

dde6119

fix empty json behavior

4495be2

nits

1162d75

cleaning

f3d5d36

centdix force-pushed the fg/arraymessages branch from ab00ab0 to f3d5d36 Compare December 16, 2025 23:01

github-actions bot and others added 3 commits December 16, 2025 23:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(aiagent): allow giving messages history #7395

feat(aiagent): allow giving messages history #7395

centdix commented Dec 16, 2025 •

edited

Loading

Uh oh!

cloudflare-workers-and-pages bot commented Dec 16, 2025 •

edited

Loading

Uh oh!

claude bot commented Dec 16, 2025 •

edited

Loading

Uh oh!

claude bot Dec 16, 2025

Uh oh!

claude bot Dec 16, 2025

Uh oh!

claude bot Dec 16, 2025

Uh oh!

claude bot Dec 16, 2025

Uh oh!

Uh oh!

centdix commented Dec 16, 2025

Uh oh!

claude bot commented Dec 16, 2025 •

edited

Loading

Uh oh!

centdix commented Dec 16, 2025

Uh oh!

claude bot commented Dec 16, 2025 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

feat(aiagent): allow giving messages history #7395

Are you sure you want to change the base?

feat(aiagent): allow giving messages history #7395

Conversation

centdix commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cloudflare-workers-and-pages bot commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Deploying windmill with Cloudflare Pages

Uh oh!

claude bot commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Code Review Completed ✅

Summary

Code Quality Assessment

✅ Strengths

⚠️ Issues & Recommendations

Critical Issue (Not in diff, but important)

Security Considerations

Performance Considerations

Testing Instructions

Overall Assessment

Uh oh!

claude bot Dec 16, 2025

Choose a reason for hiding this comment

Uh oh!

claude bot Dec 16, 2025

Choose a reason for hiding this comment

Uh oh!

claude bot Dec 16, 2025

Choose a reason for hiding this comment

Uh oh!

claude bot Dec 16, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

centdix commented Dec 16, 2025

Uh oh!

claude bot commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

centdix commented Dec 16, 2025

Uh oh!

claude bot commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

AI Agent Schema Changes Complete ✅

Summary of Changes

Commits

Backward Compatibility

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

centdix commented Dec 16, 2025 •

edited

Loading

cloudflare-workers-and-pages bot commented Dec 16, 2025 •

edited

Loading

claude bot commented Dec 16, 2025 •

edited

Loading

claude bot commented Dec 16, 2025 •

edited

Loading

claude bot commented Dec 16, 2025 •

edited

Loading