feat: add image support

[mike-inkeep]

• Add image handling support (without persistence to conversation history)
• Tweak image URL schema and tighten up tests

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agents-api	Ready	Preview, Comment	Feb 14, 2026 2:43am
agents-manage-ui	Ready	Preview, Comment	Feb 14, 2026 2:43am

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
agents-docs	Skipped		Feb 14, 2026 2:43am

[changeset-bot]

⚠️ No Changeset found

Latest commit: 633f5c0

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes no changesets

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[claude]

MAJOR Type guard uses inline type instead of imported ImageContentItem

The type predicate re-declares the type inline instead of using the imported ImageContentItem type from ../types/chat. This creates maintenance risk if the type changes.

Suggested change

	export const isImageContentItem = (
	item: ContentItem
	): item is { type: 'image_url'; image_url: { url: string; detail?: 'auto' | 'low' | 'high' } } => {
	return (
	item.type === 'image_url' &&
	'image_url' in item &&
	imageUrlSchema.safeParse(item.image_url?.url).success
	);
	};
	export const isImageContentItem = (
	item: ContentItem
	): item is ImageContentItem => {
	return (
	item.type === 'image_url' &&
	'image_url' in item &&
	item.image_url != null &&
	imageUrlSchema.safeParse(item.image_url.url).success
	);
	};

[miles-kt-inkeep]

@inkeep docs updates?

[claude]

PR Review Summary

3 Key Findings | Risk: Medium

🟠⚠️ Major (3) 🟠⚠️

🟠 1) message-parts.ts Missing dedicated unit tests for new utility module

files: agents-api/src/domains/run/utils/message-parts.ts

Issue: The new message-parts.ts utility module (149 lines) exports 3 critical functions (extractTextFromParts, getMessagePartsFromOpenAIContent, getMessagePartsFromVercelContent) with zero dedicated unit tests. These functions handle content format conversion between OpenAI/Vercel formats and the internal A2A Part[] format, including URL validation, data URI parsing, and error logging for dropped parts.

Why: This has been flagged in multiple prior reviews but remains unaddressed. Without direct unit tests, several failure modes could go undetected: malformed base64 data passing validation, invalid URIs not being rejected, and valid content being silently dropped. Per AGENTS.md, all new features require comprehensive unit tests. The codebase has a consistent pattern of utility tests (see model-resolver.test.ts, stream-helpers.test.ts).

Fix: Create agents-api/src/__tests__/run/utils/message-parts.test.ts with test suites for:

• extractTextFromParts: empty arrays, single text part, multiple text parts joined with space, mixed part types
• getMessagePartsFromOpenAIContent: string input, array input, mixed valid/invalid content logged and dropped
• getMessagePartsFromVercelContent: backwards-compat string content, parts array, invalid parts logged and dropped
• buildFilePart (internal but critical): HTTP URLs, data URIs, invalid URIs throw

Refs:

• Existing utility test pattern
• AGENTS.md testing requirements

---

🟠 2) chatDataStream.ts:56 Schema uses z.any() for content field

files: agents-api/src/domains/run/routes/chatDataStream.ts:56

Issue: The message schema uses content: z.any() which allows arbitrary data to pass validation, bypassing type safety. While parts is validated separately, the content field accepts any value, allowing illegal states like { content: 123, parts: [...] } to reach the handler.

Why: Any value can be passed as content and will pass Zod validation. While the code handles typeof content === 'string' as a backwards-compatibility path, other invalid content types could cause runtime errors. This was also noted by @amikofalvy referencing PR #1789 for schema reuse.

Fix: Replace z.any() with a proper union that captures the valid states:

content: z.union([z.string(), z.undefined()]).optional(),

The actual content extraction happens from parts, and content is only used for backwards-compatibility string mode.

Refs:

• Prior review comment from @amikofalvy

---

🟠 3) scope: documentation Missing documentation for image support feature

scope: Documentation Site, API Reference

Issue: This PR adds image support to both chat API endpoints but no customer-facing documentation was updated. The OpenAPI schema now includes image_url content types (for OpenAI-compatible endpoint) and image parts (for Vercel data stream endpoint), but the existing Chat API guide at agents-docs/content/talk-to-your-agents/chat-api.mdx only documents text-only message content.

Why: Developers integrating with the chat APIs will not know they can send images. The current Request Body Schema example shows only {"content": "Hello!"}. Users wanting to build vision/multimodal features will miss this capability or send incorrectly formatted requests. The @miles-kt-inkeep comment asking about docs updates remains unaddressed.

Fix: Update agents-docs/content/talk-to-your-agents/chat-api.mdx to document:

1. Supported formats: PNG, JPEG, WebP (NOT GIF per team decision)
2. Examples for both endpoint formats (OpenAI image_url vs Vercel image)
3. detail parameter: Only affects OpenAI-compatible providers
4. Limitations: Images are not persisted to conversation history

Refs:

• @miles-kt-inkeep comment

🟡 Minor (1) 🟡

• 🟡 Minor: message-parts.ts:14-18 Type guard uses inline type instead of imported TextContentItem

💭 Consider (1) 💭

• 💭 Consider: message-parts.ts:29-37 Use z.discriminatedUnion() for better error messages

🕐 Pending Recommendations (1)

• 🟠 chatDataStream.ts:73 @amikofalvy suggested reusing schema definitions from PR #1789

---

💡 APPROVE WITH SUGGESTIONS

Summary: This PR introduces well-architected image support with proper URL validation, format transformation between OpenAI/Vercel/A2A protocols, and solid AI SDK integration. The previous review feedback regarding type guards, null checks, and URL validation has been addressed in the latest commits. The core implementation is sound and follows existing patterns.

The remaining gaps are: (1) missing unit tests for the new message-parts.ts utility module (repeatedly flagged), (2) the permissive z.any() schema for content, and (3) missing documentation for customers. Recommend addressing these before or shortly after merge to maintain the codebase's testing standards and ensure customers can discover and correctly use this new capability.

---

Discarded (8)

Location	Issue	Reason Discarded
Agent.ts:93-104	Local AI SDK types duplicate Vercel SDK types	Low confidence; may be intentional to avoid coupling to experimental SDK APIs
chatDataStream.ts:65-66	text field for image URL unconventional naming	Intentional — matches Vercel AI SDK v5 conventions
chat.ts:329-340	Images not persisted to conversation history	Intentional per PR description — working as designed
chat.ts:8-20	GIF rejection inconsistent between URL and data URI	Design decision deferred pending frontend capability exposure
message-parts.ts:105-110	Dropped content warning not actionable	Low priority nitpick; current logging is functional
Agent.test.ts:12	Test helper naming	Positive feedback — follows requested convention
Agent.test.ts:1831	Missing tests for all detail values	Low confidence; existing tests cover the mechanism
message-parts.ts:79-84	Function naming inconsistency (extract vs get)	Developer preference; both patterns exist in codebase

Reviewers (6)

Reviewer	Returned	Main Findings	Consider	While You're Here	Inline Comments	Pending Recs	Discarded
pr-review-standards	0	0	0	0	0	0	0
pr-review-consistency	6	0	1	0	1	0	4
pr-review-types	4	1	0	0	0	0	3
pr-review-tests	3	1	0	0	0	0	2
pr-review-product	5	0	0	0	0	1	4
pr-review-docs	1	1	0	0	0	0	0
Total	19	3	1	0	2	1	13

[claude]

🟡 MINOR: Type guard uses inline type instead of imported TextContentItem

Issue: The type predicate uses an inline intersection type { type: 'text'; text: string } & ContentItem instead of the exported TextContentItem type from ../types/chat.

Why: This creates inconsistency with isImageContentItem (lines 20-27) which correctly uses the imported ImageContentItem type. If TextContentItem changes, this guard won't be updated, creating maintenance risk.

Fix: (1-click apply)

Suggested change

	const isTextContentItem = (
	item: ContentItem
	): item is { type: 'text'; text: string } & ContentItem => {
	return item.type === 'text' && 'text' in item && typeof item.text === 'string';
	};
	const isTextContentItem = (item: ContentItem): item is TextContentItem => {
	return item.type === 'text' && 'text' in item && typeof item.text === 'string';
	};

Refs:

• isImageContentItem pattern on line 20

[claude]

💭 Consider: Use z.discriminatedUnion() for better error messages

Issue: The vercelMessageContentPartSchema uses z.union() instead of z.discriminatedUnion() on the type field.

Why: With z.union(), when validation fails, Zod attempts both branches and reports combined errors, which can be confusing. With discriminatedUnion, Zod immediately knows which branch to validate based on the type field, producing clearer error messages.

Fix:

const vercelMessageContentPartSchema = z.discriminatedUnion('type', [
  textContentPartSchema,
  imageContentPartSchema,
]);

Refs:

• chat.ts uses discriminatedUnion on line 60
• PartSchema in schemas.ts uses discriminatedUnion

[claude]

PR Review Summary

0 New Findings | Risk: Medium (unchanged from prior review)

Delta Analysis

This re-review was triggered by commit 1da57327 — a merge from main into the feature branch. The merge brought in unrelated changes (primarily SchemaProcessor for JSON schema normalization from PR #1915).

The image support code is unchanged since the previous review. All 9 PR files have no diff in the delta:

• message-parts.ts — unchanged
• chat.ts, chatDataStream.ts — unchanged
• types/chat.ts — unchanged
• Agent.ts — only change is an unrelated SchemaProcessor import
• Test files — unchanged

🕐 Pending Recommendations (3)

The following issues from the prior review remain unaddressed:

• 🟠 message-parts.ts Missing dedicated unit tests for the new utility module (149 lines, 3 exported functions)
• 🟠 chatDataStream.ts:56 Schema uses z.any() for content field — bypasses type safety
• 🟠 scope: documentation Missing documentation for image support feature (also flagged by @miles-kt-inkeep in comment)

Additionally, there are active inline suggestions that could be applied:

• 💭 message-parts.ts:29-37 Consider using z.discriminatedUnion() for better error messages
• 🟡 message-parts.ts:14-18 Type guard uses inline type instead of imported TextContentItem

---

💡 APPROVE WITH SUGGESTIONS

Summary: The image support implementation is solid and well-architected. The core functionality is working correctly with proper URL validation, format transformation between OpenAI/Vercel/A2A protocols, and AI SDK integration. However, the three major items from the prior review (unit tests, schema safety, documentation) remain unaddressed. These are important for maintainability and discoverability of this new capability. Recommend addressing before or shortly after merge.

---

Discarded (0)

No new findings to discard — delta contained no substantive changes to review.

Reviewers (0)

Reviewer	Returned	Main Findings	Consider	While You're Here	Inline Comments	Pending Recs	Discarded
Total	0	0	0	0	0	3	0

Note: No reviewers dispatched — delta analysis determined the image support code is unchanged since the prior review. This re-review surfaces pending recommendations only.

[claude]

PR Review Summary

0 New Findings | Risk: Medium (unchanged from prior review)

Delta Analysis

This re-review was triggered by commit 5b5789f6 — a merge from main into the feature branch. The delta contains:

File	Change	Relevance
.changeset/fix-monaco-layout-flickering.md	New changeset	❌ Unrelated
agents-manage-ui/src/components/editors/monaco-editor.tsx	automaticLayout: false	❌ Unrelated

The image support code is unchanged since the previous review. All 9 PR files (message-parts.ts, chat.ts, chatDataStream.ts, types/chat.ts, Agent.ts, generateTaskHandler.ts, test files) have no modifications in the delta.

🕐 Pending Recommendations (5)

The following issues from prior reviews remain unaddressed:

From Previous Review Summary:

• 🟠 message-parts.ts — Missing dedicated unit tests for new utility module (149 lines, 3 exported functions). Per AGENTS.md: "ALL new work MUST include comprehensive unit tests."
• 🟠 chatDataStream.ts:56 — Schema uses z.any() for content field, bypassing type safety
• 🟠 scope: documentation — Missing documentation for image support feature (also flagged by @miles-kt-inkeep)

Active Inline Suggestions:

• 💭 message-parts.ts:29-37 — Consider using z.discriminatedUnion() for better error messages
• 🟡 message-parts.ts:14-18 — Type guard uses inline type instead of imported TextContentItem

From Human Reviewers:

• 🟠 chatDataStream.ts:73 — @amikofalvy suggested reusing schema definitions from PR #1789

---

💡 APPROVE WITH SUGGESTIONS

Summary: The image support implementation is solid and well-architected. The core functionality provides proper URL validation, format transformation between OpenAI/Vercel/A2A protocols, and clean AI SDK integration. The delta since the last review contains only unrelated Monaco editor changes — the image support code is unchanged.

However, the three major items flagged in prior reviews remain unaddressed:

1. Unit tests — message-parts.ts (149 lines, 3 exported functions) has no dedicated test file
2. Schema safety — z.any() on line 56 of chatDataStream.ts bypasses type validation
3. Documentation — Customers have no guidance on how to use this new image capability

Recommend addressing these before or shortly after merge to maintain codebase standards and ensure customers can discover and correctly use this feature. 🎉

---

Discarded (0)

No new findings to discard — delta contained no substantive changes to review.

Reviewers (7)

Reviewer	Returned	Main Findings	Consider	While You're Here	Inline Comments	Pending Recs	Discarded
pr-review-standards	0	0	0	0	0	0	0
pr-review-consistency	0	0	0	0	0	0	0
pr-review-tests	0	0	0	0	0	1	0
pr-review-types	0	0	0	0	0	2	0
pr-review-docs	1	0	0	0	0	1	0
pr-review-product	0	0	0	0	0	0	0
pr-review-llm	0	0	0	0	0	0	0
Total	1	0	0	0	0	5	0

Note: All reviewers correctly identified that the delta contains no image support code changes. Pending recommendations reference prior review findings.

[claude]

PR Review Summary

0 New Findings | Risk: Low (unchanged from prior review, human-approved)

Delta Analysis

This re-review was triggered by commit 99b224e3 — a merge from main into the feature branch. The delta (58 commits, 458 files) brings in unrelated features from main:

Feature	Files Changed	Relevance to Image Support
Skills integration	Agent.ts, PromptConfig.ts	❌ Unrelated
Error unwrapping	Agent.ts	❌ Unrelated
Telemetry flush	chat.ts, chatDataStream.ts	❌ Unrelated
Slack Work App	Multiple packages	❌ Unrelated
Scheduled triggers	New routes/services	❌ Unrelated

The image support code is unchanged since the previous review:

• message-parts.ts — unchanged
• types/chat.ts — unchanged
• Image handling in Agent.ts — unchanged
• Test coverage — unchanged

🕐 Pending Recommendations (3)

The following items from prior reviews remain unaddressed. These are noted for visibility but do not block merge given the human approval:

Severity	Location	Issue	Prior Reference
🟠 Major	message-parts.ts	Missing dedicated unit tests for new utility module (149 lines, 3 exported functions)	Prior review
🟠 Major	chatDataStream.ts:57	Schema uses z.any() for content field — bypasses type safety	Prior review
🟠 Major	scope: documentation	Missing documentation for image support feature	@miles-kt-inkeep comment

Active Inline Suggestions (2)

These suggestions from prior reviews can still be applied:

• 💭 message-parts.ts:29-37 — Consider using z.discriminatedUnion() for better error messages
• 🟡 message-parts.ts:14-18 — Type guard uses inline type instead of imported TextContentItem

---

✅ APPROVE

Summary: The image support implementation is solid and well-architected — it provides proper URL validation, format transformation between OpenAI/Vercel/A2A protocols, and clean AI SDK integration. The delta since the last review contains only unrelated merges from main (Skills, Slack Work App, telemetry, etc.) with no changes to the image support code.

Aligning with @amikofalvy's approval. The pending recommendations (unit tests, schema safety, documentation) are noted for potential follow-up but do not block this feature from shipping. 🎉

---

Discarded (0)

No new findings to discard — delta contained no substantive changes to the image support implementation.

Reviewers (0)

Reviewer	Returned	Main Findings	Consider	While You're Here	Inline Comments	Pending Recs	Discarded
Total	0	0	0	0	0	3	0

Note: No reviewers dispatched — delta analysis determined the image support code is unchanged since the prior review. This re-review surfaces pending recommendations only.

[claude]

PR Review Summary

(0) New Issues | Risk: Low (unchanged from prior review, human-approved)

Delta Analysis

This re-review was triggered by commit 633f5c07 — a cleanup commit for generateTaskHandler.ts. The delta contains:

File	Change	Assessment
generateTaskHandler.ts	Removed duplicate data part extraction logic (now handled by Agent.generate()); improved error message	✅ Good cleanup
generateTaskHandler.test.ts	Updated expected error message assertion	✅ Test updated correctly

Changes in this delta:

1. Removed DataPart import — No longer needed since data extraction moved to Agent.ts
2. Simplified validation logic — Error condition now checks !textParts.trim() && !hasImages && !hasData which is cleaner and more accurate
3. Improved error message — Changed from "No text content found in task input" to "No content found in task input" (reflects that images and data are now valid input types)
4. Reused computed variables — Uses hasData and hasImages in logging instead of recalculating

This is a well-scoped cleanup that removes code duplication introduced when data part handling was consolidated into Agent.generate().

🕐 Pending Recommendations (3)

The following items from prior reviews remain unaddressed. These are noted for visibility but do not block merge given the human approval:

Severity	Location	Issue	Prior Reference
🟠 Major	message-parts.ts	Missing dedicated unit tests for new utility module (149 lines, 3 exported functions)	Prior review
🟠 Major	chatDataStream.ts:57	Schema uses z.any() for content field — bypasses type safety	Prior review
🟠 Major	scope: documentation	Missing documentation for image support feature	@miles-kt-inkeep comment

Active Inline Suggestions (2)

These suggestions from prior reviews can still be applied:

• 💭 message-parts.ts:29-37 — Consider using z.discriminatedUnion() for better error messages
• 🟡 message-parts.ts:14-18 — Type guard uses inline type instead of imported TextContentItem

---

✅ APPROVE

Summary: The delta is a clean refactor that removes duplicate data part extraction logic from generateTaskHandler.ts — this code was already consolidated into Agent.generate() in earlier commits. The improved error message ("No content found" instead of "No text content found") correctly reflects that images and data parts are now valid input types. Aligning with @amikofalvy's approval. 🎉

The pending recommendations (unit tests for message-parts.ts, schema safety, documentation) are noted for potential follow-up but do not block this feature from shipping.

---

Discarded (0)

No new findings to discard — delta is a straightforward cleanup.

Reviewers (0)

Reviewer	Returned	Main Findings	Consider	While You're Here	Inline Comments	Pending Recs	Discarded
Total	0	0	0	0	0	3	0

Note: No reviewers dispatched — delta analysis determined this is a trivial cleanup commit with no new logic to review. This re-review surfaces pending recommendations only.