feat: add Replicate video task adaptor

[smartchainark]

Summary

Add support for video generation through Replicate platform, enabling access to multiple video models via a single channel.

Supported Models

• kwaivgi/kling-v2.5-turbo-pro - Kling 2.5 Turbo Pro (high quality, ~$0.30/video)
• wan-video/wan-2.2-t2v-fast - Wan 2.2 T2V Fast (fast & cheap, ~$0.04/video)
• luma/ray - Luma Dream Machine
• minimax/video-01 - MiniMax Video

API Flow

1. POST /v1/models/{model}/predictions - Create prediction task
2. GET /v1/predictions/{id} - Poll for status
3. Return video URL on success

Changes

• New: relay/channel/task/replicate/adaptor.go - Task adaptor implementation (~300 lines)
• New: relay/channel/task/replicate/constants.go - Model list and channel name
• Modified: relay/relay_adaptor.go - Register Replicate task adaptor (Type 56)

Testing

• Code compiles successfully
• Follows existing adaptor patterns (vidu, gemini, kling)

---

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added Replicate integration for video generation tasks. Users can submit video generation requests through Replicate with support for multiple models, automatic status tracking, and response conversion to standard OpenAI-compatible formats. Includes built-in error handling and progress monitoring for generated content.

[coderabbitai]

Walkthrough

Adds a complete Replicate channel integration for task handling, including a new TaskAdaptor implementation that manages request construction, API communication, response parsing, and conversion to OpenAI-compatible video format, alongside supporting constants and routing wiring.

Changes

Cohort / File(s)	Summary
Replicate Adaptor Implementation relay/channel/task/replicate/adaptor.go	New TaskAdaptor struct with lifecycle methods (Init, ValidateRequestAndSetAction), request building (BuildRequestURL, BuildRequestHeader, BuildRequestBody), API interaction (DoRequest, FetchTask), response handling (DoResponse, ParseTaskResult), and data conversion (ConvertToOpenAIVideo). Includes helper functions for URL extraction and aspect ratio mapping.
Configuration Constants relay/channel/task/replicate/constants.go	Introduces ModelList with four Replicate model identifiers and ChannelName constant set to "replicate".
Routing and Registration relay/relay_adaptor.go	Wires Replicate channel support into GetAdaptor and GetTaskAdaptor dispatcher methods to instantiate appropriate Replicate adaptor instances based on API type and channel type.

Sequence Diagram(s)

sequenceDiagram
    participant Client as Client Request
    participant Adaptor as Replicate TaskAdaptor
    participant ReplicateAPI as Replicate API
    participant Response as Response Handler

    Client->>Adaptor: ValidateRequestAndSetAction()
    Adaptor->>Adaptor: Validate task request
    
    Client->>Adaptor: BuildRequestURL()
    Adaptor-->>Client: Replicate API URL
    
    Client->>Adaptor: BuildRequestHeader()
    Adaptor-->>Client: Auth + Content-Type headers
    
    Client->>Adaptor: BuildRequestBody()
    Adaptor->>Adaptor: Map size to aspect_ratio<br/>Build input payload
    Adaptor-->>Client: JSON request body
    
    Client->>Adaptor: DoRequest()
    Adaptor->>ReplicateAPI: POST with request
    ReplicateAPI-->>Adaptor: Prediction response
    
    Adaptor->>Response: DoResponse()
    Response->>Response: Parse predictionResponse<br/>Extract output URL
    Response->>Adaptor: ConvertToOpenAIVideo()
    Adaptor->>Adaptor: Map to OpenAI format
    Adaptor-->>Client: OpenAI-compatible video response

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~25 minutes

Possibly related PRs

• feat: vertex veo (#1450) #1659: Adds new channel TaskAdaptor implementations and modifies relay_adaptor.go for provider registration via GetAdaptor/GetTaskAdaptor.
• 即梦和vidu支持 OpenAI sdk生成和查询视频 #2018: Implements task adaptor response handling producing OpenAI-style video responses with ConvertToOpenAIVideo method.
• feat: add ali wan video #2141: Adds new channel task adaptors and wires them into relay_adaptor.go dispatcher functions.

Suggested reviewers

• seefs001
• Calcium-Ion
• creamlike1024

Poem

🐰 Hop, hop, hooray! A Replicate gateway,
Where Replicate flows through our relay,
Headers and bodies dance in delight,
Converting predictions to OpenAI bright,
New models await on the furry highway! 🌟

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 66.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main change: adding a Replicate video task adaptor implementation with complete request/response handling and model support.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

• 📝 Generate docstrings (stacked PR)
• 📝 Generate docstrings (commit on current branch)

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

Tip

Try Coding Plans. Let us write the prompt for your AI agent so you can ship faster (with fewer bugs).
Share your feedback on Discord.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@relay/channel/task/replicate/adaptor.go`:
- Around line 92-93: The direct type assertion req :=
v.(relaycommon.TaskSubmitReq) can panic if v is a pointer or another type;
change it to a safe check that handles both value and pointer forms and returns
an error instead of panicking: use a type switch or two-step assertion (e.g., if
r, ok := v.(relaycommon.TaskSubmitReq); ok { req = r } else if rp, ok :=
v.(*relaycommon.TaskSubmitReq); ok { req = *rp } else { return
fmt.Errorf("unexpected task_request type: %T", v) }), updating the function that
currently performs this assertion so it returns the error path when the type is
not recognized.
- Around line 262-267: The current mapping only sets openAIVideo.Error when
prediction.Status == "failed"; update the condition in the replicate result
handling (the block that assigns to openAIVideo and dto.OpenAIVideoError) to
treat canceled tasks as failures as well (e.g., check prediction.Status ==
"failed" || prediction.Status == "canceled"), and ensure the
dto.OpenAIVideoError.Message uses prediction.Error or a sensible fallback like
"task canceled" when prediction.Error is empty so canceled tasks surface an
error payload.
- Around line 130-170: In TaskAdaptor.DoResponse validate the upstream HTTP
status and that the parsed prediction.ID is present before sending 200 to the
client: check resp.StatusCode is in the 2xx range and prediction.ID is not
empty, and if either check fails produce a TaskError (use
service.TaskErrorWrapper or TaskErrorWrapperLocal with an appropriate code like
"task_failed" or "upstream_error") and return without calling c.JSON; only
create and return dto.NewOpenAIVideo (using info.PublicTaskID and
info.OriginModelName) when both the upstream status and prediction.ID indicate a
successful, pollable task.

---

ℹ️ Review info

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 982dc5c and e14727d.

📒 Files selected for processing (3)

• relay/channel/task/replicate/adaptor.go
• relay/channel/task/replicate/constants.go
• relay/relay_adaptor.go

[coderabbitai]

⚠️ Potential issue | 🟠 Major

Guard task_request assertion to prevent panic.

Line 92 can panic if upstream middleware stores a pointer/value variant. Return an error instead of crashing request handling.

Suggested fix

-	req := v.(relaycommon.TaskSubmitReq)
+	var req relaycommon.TaskSubmitReq
+	switch typed := v.(type) {
+	case relaycommon.TaskSubmitReq:
+		req = typed
+	case *relaycommon.TaskSubmitReq:
+		if typed == nil {
+			return nil, fmt.Errorf("invalid task_request type")
+		}
+		req = *typed
+	default:
+		return nil, fmt.Errorf("invalid task_request type")
+	}

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	req := v.(relaycommon.TaskSubmitReq)
	var req relaycommon.TaskSubmitReq
	switch typed := v.(type) {
	case relaycommon.TaskSubmitReq:
	req = typed
	case *relaycommon.TaskSubmitReq:
	if typed == nil {
	return nil, fmt.Errorf("invalid task_request type")
	}
	req = *typed
	default:
	return nil, fmt.Errorf("invalid task_request type")
	}

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@relay/channel/task/replicate/adaptor.go` around lines 92 - 93, The direct
type assertion req := v.(relaycommon.TaskSubmitReq) can panic if v is a pointer
or another type; change it to a safe check that handles both value and pointer
forms and returns an error instead of panicking: use a type switch or two-step
assertion (e.g., if r, ok := v.(relaycommon.TaskSubmitReq); ok { req = r } else
if rp, ok := v.(*relaycommon.TaskSubmitReq); ok { req = *rp } else { return
fmt.Errorf("unexpected task_request type: %T", v) }), updating the function that
currently performs this assertion so it returns the error path when the type is
not recognized.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

Validate upstream success before acknowledging task creation.

This path can return 200 to the client even when upstream returned non-2xx or no prediction ID, which creates unpollable tasks.

Suggested fix

 	_ = resp.Body.Close()
+
+	if resp.StatusCode < http.StatusOK || resp.StatusCode >= http.StatusMultipleChoices {
+		taskErr = service.TaskErrorWrapperLocal(
+			fmt.Errorf("replicate create prediction failed: status=%d body=%s", resp.StatusCode, string(responseBody)),
+			"upstream_request_failed",
+			http.StatusBadGateway,
+		)
+		return
+	}
 
 	var prediction predictionResponse
 	err = common.Unmarshal(responseBody, &prediction)
@@
 	// Check if task already failed
 	if prediction.Status == "failed" || prediction.Status == "canceled" {
@@
 		return
 	}
+
+	if prediction.ID == "" {
+		taskErr = service.TaskErrorWrapperLocal(
+			fmt.Errorf("replicate response missing prediction id"),
+			"invalid_upstream_response",
+			http.StatusBadGateway,
+		)
+		return
+	}

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@relay/channel/task/replicate/adaptor.go` around lines 130 - 170, In
TaskAdaptor.DoResponse validate the upstream HTTP status and that the parsed
prediction.ID is present before sending 200 to the client: check resp.StatusCode
is in the 2xx range and prediction.ID is not empty, and if either check fails
produce a TaskError (use service.TaskErrorWrapper or TaskErrorWrapperLocal with
an appropriate code like "task_failed" or "upstream_error") and return without
calling c.JSON; only create and return dto.NewOpenAIVideo (using
info.PublicTaskID and info.OriginModelName) when both the upstream status and
prediction.ID indicate a successful, pollable task.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor

Include canceled-task failures in OpenAI error payload.

Line 262 only maps failed; canceled tasks are also failure states and should surface an error message.

Suggested fix

-	// Set error info if failed
-	if prediction.Status == "failed" && prediction.Error != "" {
-		openAIVideo.Error = &dto.OpenAIVideoError{
-			Message: prediction.Error,
-			Code:    "replicate_error",
-		}
-	}
+	// Set error info for all failure states
+	if originTask.Status == model.TaskStatusFailure {
+		msg := prediction.Error
+		if msg == "" {
+			msg = originTask.Reason
+		}
+		if msg != "" {
+			openAIVideo.Error = &dto.OpenAIVideoError{
+				Message: msg,
+				Code:    "replicate_error",
+			}
+		}
+	}

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@relay/channel/task/replicate/adaptor.go` around lines 262 - 267, The current
mapping only sets openAIVideo.Error when prediction.Status == "failed"; update
the condition in the replicate result handling (the block that assigns to
openAIVideo and dto.OpenAIVideoError) to treat canceled tasks as failures as
well (e.g., check prediction.Status == "failed" || prediction.Status ==
"canceled"), and ensure the dto.OpenAIVideoError.Message uses prediction.Error
or a sensible fallback like "task canceled" when prediction.Error is empty so
canceled tasks surface an error payload.

[Calcium-Ion]

Has billing been adapted?