Skip to content

Open source AI coding agent. Designed for large projects and real world tasks.

fix for occasional model error related to empty messages. improved handling of model errors and retries

Dane Schneiderc2ce82b081a2
diff --git a/app/server/model/client.go b/app/server/model/client.go
index a10855de..e31df002 100644
--- a/app/server/model/client.go
+++ b/app/server/model/client.go
@@ -348,6 +348,8 @@ func (stream *StreamReader[T]) Recv() (*T, error) {
// Extract the data
data := strings.TrimPrefix(line, "data: ")
+ // log.Println("\n\n--- stream data:\n", data, "\n\n")
+
// Check for stream completion
if data == "[DONE]" {
return nil, io.EOF
diff --git a/app/server/model/model_request.go b/app/server/model/model_request.go
index eb74308b..7f0b4ee8 100644
--- a/app/server/model/model_request.go
+++ b/app/server/model/model_request.go
@@ -66,6 +66,8 @@ func ModelRequest(
return nil, fmt.Errorf("purpose is required")
}
+ messages = FilterEmptyMessages(messages)
+
inputTokensEstimate := GetMessagesTokenEstimate(messages...) + TokensPerRequest
config := modelConfig.GetRoleForInputTokens(inputTokensEstimate)
@@ -183,3 +185,22 @@ func ModelRequest(
return res, nil
}
+
+func FilterEmptyMessages(messages []types.ExtendedChatMessage) []types.ExtendedChatMessage {
+ filteredMessages := []types.ExtendedChatMessage{}
+ for _, message := range messages {
+ var content []types.ExtendedChatMessagePart
+ for _, part := range message.Content {
+ if part.Type != openai.ChatMessagePartTypeText || part.Text != "" {
+ content = append(content, part)
+ }
+ }
+ if len(content) > 0 {
+ filteredMessages = append(filteredMessages, types.ExtendedChatMessage{
+ Role: message.Role,
+ Content: content,
+ })
+ }
+ }
+ return filteredMessages
+}
diff --git a/app/server/model/plan/tell_context.go b/app/server/model/plan/tell_context.go
index 41373f4b..7b5be8bb 100644
--- a/app/server/model/plan/tell_context.go
+++ b/app/server/model/plan/tell_context.go
@@ -55,6 +55,14 @@ func (state *activeTellStreamState) formatModelContext(params formatModelContext
addedFilesSet := map[string]bool{}
uses := map[string]bool{}
+
+ // log.Println("Tell plan - formatModelContext - state.currentSubtask:\n", spew.Sdump(state.currentSubtask))
+ // if state.currentSubtask != nil {
+ // log.Println("Tell plan - formatModelContext - state.currentSubtask.UsesFiles:\n", spew.Sdump(state.currentSubtask.UsesFiles))
+ // }
+ // log.Println("Tell plan - formatModelContext - currentStage.TellStage:\n", currentStage.TellStage)
+ // log.Println("Tell plan - formatModelContext - smartContextEnabled:\n", smartContextEnabled)
+
if currentStage.TellStage == shared.TellStageImplementation && smartContextEnabled && state.currentSubtask != nil {
log.Println("Tell plan - formatModelContext - implementation stage - smart context enabled for current subtask")
for _, path := range state.currentSubtask.UsesFiles {
diff --git a/app/server/model/plan/tell_exec.go b/app/server/model/plan/tell_exec.go
index 8460b470..f80e0478 100644
--- a/app/server/model/plan/tell_exec.go
+++ b/app/server/model/plan/tell_exec.go
@@ -149,7 +149,7 @@ func execTellPlan(params execTellPlanParams) {
}
log.Println("execTellPlan - Loading tell plan")
- err = state.loadTellPlan()
+ err = state.loadTellPlan(params.numErrorRetry > 0)
if err != nil {
return
}
@@ -293,7 +293,7 @@ func execTellPlan(params execTellPlanParams) {
return
}
- // log.Println("messages:\n\n", spew.Sdump(state.messages))
+ log.Println("messages:\n\n", spew.Sdump(state.messages))
// log.Println("promptMessage:", spew.Sdump(promptMessage))
@@ -349,6 +349,9 @@ func execTellPlan(params execTellPlanParams) {
return
}
+ // filter out any messages that are empty
+ state.messages = model.FilterEmptyMessages(state.messages)
+
log.Printf("\n\nMessages: %d\n", len(state.messages))
// for _, message := range state.messages {
// log.Printf("%s: %v\n", message.Role, message.Content)
diff --git a/app/server/model/plan/tell_load.go b/app/server/model/plan/tell_load.go
index 28d50101..e556a10f 100644
--- a/app/server/model/plan/tell_load.go
+++ b/app/server/model/plan/tell_load.go
@@ -15,7 +15,7 @@ import (
"github.com/sashabaranov/go-openai"
)
-func (state *activeTellStreamState) loadTellPlan() error {
+func (state *activeTellStreamState) loadTellPlan(isRetry bool) error {
clients := state.clients
req := state.req
auth := state.auth
@@ -154,7 +154,7 @@ func (state *activeTellStreamState) loadTellPlan() error {
innerErrCh := make(chan error)
go func() {
- if iteration == 0 && missingFileResponse == "" && !req.IsUserContinue {
+ if iteration == 0 && missingFileResponse == "" && !req.IsUserContinue && !isRetry {
num := len(convo) + 1
log.Printf("[TellLoad] storing user message | len(convo): %d | num: %d\n", len(convo), num)
diff --git a/app/server/model/plan/tell_stream_error.go b/app/server/model/plan/tell_stream_error.go
index 281cedb3..126c531b 100644
--- a/app/server/model/plan/tell_stream_error.go
+++ b/app/server/model/plan/tell_stream_error.go
@@ -53,6 +53,7 @@ func (state *activeTellStreamState) onError(params onErrorParams) onErrorResult
canRetry := params.canRetry
if canRetry {
+ log.Println("tellStream onError - canRetry", canRetry)
if numRetries >= NumTellStreamRetries {
log.Printf("tellStream onError - Max retries reached for plan ID %s on branch %s\n", planId, branch)
@@ -61,6 +62,7 @@ func (state *activeTellStreamState) onError(params onErrorParams) onErrorResult
}
if canRetry {
+ log.Println("tellStream onError - retrying stream")
// stop stream via context (ensures we stop child streams too)
active.CancelModelStreamFn()
@@ -81,6 +83,7 @@ func (state *activeTellStreamState) onError(params onErrorParams) onErrorResult
}
storeDescAndReply := func() error {
+ log.Println("tellStream onError - storing desc and reply")
ctx, cancelFn := context.WithTimeout(shutdown.ShutdownCtx, 5*time.Second)
err := db.ExecRepoOperation(db.ExecRepoOperationParams{
@@ -161,7 +164,9 @@ func (state *activeTellStreamState) onError(params onErrorParams) onErrorResult
return nil
}
- storeDescAndReply() // best effort to store description and reply, ignore errors
+ if active.CurrentReplyContent != "" {
+ storeDescAndReply() // best effort to store description and reply, ignore errors
+ }
if params.streamApiErr != nil {
active.StreamDoneCh <- params.streamApiErr
diff --git a/app/server/model/plan/tell_stream_main.go b/app/server/model/plan/tell_stream_main.go
index a97d61a5..10588af3 100644
--- a/app/server/model/plan/tell_stream_main.go
+++ b/app/server/model/plan/tell_stream_main.go
@@ -58,6 +58,9 @@ func (state *activeTellStreamState) listenStream(stream *model.ExtendedChatCompl
defer timer.Stop()
streamFinished := false
+ modelProvider := state.modelConfig.BaseModelConfig.Provider
+ modelName := state.modelConfig.BaseModelConfig.ModelName
+
mainLoop:
for {
select {
@@ -109,9 +112,9 @@ mainLoop:
var msg string
if active.CurrentReplyContent == "" {
- msg = "The AI model didn't respond: %v"
+ msg = fmt.Sprintf("The AI model (%s/%s) didn't respond: %v", modelProvider, modelName, err)
} else {
- msg = "The AI model stopped responding: %v"
+ msg = fmt.Sprintf("The AI model (%s/%s) stopped responding: %v", modelProvider, modelName, err)
}
state.onError(onErrorParams{
streamErr: fmt.Errorf(msg, err),
@@ -132,6 +135,21 @@ mainLoop:
state.firstTokenAt = time.Now()
}
+ if response.Error != nil {
+ log.Println("listenStream - stream finished with error", spew.Sdump(response.Error))
+ res := state.onError(onErrorParams{
+ streamErr: fmt.Errorf("The AI model (%s/%s) stopped streaming with error code %d: %s", modelProvider, modelName, response.Error.Code, response.Error.Message),
+ storeDesc: true,
+ canRetry: active.CurrentReplyContent == "" && response.Error.Code != 429,
+ })
+ if res.shouldReturn {
+ return
+ }
+ if res.shouldContinueMainLoop {
+ continue mainLoop
+ }
+ }
+
if len(response.Choices) == 0 {
if response.Usage != nil {
state.handleUsageChunk(response.Usage)
@@ -185,10 +203,12 @@ mainLoop:
log.Println("Finish reason: ", choice.FinishReason)
if choice.FinishReason == "error" {
+ log.Println("Model stream finished with error")
+
res := state.onError(onErrorParams{
- streamErr: fmt.Errorf("the underlying LLM stopped with an error status | This means the LLM is not responding."),
+ streamErr: fmt.Errorf("The AI model (%s/%s) stopped streaming with an error status", modelProvider, modelName),
storeDesc: true,
- canRetry: true,
+ canRetry: active.CurrentReplyContent == "",
})
if res.shouldReturn {
return
diff --git a/app/server/types/message.go b/app/server/types/message.go
index a5434919..fc5df675 100644
--- a/app/server/types/message.go
+++ b/app/server/types/message.go
@@ -200,6 +200,15 @@ type ExtendedChatCompletionStreamResponse struct {
// When present, it contains a null value except for the last chunk which contains the token usage statistics
// for the entire request.
Usage *openai.Usage `json:"usage,omitempty"`
+ Error *struct {
+ Message string `json:"message"`
+ Code int `json:"code"`
+ Metadata struct {
+ Headers map[string]string `json:"headers,omitempty"`
+ ProviderName string `json:"provider_name,omitempty"`
+ } `json:"metadata,omitempty"`
+ UserId string `json:"user_id,omitempty"`
+ } `json:"error,omitempty"`
}
// ModelResponse holds both the accumulated content and final usage information from a streaming completion request
diff --git a/app/shared/convo_message.go b/app/shared/convo_message.go
index b644da51..9d2b4199 100644
--- a/app/shared/convo_message.go
+++ b/app/shared/convo_message.go
@@ -25,7 +25,7 @@ func (f *ConvoMessageFlags) GetReplyTags() []string {
}
if f.HasError {
- replyTags = append(replyTags, "🚨 Finished With Errors")
+ replyTags = append(replyTags, "🚨 Error")
}
return replyTags