victorvargas/rony-llm-agent
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix(openai): make streaming actually work; honor configured model

Browse source 
The Stream() path was broken end to end:
- requests always went out with "stream": false, so the SSE parser found
  no data lines and every stream ended empty
- Config.Model was discarded at construction, and the agent loop never
  sets req.Model, so requests carried an empty model (hard API error)
- tool-call deltas were ignored entirely: the agent never executed tools
  over a stream with this provider (which also backs the ollama type)
- usage was neither requested nor parsed, so token tracking stayed at 0

Now mirrors the proven llamacpp client: stream flag + stream_options
.include_usage, per-index tool-call fragment accumulation flushed on
finish_reason, usage passthrough, a 4MB SSE scanner buffer (64KB default
kills the stream on large tool arguments), and an empty-choices guard in
toResponse instead of a panic.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Victor Hugo Vargas Servin 2026-07-12 16:14:01 -07:00
parent eea51e30d1
commit 838eef642a
1 changed files with 166 additions and 53 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

133
pkg/llm/providers/openai/client.go
View file

@ -2,12 +2,13 @@ package openai
import (
"bufio"
"bytes"
"context"
"encoding/json"
"fmt"
"io"
"net/http"
"iter"
"net/http"
"strings"
"github.com/VictorVargas/rony-llm-agent/pkg/llm"
@ -23,6 +24,7 @@ type Config struct {
// Client implements llm.LLMClient for OpenAI.
type Client struct {
apiKey string
model string
baseURL string
http *http.Client
}
@ -40,6 +42,7 @@ func New(cfg Config) (*Client, error) {
return &Client{
apiKey: cfg.APIKey,
model: cfg.Model,
baseURL: baseURL,
http: http.DefaultClient,
}, nil
@ -48,7 +51,7 @@ func New(cfg Config) (*Client, error) {
func (c *Client) Generate(ctx context.Context, req llm.CompletionRequest) (llm.CompletionResponse, error) {
endpoint := c.baseURL + "/chat/completions"
payload, err := c.buildRequest(req)
payload, err := c.buildRequest(req, false)
if err != nil {
return llm.CompletionResponse{}, fmt.Errorf("building request: %w", err)
}
@ -76,14 +79,14 @@ func (c *Client) Generate(ctx context.Context, req llm.CompletionRequest) (llm.C
return llm.CompletionResponse{}, fmt.Errorf("decoding response: %w", err)
}
return c.toResponse(apiResp), nil
return c.toResponse(apiResp)
}
func (c *Client) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq2[llm.StreamChunk, error] {
return func(yield func(llm.StreamChunk, error) bool) {
endpoint := c.baseURL + "/chat/completions"
payload, err := c.buildRequest(req)
payload, err := c.buildRequest(req, true)
if err != nil {
yield(llm.StreamChunk{}, fmt.Errorf("building request: %w", err))
return
@ -111,7 +114,45 @@ func (c *Client) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq
return
}
// toolCallAccum buffers one tool call's fragments as they stream in:
// the SSE format sends the id/name in the first delta for a given
// tool-call index and the (potentially large) arguments JSON in
// pieces across many subsequent deltas, so it can't be handed to a
// tool handler until it's fully assembled. Same accumulation the
// llamacpp client does — without it, tool calls made over a stream
// were silently dropped and the agent loop never executed them.
type toolCallAccum struct {
id string
name string
args strings.Builder
}
toolCallFrags := map[int]*toolCallAccum{}
var toolCallOrder []int
flushToolCalls := func() []llm.ToolCall {
if len(toolCallOrder) == 0 {
return nil
}
calls := make([]llm.ToolCall, 0, len(toolCallOrder))
for _, idx := range toolCallOrder {
frag := toolCallFrags[idx]
calls = append(calls, llm.ToolCall{
ID: frag.id,
Name: frag.name,
Arguments: json.RawMessage(frag.args.String()),
})
}
toolCallFrags = map[int]*toolCallAccum{}
toolCallOrder = nil
return calls
}
scanner := bufio.NewScanner(resp.Body)
// A single SSE line can exceed bufio.Scanner's 64KB default cap
// (e.g. a large tool-call arguments delta), which would kill the
// stream with "token too long" — same headroom the anthropic
// client already reserves.
scanner.Buffer(make([]byte, 0, 64*1024), 4*1024*1024)
for scanner.Scan() {
line := scanner.Text()
if !strings.HasPrefix(line, "data: ") {
@ -128,13 +169,61 @@ func (c *Client) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq
return
}
var usage llm.TokenUsage
if event.Usage != nil {
usage = llm.TokenUsage{
InputTokens: event.Usage.PromptTokens,
OutputTokens: event.Usage.CompletionTokens,
TotalTokens: event.Usage.TotalTokens,
}
}
if len(event.Choices) == 0 {
// The usage-only event (per stream_options.include_usage)
// carries no choices, so it needs its own chunk.
if event.Usage != nil {
if !yield(llm.StreamChunk{Usage: usage}, nil) {
return
}
}
continue
}
for _, choice := range event.Choices {
hasFragment := len(choice.Delta.ToolCalls) > 0
for _, tc := range choice.Delta.ToolCalls {
frag, ok := toolCallFrags[tc.Index]
if !ok {
frag = &toolCallAccum{}
toolCallFrags[tc.Index] = frag
toolCallOrder = append(toolCallOrder, tc.Index)
}
if tc.ID != "" {
frag.id = tc.ID
}
if tc.Function.Name != "" {
frag.name = tc.Function.Name
}
frag.args.WriteString(tc.Function.Arguments)
}
chunk := llm.StreamChunk{
Delta: choice.Delta.Content,
Usage: usage,
}
if choice.FinishReason != "" {
chunk.FinishReason = choice.FinishReason
chunk.ToolCalls = flushToolCalls()
}
// A fragment-only event (a piece of a tool call's streamed
// arguments, with nothing else in this delta) has nothing
// yet for the agent loop to act on: it was buffered above,
// so skip yielding an empty chunk for it.
if hasFragment && chunk.Delta == "" && chunk.FinishReason == "" {
continue
}
if !yield(chunk, nil) {
return
}
@ -160,7 +249,7 @@ func (c *Client) Capabilities() llm.ProviderCapabilities {
}
// buildRequest converts an llm.CompletionRequest to the OpenAI API format.
func (c *Client) buildRequest(req llm.CompletionRequest) (io.Reader, error) {
func (c *Client) buildRequest(req llm.CompletionRequest, stream bool) (io.Reader, error) {
// Convert messages to OpenAI format
messages := make([]openaiMessage, len(req.Messages))
for i, m := range req.Messages {
@ -195,10 +284,25 @@ func (c *Client) buildRequest(req llm.CompletionRequest) (io.Reader, error) {
tools[i] = tool
}
// The per-request model wins when set; otherwise fall back to the
// client's configured one (Config.Model used to be discarded entirely,
// so every request went out with an empty model — a hard API error on
// OpenAI, and the agent loop never sets req.Model).
model := req.Model
if model == "" {
model = c.model
}
openaiReq := openaiChatRequest{
Model: req.Model,
Model: model,
Messages: messages,
Stream: false,
Stream: stream,
}
if stream {
// Ask for a final SSE event carrying token usage (OpenAI-style
// streaming omits it otherwise), so callers can track real token
// counts per turn instead of always seeing zero.
openaiReq.StreamOptions = &openaiStreamOptions{IncludeUsage: true}
}
if len(tools) > 0 {
openaiReq.Tools = tools
@ -222,11 +326,14 @@ func (c *Client) buildRequest(req llm.CompletionRequest) (io.Reader, error) {
if err != nil {
return nil, fmt.Errorf("marshaling request: %w", err)
}
return strings.NewReader(string(data)), nil
return bytes.NewReader(data), nil
}
// toResponse converts an OpenAI API response to our CompletionResponse.
func (c *Client) toResponse(resp openaiChatResponse) llm.CompletionResponse {
func (c *Client) toResponse(resp openaiChatResponse) (llm.CompletionResponse, error) {
if len(resp.Choices) == 0 {
return llm.CompletionResponse{}, fmt.Errorf("openai: response contained no choices")
}
choice := resp.Choices[0]
result := llm.CompletionResponse{
ID: resp.ID,
@ -249,7 +356,7 @@ func (c *Client) toResponse(resp openaiChatResponse) llm.CompletionResponse {
TotalTokens: resp.Usage.TotalTokens,
}
return result
return result, nil
}
// OpenAI API types
@ -263,6 +370,11 @@ type openaiChatRequest struct {
MaxTokens *int `json:"max_tokens,omitempty"`
Stop []string `json:"stop,omitempty"`
Stream bool `json:"stream"`
StreamOptions *openaiStreamOptions `json:"stream_options,omitempty"`
}
type openaiStreamOptions struct {
IncludeUsage bool `json:"include_usage"`
}
type openaiMessage struct {
@ -319,6 +431,7 @@ type openaiUsage struct {
type openaiStreamEvent struct {
ID string `json:"id"`
Choices []openaiStreamChoice `json:"choices"`
Usage *openaiUsage `json:"usage"`
}
type openaiStreamChoice struct {