mirror of
https://github.com/getcompanion-ai/co-mono.git
synced 2026-04-15 19:05:11 +00:00
refactor(ai): Update OpenAI Completions provider to new content block API
This commit is contained in:
Mario Zechner
parent
7c8cdacc09
commit
a72e6d08d4
5 changed files with 209 additions and 86 deletions
|
|
@ -6,7 +6,17 @@ import type {
|
|||
Tool,
|
||||
} from "@anthropic-ai/sdk/resources/messages.js";
|
||||
import { calculateCost } from "../models.js";
|
||||
import type { AssistantMessage, Context, LLM, LLMOptions, Message, Model, StopReason, Usage } from "../types.js";
|
||||
import type {
|
||||
AssistantMessage,
|
||||
Context,
|
||||
LLM,
|
||||
LLMOptions,
|
||||
Message,
|
||||
Model,
|
||||
StopReason,
|
||||
ToolCall,
|
||||
Usage,
|
||||
} from "../types.js";
|
||||
|
||||
export interface AnthropicLLMOptions extends LLMOptions {
|
||||
thinking?: {
|
||||
|
|
@ -119,8 +129,9 @@ export class AnthropicLLM implements LLM<AnthropicLLMOptions> {
|
|||
},
|
||||
);
|
||||
|
||||
let blockType: "text" | "thinking" | "other" = "other";
|
||||
let blockType: "text" | "thinking" | "toolUse" | "other" = "other";
|
||||
let blockContent = "";
|
||||
let toolCall: (ToolCall & { partialJson: string }) | null = null;
|
||||
for await (const event of stream) {
|
||||
if (event.type === "content_block_start") {
|
||||
if (event.content_block.type === "text") {
|
||||
|
|
@ -131,6 +142,17 @@ export class AnthropicLLM implements LLM<AnthropicLLMOptions> {
|
|||
blockType = "thinking";
|
||||
blockContent = "";
|
||||
options?.onEvent?.({ type: "thinking_start" });
|
||||
} else if (event.content_block.type === "tool_use") {
|
||||
// We wait for the full tool use to be streamed to send the event
|
||||
toolCall = {
|
||||
type: "toolCall",
|
||||
id: event.content_block.id,
|
||||
name: event.content_block.name,
|
||||
arguments: event.content_block.input as Record<string, any>,
|
||||
partialJson: "",
|
||||
};
|
||||
blockType = "toolUse";
|
||||
blockContent = "";
|
||||
} else {
|
||||
blockType = "other";
|
||||
blockContent = "";
|
||||
|
|
@ -145,12 +167,24 @@ export class AnthropicLLM implements LLM<AnthropicLLMOptions> {
|
|||
options?.onEvent?.({ type: "thinking_delta", content: blockContent, delta: event.delta.thinking });
|
||||
blockContent += event.delta.thinking;
|
||||
}
|
||||
if (event.delta.type === "input_json_delta") {
|
||||
toolCall!.partialJson += event.delta.partial_json;
|
||||
}
|
||||
}
|
||||
if (event.type === "content_block_stop") {
|
||||
if (blockType === "text") {
|
||||
options?.onEvent?.({ type: "text_end", content: blockContent });
|
||||
} else if (blockType === "thinking") {
|
||||
options?.onEvent?.({ type: "thinking_end", content: blockContent });
|
||||
} else if (blockType === "toolUse") {
|
||||
const finalToolCall: ToolCall = {
|
||||
type: "toolCall",
|
||||
id: toolCall!.id,
|
||||
name: toolCall!.name,
|
||||
arguments: toolCall!.partialJson ? JSON.parse(toolCall!.partialJson) : toolCall!.arguments,
|
||||
};
|
||||
toolCall = null;
|
||||
options?.onEvent?.({ type: "toolCall", toolCall: finalToolCall });
|
||||
}
|
||||
blockType = "other";
|
||||
}
|
||||
|
|
@ -194,16 +228,19 @@ export class AnthropicLLM implements LLM<AnthropicLLMOptions> {
|
|||
};
|
||||
calculateCost(this.modelInfo, usage);
|
||||
|
||||
return {
|
||||
const output = {
|
||||
role: "assistant",
|
||||
content: blocks,
|
||||
provider: this.modelInfo.provider,
|
||||
model: this.modelInfo.id,
|
||||
usage,
|
||||
stopReason: this.mapStopReason(msg.stop_reason),
|
||||
};
|
||||
} satisfies AssistantMessage;
|
||||
options?.onEvent?.({ type: "done", reason: output.stopReason, message: output });
|
||||
|
||||
return output;
|
||||
} catch (error) {
|
||||
return {
|
||||
const output = {
|
||||
role: "assistant",
|
||||
content: [],
|
||||
provider: this.modelInfo.provider,
|
||||
|
|
@ -216,8 +253,10 @@ export class AnthropicLLM implements LLM<AnthropicLLMOptions> {
|
|||
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
|
||||
},
|
||||
stopReason: "error",
|
||||
error: error instanceof Error ? error.message : String(error),
|
||||
};
|
||||
error: error instanceof Error ? error.message : JSON.stringify(error),
|
||||
} satisfies AssistantMessage;
|
||||
options?.onEvent?.({ type: "error", error: output.error });
|
||||
return output;
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -148,7 +148,11 @@ export class GoogleLLM implements LLM<GoogleLLMOptions> {
|
|||
if (currentBlock.type === "thinking") {
|
||||
currentBlock.thinking += part.text;
|
||||
currentBlock.thinkingSignature = part.thoughtSignature;
|
||||
options?.onEvent?.({type: "thinking_delta", content: currentBlock.thinking, delta: part.text });
|
||||
options?.onEvent?.({
|
||||
type: "thinking_delta",
|
||||
content: currentBlock.thinking,
|
||||
delta: part.text,
|
||||
});
|
||||
} else {
|
||||
currentBlock.text += part.text;
|
||||
options?.onEvent?.({ type: "text_delta", content: currentBlock.text, delta: part.text });
|
||||
|
|
|
|||
|
|
@ -15,6 +15,8 @@ import type {
|
|||
Message,
|
||||
Model,
|
||||
StopReason,
|
||||
TextContent,
|
||||
ThinkingContent,
|
||||
Tool,
|
||||
ToolCall,
|
||||
Usage,
|
||||
|
|
@ -90,10 +92,8 @@ export class OpenAICompletionsLLM implements LLM<OpenAICompletionsLLMOptions> {
|
|||
signal: options?.signal,
|
||||
});
|
||||
|
||||
let content = "";
|
||||
let reasoningContent = "";
|
||||
let reasoningField: "reasoning" | "reasoning_content" | null = null;
|
||||
const parsedToolCalls: { id: string; name: string; arguments: string }[] = [];
|
||||
const blocks: AssistantMessage["content"] = [];
|
||||
let currentBlock: TextContent | ThinkingContent | (ToolCall & { partialArgs?: string }) | null = null;
|
||||
let usage: Usage = {
|
||||
input: 0,
|
||||
output: 0,
|
||||
|
|
@ -102,7 +102,6 @@ export class OpenAICompletionsLLM implements LLM<OpenAICompletionsLLMOptions> {
|
|||
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
|
||||
};
|
||||
let finishReason: ChatCompletionChunk.Choice["finish_reason"] | null = null;
|
||||
let blockType: "text" | "thinking" | null = null;
|
||||
for await (const chunk of stream) {
|
||||
if (chunk.usage) {
|
||||
usage = {
|
||||
|
|
@ -132,13 +131,32 @@ export class OpenAICompletionsLLM implements LLM<OpenAICompletionsLLMOptions> {
|
|||
choice.delta.content !== undefined &&
|
||||
choice.delta.content.length > 0
|
||||
) {
|
||||
if (blockType === "thinking") {
|
||||
options?.onThinking?.("", true);
|
||||
blockType = null;
|
||||
// Check if we need to switch to text block
|
||||
if (!currentBlock || currentBlock.type !== "text") {
|
||||
// Save current block if exists
|
||||
if (currentBlock) {
|
||||
if (currentBlock.type === "thinking") {
|
||||
options?.onEvent?.({ type: "thinking_end", content: currentBlock.thinking });
|
||||
} else if (currentBlock.type === "toolCall") {
|
||||
currentBlock.arguments = JSON.parse(currentBlock.partialArgs || "{}");
|
||||
delete currentBlock.partialArgs;
|
||||
options?.onEvent?.({ type: "toolCall", toolCall: currentBlock as ToolCall });
|
||||
}
|
||||
blocks.push(currentBlock);
|
||||
}
|
||||
// Start new text block
|
||||
currentBlock = { type: "text", text: "" };
|
||||
options?.onEvent?.({ type: "text_start" });
|
||||
}
|
||||
// Append to text block
|
||||
if (currentBlock.type === "text") {
|
||||
options?.onEvent?.({
|
||||
type: "text_delta",
|
||||
content: currentBlock.text,
|
||||
delta: choice.delta.content,
|
||||
});
|
||||
currentBlock.text += choice.delta.content;
|
||||
}
|
||||
content += choice.delta.content;
|
||||
options?.onText?.(choice.delta.content, false);
|
||||
blockType = "text";
|
||||
}
|
||||
|
||||
// Handle reasoning_content field
|
||||
|
|
@ -146,55 +164,98 @@ export class OpenAICompletionsLLM implements LLM<OpenAICompletionsLLMOptions> {
|
|||
(choice.delta as any).reasoning_content !== null &&
|
||||
(choice.delta as any).reasoning_content !== undefined
|
||||
) {
|
||||
if (blockType === "text") {
|
||||
options?.onText?.("", true);
|
||||
blockType = null;
|
||||
// Check if we need to switch to thinking block
|
||||
if (!currentBlock || currentBlock.type !== "thinking") {
|
||||
// Save current block if exists
|
||||
if (currentBlock) {
|
||||
if (currentBlock.type === "text") {
|
||||
options?.onEvent?.({ type: "text_end", content: currentBlock.text });
|
||||
} else if (currentBlock.type === "toolCall") {
|
||||
currentBlock.arguments = JSON.parse(currentBlock.partialArgs || "{}");
|
||||
delete currentBlock.partialArgs;
|
||||
options?.onEvent?.({ type: "toolCall", toolCall: currentBlock as ToolCall });
|
||||
}
|
||||
blocks.push(currentBlock);
|
||||
}
|
||||
// Start new thinking block
|
||||
currentBlock = { type: "thinking", thinking: "", thinkingSignature: "reasoning_content" };
|
||||
options?.onEvent?.({ type: "thinking_start" });
|
||||
}
|
||||
// Append to thinking block
|
||||
if (currentBlock.type === "thinking") {
|
||||
const delta = (choice.delta as any).reasoning_content;
|
||||
options?.onEvent?.({ type: "thinking_delta", content: currentBlock.thinking, delta });
|
||||
currentBlock.thinking += delta;
|
||||
}
|
||||
reasoningContent += (choice.delta as any).reasoning_content;
|
||||
reasoningField = "reasoning_content";
|
||||
options?.onThinking?.((choice.delta as any).reasoning_content, false);
|
||||
blockType = "thinking";
|
||||
}
|
||||
|
||||
// Handle reasoning field
|
||||
if ((choice.delta as any).reasoning !== null && (choice.delta as any).reasoning !== undefined) {
|
||||
if (blockType === "text") {
|
||||
options?.onText?.("", true);
|
||||
blockType = null;
|
||||
// Check if we need to switch to thinking block
|
||||
if (!currentBlock || currentBlock.type !== "thinking") {
|
||||
// Save current block if exists
|
||||
if (currentBlock) {
|
||||
if (currentBlock.type === "text") {
|
||||
options?.onEvent?.({ type: "text_end", content: currentBlock.text });
|
||||
} else if (currentBlock.type === "toolCall") {
|
||||
currentBlock.arguments = JSON.parse(currentBlock.partialArgs || "{}");
|
||||
delete currentBlock.partialArgs;
|
||||
options?.onEvent?.({ type: "toolCall", toolCall: currentBlock as ToolCall });
|
||||
}
|
||||
blocks.push(currentBlock);
|
||||
}
|
||||
// Start new thinking block
|
||||
currentBlock = { type: "thinking", thinking: "", thinkingSignature: "reasoning" };
|
||||
options?.onEvent?.({ type: "thinking_start" });
|
||||
}
|
||||
// Append to thinking block
|
||||
if (currentBlock.type === "thinking") {
|
||||
const delta = (choice.delta as any).reasoning;
|
||||
options?.onEvent?.({ type: "thinking_delta", content: currentBlock.thinking, delta });
|
||||
currentBlock.thinking += delta;
|
||||
}
|
||||
reasoningContent += (choice.delta as any).reasoning;
|
||||
reasoningField = "reasoning";
|
||||
options?.onThinking?.((choice.delta as any).reasoning, false);
|
||||
blockType = "thinking";
|
||||
}
|
||||
|
||||
// Handle tool calls
|
||||
if (choice?.delta?.tool_calls) {
|
||||
if (blockType === "text") {
|
||||
options?.onText?.("", true);
|
||||
blockType = null;
|
||||
}
|
||||
if (blockType === "thinking") {
|
||||
options?.onThinking?.("", true);
|
||||
blockType = null;
|
||||
}
|
||||
for (const toolCall of choice.delta.tool_calls) {
|
||||
// Check if we need a new tool call block
|
||||
if (
|
||||
parsedToolCalls.length === 0 ||
|
||||
(toolCall.id !== undefined && parsedToolCalls[parsedToolCalls.length - 1].id !== toolCall.id)
|
||||
!currentBlock ||
|
||||
currentBlock.type !== "toolCall" ||
|
||||
(toolCall.id && currentBlock.id !== toolCall.id)
|
||||
) {
|
||||
parsedToolCalls.push({
|
||||
// Save current block if exists
|
||||
if (currentBlock) {
|
||||
if (currentBlock.type === "text") {
|
||||
options?.onEvent?.({ type: "text_end", content: currentBlock.text });
|
||||
} else if (currentBlock.type === "thinking") {
|
||||
options?.onEvent?.({ type: "thinking_end", content: currentBlock.thinking });
|
||||
} else if (currentBlock.type === "toolCall") {
|
||||
currentBlock.arguments = JSON.parse(currentBlock.partialArgs || "{}");
|
||||
delete currentBlock.partialArgs;
|
||||
options?.onEvent?.({ type: "toolCall", toolCall: currentBlock as ToolCall });
|
||||
}
|
||||
blocks.push(currentBlock);
|
||||
}
|
||||
|
||||
// Start new tool call block
|
||||
currentBlock = {
|
||||
type: "toolCall",
|
||||
id: toolCall.id || "",
|
||||
name: toolCall.function?.name || "",
|
||||
arguments: "",
|
||||
});
|
||||
arguments: {},
|
||||
partialArgs: "",
|
||||
};
|
||||
}
|
||||
|
||||
const current = parsedToolCalls[parsedToolCalls.length - 1];
|
||||
if (toolCall.id) current.id = toolCall.id;
|
||||
if (toolCall.function?.name) current.name = toolCall.function.name;
|
||||
if (toolCall.function?.arguments) {
|
||||
current.arguments += toolCall.function.arguments;
|
||||
// Accumulate tool call data
|
||||
if (currentBlock.type === "toolCall") {
|
||||
if (toolCall.id) currentBlock.id = toolCall.id;
|
||||
if (toolCall.function?.name) currentBlock.name = toolCall.function.name;
|
||||
if (toolCall.function?.arguments) {
|
||||
currentBlock.partialArgs += toolCall.function.arguments;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -202,42 +263,41 @@ export class OpenAICompletionsLLM implements LLM<OpenAICompletionsLLMOptions> {
|
|||
|
||||
// Capture finish reason
|
||||
if (choice.finish_reason) {
|
||||
if (blockType === "text") {
|
||||
options?.onText?.("", true);
|
||||
blockType = null;
|
||||
}
|
||||
if (blockType === "thinking") {
|
||||
options?.onThinking?.("", true);
|
||||
blockType = null;
|
||||
}
|
||||
finishReason = choice.finish_reason;
|
||||
}
|
||||
}
|
||||
|
||||
// Convert tool calls map to array
|
||||
const toolCalls: ToolCall[] = parsedToolCalls.map((tc) => ({
|
||||
id: tc.id,
|
||||
name: tc.name,
|
||||
arguments: JSON.parse(tc.arguments),
|
||||
}));
|
||||
// Save final block if exists
|
||||
if (currentBlock) {
|
||||
if (currentBlock.type === "text") {
|
||||
options?.onEvent?.({ type: "text_end", content: currentBlock.text });
|
||||
} else if (currentBlock.type === "thinking") {
|
||||
options?.onEvent?.({ type: "thinking_end", content: currentBlock.thinking });
|
||||
} else if (currentBlock.type === "toolCall") {
|
||||
currentBlock.arguments = JSON.parse(currentBlock.partialArgs || "{}");
|
||||
delete currentBlock.partialArgs;
|
||||
options?.onEvent?.({ type: "toolCall", toolCall: currentBlock as ToolCall });
|
||||
}
|
||||
blocks.push(currentBlock);
|
||||
}
|
||||
|
||||
// Calculate cost
|
||||
calculateCost(this.modelInfo, usage);
|
||||
|
||||
return {
|
||||
const output = {
|
||||
role: "assistant",
|
||||
content: content || undefined,
|
||||
thinking: reasoningContent || undefined,
|
||||
thinkingSignature: reasoningField || undefined,
|
||||
toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
|
||||
content: blocks,
|
||||
provider: this.modelInfo.provider,
|
||||
model: this.modelInfo.id,
|
||||
usage,
|
||||
stopReason: this.mapStopReason(finishReason),
|
||||
};
|
||||
} satisfies AssistantMessage;
|
||||
options?.onEvent?.({ type: "done", reason: output.stopReason, message: output });
|
||||
return output;
|
||||
} catch (error) {
|
||||
return {
|
||||
const output = {
|
||||
role: "assistant",
|
||||
content: [],
|
||||
provider: this.modelInfo.provider,
|
||||
model: this.modelInfo.id,
|
||||
usage: {
|
||||
|
|
@ -249,7 +309,9 @@ export class OpenAICompletionsLLM implements LLM<OpenAICompletionsLLMOptions> {
|
|||
},
|
||||
stopReason: "error",
|
||||
error: error instanceof Error ? error.message : String(error),
|
||||
};
|
||||
} satisfies AssistantMessage;
|
||||
options?.onEvent?.({ type: "error", error: output.error || "Unknown error" });
|
||||
return output;
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -302,16 +364,29 @@ export class OpenAICompletionsLLM implements LLM<OpenAICompletionsLLMOptions> {
|
|||
} else if (msg.role === "assistant") {
|
||||
const assistantMsg: ChatCompletionMessageParam = {
|
||||
role: "assistant",
|
||||
content: msg.content || null,
|
||||
content: null,
|
||||
};
|
||||
|
||||
// LLama.cpp server + gpt-oss
|
||||
if (msg.thinking && msg.thinkingSignature && msg.thinkingSignature.length > 0) {
|
||||
(assistantMsg as any)[msg.thinkingSignature] = msg.thinking;
|
||||
// Build content from blocks
|
||||
const textBlocks = msg.content.filter((b) => b.type === "text") as TextContent[];
|
||||
if (textBlocks.length > 0) {
|
||||
assistantMsg.content = textBlocks.map((b) => b.text).join("");
|
||||
}
|
||||
|
||||
if (msg.toolCalls) {
|
||||
assistantMsg.tool_calls = msg.toolCalls.map((tc) => ({
|
||||
// Handle thinking blocks for llama.cpp server + gpt-oss
|
||||
const thinkingBlocks = msg.content.filter((b) => b.type === "thinking") as ThinkingContent[];
|
||||
if (thinkingBlocks.length > 0) {
|
||||
// Use the signature from the first thinking block if available
|
||||
const signature = thinkingBlocks[0].thinkingSignature;
|
||||
if (signature && signature.length > 0) {
|
||||
(assistantMsg as any)[signature] = thinkingBlocks.map((b) => b.thinking).join("");
|
||||
}
|
||||
}
|
||||
|
||||
// Handle tool calls
|
||||
const toolCalls = msg.content.filter((b) => b.type === "toolCall") as ToolCall[];
|
||||
if (toolCalls.length > 0) {
|
||||
assistantMsg.tool_calls = toolCalls.map((tc) => ({
|
||||
id: tc.id,
|
||||
type: "function" as const,
|
||||
function: {
|
||||
|
|
|
|||
|
|
@ -69,8 +69,9 @@ export interface AssistantMessage {
|
|||
|
||||
export interface ToolResultMessage {
|
||||
role: "toolResult";
|
||||
content: string;
|
||||
toolCallId: string;
|
||||
toolName: string;
|
||||
content: string;
|
||||
isError: boolean;
|
||||
}
|
||||
|
||||
|
|
@ -97,9 +98,8 @@ export type AssistantMessageEvent =
|
|||
| { type: "thinking_delta"; content: string; delta: string }
|
||||
| { type: "thinking_end"; content: string }
|
||||
| { type: "toolCall"; toolCall: ToolCall }
|
||||
| { type: "usage"; usage: Usage }
|
||||
| { type: "done"; reason: StopReason; message: AssistantMessage }
|
||||
| { type: "error"; error: Error };
|
||||
| { type: "error"; error: string };
|
||||
|
||||
// Model interface for the unified model system
|
||||
export interface Model {
|
||||
|
|
|
|||
|
|
@ -47,7 +47,7 @@ async function basicTextGeneration<T extends LLMOptions>(llm: LLM<T>) {
|
|||
expect(response.usage.input).toBeGreaterThan(0);
|
||||
expect(response.usage.output).toBeGreaterThan(0);
|
||||
expect(response.error).toBeFalsy();
|
||||
expect(response.content.map(b => b.type == "text" ? b.text : "").join("\n")).toContain("Hello test successful");
|
||||
expect(response.content.map(b => b.type == "text" ? b.text : "").join("")).toContain("Hello test successful");
|
||||
|
||||
context.messages.push(response);
|
||||
context.messages.push({ role: "user", content: "Now say 'Goodbye test successful'" });
|
||||
|
|
@ -56,10 +56,10 @@ async function basicTextGeneration<T extends LLMOptions>(llm: LLM<T>) {
|
|||
|
||||
expect(secondResponse.role).toBe("assistant");
|
||||
expect(secondResponse.content).toBeTruthy();
|
||||
expect(secondResponse.usage.input).toBeGreaterThan(0);
|
||||
expect(secondResponse.usage.input + secondResponse.usage.cacheRead).toBeGreaterThan(0);
|
||||
expect(secondResponse.usage.output).toBeGreaterThan(0);
|
||||
expect(secondResponse.error).toBeFalsy();
|
||||
expect(secondResponse.content.map(b => b.type == "text" ? b.text : "").join("\n")).toContain("Goodbye test successful");
|
||||
expect(secondResponse.content.map(b => b.type == "text" ? b.text : "").join("")).toContain("Goodbye test successful");
|
||||
}
|
||||
|
||||
async function handleToolCall<T extends LLMOptions>(llm: LLM<T>) {
|
||||
|
|
@ -225,8 +225,9 @@ async function multiTurn<T extends LLMOptions>(llm: LLM<T>, thinkingOptions: T)
|
|||
// Add tool result to context
|
||||
context.messages.push({
|
||||
role: "toolResult",
|
||||
content: `${result}`,
|
||||
toolCallId: block.id,
|
||||
toolName: block.name,
|
||||
content: `${result}`,
|
||||
isError: false
|
||||
});
|
||||
}
|
||||
|
|
@ -275,6 +276,10 @@ describe("AI Providers E2E Tests", () => {
|
|||
it("should handle multi-turn with thinking and tools", async () => {
|
||||
await multiTurn(llm, {thinking: { enabled: true, budgetTokens: 2048 }});
|
||||
});
|
||||
|
||||
it("should handle image input", async () => {
|
||||
await handleImage(llm);
|
||||
});
|
||||
});
|
||||
|
||||
describe.skipIf(!process.env.OPENAI_API_KEY)("OpenAI Completions Provider", () => {
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue