1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
|
import type { ChatMessage, Chunk } from "@dispatch/kernel";
/** A text part within a multimodal OpenAI content array. */
export interface OpenAITextPart {
readonly type: "text";
readonly text: string;
}
/** An image part within a multimodal OpenAI content array (OpenAI vision format). */
export interface OpenAIImagePart {
readonly type: "image_url";
readonly image_url: { readonly url: string };
}
/**
* A part of a multimodal message content array. When a message has mixed text
* and image chunks, the content is serialized as an array of these parts
* (OpenAI's vision format). Plain-text messages keep a string `content` for
* byte-stability with providers that only accept strings.
*/
export type OpenAIContentPart = OpenAITextPart | OpenAIImagePart;
export interface OpenAIMessage {
readonly role: "system" | "user" | "assistant" | "tool";
readonly content: string | null | readonly OpenAIContentPart[];
readonly tool_calls?: readonly OpenAIToolCall[];
readonly tool_call_id?: string;
}
export interface OpenAIToolCall {
readonly id: string;
readonly type: "function";
readonly function: { readonly name: string; readonly arguments: string };
}
export function convertMessages(messages: readonly ChatMessage[]): OpenAIMessage[] {
const result: OpenAIMessage[] = [];
for (const msg of messages) {
const converted = convertMessage(msg);
for (const m of converted) {
result.push(m);
}
}
return result;
}
function convertMessage(msg: ChatMessage): OpenAIMessage[] {
switch (msg.role) {
case "system":
return [convertSystemMessage(msg)];
case "user":
return [convertUserMessage(msg)];
case "assistant":
return [convertAssistantMessage(msg)];
case "tool":
return convertToolResultMessages(msg);
}
}
function convertSystemMessage(msg: ChatMessage): OpenAIMessage {
const text = msg.chunks
.filter(
(c): c is Extract<Chunk, { type: "text" | "system" }> =>
c.type === "text" || c.type === "system",
)
.map((c) => c.text)
.join("");
return { role: "system", content: text };
}
function convertUserMessage(msg: ChatMessage): OpenAIMessage {
// If the message has image chunks, serialize as a multimodal content array
// (OpenAI vision format): text parts + image_url parts in chunk order.
// Plain text-only messages keep a string `content` for byte-stability with
// providers that only accept a string (and to keep prompt-cache prefixes
// unchanged for the common no-image case).
const hasImage = msg.chunks.some((c) => c.type === "image");
if (hasImage) {
const parts: OpenAIContentPart[] = [];
for (const chunk of msg.chunks) {
if (chunk.type === "text") {
if (chunk.text.length > 0) {
parts.push({ type: "text", text: chunk.text });
}
} else if (chunk.type === "image") {
parts.push({ type: "image_url", image_url: { url: chunk.url } });
}
// Non-text/non-image chunks (tool-call, thinking, etc.) are not part of a
// user message's provider content and are skipped here.
}
// An image-only message (no text) still needs at least the image part.
return { role: "user", content: parts.length > 0 ? parts : "" };
}
const text = msg.chunks
.filter((c): c is Extract<Chunk, { type: "text" }> => c.type === "text")
.map((c) => c.text)
.join("");
return { role: "user", content: text };
}
function convertAssistantMessage(msg: ChatMessage): OpenAIMessage {
const textChunks = msg.chunks.filter(
(c): c is Extract<Chunk, { type: "text" | "thinking" }> =>
c.type === "text" || c.type === "thinking",
);
const content = textChunks.map((c) => c.text).join("");
const toolCalls = msg.chunks
.filter((c): c is Extract<Chunk, { type: "tool-call" }> => c.type === "tool-call")
.map(
(c): OpenAIToolCall => ({
id: c.toolCallId,
type: "function",
function: {
name: c.toolName,
arguments: serializeToolArguments(c.input),
},
}),
);
if (toolCalls.length > 0) {
return {
role: "assistant",
content: content || null,
tool_calls: toolCalls,
};
}
return { role: "assistant", content };
}
function convertToolResultMessages(msg: ChatMessage): OpenAIMessage[] {
return msg.chunks
.filter((c): c is Extract<Chunk, { type: "tool-result" }> => c.type === "tool-result")
.map(
(c): OpenAIMessage => ({
role: "tool",
content: c.content,
tool_call_id: c.toolCallId,
}),
);
}
/**
* Serialize a tool-call's `input` into a JSON string the provider will accept.
*
* The OpenAI `arguments` field MUST be a valid JSON string. A broken chat can
* have a tool-call whose `input` is a raw malformed-JSON string (the model
* emitted broken JSON as the tool arguments and it was stored verbatim).
* Passing that string straight through makes the provider 400
* `unexpected character` on EVERY continuation, bricking the chat.
*
* - object input → `JSON.stringify(input)` (regression, unchanged shape).
* - string input that is valid JSON → re-serialized to canonical JSON.
* - string input that fails to parse → a valid fallback object preserving a
* truncated hint of the original, so the chat can continue (the model sees
* its tool-call had no usable args and adjusts).
*
* Pure: input → output, no I/O.
*/
function serializeToolArguments(input: unknown): string {
if (typeof input === "string") {
try {
return JSON.stringify(JSON.parse(input));
} catch {
return JSON.stringify({ _malformed_arguments: input.slice(0, 200) });
}
}
return JSON.stringify(input);
}
|
