mirror of
https://github.com/binaricat/Netcatty.git
synced 2026-11-20 23:20:46 +00:00
Typed request compression ran before every Catty request (routine pre-turn and every step) and replaced image/file parts with a text placeholder, so attached images never reached the model. preserveAttachments now keeps image attachments on the routine pre-turn and step paths. Non-image files stay placeholders (chat providers reject e.g. text/plain; the model reads them via list_attachments/read_attachment). The explicit 413 retry still trades attachments for placeholders. Verified in the real Electron app: image reaches the provider as image_url (fake OpenAI-compatible server and DeepSeek deepseek-flash, which described the test image correctly); text-file turns still succeed via attachment tools; request prefix stability across turns unchanged vs main (DeepSeek cache hits ~98% on follow-up turns). Fixes #3764 Co-authored-by: yangfan <1413737839@qq.com>
167 lines
6.0 KiB
TypeScript
167 lines
6.0 KiB
TypeScript
import test from "node:test";
|
|
import assert from "node:assert/strict";
|
|
import type { ModelMessage } from "ai";
|
|
|
|
import {
|
|
compressMessagesForRequestTooLargeRetry,
|
|
compressVerboseText,
|
|
truncateTextWithHeadAndTail,
|
|
} from "./requestPayloadCompression.ts";
|
|
|
|
test("compressVerboseText collapses repeated blank lines and duplicate runs", () => {
|
|
const input = "line1\n\n\n\n\nline2\nsame\nsame\nsame\nsame\nline3";
|
|
const output = compressVerboseText(input);
|
|
assert.match(output, /line1\n\n\nline2/);
|
|
assert.ok(output.split("\nsame\n").length <= 3);
|
|
});
|
|
|
|
test("compressVerboseText normalizes terminal control noise and huge payload lines", () => {
|
|
const ansiProgress = "\u001b[31mstarting\u001b[0m\r10%\r20%\r100%\n";
|
|
const longLine = "x".repeat(8_000);
|
|
const base64 = "QUJDREVGR0hJSktMTU5PUFFSU1RVVldYWVo=".repeat(100);
|
|
const output = compressVerboseText(`${ansiProgress}${longLine}\n${base64}`);
|
|
|
|
assert.equal(output.includes("\u001b["), false);
|
|
assert.doesNotMatch(output, /10%|20%/);
|
|
assert.match(output, /100%/);
|
|
assert.match(output, /long line shortened/);
|
|
assert.match(output, /base64-like payload omitted/);
|
|
assert.ok(output.length < 5_000);
|
|
});
|
|
|
|
test("compressVerboseText preserves CRLF-delimited lines", () => {
|
|
assert.equal(compressVerboseText("alpha\r\nbeta\r\n"), "alpha\nbeta\n");
|
|
});
|
|
|
|
test("truncateTextWithHeadAndTail keeps both ends of long terminal output", () => {
|
|
const value = `${"A".repeat(500)}${"B".repeat(20_000)}${"C".repeat(500)}`;
|
|
const truncated = truncateTextWithHeadAndTail(value, 2_000);
|
|
assert.ok(truncated.startsWith("AAA"));
|
|
assert.ok(truncated.includes("[... output truncated for request size ...]"));
|
|
assert.ok(truncated.endsWith("CCC"));
|
|
assert.ok(truncated.length <= 2_000);
|
|
});
|
|
|
|
test("compressMessagesForRequestTooLargeRetry compresses messages without enforcing a byte budget", () => {
|
|
const messages: ModelMessage[] = [
|
|
{ role: "user", content: "run build" },
|
|
{
|
|
role: "tool",
|
|
content: [{
|
|
type: "tool-result",
|
|
toolCallId: "call-1",
|
|
toolName: "terminal_execute",
|
|
output: { type: "text", value: "X".repeat(200_000) },
|
|
}],
|
|
},
|
|
{
|
|
role: "user",
|
|
content: [
|
|
{ type: "text", text: "please inspect this image" },
|
|
{ type: "image", image: "A".repeat(1_000_000), mediaType: "image/png" },
|
|
],
|
|
},
|
|
];
|
|
|
|
const result = compressMessagesForRequestTooLargeRetry(messages);
|
|
|
|
assert.equal(result.didAdjust, true);
|
|
assert.deepEqual(Object.keys(result).sort(), ["didAdjust", "messages"]);
|
|
assert.equal(result.messages.length, messages.length);
|
|
|
|
const toolContent = result.messages[1].content;
|
|
assert.ok(Array.isArray(toolContent));
|
|
const toolPart = toolContent[0] as { output?: { value?: string } };
|
|
assert.ok((toolPart.output?.value?.length ?? 0) < 5_000);
|
|
|
|
const userContent = result.messages[2].content;
|
|
assert.ok(Array.isArray(userContent));
|
|
assert.deepEqual(userContent[1], {
|
|
type: "text",
|
|
text: "[image attachment omitted to keep the AI request small: mediaType=image/png, 1000000 chars]",
|
|
});
|
|
});
|
|
|
|
test("compressMessagesForRequestTooLargeRetry reports no adjustment for compact messages", () => {
|
|
const messages: ModelMessage[] = [{ role: "user", content: "hello" }];
|
|
|
|
const result = compressMessagesForRequestTooLargeRetry(messages);
|
|
|
|
assert.equal(result.didAdjust, false);
|
|
assert.deepEqual(result.messages, messages);
|
|
});
|
|
|
|
test("compressMessagesForRequestTooLargeRetry keeps attachments when preserveAttachments is set", () => {
|
|
const imageData = "A".repeat(300_000);
|
|
const fileData = Buffer.from("label,hostname\nprod,prod.example.com\n").toString("base64");
|
|
const messages: ModelMessage[] = [
|
|
{
|
|
role: "user",
|
|
content: [
|
|
{ type: "text", text: "what does this screenshot say?" },
|
|
{ type: "file", data: imageData, mediaType: "image/png" },
|
|
],
|
|
},
|
|
{
|
|
role: "user",
|
|
content: [
|
|
{ type: "text", text: "import these hosts" },
|
|
{ type: "file", data: fileData, mediaType: "text/csv", filename: "hosts.csv" },
|
|
],
|
|
},
|
|
{
|
|
role: "tool",
|
|
content: [{
|
|
type: "tool-result",
|
|
toolCallId: "call-1",
|
|
toolName: "terminal_execute",
|
|
output: { type: "text", value: "X".repeat(200_000) },
|
|
}],
|
|
},
|
|
];
|
|
|
|
const result = compressMessagesForRequestTooLargeRetry(messages, { preserveAttachments: true });
|
|
|
|
// Text and tool output are still reduced. Images stay; other files become placeholders.
|
|
assert.equal(result.didAdjust, true);
|
|
|
|
const imageContent = result.messages[0].content;
|
|
assert.ok(Array.isArray(imageContent));
|
|
assert.deepEqual(imageContent[1], messages[0].content[1]);
|
|
assert.doesNotMatch(JSON.stringify(imageContent), /attachment omitted to keep the AI request small/);
|
|
|
|
const fileContent = result.messages[1].content;
|
|
assert.ok(Array.isArray(fileContent));
|
|
assert.deepEqual(fileContent[1], {
|
|
type: "text",
|
|
text: `[file attachment omitted to keep the AI request small: filename=hosts.csv, mediaType=text/csv, ${fileData.length} chars]`,
|
|
});
|
|
|
|
const toolContent = result.messages[2].content;
|
|
assert.ok(Array.isArray(toolContent));
|
|
const toolPart = toolContent[0] as { output?: { value?: string } };
|
|
assert.ok((toolPart.output?.value?.length ?? 0) < 5_000);
|
|
});
|
|
|
|
test("compressMessagesForRequestTooLargeRetry still placeholders text/plain files when preserveAttachments is set", () => {
|
|
const text = "hello from a text file";
|
|
const messages: ModelMessage[] = [
|
|
{
|
|
role: "user",
|
|
content: [
|
|
{ type: "text", text: "read this note" },
|
|
{ type: "file", data: text, mediaType: "text/plain", filename: "note.txt" },
|
|
],
|
|
},
|
|
];
|
|
|
|
const result = compressMessagesForRequestTooLargeRetry(messages, { preserveAttachments: true });
|
|
const content = result.messages[0].content;
|
|
assert.ok(Array.isArray(content));
|
|
assert.deepEqual(content[0], { type: "text", text: "read this note" });
|
|
assert.deepEqual(content[1], {
|
|
type: "text",
|
|
text: `[file attachment omitted to keep the AI request small: filename=note.txt, mediaType=text/plain, ${text.length} chars]`,
|
|
});
|
|
});
|