Skip to content

Commit 01466a7

Browse files
fix(cursor): estimate pruned context payload
Co-authored-by: Copilot App <[email protected]>
1 parent 5335320 commit 01466a7

6 files changed

Lines changed: 95 additions & 28 deletions

File tree

‎src/adapters/cursor/protobuf-request.ts‎

Lines changed: 31 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -77,13 +77,10 @@ function buildRequestContext() {
7777
});
7878
}
7979

80-
function jsonBlob(value: unknown): Uint8Array {
81-
return encoder.encode(JSON.stringify(value));
82-
}
83-
8480
type StoredRootBlob = {
8581
id: Uint8Array;
8682
byteLength: number;
83+
serialized: string;
8784
role: "system" | "user" | "assistant" | "toolResult";
8885
messageIndex?: number;
8986
/** Original JSON text payload used when an active tool result must be truncated to fit. */
@@ -95,10 +92,12 @@ function storedRootBlob(
9592
role: StoredRootBlob["role"],
9693
opts?: { messageIndex?: number; text?: string },
9794
): StoredRootBlob {
98-
const data = jsonBlob(value);
95+
const serialized = JSON.stringify(value);
96+
const data = encoder.encode(serialized);
9997
return {
10098
id: storeCursorBlob(data),
10199
byteLength: data.byteLength,
100+
serialized,
102101
role,
103102
...(opts?.messageIndex !== undefined ? { messageIndex: opts.messageIndex } : {}),
104103
...(opts?.text !== undefined ? { text: opts.text } : {}),
@@ -165,6 +164,7 @@ function rootPromptMessages(request: CursorRunRequest): {
165164
ids: Uint8Array[];
166165
byteLength: number;
167166
historyMessageStart: number;
167+
serialized: string[];
168168
} {
169169
const entries = systemPromptBlobs(request);
170170
const systemEntryCount = entries.length;
@@ -174,6 +174,7 @@ function rootPromptMessages(request: CursorRunRequest): {
174174
ids: entries.map(entry => entry.id),
175175
byteLength: entries.reduce((sum, entry) => sum + entry.byteLength, 0),
176176
historyMessageStart: 0,
177+
serialized: entries.map(entry => entry.serialized),
177178
};
178179
}
179180

@@ -290,6 +291,7 @@ function rootPromptMessages(request: CursorRunRequest): {
290291
ids: selected.map(entry => entry.id),
291292
byteLength: selected.reduce((sum, entry) => sum + entry.byteLength, 0),
292293
historyMessageStart,
294+
serialized: selected.map(entry => entry.serialized),
293295
};
294296
}
295297

@@ -504,25 +506,37 @@ export function activePromptText(request: CursorRunRequest): string {
504506
return last?.role === "tool" ? last.content : "";
505507
}
506508

509+
function modelVisibleRequestPayload(request: CursorRunRequest) {
510+
const rawText = activePromptText(request);
511+
const lastRole = request.messages.at(-1)?.role;
512+
const text = lastRole === "user" || lastRole === "developer"
513+
? appendCursorShellAliasHint(request.tools, appendCursorGenericToolUseHint(request.tools, rawText))
514+
: rawText;
515+
return {
516+
text,
517+
roots: rootPromptMessages(request),
518+
mcpToolDefs: buildCursorToolDefinitions(
519+
cursorToolsForActivePrompt(request.tools, rawText, request.toolChoice),
520+
request.toolChoice,
521+
),
522+
};
523+
}
524+
507525
/**
508526
* Conservative request-local fallback when Cursor emits neither a checkpoint nor a previously
509527
* observed context total. This is deliberately not persisted: a later authoritative checkpoint
510528
* remains the only source carried across turns.
511529
*/
512530
export function estimateCursorInputTokens(request: CursorRunRequest): number {
513-
return estimateTokens(JSON.stringify({
514-
system: request.system,
515-
messages: request.rawMessages ?? request.messages,
516-
tools: request.tools,
517-
}), request.modelId);
531+
const payload = modelVisibleRequestPayload(request);
532+
const lastRawIsToolResult = request.rawMessages?.at(-1)?.role === "toolResult";
533+
const action = !lastRawIsToolResult && payload.text.trim().length > 0 ? payload.text : "";
534+
return estimateTokens(JSON.stringify({ roots: payload.roots.serialized, action, tools: payload.mcpToolDefs }), request.modelId);
518535
}
519536

520537
export function encodeCursorRunRequest(request: CursorRunRequest): Uint8Array {
521-
const rawText = activePromptText(request);
522-
const lastRole = request.messages.at(-1)?.role;
523-
const text = lastRole === "user" || lastRole === "developer"
524-
? appendCursorShellAliasHint(request.tools, appendCursorGenericToolUseHint(request.tools, rawText))
525-
: rawText;
538+
const payload = modelVisibleRequestPayload(request);
539+
const text = payload.text;
526540
// Tool-result-only turns resume the remembered Cursor conversation with results in history.
527541
const lastRawIsToolResult = request.rawMessages?.at(-1)?.role === "toolResult";
528542
const actionCase = !lastRawIsToolResult && text.trim().length > 0
@@ -547,7 +561,7 @@ export function encodeCursorRunRequest(request: CursorRunRequest): Uint8Array {
547561
}),
548562
},
549563
});
550-
const rootPromptMessagesState = rootPromptMessages(request);
564+
const rootPromptMessagesState = payload.roots;
551565
const rootPromptMessageIds = rootPromptMessagesState.ids;
552566
const turnIds = conversationTurns(request, rootPromptMessagesState.historyMessageStart);
553567
debugProviderDiagnostic("cursor", "run-request", {
@@ -612,11 +626,7 @@ export function encodeCursorRunRequest(request: CursorRunRequest): Uint8Array {
612626
// the event-state `clientToolNames` use (live-transport.ts). Advertising the raw `request.tools`
613627
// here would let mcp_tools expose a tool that the event state does not recognize for a generic
614628
// tool-count prompt, so a call to it would be rejected as an unknown Responses tool.
615-
...(() => {
616-
const visibleTools = cursorToolsForActivePrompt(request.tools, activePromptText(request), request.toolChoice);
617-
const mcpToolDefs = buildCursorToolDefinitions(visibleTools, request.toolChoice);
618-
return mcpToolDefs.length > 0 ? { mcpTools: create(McpToolsSchema, { mcpTools: mcpToolDefs }) } : {};
619-
})(),
629+
...(payload.mcpToolDefs.length > 0 ? { mcpTools: create(McpToolsSchema, { mcpTools: payload.mcpToolDefs }) } : {}),
620630
});
621631

622632
const message = create(AgentClientMessageSchema, {

‎src/lib/token-estimate.ts‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -21,7 +21,7 @@ const DEFAULT_CHARS_PER_TOKEN = 4;
2121
*/
2222
const KIRO_CHARS_PER_TOKEN = 3.5;
2323

24-
const KIRO_MODEL_PREFIXES = ["kiro", "claude", "deepseek", "minimax", "glm", "qwen"];
24+
const KIRO_MODEL_PREFIXES = ["kiro", "claude", "deepseek", "minimax", "glm", "qwen", "grok", "xai/"];
2525

2626
/** Model-aware chars-per-token ratio. Unknown models fall back to the generic English ratio. */
2727
export function charsPerToken(modelId?: string): number {

‎tests/cursor-blob.test.ts‎

Lines changed: 43 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,12 +1,14 @@
11
import { describe, expect, test } from "bun:test";
22
import { createHash } from "node:crypto";
33
import { create, fromBinary } from "@bufbuild/protobuf";
4+
import { estimateTokens } from "../src/lib/token-estimate";
45
import { handleCursorNativeKv, storeCursorBlob } from "../src/adapters/cursor/native-exec";
56
import {
67
CURSOR_EXTERNAL_ROOT_BYTE_LIMIT,
78
CURSOR_EXTERNAL_ROOT_BLOB_LIMIT,
89
CURSOR_ROUTING_LEVEL_PARAMETER_ID,
910
encodeCursorRunRequest,
11+
estimateCursorInputTokens,
1012
} from "../src/adapters/cursor/protobuf-request";
1113
import {
1214
AgentClientMessageSchema,
@@ -146,6 +148,47 @@ describe("Cursor blob handshake", () => {
146148
expect(rootRoles.find(role => role !== "system")).toBe("user");
147149
});
148150

151+
test("estimates only the externally replayed root history and filtered tool catalog", () => {
152+
const current = { role: "user" as const, content: "Use any 10 tools", timestamp: 10_000 };
153+
const retained = [
154+
{ role: "user" as const, content: "newest retained user", timestamp: 9_998 },
155+
{ role: "assistant" as const, model: "cursor/grok-4.5", content: [{ type: "text" as const, text: "newest retained assistant" }], timestamp: 9_999 },
156+
current,
157+
];
158+
const dropped = Array.from({ length: 200 }, (_, index) => ({
159+
role: "user" as const,
160+
content: `old-${index}:${"x".repeat(4_000)}`,
161+
timestamp: index,
162+
}));
163+
const exec = { name: "exec_command", description: "Run a command", parameters: {} };
164+
const filtered = { name: "filtered", description: "z".repeat(50_000), parameters: {} };
165+
const common = {
166+
modelId: "grok-4.5",
167+
conversationId: "c-estimate",
168+
system: ["system"],
169+
messages: [{ role: "user" as const, content: current.content }],
170+
tools: [exec, filtered],
171+
};
172+
173+
const request = { ...common, rawMessages: [...dropped, ...retained] };
174+
const bytes = encodeCursorRunRequest(request);
175+
const msg = fromBinary(AgentClientMessageSchema, bytes);
176+
const run = msg.message.case === "runRequest" ? msg.message.value : undefined;
177+
const payload = JSON.stringify({
178+
roots: (run?.conversationState?.rootPromptMessagesJson ?? [])
179+
.map(blobId => new TextDecoder().decode(blobData(blobId))),
180+
action: actionText(bytes),
181+
tools: run?.mcpTools?.mcpTools ?? [],
182+
});
183+
184+
expect(estimateCursorInputTokens(request)).toBe(estimateTokens(payload, request.modelId));
185+
expect(estimateCursorInputTokens(request)).toBeLessThan(estimateTokens(JSON.stringify({
186+
roots: [...dropped, ...retained],
187+
action: current.content,
188+
tools: [exec, filtered],
189+
}), request.modelId));
190+
});
191+
149192
test("preserves oversized active tool result under external byte budget", () => {
150193
const huge = "y".repeat(CURSOR_EXTERNAL_ROOT_BYTE_LIMIT);
151194
const bytes = encodeCursorRunRequest({

‎tests/cursor-interaction-query.test.ts‎

Lines changed: 10 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -171,12 +171,19 @@ describe("partialUsageFromEventState", () => {
171171
test("uses the request-local estimate after output when no checkpoint or carry exists", async () => {
172172
const { partialUsageFromEventState } = await import("../src/adapters/cursor/live-transport");
173173
const { createCursorProtobufEventState } = await import("../src/adapters/cursor/protobuf-events");
174-
const state = createCursorProtobufEventState({ estimatedInputTokens: 1_200 });
174+
const { estimateCursorInputTokens } = await import("../src/adapters/cursor/protobuf-request");
175+
const estimatedInputTokens = estimateCursorInputTokens({
176+
modelId: "grok-4.5",
177+
conversationId: "partial-without-checkpoint",
178+
system: ["You are helpful."],
179+
messages: [{ role: "user", content: "Read this JSON: {\\\"path\\\":\\\"src/app.ts\\\"}" }],
180+
});
181+
const state = createCursorProtobufEventState({ estimatedInputTokens });
175182
state.usage.outputTokens = 7;
176183
expect(partialUsageFromEventState(state)).toEqual({
177-
inputTokens: 1_200,
184+
inputTokens: estimatedInputTokens,
178185
outputTokens: 7,
179-
totalTokens: 1_207,
186+
totalTokens: estimatedInputTokens + 7,
180187
estimated: true,
181188
});
182189
});

‎tests/cursor-protobuf-events.test.ts‎

Lines changed: 9 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -19,6 +19,7 @@ import {
1919
finalizeTurnEvents,
2020
mapCursorProtobufServerMessage,
2121
} from "../src/adapters/cursor/protobuf-events";
22+
import { estimateCursorInputTokens } from "../src/adapters/cursor/protobuf-request";
2223

2324
const encoder = new TextEncoder();
2425

@@ -414,11 +415,17 @@ describe("Cursor protobuf tool-call events", () => {
414415
});
415416

416417
test("uses a request-local input estimate when no checkpoint or carry-forward context exists", () => {
417-
const state = createCursorProtobufEventState({ estimatedInputTokens: 1_200 });
418+
const estimatedInputTokens = estimateCursorInputTokens({
419+
modelId: "grok-4.5",
420+
conversationId: "restart-without-checkpoint",
421+
system: ["You are helpful."],
422+
messages: [{ role: "user", content: "Explain this JSON: {\\\"path\\\":\\\"src/app.ts\\\"}" }],
423+
});
424+
const state = createCursorProtobufEventState({ estimatedInputTokens });
418425
state.usage.outputTokens = 7;
419426

420427
expect(finalizeTurnEvents(state)).toEqual([
421-
{ type: "done", usage: { inputTokens: 1_200, outputTokens: 7, totalTokens: 1_207, estimated: true } },
428+
{ type: "done", usage: { inputTokens: estimatedInputTokens, outputTokens: 7, totalTokens: estimatedInputTokens + 7, estimated: true } },
422429
]);
423430
});
424431

‎tests/token-estimate.test.ts‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -25,7 +25,7 @@ describe("token-estimate sidecar", () => {
2525
});
2626

2727
test("kiro text models use the 3.5 ratio", () => {
28-
for (const m of ["kiro-auto", "claude-opus-4.8", "claude-opus-4.5", "deepseek-3.2", "minimax-m2.5", "minimax-m2.1", "glm-5", "qwen3-coder-next"]) {
28+
for (const m of ["kiro-auto", "claude-opus-4.8", "claude-opus-4.5", "deepseek-3.2", "minimax-m2.5", "minimax-m2.1", "glm-5", "qwen3-coder-next", "grok-4.5", "xai/grok-4.5"]) {
2929
expect(charsPerToken(m)).toBe(3.5);
3030
}
3131
});

0 commit comments

Comments
 (0)