Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions apps/server/scripts/acp-mock-agent.ts
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,7 @@ const emitOverlappingXAiPromptCompleteOutOfOrder =
const failPrompt = process.env.T3_ACP_FAIL_PROMPT === "1";
const failSetConfigOption = process.env.T3_ACP_FAIL_SET_CONFIG_OPTION === "1";
const exitOnSetConfigOption = process.env.T3_ACP_EXIT_ON_SET_CONFIG_OPTION === "1";
const emitContextUsage = process.env.T3_ACP_EMIT_CONTEXT_USAGE === "1";
const promptResponseText = process.env.T3_ACP_PROMPT_RESPONSE_TEXT;
const initialGrokReasoningEffort =
process.env.T3_ACP_INITIAL_GROK_REASONING_EFFORT?.trim() || undefined;
Expand Down Expand Up @@ -1061,6 +1062,7 @@ const program = Effect.gen(function* () {

yield* agent.client.sessionUpdate({
sessionId: requestedSessionId,
...(emitContextUsage ? { _meta: { totalTokens: 1615 } } : {}),
update: {
sessionUpdate: "agent_message_chunk",
content: { type: "text", text: promptResponseText ?? "hello from mock" },
Expand Down
22 changes: 22 additions & 0 deletions apps/server/src/provider/Layers/CursorAdapter.ts
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,7 @@ import {
makeAcpPlanUpdatedEvent,
makeAcpRequestOpenedEvent,
makeAcpRequestResolvedEvent,
makeAcpTokenUsageEvent,
makeAcpToolCallEvent,
} from "../acp/AcpCoreRuntimeEvents.ts";
import {
Expand Down Expand Up @@ -867,6 +868,27 @@ export function makeCursorAdapter(
}),
);
return;
case "UsageUpdated": {
yield* logNative(
ctx.threadId,
"session/update",
event.rawPayload,
"acp.jsonrpc",
);
const usageEvent = makeAcpTokenUsageEvent({
stamp: yield* makeEventStamp(),
provider: PROVIDER,
threadId: ctx.threadId,
turnId: ctx.activeTurnId,
used: event.used,
size: event.size,
rawPayload: event.rawPayload,
});
if (usageEvent) {
yield* offerRuntimeEvent(usageEvent);
}
return;
}
}
}),
),
Expand Down
130 changes: 130 additions & 0 deletions apps/server/src/provider/Layers/GrokAdapter.test.ts
Original file line number Diff line number Diff line change
Expand Up @@ -27,10 +27,13 @@ import {

import { ServerConfig } from "../../config.ts";
import {
GROK_DEFAULT_CONTEXT_WINDOW,
grokContextWindowAfterModelChange,
grokPromptSettlementBelongsToContext,
isGrokEnterPlanModeToolCall,
makeGrokAdapter,
nextGrokPlanModeActive,
resolveGrokContextUsage,
selectGrokPermissionOptionId,
} from "./GrokAdapter.ts";
const decodeGrokSettings = Schema.decodeSync(GrokSettings);
Expand Down Expand Up @@ -179,6 +182,83 @@ it("prefers allow_always when Grok offers it", () => {
assert.equal(selectGrokPermissionOptionId(request, "accept"), "allow-once");
});

it("keeps occupancy snapshots when only the context window changes", () => {
assert.deepStrictEqual(
resolveGrokContextUsage({
used: 1615,
eventSize: 0,
lastUsed: 1615,
lastKnownMaxTokens: 500_000,
lastEmittedMaxTokens: 500_000,
}),
undefined,
);
assert.deepStrictEqual(
resolveGrokContextUsage({
used: 1615,
eventSize: 256_000,
lastUsed: 1615,
lastKnownMaxTokens: 500_000,
lastEmittedMaxTokens: 500_000,
}),
{ used: 1615, size: 256_000 },
);
assert.deepStrictEqual(
resolveGrokContextUsage({
used: 1615,
eventSize: 0,
lastUsed: 1615,
lastKnownMaxTokens: 256_000,
lastEmittedMaxTokens: 500_000,
}),
{ used: 1615, size: 256_000 },
);
assert.equal(
resolveGrokContextUsage({
used: 1_200,
eventSize: 0,
lastUsed: undefined,
lastKnownMaxTokens: undefined,
lastEmittedMaxTokens: undefined,
})?.size,
GROK_DEFAULT_CONTEXT_WINDOW,
);
});

it("refreshes the cached window when the Grok model changes", () => {
const windows = new Map([
["grok-4.6", 500_000],
["grok-code", 256_000],
]);
assert.equal(
grokContextWindowAfterModelChange({
previousModelId: "grok-4.6",
nextModelId: "grok-4.6",
windows,
lastKnownMaxTokens: 500_000,
}),
500_000,
);
assert.equal(
grokContextWindowAfterModelChange({
previousModelId: "grok-4.6",
nextModelId: "grok-code",
windows,
lastKnownMaxTokens: 500_000,
}),
256_000,
);
assert.equal(
grokContextWindowAfterModelChange({
previousModelId: "grok-4.6",
nextModelId: "grok-unknown",
windows,
lastKnownMaxTokens: 500_000,
}),
undefined,
);
});

it("requires a settlement to match the live Grok turn", () => {
const staleTurnId = TurnId.make("stale-turn");
const replacementTurnId = TurnId.make("replacement-turn");
Expand Down Expand Up @@ -278,6 +358,56 @@ it.layer(grokAdapterTestLayer)("GrokAdapterLive", (it) => {
}),
);

it.effect("emits context occupancy from session-update _meta.totalTokens", () =>
Effect.gen(function* () {
const threadId = ThreadId.make("grok-context-window-thread");
const wrapperPath = yield* Effect.promise(() =>
makeMockGrokWrapper({ T3_ACP_EMIT_CONTEXT_USAGE: "1" }),
);
const adapter = yield* makeTestAdapter(wrapperPath);

const runtimeEvents: ProviderRuntimeEvent[] = [];
const turnCompleted = yield* Deferred.make<void>();
const runtimeEventsFiber = yield* Stream.runForEach(adapter.streamEvents, (event) =>
Effect.sync(() => {
runtimeEvents.push(event);
}).pipe(
Effect.andThen(
event.type === "turn.completed"
? Deferred.succeed(turnCompleted, undefined)
: Effect.void,
),
),
).pipe(Effect.forkChild);

yield* adapter.startSession({
threadId,
provider: ProviderDriverKind.make("grok"),
cwd: process.cwd(),
runtimeMode: "full-access",
modelSelection: { instanceId: ProviderInstanceId.make("grok"), model: "grok-4.6" },
});

yield* adapter.sendTurn({
threadId,
input: "hello grok",
attachments: [],
});

yield* Deferred.await(turnCompleted);
yield* Fiber.interrupt(runtimeEventsFiber);

const usageEvent = runtimeEvents.find((event) => event.type === "thread.token-usage.updated");
assert.equal(usageEvent?.type, "thread.token-usage.updated");
if (usageEvent?.type === "thread.token-usage.updated") {
assert.equal(usageEvent.payload.usage.usedTokens, 1615);
assert.equal(usageEvent.payload.usage.maxTokens, 500_000);
}

yield* adapter.stopSession(threadId);
}),
);

it.effect("closes the ACP child process when a session stops", () =>
Effect.gen(function* () {
const threadId = ThreadId.make("grok-stop-session-close");
Expand Down
90 changes: 89 additions & 1 deletion apps/server/src/provider/Layers/GrokAdapter.ts
Original file line number Diff line number Diff line change
Expand Up @@ -55,12 +55,16 @@ import {
makeAcpPlanUpdatedEvent,
makeAcpRequestOpenedEvent,
makeAcpRequestResolvedEvent,
makeAcpTokenUsageEvent,
makeAcpToolCallEvent,
} from "../acp/AcpCoreRuntimeEvents.ts";
import { parsePermissionRequest } from "../acp/AcpRuntimeModel.ts";
import { makeAcpNativeLoggerFactory } from "../acp/AcpNativeLogging.ts";
import {
applyGrokAcpModelSelection,
contextWindowForGrokModelId,
contextWindowFromGrokSessionSetup,
contextWindowsFromGrokSessionModels,
currentGrokModelIdFromSessionSetup,
currentGrokReasoningEffortFromSessionSetup,
makeGrokAcpRuntime,
Expand Down Expand Up @@ -169,6 +173,10 @@ interface GrokSessionContext {
promptResponsesReady: number;
currentModelId: string | undefined;
currentReasoningEffort: string | undefined;
contextWindows: ReadonlyMap<string, number>;
lastContextTokens: number | undefined;
lastKnownMaxTokens: number | undefined;
lastEmittedMaxTokens: number | undefined;
stopped: boolean;
}

Expand Down Expand Up @@ -314,6 +322,37 @@ function selectAutoApprovedPermissionOption(
);
}

export const GROK_DEFAULT_CONTEXT_WINDOW = 500_000;

export function resolveGrokContextUsage(input: {
readonly used: number;
readonly eventSize: number;
readonly lastUsed: number | undefined;
readonly lastKnownMaxTokens: number | undefined;
readonly lastEmittedMaxTokens: number | undefined;
}): { readonly used: number; readonly size: number } | undefined {
const size =
input.eventSize > 0
? input.eventSize
: (input.lastKnownMaxTokens ?? GROK_DEFAULT_CONTEXT_WINDOW);
if (input.lastUsed === input.used && input.lastEmittedMaxTokens === size) {
return undefined;
}
return { used: input.used, size };
}

export function grokContextWindowAfterModelChange(input: {
readonly previousModelId: string | undefined;
readonly nextModelId: string | undefined;
readonly windows: ReadonlyMap<string, number>;
readonly lastKnownMaxTokens: number | undefined;
}): number | undefined {
if (input.nextModelId === input.previousModelId) {
return input.lastKnownMaxTokens;
}
return contextWindowForGrokModelId(input.windows, input.nextModelId);
}

function completedStopReasonFromPromptResponse(
response: EffectAcpSchema.PromptResponse | undefined,
): EffectAcpSchema.StopReason | null {
Expand Down Expand Up @@ -1300,6 +1339,13 @@ export function makeGrokAdapter(grokSettings: GrokSettings, options?: GrokAdapte
requestedStartReasoningEffort !== undefined
? normalizeGrokReasoningEffort(requestedStartReasoningEffort)
: currentStartReasoningEffort,
contextWindows: contextWindowsFromGrokSessionModels(started.sessionSetupResult.models),
lastContextTokens: undefined,
lastKnownMaxTokens: contextWindowFromGrokSessionSetup(
Comment thread
macroscopeapp[bot] marked this conversation as resolved.
started.sessionSetupResult,
boundModelId,
),
lastEmittedMaxTokens: undefined,
stopped: false,
};

Expand All @@ -1313,7 +1359,8 @@ export function makeGrokAdapter(grokSettings: GrokSettings, options?: GrokAdapte
if (
event._tag === "PlanUpdated" ||
event._tag === "ToolCallUpdated" ||
event._tag === "ContentDelta"
event._tag === "ContentDelta" ||
event._tag === "UsageUpdated"
) {
yield* logNative(ctx.threadId, "session/update", event.rawPayload);
}
Expand All @@ -1322,6 +1369,41 @@ export function makeGrokAdapter(grokSettings: GrokSettings, options?: GrokAdapte
return;
}

if (event._tag === "UsageUpdated") {
const usageTurnId = resolveNotificationTurnId(ctx);
if (usageTurnId !== undefined && ctx.interruptedTurnIds.has(usageTurnId)) {

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Medium Layers/GrokAdapter.ts:1374

A queued UsageUpdated from a cancelled prompt is emitted as thread.token-usage.updated after interruptTurn clears activeTurnId, so stale usage overwrites the composer context meter. Treat usageTurnId === undefined as stale here, in addition to checking interruptedTurnIds.

🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/server/src/provider/Layers/GrokAdapter.ts around line 1374:

A queued `UsageUpdated` from a cancelled prompt is emitted as `thread.token-usage.updated` after `interruptTurn` clears `activeTurnId`, so stale usage overwrites the composer context meter. Treat `usageTurnId === undefined` as stale here, in addition to checking `interruptedTurnIds`.

return;
}
const snapshot = resolveGrokContextUsage({
used: event.used,
eventSize: event.size,
lastUsed: ctx.lastContextTokens,
lastKnownMaxTokens: ctx.lastKnownMaxTokens,
lastEmittedMaxTokens: ctx.lastEmittedMaxTokens,
});
if (!snapshot) {
return;
}
ctx.lastContextTokens = snapshot.used;
if (event.size > 0) {
ctx.lastKnownMaxTokens = event.size;
}
ctx.lastEmittedMaxTokens = snapshot.size;
const usageEvent = makeAcpTokenUsageEvent({
stamp: yield* makeEventStamp(),
provider: PROVIDER,
threadId: ctx.threadId,
turnId: usageTurnId,
used: snapshot.used,
size: snapshot.size,
rawPayload: event.rawPayload,
});
if (usageEvent) {
yield* offerRuntimeEvent(usageEvent);
}
return;
}

const notificationTurnId = resolveNotificationTurnId(ctx);
if (
notificationTurnId === undefined ||
Expand Down Expand Up @@ -1574,6 +1656,12 @@ export function makeGrokAdapter(grokSettings: GrokSettings, options?: GrokAdapte
mapError: (cause) =>
mapAcpToAdapterError(PROVIDER, input.threadId, "session/set_model", cause),
});
ctx.lastKnownMaxTokens = grokContextWindowAfterModelChange({
previousModelId: ctx.currentModelId,
nextModelId: currentModelId,
windows: ctx.contextWindows,
lastKnownMaxTokens: ctx.lastKnownMaxTokens,
});
ctx.currentModelId = currentModelId;
if (requestedTurnReasoningEffort !== undefined) {
ctx.currentReasoningEffort = normalizeGrokReasoningEffort(
Expand Down
33 changes: 33 additions & 0 deletions apps/server/src/provider/acp/AcpCoreRuntimeEvents.test.ts
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@ import {
makeAcpPlanUpdatedEvent,
makeAcpRequestOpenedEvent,
makeAcpRequestResolvedEvent,
makeAcpTokenUsageEvent,
makeAcpToolCallEvent,
} from "./AcpCoreRuntimeEvents.ts";

Expand Down Expand Up @@ -191,5 +192,37 @@ describe("AcpCoreRuntimeEvents", () => {
status: "inProgress",
},
});

expect(
makeAcpTokenUsageEvent({
stamp,
provider: ProviderDriverKind.make("grok"),
threadId: "thread-1" as never,
turnId,
used: 12_345,
size: 200_000,
rawPayload: { sessionUpdate: "usage_update" },
}),
).toMatchObject({
type: "thread.token-usage.updated",
payload: {
usage: {
usedTokens: 12_345,
maxTokens: 200_000,
},
},
});

expect(
makeAcpTokenUsageEvent({
stamp,
provider: ProviderDriverKind.make("grok"),
threadId: "thread-1" as never,
turnId,
used: 0,
size: 200_000,
rawPayload: { sessionUpdate: "usage_update" },
}),
).toBeNull();
});
});
Loading
Loading