@@ -6,10 +6,20 @@ import {
66 annotateAnthropicCacheBreakpoints,
77 applyStepPruning,
88 compactToolResultPayloads,
9+ getMessageStats,
910 } from "./turn-context.ts";
1011 import { prepareTurnMessages } from "./turn-prepare-messages.ts";
1112 import { buildTurnProviderOptions } from "./turn-provider-options.ts";
1213
14+export interface StepPruneRecord {
15+ removedMessageCount: number;
16+ removedBytes: number;
17+ beforeMessageCount: number;
18+ afterMessageCount: number;
19+ beforeTotalBytes: number;
20+ afterTotalBytes: number;
21+}
22+
1323 type StreamTextOptions = Parameters<typeof streamText>[0];
1424 type CoreMessage = NonNullable<StreamTextOptions["messages"]>[number];
1525 type CoreModel = StreamTextOptions["model"];
@@ -54,6 +64,7 @@ interface BuildStreamTextRequestInput {
5464 onStepFinish: NonNullable<StreamTextOptions["onStepFinish"]>;
5565 signal: AbortSignal | undefined;
5666 providerOptions: Record<string, unknown>;
67+ stepPruneQueue: StepPruneRecord[];
5768 }
5869
5970 export function buildStreamTextRequest(
@@ -87,10 +98,24 @@ export function buildStreamTextRequest(
8798 }
8899 // Steps 1+: use step-aware pruning that preserves the cached
89100 // prefix and keeps the model's chain-of-thought.
101+ const preCount = messages.length;
90102 const pruned = applyStepPruning(
91103 messages as CoreMessage[],
92104 initialMessageCount,
93105 );
106+ const postCount = pruned.length;
107+ if (postCount < preCount) {
108+ const pre = getMessageStats(messages as CoreMessage[]);
109+ const post = getMessageStats(pruned);
110+ input.stepPruneQueue.push({
111+ beforeMessageCount: pre.messageCount,
112+ afterMessageCount: post.messageCount,
113+ removedMessageCount: pre.messageCount - post.messageCount,
114+ beforeTotalBytes: pre.totalBytes,
115+ afterTotalBytes: post.totalBytes,
116+ removedBytes: pre.totalBytes - post.totalBytes,
117+ });
118+ }
94119 const compacted = compactToolResultPayloads(pruned);
95120 const final = isAnthropic
96121 ? annotateAnthropicCacheBreakpoints(compacted)