feat(agent): opt-in tool-result context pruning
This commit is contained in:
Max Sumrall
committed by
Peter Steinberger
Peter Steinberger
parent
937e0265a3
commit
eeaa6ea46f
310
src/agents/pi-extensions/context-pruning/pruner.ts
Normal file
310
src/agents/pi-extensions/context-pruning/pruner.ts
Normal file
@@ -0,0 +1,310 @@
|
||||
import type { AgentMessage } from "@mariozechner/pi-agent-core";
|
||||
import type {
|
||||
ImageContent,
|
||||
TextContent,
|
||||
ToolResultMessage,
|
||||
} from "@mariozechner/pi-ai";
|
||||
import type { ExtensionContext } from "@mariozechner/pi-coding-agent";
|
||||
|
||||
import type { EffectiveContextPruningSettings } from "./settings.js";
|
||||
import { makeToolPrunablePredicate } from "./tools.js";
|
||||
|
||||
const CHARS_PER_TOKEN_ESTIMATE = 4;
|
||||
// We currently skip pruning tool results that contain images. Still, we count them (approx.) so
|
||||
// we start trimming prunable tool results earlier when image-heavy context is consuming the window.
|
||||
const IMAGE_CHAR_ESTIMATE = 8_000;
|
||||
|
||||
function asText(text: string): TextContent {
|
||||
return { type: "text", text };
|
||||
}
|
||||
|
||||
function collectTextSegments(
|
||||
content: ReadonlyArray<TextContent | ImageContent>,
|
||||
): string[] {
|
||||
const parts: string[] = [];
|
||||
for (const block of content) {
|
||||
if (block.type === "text") parts.push(block.text);
|
||||
}
|
||||
return parts;
|
||||
}
|
||||
|
||||
function estimateJoinedTextLength(parts: string[]): number {
|
||||
if (parts.length === 0) return 0;
|
||||
let len = 0;
|
||||
for (const p of parts) len += p.length;
|
||||
// Joined with "\n" separators between blocks.
|
||||
len += Math.max(0, parts.length - 1);
|
||||
return len;
|
||||
}
|
||||
|
||||
function takeHeadFromJoinedText(parts: string[], maxChars: number): string {
|
||||
if (maxChars <= 0 || parts.length === 0) return "";
|
||||
let remaining = maxChars;
|
||||
let out = "";
|
||||
for (let i = 0; i < parts.length && remaining > 0; i++) {
|
||||
if (i > 0) {
|
||||
out += "\n";
|
||||
remaining -= 1;
|
||||
if (remaining <= 0) break;
|
||||
}
|
||||
const p = parts[i];
|
||||
if (p.length <= remaining) {
|
||||
out += p;
|
||||
remaining -= p.length;
|
||||
} else {
|
||||
out += p.slice(0, remaining);
|
||||
remaining = 0;
|
||||
}
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
function takeTailFromJoinedText(parts: string[], maxChars: number): string {
|
||||
if (maxChars <= 0 || parts.length === 0) return "";
|
||||
let remaining = maxChars;
|
||||
const out: string[] = [];
|
||||
for (let i = parts.length - 1; i >= 0 && remaining > 0; i--) {
|
||||
const p = parts[i];
|
||||
if (p.length <= remaining) {
|
||||
out.push(p);
|
||||
remaining -= p.length;
|
||||
} else {
|
||||
out.push(p.slice(p.length - remaining));
|
||||
remaining = 0;
|
||||
break;
|
||||
}
|
||||
if (remaining > 0 && i > 0) {
|
||||
out.push("\n");
|
||||
remaining -= 1;
|
||||
}
|
||||
}
|
||||
out.reverse();
|
||||
return out.join("");
|
||||
}
|
||||
|
||||
function hasImageBlocks(
|
||||
content: ReadonlyArray<TextContent | ImageContent>,
|
||||
): boolean {
|
||||
for (const block of content) {
|
||||
if (block.type === "image") return true;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
function estimateMessageChars(message: AgentMessage): number {
|
||||
if (message.role === "user") {
|
||||
const content = message.content;
|
||||
if (typeof content === "string") return content.length;
|
||||
let chars = 0;
|
||||
for (const b of content) {
|
||||
if (b.type === "text") chars += b.text.length;
|
||||
if (b.type === "image") chars += IMAGE_CHAR_ESTIMATE;
|
||||
}
|
||||
return chars;
|
||||
}
|
||||
|
||||
if (message.role === "assistant") {
|
||||
let chars = 0;
|
||||
for (const b of message.content) {
|
||||
if (b.type === "text") chars += b.text.length;
|
||||
if (b.type === "thinking") chars += b.thinking.length;
|
||||
if (b.type === "toolCall") {
|
||||
try {
|
||||
chars += JSON.stringify(b.arguments ?? {}).length;
|
||||
} catch {
|
||||
chars += 128;
|
||||
}
|
||||
}
|
||||
}
|
||||
return chars;
|
||||
}
|
||||
|
||||
if (message.role === "toolResult") {
|
||||
let chars = 0;
|
||||
for (const b of message.content) {
|
||||
if (b.type === "text") chars += b.text.length;
|
||||
if (b.type === "image") chars += IMAGE_CHAR_ESTIMATE;
|
||||
}
|
||||
return chars;
|
||||
}
|
||||
|
||||
return 256;
|
||||
}
|
||||
|
||||
function estimateContextChars(messages: AgentMessage[]): number {
|
||||
return messages.reduce((sum, m) => sum + estimateMessageChars(m), 0);
|
||||
}
|
||||
|
||||
function findAssistantCutoffIndex(
|
||||
messages: AgentMessage[],
|
||||
keepLastAssistants: number,
|
||||
): number | null {
|
||||
// keepLastAssistants <= 0 => everything is potentially prunable.
|
||||
if (keepLastAssistants <= 0) return messages.length;
|
||||
|
||||
let remaining = keepLastAssistants;
|
||||
for (let i = messages.length - 1; i >= 0; i--) {
|
||||
if (messages[i]?.role !== "assistant") continue;
|
||||
remaining--;
|
||||
if (remaining === 0) return i;
|
||||
}
|
||||
|
||||
// Not enough assistant messages to establish a protected tail.
|
||||
return null;
|
||||
}
|
||||
|
||||
function softTrimToolResultMessage(params: {
|
||||
msg: ToolResultMessage;
|
||||
settings: EffectiveContextPruningSettings;
|
||||
}): ToolResultMessage | null {
|
||||
const { msg, settings } = params;
|
||||
// Ignore image tool results for now: these are often directly relevant and hard to partially prune safely.
|
||||
if (hasImageBlocks(msg.content)) return null;
|
||||
|
||||
const parts = collectTextSegments(msg.content);
|
||||
const rawLen = estimateJoinedTextLength(parts);
|
||||
if (rawLen <= settings.softTrim.maxChars) return null;
|
||||
|
||||
const headChars = Math.max(0, settings.softTrim.headChars);
|
||||
const tailChars = Math.max(0, settings.softTrim.tailChars);
|
||||
if (headChars + tailChars >= rawLen) return null;
|
||||
|
||||
const head = takeHeadFromJoinedText(parts, headChars);
|
||||
const tail = takeTailFromJoinedText(parts, tailChars);
|
||||
const trimmed = `${head}
|
||||
...
|
||||
${tail}`;
|
||||
|
||||
const note = `
|
||||
|
||||
[Tool result trimmed: kept first ${headChars} chars and last ${tailChars} chars of ${rawLen} chars.]`;
|
||||
|
||||
return { ...msg, content: [asText(trimmed + note)] };
|
||||
}
|
||||
|
||||
export function pruneContextMessages(params: {
|
||||
messages: AgentMessage[];
|
||||
settings: EffectiveContextPruningSettings;
|
||||
ctx: Pick<ExtensionContext, "model">;
|
||||
isToolPrunable?: (toolName: string) => boolean;
|
||||
contextWindowTokensOverride?: number;
|
||||
}): AgentMessage[] {
|
||||
const { messages, settings, ctx } = params;
|
||||
const contextWindowTokens =
|
||||
typeof params.contextWindowTokensOverride === "number" &&
|
||||
Number.isFinite(params.contextWindowTokensOverride) &&
|
||||
params.contextWindowTokensOverride > 0
|
||||
? params.contextWindowTokensOverride
|
||||
: ctx.model?.contextWindow;
|
||||
if (!contextWindowTokens || contextWindowTokens <= 0) return messages;
|
||||
|
||||
const charWindow = contextWindowTokens * CHARS_PER_TOKEN_ESTIMATE;
|
||||
if (charWindow <= 0) return messages;
|
||||
|
||||
const cutoffIndex = findAssistantCutoffIndex(
|
||||
messages,
|
||||
settings.keepLastAssistants,
|
||||
);
|
||||
if (cutoffIndex === null) return messages;
|
||||
|
||||
const isToolPrunable =
|
||||
params.isToolPrunable ?? makeToolPrunablePredicate(settings.tools);
|
||||
|
||||
if (settings.mode === "aggressive") {
|
||||
let next: AgentMessage[] | null = null;
|
||||
|
||||
for (let i = 0; i < cutoffIndex; i++) {
|
||||
const msg = messages[i];
|
||||
if (!msg || msg.role !== "toolResult") continue;
|
||||
if (!isToolPrunable(msg.toolName)) continue;
|
||||
if (hasImageBlocks(msg.content)) {
|
||||
continue;
|
||||
}
|
||||
|
||||
const alreadyCleared =
|
||||
msg.content.length === 1 &&
|
||||
msg.content[0]?.type === "text" &&
|
||||
msg.content[0].text === settings.hardClear.placeholder;
|
||||
if (alreadyCleared) continue;
|
||||
|
||||
const cleared: ToolResultMessage = {
|
||||
...msg,
|
||||
content: [asText(settings.hardClear.placeholder)],
|
||||
};
|
||||
if (!next) next = messages.slice();
|
||||
next[i] = cleared as unknown as AgentMessage;
|
||||
}
|
||||
|
||||
return next ?? messages;
|
||||
}
|
||||
|
||||
const totalCharsBefore = estimateContextChars(messages);
|
||||
let totalChars = totalCharsBefore;
|
||||
let ratio = totalChars / charWindow;
|
||||
if (ratio < settings.softTrimRatio) {
|
||||
return messages;
|
||||
}
|
||||
|
||||
const prunableToolIndexes: number[] = [];
|
||||
let next: AgentMessage[] | null = null;
|
||||
|
||||
for (let i = 0; i < cutoffIndex; i++) {
|
||||
const msg = messages[i];
|
||||
if (!msg || msg.role !== "toolResult") continue;
|
||||
if (!isToolPrunable(msg.toolName)) continue;
|
||||
if (hasImageBlocks(msg.content)) {
|
||||
continue;
|
||||
}
|
||||
prunableToolIndexes.push(i);
|
||||
|
||||
const updated = softTrimToolResultMessage({
|
||||
msg: msg as unknown as ToolResultMessage,
|
||||
settings,
|
||||
});
|
||||
if (!updated) continue;
|
||||
|
||||
const beforeChars = estimateMessageChars(msg);
|
||||
const afterChars = estimateMessageChars(updated as unknown as AgentMessage);
|
||||
totalChars += afterChars - beforeChars;
|
||||
if (!next) next = messages.slice();
|
||||
next[i] = updated as unknown as AgentMessage;
|
||||
}
|
||||
|
||||
const outputAfterSoftTrim = next ?? messages;
|
||||
ratio = totalChars / charWindow;
|
||||
if (ratio < settings.hardClearRatio) {
|
||||
return outputAfterSoftTrim;
|
||||
}
|
||||
if (!settings.hardClear.enabled) {
|
||||
return outputAfterSoftTrim;
|
||||
}
|
||||
|
||||
let prunableToolChars = 0;
|
||||
for (const i of prunableToolIndexes) {
|
||||
const msg = outputAfterSoftTrim[i];
|
||||
if (!msg || msg.role !== "toolResult") continue;
|
||||
prunableToolChars += estimateMessageChars(msg);
|
||||
}
|
||||
if (prunableToolChars < settings.minPrunableToolChars) {
|
||||
return outputAfterSoftTrim;
|
||||
}
|
||||
|
||||
for (const i of prunableToolIndexes) {
|
||||
if (ratio < settings.hardClearRatio) break;
|
||||
const msg = (next ?? messages)[i];
|
||||
if (!msg || msg.role !== "toolResult") continue;
|
||||
|
||||
const beforeChars = estimateMessageChars(msg);
|
||||
const cleared: ToolResultMessage = {
|
||||
...msg,
|
||||
content: [asText(settings.hardClear.placeholder)],
|
||||
};
|
||||
if (!next) next = messages.slice();
|
||||
next[i] = cleared as unknown as AgentMessage;
|
||||
const afterChars = estimateMessageChars(cleared as unknown as AgentMessage);
|
||||
totalChars += afterChars - beforeChars;
|
||||
ratio = totalChars / charWindow;
|
||||
}
|
||||
|
||||
return next ?? messages;
|
||||
}
|
||||
Reference in New Issue
Block a user