Spaces:
Running
Running
File size: 8,483 Bytes
2a2730a 29b5d55 14ba239 2a2730a ea8c7bd 84c554d 2a2730a 29b5d55 2a2730a 84c554d 2a2730a 84c554d 2a2730a 14ba239 2a2730a 0b469f5 2a2730a 0b469f5 b149c75 2a2730a 14ba239 2a2730a 14ba239 2a2730a 0b469f5 2a2730a 0b469f5 2a2730a 0b469f5 2a2730a 0b469f5 2a2730a 0b469f5 2a2730a 69ad73f 84c554d ea8c7bd 69ad73f 84c554d 69ad73f ea8c7bd 2a2730a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 | /**
* HeadroomContextEngine β ContextEngine implementation for OpenClaw.
*
* Compresses tool outputs and conversation context using the Headroom proxy.
* Zero LLM calls β all compression is algorithmic (SmartCrusher, ContentRouter, etc.)
*/
/* eslint-disable @typescript-eslint/no-explicit-any */
import { compress } from "headroom-ai";
import { ProxyManager, defaultLogger, type ProxyManagerConfig, type ProxyManagerLogger } from "./proxy-manager.js";
import { agentToOpenAI, normalizeAgentMessages, openAIToAgent } from "./convert.js";
export interface HeadroomEngineConfig extends ProxyManagerConfig {
enabled?: boolean;
}
export class HeadroomContextEngine {
readonly info = {
id: "headroom",
name: "Headroom Context Compression",
version: "0.1.0",
ownsCompaction: true,
};
private proxyManager: ProxyManager;
private proxyUrl: string | null = null;
private config: HeadroomEngineConfig;
private logger: ProxyManagerLogger;
private proxyReadyListeners = new Set<(proxyUrl: string) => void | Promise<void>>();
private proxyStartupPromise: Promise<string> | null = null;
private stats = {
totalCompressions: 0,
totalTokensSaved: 0,
totalTokensBefore: 0,
compactions: 0,
};
constructor(config: HeadroomEngineConfig = {}, logger?: ProxyManagerLogger) {
this.config = config;
this.logger = logger ?? defaultLogger;
this.proxyManager = new ProxyManager(config, this.logger);
}
// === ContextEngine Lifecycle ===
async bootstrap(params: {
sessionId: string;
sessionKey?: string;
sessionFile: string;
}): Promise<{ bootstrapped: boolean; reason?: string }> {
if (this.config.enabled === false) {
return { bootstrapped: false, reason: "disabled" };
}
this.ensureProxyStarted();
return { bootstrapped: true, reason: "proxy startup scheduled" };
}
async ingest(params: {
sessionId: string;
message: any;
isHeartbeat?: boolean;
}): Promise<{ ingested: boolean }> {
// No-op: OpenClaw's runtime stores messages. We don't need a separate store.
return { ingested: true };
}
async ingestBatch?(params: {
sessionId: string;
messages: any[];
isHeartbeat?: boolean;
}): Promise<{ ingestedCount: number }> {
return { ingestedCount: params.messages.length };
}
/**
* Assemble context for the model β THE CORE HOOK.
*
* Converts AgentMessage[] β OpenAI format β compress() β AgentMessage[]
*/
async assemble(params: {
sessionId: string;
messages: any[];
tokenBudget?: number;
model?: string;
prompt?: string;
}): Promise<{
messages: any[];
estimatedTokens: number;
systemPromptAddition?: string;
}> {
if (!this.proxyUrl || this.config.enabled === false) {
this.ensureProxyStarted();
// Fallback: return messages unchanged
return { messages: normalizeAgentMessages(params.messages), estimatedTokens: 0 };
}
try {
// Convert AgentMessage β OpenAI format
const openaiMessages = agentToOpenAI(params.messages);
// Compress via proxy β pass tokenBudget so RollingWindow enforces it
const result = await compress(openaiMessages, {
model: params.model ?? "claude-sonnet-4-5",
baseUrl: this.proxyUrl,
fallback: true,
tokenBudget: params.tokenBudget,
} as any);
if (!result.compressed || result.tokensSaved === 0) {
return {
messages: normalizeAgentMessages(params.messages),
estimatedTokens: result.tokensBefore,
};
}
// Convert back to AgentMessage format
const compressedAgentMessages = openAIToAgent(result.messages);
// Track stats
this.stats.totalCompressions++;
this.stats.totalTokensSaved += result.tokensSaved;
this.stats.totalTokensBefore += result.tokensBefore;
this.logger.debug(
`Assembled: ${result.tokensBefore} β ${result.tokensAfter} tokens (saved ${result.tokensSaved})`,
);
return {
messages: compressedAgentMessages,
estimatedTokens: result.tokensAfter,
systemPromptAddition:
result.tokensSaved > 100
? `[Context compressed by Headroom: ${result.tokensSaved} tokens saved. Use headroom_retrieve with the hash to get full details.]`
: undefined,
};
} catch (error) {
this.logger.error(`Assemble failed: ${error}`);
// Graceful fallback: return original messages
return { messages: normalizeAgentMessages(params.messages), estimatedTokens: 0 };
}
}
/**
* Compact context β zero-cost alternative to LLM summarization.
*
* Calls compress() with the token budget, which triggers:
* - SmartCrusher: aggressive JSON compression (70-90% on tool outputs)
* - Kompress: ModernBERT text compression (40-60% on assistant text)
* - RollingWindow: drops oldest messages if still over budget
* - CCR: stores originals for retrieval via headroom_retrieve tool
*
* Zero LLM calls. All algorithmic.
*/
async compact(params: {
sessionId: string;
sessionFile: string;
tokenBudget?: number;
force?: boolean;
runtimeContext?: any;
}): Promise<{
ok: boolean;
compacted: boolean;
reason?: string;
result?: {
tokensBefore: number;
tokensAfter?: number;
};
}> {
if (!this.proxyUrl) {
return { ok: false, compacted: false, reason: "Proxy not available" };
}
// Read current messages from session file if available
// For now, compact() works in tandem with assemble() β the next assemble()
// call will compress with the token budget. When compact() is called
// independently, we report success since our pipeline handles it.
//
// TODO: Read session file, extract messages, call compress() with tokenBudget,
// write back compacted messages.
this.stats.compactions++;
this.logger.info(
`Compact called (budget: ${params.tokenBudget ?? "none"}, force: ${params.force ?? false})`,
);
return {
ok: true,
compacted: true,
reason: "Headroom applies SmartCrusher + Kompress + RollingWindow on next assemble()",
};
}
async afterTurn?(params: {
sessionId: string;
messages: any[];
prePromptMessageCount: number;
isHeartbeat?: boolean;
}): Promise<void> {
// Optional: could log stats or trigger learning
}
async prepareSubagentSpawn?(params: {
parentSessionKey: string;
childSessionKey: string;
ttlMs?: number;
}): Promise<{ rollback: () => Promise<void> } | undefined> {
// Subagent context is compressed naturally via assemble()
return undefined;
}
async onSubagentEnded?(params: {
childSessionKey: string;
reason: string;
}): Promise<void> {
// No-op
}
async dispose(): Promise<void> {
await this.proxyManager.stop();
this.logger.info(
`Engine disposed. Stats: ${this.stats.totalCompressions} compressions, ` +
`${this.stats.totalTokensSaved} tokens saved`,
);
}
// --- Public API ---
getStats() {
return { ...this.stats };
}
getProxyUrl(): string | null {
return this.proxyUrl;
}
ensureProxyStarted(): void {
if (this.config.enabled === false || this.proxyUrl || this.proxyStartupPromise) {
return;
}
this.proxyStartupPromise = this.proxyManager
.start()
.then(async (proxyUrl) => {
this.proxyUrl = proxyUrl;
await this.notifyProxyReady(proxyUrl);
this.logger.info(`Headroom proxy ready at ${proxyUrl}`);
return proxyUrl;
})
.catch((error) => {
this.logger.warn(`Headroom proxy unavailable: ${error}`);
throw error;
})
.finally(() => {
this.proxyStartupPromise = null;
});
}
onProxyReady(listener: (proxyUrl: string) => void | Promise<void>): () => void {
this.proxyReadyListeners.add(listener);
return () => {
this.proxyReadyListeners.delete(listener);
};
}
async ensureProxyUrl(): Promise<string> {
if (this.proxyUrl) {
return this.proxyUrl;
}
this.ensureProxyStarted();
if (!this.proxyStartupPromise) {
throw new Error("Headroom proxy startup is disabled");
}
return this.proxyStartupPromise;
}
private async notifyProxyReady(proxyUrl: string): Promise<void> {
for (const listener of this.proxyReadyListeners) {
await listener(proxyUrl);
}
}
}
|