File size: 13,176 Bytes
391c43e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 | // lib/llm/core/context-manager.ts
// Portable context manager β no browser imports, no VFS, no Next.js, no configManager.
import type {
Message,
ParsedResponse,
ToolResult,
ProviderAdapter,
CompactionConfig,
ContextManager,
ContentBlock,
UsageInfo,
} from './types';
export class ContextManagerImpl implements ContextManager {
private messages: Message[] = [];
private compactionCount = 0;
onMessageAdded?: (message: Message) => void;
onMessagesReplaced?: (newMessages: Message[]) => void;
constructor(private config: CompactionConfig) {}
getMessages(): Message[] {
return this.messages;
}
getCompactionCount(): number {
return this.compactionCount;
}
setSystemPrompt(prompt: string): void {
if (this.messages.length > 0 && this.messages[0].role === 'system') {
this.messages[0] = { role: 'system', content: prompt };
} else {
this.messages.unshift({ role: 'system', content: prompt });
}
}
addUserMessage(content: string | ContentBlock[]): void {
const msg: Message = { role: 'user', content };
this.messages.push(msg);
this.onMessageAdded?.(msg);
}
addAssistantTurn(response: ParsedResponse): void {
// Sanitize tool call arguments: providers reject invalid JSON in history
if (response.toolCalls?.length) {
for (const tc of response.toolCalls) {
if (tc.function?.arguments) {
try { JSON.parse(tc.function.arguments); } catch {
tc.function.arguments = '{}';
}
}
}
}
const msg: Message = {
role: 'assistant',
content: response.content || '',
...(response.toolCalls?.length ? { tool_calls: response.toolCalls } : {}),
...(response.reasoningDetails?.length ? { reasoning_details: response.reasoningDetails } : {}),
};
this.messages.push(msg);
this.onMessageAdded?.(msg);
}
addToolResults(results: ToolResult[]): void {
for (const r of results) {
const msg: Message = { role: 'tool', content: r.content, tool_call_id: r.tool_call_id };
this.messages.push(msg);
this.onMessageAdded?.(msg);
}
}
importMessages(messages: Message[]): void {
this.messages = [...messages];
}
getTokenEstimate(): number {
return this.messages.reduce((sum, m) => sum + ContextManagerImpl.estimateMessageTokens(m), 0);
}
needsCompaction(tokenCount: number): boolean {
return tokenCount >= this.config.threshold;
}
/**
* Returns messages with orphan tool calls repaired.
* Operates on a copy; never mutates the persistent conversation history.
*/
getSanitizedMessages(): Message[] {
return this.repairOrphanToolCalls(this.messages);
}
/**
* Compact the conversation by summarizing older messages and keeping recent ones.
*
* Strategy:
* - System prompt: replaced with opts.freshSystemPrompt
* - Older ~80% of non-system messages: sent for summarization via provider
* - Recent ~20% of non-system messages: kept verbatim
* - Summary output capped at ~10% of contextLength
*/
async compact(
provider: ProviderAdapter,
opts?: { freshSystemPrompt?: string; projectContext?: string; signal?: AbortSignal }
): Promise<UsageInfo | undefined> {
// 1. Separate system messages from conversation messages
const systemMessages = this.messages.filter(m => m.role === 'system');
const nonSystemMessages = this.messages.filter(m => m.role !== 'system');
if (nonSystemMessages.length < 3) {
return undefined;
}
// 2. Group non-system messages into "turns" (assistant + its tool results).
// A turn is: one assistant message (possibly with tool_calls) + all following
// tool-role messages that belong to it. User messages are standalone turns.
// This ensures we never orphan a tool result from its assistant message.
interface Turn { messages: Message[]; tokens: number }
const turns: Turn[] = [];
let currentTurn: Turn | null = null;
for (const msg of nonSystemMessages) {
if (msg.role === 'tool') {
// Tool results attach to the current turn (started by assistant)
if (currentTurn) {
const t = ContextManagerImpl.estimateMessageTokens(msg);
currentTurn.messages.push(msg);
currentTurn.tokens += t;
}
} else {
// assistant or user β starts a new turn
if (currentTurn) turns.push(currentTurn);
const t = ContextManagerImpl.estimateMessageTokens(msg);
currentTurn = { messages: [msg], tokens: t };
}
}
if (currentTurn) turns.push(currentTurn);
if (turns.length < 2) {
return undefined;
}
// Walk backwards from the end, keeping whole turns within the recent budget.
// Always keep at least 1 turn.
const recentTokenBudget = Math.round(this.config.contextLength * this.config.recentKeepRatio);
let recentTokens = 0;
let recentTurnCount = 0;
for (let i = turns.length - 1; i >= 0; i--) {
if (recentTurnCount >= 1 && recentTokens + turns[i].tokens > recentTokenBudget) {
break;
}
recentTokens += turns[i].tokens;
recentTurnCount++;
}
const splitTurnIndex = turns.length - recentTurnCount;
if (splitTurnIndex <= 0) {
return undefined;
}
const olderMessages = turns.slice(0, splitTurnIndex).flatMap(t => t.messages);
const recentMessages = turns.slice(splitTurnIndex).flatMap(t => t.messages);
// 3. Convert older messages to plain text for summarization.
// Models hallucinate tool calls when they see tool_calls/tool messages in history,
// even without tool definitions. Flatten everything to user/assistant text.
const flattenedMessages: Message[] = [];
for (const msg of olderMessages) {
if (msg.role === 'assistant') {
// Merge tool call info into text content
let text = typeof msg.content === 'string' ? msg.content : '';
if (msg.tool_calls) {
for (const tc of msg.tool_calls) {
const args = tc.function?.arguments || '';
// Truncate very large tool args (file contents) to save tokens
const truncatedArgs = args.length > 500 ? args.slice(0, 500) + '...[truncated]' : args;
text += `\n[Called ${tc.function?.name}(${truncatedArgs})]`;
}
}
if (text.trim()) {
flattenedMessages.push({ role: 'assistant', content: text.trim() });
}
} else if (msg.role === 'tool') {
// Convert tool result to user message (tools role confuses models without tool defs)
const content = typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content);
const truncated = content.length > 500 ? content.slice(0, 500) + '...[truncated]' : content;
if (truncated.trim()) {
flattenedMessages.push({ role: 'user', content: `[Tool result: ${truncated.trim()}]` });
}
} else {
flattenedMessages.push(msg);
}
}
// Merge consecutive same-role messages (some APIs reject adjacent same-role)
const mergedMessages: Message[] = [];
for (const msg of flattenedMessages) {
const last = mergedMessages[mergedMessages.length - 1];
if (last && last.role === msg.role && typeof last.content === 'string' && typeof msg.content === 'string') {
last.content += '\n' + msg.content;
} else {
mergedMessages.push({ ...msg });
}
}
// Extract previous compaction summary for iterative awareness
let previousSummary: string | undefined;
for (const msg of olderMessages) {
if (msg.role === 'assistant' && msg.metadata?.isCompactSummary) {
previousSummary = typeof msg.content === 'string'
? msg.content.replace(/^Here is a summary of the conversation so far:\n\n/, '')
: undefined;
}
}
// Build the compaction request messages
const compactionMessages: Message[] = [
...systemMessages,
...mergedMessages,
{ role: 'user', content: this.config.buildCompactionPrompt(previousSummary) },
];
const summaryMaxTokens = Math.min(
16384,
Math.max(256, Math.round(this.config.contextLength * this.config.summaryTokenRatio))
);
// Call provider for summarization (silent β no progress events in main chat)
const result = await provider.call({
messages: compactionMessages,
maxTokens: summaryMaxTokens,
signal: opts?.signal,
silent: true,
});
const summary = result.content || '';
if (!summary) {
return undefined;
}
// 4. Rebuild conversation:
// [fresh system prompt] + [project context as user msg] + [summary as assistant] + [recent messages]
let freshFromConfig: { systemPrompt?: string; projectContext?: string } = {};
if ((!opts?.freshSystemPrompt || !opts?.projectContext) && this.config.getFreshContext) {
try {
freshFromConfig = await this.config.getFreshContext();
} catch { /* fall back to stale system prompt */ }
}
const freshSystemPrompt = opts?.freshSystemPrompt
|| freshFromConfig.systemPrompt
|| (systemMessages[0] && typeof systemMessages[0].content === 'string' ? systemMessages[0].content : '');
const summaryContent = `Here is a summary of the conversation so far:\n\n${summary}`;
const projectContext = opts?.projectContext ?? freshFromConfig.projectContext;
const contextUserContent = projectContext
? `${projectContext}\n\nThe earlier conversation was compacted into the summary below.`
: 'The earlier conversation was compacted into the summary below.';
this.messages = [
{ role: 'system', content: freshSystemPrompt },
{ role: 'user', content: contextUserContent },
{ role: 'assistant', content: summaryContent, metadata: { isCompactSummary: true } },
...recentMessages,
];
this.compactionCount++;
this.onMessagesReplaced?.(this.messages);
return result.usage;
}
/**
* Repair orphan tool calls in a message array.
*
* Two common issues fixed:
* (1) tool_calls with empty arguments β provider sent a malformed call.
* Drop the offending entry. If the message ends up empty, drop it.
* (2) tool_calls with no matching tool result β call was cancelled mid-flight.
* Synthesize a role:'tool' placeholder so the sequence is well-formed.
*
* Operates on a copy; never mutates the persistent conversation history.
*/
private repairOrphanToolCalls(messages: Message[]): Message[] {
const out: Message[] = [];
for (let i = 0; i < messages.length; i++) {
const msg = messages[i];
if (msg.role !== 'assistant' || !msg.tool_calls || msg.tool_calls.length === 0) {
out.push(msg);
continue;
}
// Filter out tool calls with empty arguments and repair invalid JSON args
const validCalls = msg.tool_calls.filter(tc => {
const args = tc.function?.arguments;
if (typeof args !== 'string' || args.trim() === '') return false;
try {
JSON.parse(args);
} catch {
// Truncated/malformed args β replace with empty object so the
// conversation history stays valid JSON for every provider.
tc.function.arguments = '{}';
}
return true;
});
const contentEmpty = typeof msg.content === 'string'
? msg.content.trim() === ''
: !msg.content || (Array.isArray(msg.content) && msg.content.length === 0);
if (validCalls.length === 0) {
// No valid tool calls remain
if (contentEmpty) {
// Drop the entire message
continue;
}
// Keep message without tool_calls
const { tool_calls: _, ...rest } = msg;
out.push(rest);
continue;
}
// Push assistant with only valid calls
out.push({ ...msg, tool_calls: validCalls });
// Check which tool calls have matching results in subsequent messages
const matchedIds = new Set<string>();
for (let j = i + 1; j < messages.length; j++) {
const next = messages[j];
if (next.role === 'assistant') break;
if (next.role === 'tool' && next.tool_call_id) {
matchedIds.add(next.tool_call_id);
}
}
// Inject synthetic results for unmatched calls
for (const tc of validCalls) {
if (!matchedIds.has(tc.id)) {
out.push({
role: 'tool',
tool_call_id: tc.id,
content: 'No result β call was cancelled or aborted before completion.',
});
}
}
}
return out;
}
/**
* Estimate token count of a message (content + tool call arguments).
* Uses char/3.5 heuristic β fast and reasonable for context budgeting.
*/
static estimateMessageTokens(msg: Message): number {
const contentLen = typeof msg.content === 'string' ? msg.content.length : JSON.stringify(msg.content).length;
const argsLen = msg.tool_calls?.reduce((s, tc) => s + (tc.function?.arguments?.length ?? 0), 0) ?? 0;
return Math.round((contentLen + argsLen) / 3.5);
}
}
|