thatgfsj-code 2.2.9 → 3.0.1

package/src/session/compactor.ts

@@ -1,103 +1,103 @@
-/**
- * Context Compactor - Progressive compression for conversation history
- * Migrated from old src/core/context-compactor.ts
- *
- * Strategy:
- * 1. Within limit: no compression
- * 2. Beyond limit: keep system + recent N turns + summarize middle
- */
-
-import type { ChatMessage } from '../types.js';
-
-export interface CompactorConfig {
-  /** Max messages before compression kicks in */
-  maxMessages?: number;
-  /** How many recent turns to always preserve */
-  preserveRecent?: number;
-}
-
-export interface CompressionResult {
-  originalCount: number;
-  compactedCount: number;
-  removedCount: number;
-}
-
-export class ContextCompactor {
-  private maxMessages: number;
-  private preserveRecent: number;
-
-  constructor(config: CompactorConfig = {}) {
-    this.maxMessages = config.maxMessages ?? 50;
-    this.preserveRecent = config.preserveRecent ?? 10;
-  }
-
-  /**
-   * Compact messages if needed
-   */
-  compact(messages: ChatMessage[]): { compacted: ChatMessage[]; result: CompressionResult } {
-    if (messages.length <= this.maxMessages) {
-      return {
-        compacted: messages,
-        result: { originalCount: messages.length, compactedCount: messages.length, removedCount: 0 },
-      };
-    }
-
-    const systemMsg = messages.find(m => m.role === 'system');
-    const others = messages.filter(m => m.role !== 'system');
-
-    if (others.length <= this.preserveRecent) {
-      return {
-        compacted: messages,
-        result: { originalCount: messages.length, compactedCount: messages.length, removedCount: 0 },
-      };
-    }
-
-    // Keep recent N turns
-    const recent = others.slice(-this.preserveRecent);
-
-    // Summarize the middle portion
-    const middle = others.slice(0, -this.preserveRecent);
-    const summary: ChatMessage = {
-      role: 'system',
-      content: `[Earlier conversation summary: ${middle.length} messages covering ${this.extractTopics(middle)}]`,
-    };
-
-    const compacted = systemMsg
-      ? [systemMsg, summary, ...recent]
-      : [summary, ...recent];
-
-    return {
-      compacted,
-      result: {
-        originalCount: messages.length,
-        compactedCount: compacted.length,
-        removedCount: messages.length - compacted.length,
-      },
-    };
-  }
-
-  /**
-   * Auto-compact if messages exceed limit
-   */
-  autoCompact(messages: ChatMessage[]): { compacted: ChatMessage[]; result: CompressionResult } {
-    return this.compact(messages);
-  }
-
-  /**
-   * Estimate token count (rough: ~4 chars per token)
-   */
-  estimateTokens(messages: ChatMessage[]): number {
-    return messages.reduce((sum, m) => sum + Math.ceil(m.content.length / 4) + 10, 0);
-  }
-
-  /**
-   * Extract topic summary from messages
-   */
-  private extractTopics(msgs: ChatMessage[]): string {
-    const topics = msgs
-      .filter(m => m.role === 'user')
-      .map(m => m.content.slice(0, 30))
-      .slice(0, 3);
-    return topics.join(', ') || 'various topics';
-  }
-}
+/**
+ * Context Compactor - Progressive compression for conversation history
+ * Migrated from old src/core/context-compactor.ts
+ *
+ * Strategy:
+ * 1. Within limit: no compression
+ * 2. Beyond limit: keep system + recent N turns + summarize middle
+ */
+
+import type { ChatMessage } from '../types.js';
+
+export interface CompactorConfig {
+  /** Max messages before compression kicks in */
+  maxMessages?: number;
+  /** How many recent turns to always preserve */
+  preserveRecent?: number;
+}
+
+export interface CompressionResult {
+  originalCount: number;
+  compactedCount: number;
+  removedCount: number;
+}
+
+export class ContextCompactor {
+  private maxMessages: number;
+  private preserveRecent: number;
+
+  constructor(config: CompactorConfig = {}) {
+    this.maxMessages = config.maxMessages ?? 50;
+    this.preserveRecent = config.preserveRecent ?? 10;
+  }
+
+  /**
+   * Compact messages if needed
+   */
+  compact(messages: ChatMessage[]): { compacted: ChatMessage[]; result: CompressionResult } {
+    if (messages.length <= this.maxMessages) {
+      return {
+        compacted: messages,
+        result: { originalCount: messages.length, compactedCount: messages.length, removedCount: 0 },
+      };
+    }
+
+    const systemMsg = messages.find(m => m.role === 'system');
+    const others = messages.filter(m => m.role !== 'system');
+
+    if (others.length <= this.preserveRecent) {
+      return {
+        compacted: messages,
+        result: { originalCount: messages.length, compactedCount: messages.length, removedCount: 0 },
+      };
+    }
+
+    // Keep recent N turns
+    const recent = others.slice(-this.preserveRecent);
+
+    // Summarize the middle portion
+    const middle = others.slice(0, -this.preserveRecent);
+    const summary: ChatMessage = {
+      role: 'system',
+      content: `[Earlier conversation summary: ${middle.length} messages covering ${this.extractTopics(middle)}]`,
+    };
+
+    const compacted = systemMsg
+      ? [systemMsg, summary, ...recent]
+      : [summary, ...recent];
+
+    return {
+      compacted,
+      result: {
+        originalCount: messages.length,
+        compactedCount: compacted.length,
+        removedCount: messages.length - compacted.length,
+      },
+    };
+  }
+
+  /**
+   * Auto-compact if messages exceed limit
+   */
+  autoCompact(messages: ChatMessage[]): { compacted: ChatMessage[]; result: CompressionResult } {
+    return this.compact(messages);
+  }
+
+  /**
+   * Estimate token count (rough: ~4 chars per token)
+   */
+  estimateTokens(messages: ChatMessage[]): number {
+    return messages.reduce((sum, m) => sum + Math.ceil(m.content.length / 4) + 10, 0);
+  }
+
+  /**
+   * Extract topic summary from messages
+   */
+  private extractTopics(msgs: ChatMessage[]): string {
+    const topics = msgs
+      .filter(m => m.role === 'user')
+      .map(m => m.content.slice(0, 30))
+      .slice(0, 3);
+    return topics.join(', ') || 'various topics';
+  }
+}

package/src/session/index.ts

@@ -1,168 +1,181 @@
-/**
- * Session Manager - Manages conversation history with auto-compaction
- */
-
-import type { ChatMessage } from '../types.js';
-import { ContextCompactor } from './compactor.js';
-
-/**
- * v2.2.4 (port from v2.1.0): patterns that, if found in an assistant
- * message, indicate the message is a truncated/aborted response that
- * should NOT be persisted into history. Without this filter, the next
- * turn's LLM sees the marker and starts echoing it back, creating a
- * self-reinforcing "[已中断]" hallucination loop.
- *
- * v2.2.6.1 tightening (smoke-test driven): the design here is a
- * two-tier check:
- *   1. STRONG markers — unambiguous pollution, always drop:
- *      - `[已中断]` bracketed (the model emits this exact form)
- *      - `[interrupted]` bracketed English equivalent
- *   2. WEAK markers — only drop if the message is also short (<200
- *      chars). Long messages mentioning "response truncated" are
- *      legitimate conversation (e.g. user complaints about past
- *      behavior), not pollution.
- *
- * Earlier drafts matched bare `已中断` substrings and "response
- * truncated" anywhere in the message — both were too greedy and
- * dropped legitimate Chinese/English conversation.
- */
-const POLLUTION_STRONG: RegExp[] = [
-  /\[已中断\]/,
-  /\[interrupted\]/i,
-];
-
-const POLLUTION_WEAK: RegExp[] = [
-  /^\s*[\*#>\-`]*\s*\[已中断\]/m,
-  /^\s*[\*#>\-`]*\s*\[interrupted\]/im,
-  /\bresponse (was )?(truncated|cut off|interrupted)\b/i,
-  /\boutput (was )?(truncated|cut off|interrupted)\b/i,
-];
-
-function looksPolluted(content: string): boolean {
-  if (!content) return false;
-  // Common gates (apply to both tiers):
-  //   - Length: model got cut off mid-stream → pollution is short.
-  //     A long message with [已中断] in the middle is the model
-  //     legitimately referencing the marker, not pollution.
-  //   - Question: ends with '?' → user complaining about past behavior.
-  //   - Temporal: contains "last time" / "earlier" / etc. → past tense.
-  if (content.length >= 200) return false;
-  if (/\?\s*$/.test(content.trim())) return false;
-  if (/\b(last time|earlier|before|previously|yesterday)\b/i.test(content)) return false;
-  // STRONG markers — bracketed truncation markers, unambiguous.
-  if (POLLUTION_STRONG.some(p => p.test(content))) return true;
-  // WEAK markers — bare "response truncated" phrases.
-  if (POLLUTION_WEAK.some(p => p.test(content))) return true;
-  return false;
-}
-
-export class SessionManager {
-  private messages: ChatMessage[] = [];
-  private sessionId: string;
-  private createdAt: Date;
-  private compactor: ContextCompactor;
-  private maxMessages: number;
-  /** v2.2.4: counter for messages dropped by the pollution filter. */
-  private droppedCount: number = 0;
-
-  constructor(maxMessages = 50) {
-    this.maxMessages = maxMessages;
-    this.sessionId = `session_${Date.now()}_${Math.random().toString(36).substring(2, 9)}`;
-    this.createdAt = new Date();
-    this.compactor = new ContextCompactor({ maxMessages });
-  }
-
-  addMessage(role: ChatMessage['role'], content: string, extras?: Partial<ChatMessage>): void {
-    this.messages.push({ role, content, ...extras });
-    this.autoCompact();
-  }
-
-  /**
-   * v2.2.4 (port from v2.1.0): same as addMessage but returns false
-   * (and skips the push) if the message content matches a known
-   * truncation/abort pollution pattern. Use this for assistant
-   * messages whose stream might have been aborted.
-   */
-  addMessageSafe(role: ChatMessage['role'], content: string, extras?: Partial<ChatMessage>): boolean {
-    if (role === 'assistant' && looksPolluted(content)) {
-      this.droppedCount++;
-      return false;
-    }
-    this.addMessage(role, content, extras);
-    return true;
-  }
-
-  /** v2.2.4: total messages dropped by addMessageSafe since session start. */
-  getDroppedCount(): number {
-    return this.droppedCount;
-  }
-
-  getMessages(): ChatMessage[] {
-    return [...this.messages];
-  }
-
-  getMessageCount(): number {
-    return this.messages.length;
-  }
-
-  clear(): void {
-    this.messages = [];
-  }
-
-  /**
-   * Auto-compact when messages exceed max.
-   * Keeps system messages + recent half. One summary replaces the rest.
-   */
-  private autoCompact(): void {
-    if (this.messages.length <= this.maxMessages) return;
-
-    const systemMsgs = this.messages.filter(m => m.role === 'system');
-    const others = this.messages.filter(m => m.role !== 'system');
-    const keepCount = Math.floor(this.maxMessages * 0.5);
-    const recent = others.slice(-keepCount);
-    const removed = others.length - keepCount;
-
-    if (removed <= 0) return;
-
-    // Remove old summaries before adding new one
-    const cleanSystem = systemMsgs.filter(m => !m.content.startsWith('[Earlier conversation'));
-    const summary: ChatMessage = {
-      role: 'system',
-      content: `[Earlier conversation: ${removed} messages compacted to save context]`,
-    };
-    this.messages = [...cleanSystem, summary, ...recent];
-  }
-
-  truncate(maxMessages?: number): void {
-    if (maxMessages) {
-      // v2.2.7 edge fix: preserveRecent must be <= maxMessages - 1,
-      // otherwise the compactor's "always keep recent N" branch
-      // overrides the trim and no messages actually get dropped.
-      // Set preserveRecent to maxMessages-1 so anything beyond the
-      // last maxMessages-1 user/assistant msgs gets summarized.
-      this.maxMessages = maxMessages;
-      this.compactor = new ContextCompactor({
-        maxMessages,
-        preserveRecent: Math.max(1, maxMessages - 1),
-      });
-    }
-    const { compacted } = this.compactor.compact(this.messages);
-    this.messages = compacted;
-  }
-
-  getId(): string {
-    return this.sessionId;
-  }
-
-  getInfo(): { id: string; messageCount: number; createdAt: Date } {
-    return {
-      id: this.sessionId,
-      messageCount: this.messages.length,
-      createdAt: this.createdAt,
-    };
-  }
-}
-
-export { ContextCompactor } from './compactor.js';
-export type { CompactorConfig, CompressionResult } from './compactor.js';
-export * from './message.js';
+/**
+ * Session Manager - Manages conversation history with auto-compaction
+ */
+
+import type { ChatMessage } from '../types.js';
+import { ContextCompactor } from './compactor.js';
+
+/**
+ * v2.2.4 (port from v2.1.0): patterns that, if found in an assistant
+ * message, indicate the message is a truncated/aborted response that
+ * should NOT be persisted into history. Without this filter, the next
+ * turn's LLM sees the marker and starts echoing it back, creating a
+ * self-reinforcing "[已中断]" hallucination loop.
+ *
+ * v2.2.6.1 tightening (smoke-test driven): the design here is a
+ * two-tier check:
+ *   1. STRONG markers — unambiguous pollution, always drop:
+ *      - `[已中断]` bracketed (the model emits this exact form)
+ *      - `[interrupted]` bracketed English equivalent
+ *   2. WEAK markers — only drop if the message is also short (<200
+ *      chars). Long messages mentioning "response truncated" are
+ *      legitimate conversation (e.g. user complaints about past
+ *      behavior), not pollution.
+ *
+ * Earlier drafts matched bare `已中断` substrings and "response
+ * truncated" anywhere in the message — both were too greedy and
+ * dropped legitimate Chinese/English conversation.
+ */
+const POLLUTION_STRONG: RegExp[] = [
+  /\[已中断\]/,
+  /\[interrupted\]/i,
+];
+
+const POLLUTION_WEAK: RegExp[] = [
+  /^\s*[\*#>\-`]*\s*\[已中断\]/m,
+  /^\s*[\*#>\-`]*\s*\[interrupted\]/im,
+  /\bresponse (was )?(truncated|cut off|interrupted)\b/i,
+  /\boutput (was )?(truncated|cut off|interrupted)\b/i,
+];
+
+function looksPolluted(content: string): boolean {
+  if (!content) return false;
+  if (POLLUTION_STRONG.some(p => p.test(content))) return true;
+  // WEAK pattern gate:
+  //   - Message must be short (model got cut off mid-stream, so it
+  //     can't be long)
+  //   - Message must NOT end with '?' (questions are user complaints
+  //     about past behavior, not pollution)
+  //   - Message must NOT contain "last time" / "earlier" / "before"
+  //     (temporal references indicate user discussing past behavior)
+  if (content.length >= 200) return false;
+  if (/\?\s*$/.test(content.trim())) return false;
+  if (/\b(last time|earlier|before|previously|yesterday)\b/i.test(content)) return false;
+  if (POLLUTION_WEAK.some(p => p.test(content))) return true;
+  return false;
+}
+
+export class SessionManager {
+  private messages: ChatMessage[] = [];
+  private sessionId: string;
+  private createdAt: Date;
+  private compactor: ContextCompactor;
+  private maxMessages: number;
+  /** v2.2.4: counter for messages dropped by the pollution filter. */
+  private droppedCount: number = 0;
+  /**
+   * v3.0.0: callback invoked when the message count exceeds maxMessages.
+   * Reasonix principle: do not mutate messages to compact; instead, signal
+   * the TUI to suggest the user call /new (start a fresh session) so that
+   * the upstream prompt cache keeps a stable prefix. NWT auto-logs every
+   * meaningful event, so no history is lost — it just moves out of the
+   * live cache window.
+   */
+  public onSuggestNewSession?: (info: { currentLength: number; max: number }) => void;
+
+  constructor(maxMessages = 50, options?: { onSuggestNewSession?: SessionManager['onSuggestNewSession'] }) {
+    this.maxMessages = maxMessages;
+    this.sessionId = `session_${Date.now()}_${Math.random().toString(36).substring(2, 9)}`;
+    this.createdAt = new Date();
+    this.compactor = new ContextCompactor({ maxMessages });
+    this.onSuggestNewSession = options?.onSuggestNewSession;
+  }
+
+  addMessage(role: ChatMessage['role'], content: string, extras?: Partial<ChatMessage>): void {
+    this.messages.push({ role, content, ...extras });
+    this.autoCompact();
+  }
+
+  /**
+   * v2.2.4 (port from v2.1.0): same as addMessage but returns false
+   * (and skips the push) if the message content matches a known
+   * truncation/abort pollution pattern. Use this for assistant
+   * messages whose stream might have been aborted.
+   */
+  addMessageSafe(role: ChatMessage['role'], content: string, extras?: Partial<ChatMessage>): boolean {
+    if (role === 'assistant' && looksPolluted(content)) {
+      this.droppedCount++;
+      return false;
+    }
+    this.addMessage(role, content, extras);
+    return true;
+  }
+
+  /** v2.2.4: total messages dropped by addMessageSafe since session start. */
+  getDroppedCount(): number {
+    return this.droppedCount;
+  }
+
+  getMessages(): ChatMessage[] {
+    return [...this.messages];
+  }
+
+  getMessageCount(): number {
+    return this.messages.length;
+  }
+
+  clear(): void {
+    this.messages = [];
+  }
+
+  /**
+   * v3.0.0: "auto-compact" is now a no-op on `messages`. We instead invoke
+   * the registered `onSuggestNewSession` callback (if any) so the TUI can
+   * surface a "context is getting long, consider /new" hint. This keeps
+   * the messages array byte-stable across rounds, which is critical for
+   * prompt-cache hit-rate: any reordering or summary insertion would
+   * invalidate every cache breakpoint the upstream provider has set.
+   *
+   * Rationale (Reasonix-style): when context exceeds the model window or
+   * the configured maxMessages threshold, the right action is to start a
+   * NEW session, not to silently truncate or summarize. NWT auto-logs
+   * every meaningful event, so the user does not lose context — it just
+   * moves out of the live cache window. If the user wants a longer
+   * session, they can bump maxMessages in `~/.thatgfsj/config.json`.
+   *
+   * The original "fake compact" logic (insert `[Earlier conversation...]`
+   * summary) is preserved as `truncate()` for callers that explicitly
+   * want to compact (e.g. /new command hard-reset).
+   */
+  private autoCompact(): void {
+    if (this.messages.length <= this.maxMessages) return;
+
+    // Notify the TUI. The default TUI behavior is to render a toast:
+    //   ⚠️ 上下文超长（{currentLength}/{max}），建议调 /new 开新会话
+    //            (NWT 已自动归档历史)
+    try {
+      this.onSuggestNewSession?.({
+        currentLength: this.messages.length,
+        max: this.maxMessages,
+      });
+    } catch {
+      // Callback failure must not break the chat loop.
+    }
+    // Intentionally do NOT mutate this.messages. Prefix stability matters more.
+  }
+
+  truncate(maxMessages?: number): void {
+    if (maxMessages) {
+      this.maxMessages = maxMessages;
+      this.compactor = new ContextCompactor({ maxMessages });
+    }
+    const { compacted } = this.compactor.compact(this.messages);
+    this.messages = compacted;
+  }
+
+  getId(): string {
+    return this.sessionId;
+  }
+
+  getInfo(): { id: string; messageCount: number; createdAt: Date } {
+    return {
+      id: this.sessionId,
+      messageCount: this.messages.length,
+      createdAt: this.createdAt,
+    };
+  }
+}
+
+export { ContextCompactor } from './compactor.js';
+export type { CompactorConfig, CompressionResult } from './compactor.js';
+export * from './message.js';

package/src/session/message.ts

@@ -1,42 +1,42 @@
-/**
- * Message types and utilities
- */
-
-import type { ChatMessage, Role } from '../types.js';
-
-export { ChatMessage, Role };
-
-/**
- * Create a message
- */
-export function createMessage(role: Role, content: string, extras?: Partial<ChatMessage>): ChatMessage {
-  return { role, content, ...extras };
-}
-
-/**
- * Create a system message
- */
-export function systemMessage(content: string): ChatMessage {
-  return { role: 'system', content };
-}
-
-/**
- * Create a user message
- */
-export function userMessage(content: string): ChatMessage {
-  return { role: 'user', content };
-}
-
-/**
- * Create an assistant message
- */
-export function assistantMessage(content: string): ChatMessage {
-  return { role: 'assistant', content };
-}
-
-/**
- * Create a tool result message
- */
-export function toolMessage(content: string, toolCallId: string, name: string): ChatMessage {
-  return { role: 'tool', content, tool_call_id: toolCallId, name };
-}
+/**
+ * Message types and utilities
+ */
+
+import type { ChatMessage, Role } from '../types.js';
+
+export { ChatMessage, Role };
+
+/**
+ * Create a message
+ */
+export function createMessage(role: Role, content: string, extras?: Partial<ChatMessage>): ChatMessage {
+  return { role, content, ...extras };
+}
+
+/**
+ * Create a system message
+ */
+export function systemMessage(content: string): ChatMessage {
+  return { role: 'system', content };
+}
+
+/**
+ * Create a user message
+ */
+export function userMessage(content: string): ChatMessage {
+  return { role: 'user', content };
+}
+
+/**
+ * Create an assistant message
+ */
+export function assistantMessage(content: string): ChatMessage {
+  return { role: 'assistant', content };
+}
+
+/**
+ * Create a tool result message
+ */
+export function toolMessage(content: string, toolCallId: string, name: string): ChatMessage {
+  return { role: 'tool', content, tool_call_id: toolCallId, name };
+}