lcode-agent 0.1.0__py3-none-any.whl

lcode/compress.py

@@ -0,0 +1,80 @@
+"""上下文达到模型窗口的 auto-compact 百分比时,把旧对话折成摘要。
+
+和 Grok Build 一样:磁盘保留全文,发给模型的是摘要 + 近期原文。
+"""
+
+from __future__ import annotations
+
+from lcode.protocol import ask
+from lcode.store import Store, _est_tokens
+
+_SUMMARIZE_HINT = (
+    "把下面的对话和旧摘要压缩成后续助手必须记住的上下文。"
+    "保留目标、已确定的事实、约定、人名地名和数字,去掉寒暄与重复。"
+    "只用中文正文,不要标题,不要解释你在做什么。\n\n"
+)
+
+
+def _clip(text: str, limit: int) -> str:
+    if _est_tokens(text) <= limit:
+        return text
+    # 粗估:汉字约 1 token,多留一点余量按字符切
+    cap = max(32, limit)
+    if len(text) <= cap:
+        return text
+    return text[:cap].rstrip() + "…"
+
+
+def extractive_summary(old_summary: str, prefix: list, limit: int = 8000) -> str:
+    """不调模型也能压:每条留一句,再卡总长。"""
+    parts: list[str] = []
+    if old_summary.strip():
+        parts.append("【旧摘要】\n" + old_summary.strip())
+    for msg in prefix:
+        role = getattr(msg, "role", None) or msg.get("role")
+        content = getattr(msg, "content", None) or msg.get("content") or ""
+        content = " ".join(str(content).split())
+        if not content:
+            continue
+        if role == "user":
+            who = "用户"
+        elif role == "tool":
+            who = "工具"
+        else:
+            who = "助手"
+        line = _clip(f"{who}: {content}", 160)
+        parts.append(line)
+    return _clip("\n".join(parts), limit)
+
+
+async def llm_summary(blob: str) -> str:
+    try:
+        text = await ask(_SUMMARIZE_HINT + blob)
+    except Exception:
+        return ""
+    return (text or "").strip()
+
+
+async def prepare_context(
+    store: Store,
+    session_id: str,
+    *,
+    force: bool = False,
+    hint: str = "",
+) -> list[dict[str, str]]:
+    """若达到窗口百分比阈值(或手动 /zip),压缩旧段后再返回要发给模型的 messages。"""
+    if not force and not store.needs_compress(session_id):
+        return store.context_for_model(session_id)
+
+    prefix, _tail = store.split_overflow(session_id, force=force)
+    if not prefix:
+        return store.context_for_model(session_id)
+
+    budget = store.summary_budget()
+    draft = extractive_summary(store._summary_of(session_id), prefix, budget)
+    if hint.strip():
+        draft = f"用户要求保留：{hint.strip()}\n\n{draft}"
+    polished = await llm_summary(draft)
+    summary = _clip(polished or draft, budget)
+    store.apply_summary(session_id, summary, prefix[-1].id)
+    return store.context_for_model(session_id)

lcode/default_config.json

@@ -0,0 +1,10 @@
+{
+  "name": "Lcode",
+  "version": "0.1.0",
+  "description": "一个基于 Textual 的终端 TUI agent",
+  "url": "",
+  "requestType": "completions",
+  "model": "",
+  "models": [],
+  "autoCompactPercent": 85
+}