feat(v2): T-M3 采纳 llmrouter pinch 三档裁剪——Architect 渲染超预算按相关性分档保留

- Workspace.render_for_architect：超 1200 token 时对 archive 行按与 query+goal
  的字符 2-gram Dice 相似度分档——sim<0.25 先丢、sim<0.55 截断为 40% 摘要、
  仍超限从低到高继续丢；保留行维持原插入序（前缀稳定），最近 4 行下限不变
  （相比旧'从最旧整段丢'，高相关事实在预算内留存更久）
- 顺带消除两项旧债：压缩态与首次渲染共用同一 parts 结构（格式漂移）、
  裁剪循环不再每轮 deepcopy 全文档
- 修复本轮引入的缺陷：会话轮次 tool_calls 为 int 计数时清单提取迭代崩溃
  （test_session_multi_turn 抓出，已加类型防御）
- 新增 tests/test_pinch_trim.py 4 项；全量 244 passed ×2（基线 230）

gateway/agent.py

@@ -461,7 +461,7 @@ class AgentService:
         files = sorted({
             str((a.get("arguments") or {}).get("path") or "").strip()
             for t in session.data.get("turns", [])
-            for a in (t.get("tool_calls") or [])
+            for a in (t.get("tool_calls") if isinstance(t.get("tool_calls"), list) else [])
             if isinstance(a, dict)
             and str(a.get("name") or "") in _FILE_TOOLS
             and (a.get("arguments") or {}).get("path")

router_system/workspace.py

@@ -450,15 +450,17 @@ class Workspace:
         for p in d.get("progress", [])[-5:]:
             parts.append(f"{p.get('step')} [{p.get('status')}] {_clip(p.get('summary', ''), 40)}")
         parts.append("== archive ==")
-        for line in d.get("archive", [])[-8:]:
+        archive = d.get("archive", []) or []
-            parts.append(line)
+        shown = list(archive[-8:])
-        # token 预算：超限先截断最旧 archive（已只保留 3 条 decisions）
+        parts.extend(shown)
+        # token 预算：超限按 pinch 三档裁剪 archive（T-M3，采纳 llmrouter pinch
+        # keep/summarize/drop 分档思想），决策见 _trim_archive_lines
         out = "\n".join(parts)
-        while estimate_tokens(out) > 1200 and len(d.get("archive", [])) > 4:
+        if estimate_tokens(out) <= 1200 or len(shown) <= 4:
-            d = copy.deepcopy(d)
+            return out  # 预算内 / 已在下限（与旧实现的 4 行下限一致）
-            d["archive"] = d["archive"][4:]
+        head = parts[:-len(shown)]
-            out = "\n".join(_rerender(self, d))
+        ref_text = d.get("query", "") + " " + (d.get("brief") or {}).get("goal", "")
-        return out
+        return _trim_archive_lines(head, archive, shown, ref_text)
 
     def render_for_worker(self, step_id: str,
                           artifact_text: Optional[str] = None) -> str:
@@ -525,24 +527,69 @@ class Workspace:
         return self.data
 
 
-def _rerender(ws: "Workspace", d: Dict[str, Any]) -> List[str]:
+def _bigrams(text: str) -> Dict[str, int]:
-    """用裁剪后的数据重建 Architect 渲染（供超限压缩内部用）。"""
+    """字符 2-gram 计数（去空白、小写；零依赖轻量相关性度量）。"""
-    parts: List[str] = []
+    t = "".join(text.lower().split())
-    m = d["meta"]
+    out: Dict[str, int] = {}
-    parts.append("== meta ==")
+    for i in range(len(t) - 1):
-    parts.append(f"status={m['status']} round={m['round']}")
+        g = t[i:i + 2]
-    parts.append("== query ==")
+        out[g] = out.get(g, 0) + 1
-    parts.append(_clip(d["query"], LIMITS["query_truncate"]))
+    return out
-    parts.append("== issues ==")
+
-    for iss in d.get("issues", []):
+
-        parts.append(f"{iss['id']} step={iss.get('step')} ask={_clip(iss.get('ask', ''), 80)}")
+def _dice(a: Dict[str, int], b: Dict[str, int]) -> float:
-    parts.append("== decisions(最近3) ==")
+    """Dice 系数：2 * 交集 / (|a| + |b|)，空集返回 0。"""
-    for dec in d.get("decisions", [])[-3:]:
+    if not a or not b:
-        parts.append(f"ref={dec.get('ref')} reply={_clip(dec.get('reply', ''), 80)}")
+        return 0.0
-    parts.append("== progress ==")
+    inter = sum(min(v, b.get(k, 0)) for k, v in a.items())
-    for p in d.get("progress", [])[-5:]:
+    return 2.0 * inter / (sum(a.values()) + sum(b.values()))
-        parts.append(f"{p.get('step')} [{p.get('status')}] {_clip(p.get('summary', ''), 40)}")
+
-    parts.append("== archive ==")
+
-    for line in d.get("archive", [])[-6:]:
+def _trim_archive_lines(head: List[str], archive: List[str],
-        parts.append(line)
+                        shown: List[str], ref_text: str) -> str:
-    return parts
+    """Architect 渲染超预算时的 archive 三档裁剪（T-M3，采纳 llmrouter pinch
+    keep/summarize/drop 分档思想）。
+
+    相关度 = 行与 query+goal 的字符 2-gram Dice 系数：
+      第一档：丢 sim<0.25 的行（最低相关先丢）；
+      第二档：把 sim<0.55 的行截断为 40% 摘要；
+      第三档：仍超限则从低相关到高相关继续丢。
+    保留行维持原插入序（前缀稳定）；任何情况下至少保留 4 行，
+    达到下限仍超限则接受溢出（与旧实现一致）。
+    相比旧"从最旧起整段丢弃"，高相关事实在预算内留存得更久；
+    且压缩态与首次渲染共用同一 parts 结构（消除格式漂移与每轮 deepcopy）。
+    """
+    ref = _bigrams(ref_text)
+    start = len(archive) - len(shown)
+    idx = list(range(start, len(archive)))
+    sims = {i: _dice(_bigrams(archive[i]), ref) for i in idx}
+    order = sorted(idx, key=lambda i: (sims[i], i))
+
+    kept = set(idx)
+    clips: Dict[int, str] = {}
+
+    def _over_budget() -> bool:
+        lines = [clips.get(i, archive[i]) for i in sorted(kept)]
+        return estimate_tokens("\n".join(head + lines)) > 1200
+
+    # 第一档：丢低相关
+    for i in order:
+        if not _over_budget() or len(kept) <= 4:
+            break
+        if sims[i] < 0.25:
+            kept.discard(i)
+    # 第二档：中相关截断为摘要（40% 长度）
+    for i in order:
+        if not _over_budget():
+            break
+        if i in kept and sims[i] < 0.55:
+            clips[i] = _clip(archive[i], max(20, int(len(archive[i]) * 0.4)))
+    # 第三档：仍超限从低到高继续丢
+    for i in order:
+        if not _over_budget() or len(kept) <= 4:
+            break
+        kept.discard(i)
+        clips.pop(i, None)
+
+    lines = [clips.get(i, archive[i]) for i in sorted(kept)]
+    return "\n".join(head + lines)

tests/test_pinch_trim.py

@@ -0,0 +1,60 @@
+"""Architect 渲染三档裁剪（T-M3，采纳 llmrouter pinch keep/summarize/drop 设计）。"""
+from router_system.workspace import Workspace
+
+_GOAL = "用 Python 实现快速排序并分析时间复杂度"
+
+_BRIEF = {
+    "goal": _GOAL, "constraints": ["标准库"], "tags": ["code"],
+    "acceptance": [{"id": "a1", "check": "可运行", "machine_checkable": True}],
+    "plan": [{"id": "s1", "task": "实现", "deps": [], "done_criteria": "过"}],
+}
+
+
+def _ws_with_archive():
+    ws = Workspace.new(request_id="abc123def456", query=_GOAL,
+                       api_token_cap=8000, rounds_cap=6)
+    ws.apply_brief(_BRIEF)
+    relevant = ("s1: 完成快速排序实现，基准选取三数取中，递归深度优化，"
+                "平均时间复杂度 O(n log n) 分析见附件。" + "快排边界处理细节" * 50)
+    lines = [f"sX{i}: 无关行{i}，今日食堂菜单有红烧肉、番茄炒蛋与清炒时蔬，"
+             f"周末计划去郊外徒步露营并整理旅行照片。" + "生活琐事记录" * 50
+             + f"独特结尾标记{i}号"  # 唯一尾串：判断"该行是否还被输出"
+             for i in range(7)]
+    # 相关行放在中间：不在最新 4 行下限保护区内，检验相关性而非新近度
+    ws._data["archive"] = lines[:3] + [relevant] + lines[3:]
+    return ws, relevant, lines
+
+
+def test_trim_keeps_relevant_line_full():
+    """高相关行全文保留，低相关行被丢弃/截断（尾段不再出现在输出）。"""
+    ws, relevant, lines = _ws_with_archive()
+    out = ws.render_for_architect()
+    # 相关行全文（含尾段独特内容）仍在
+    assert relevant[-40:] in out
+    # 至少一条低相关行的尾段消失（被丢或被截为摘要）
+    assert lines[0][-40:] not in out
+
+
+def test_trim_respects_floor_of_four_lines():
+    """任何情况下至少保留 4 行（与旧实现下限一致）。"""
+    ws, _, lines = _ws_with_archive()
+    out = ws.render_for_architect()
+    archive_shown = out.split("== archive ==")[1].strip().splitlines()
+    assert len(archive_shown) >= 4
+
+
+def test_render_deterministic():
+    """同输入两次渲染逐字节一致（裁剪决策确定性）。"""
+    ws1, _, _ = _ws_with_archive()
+    ws2, _, _ = _ws_with_archive()
+    assert ws1.render_for_architect() == ws2.render_for_architect()
+
+
+def test_under_budget_untouched():
+    """预算内不触发裁剪：archive 行原样出现、无截断省略号。"""
+    ws = Workspace.new(request_id="abc123def456", query="写个快排",
+                       api_token_cap=8000, rounds_cap=6)
+    ws.apply_brief(_BRIEF)
+    ws._data["archive"] = ["s1: 完成（产出: a://s1.py）"]
+    out = ws.render_for_architect()
+    assert "s1: 完成（产出: a://s1.py）" in out