feat(v2): T10 rollup + prefix cache 前缀稳定性（--cache-reuse 对齐）

router_system/workspace.py

@@ -501,6 +501,23 @@ class Workspace:
             data = json.load(f)
         return cls(data)
 
+    def prefix_signature(self) -> str:
+        """返回稳定前缀的签名（T10 prefix cache）。
+
+        交流文本的"恒定位于前部"的部分（version + request_id + query + meta + brief）
+        应不随 progress/issues/decisions 追加而变化，从而使 llama-server 的
+        --cache-reuse 能命中该前缀、降低 prefill 开销。用紧凑 JSON 的哈希度量稳定性。
+        """
+        stable = {
+            "version": self._data.get("version"),
+            "request_id": self._data.get("request_id"),
+            "query": self._data.get("query"),
+            "brief": self._data.get("brief"),
+        }
+        import hashlib
+        s = json.dumps(stable, ensure_ascii=False, sort_keys=True)
+        return hashlib.sha256(s.encode("utf-8")).hexdigest()[:16]
+
     def to_dict(self) -> Dict[str, Any]:
         return self.data
 

tests/test_prefix_cache.py

@@ -0,0 +1,72 @@
+"""T10 rollup + prefix cache 前缀稳定性测试。
+
+验证交流文本的"恒定前缀"（version/request_id/query/meta/brief）在写入
+progress/issues/decisions 后保持不变 —— 这是 llama-server --cache-reuse
+命中、降低 prefill 开销的前提（D2 / 6.2 / T10）。
+"""
+import json
+
+from router_system.workspace import Workspace
+
+BRIEF = {
+    "goal": "实现快排",
+    "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": "可运行"}],
+}
+
+
+def _ws():
+    ws = Workspace.new("abc123def456", "写个快排")
+    ws.apply_brief(BRIEF)
+    return ws
+
+
+def test_prefix_stable_across_writes():
+    ws = _ws()
+    sig0 = ws.prefix_signature()
+    ws.add_progress("s1", "done", "完成", "a://s1.py")
+    ws.add_issue("s1", "a://f.py#L1", "obs", "exp", "try", "ask")
+    ws.add_decision("i1", "reply")
+    ws.mark_round()
+    ws.add_budget(input_tokens=100, output_tokens=20)
+    assert ws.prefix_signature() == sig0  # 前缀不随写操作变化
+
+
+def test_prefix_changes_with_brief():
+    ws = _ws()
+    sig = ws.prefix_signature()
+    ws2 = Workspace.new("abc123def456", "写个快排")
+    ws2.apply_brief({**BRIEF, "goal": "不同的目标"})
+    assert ws2.prefix_signature() != sig
+
+
+def test_prefix_changes_with_query():
+    ws = _ws()
+    sig = ws.prefix_signature()
+    ws2 = Workspace.new("abc123def456", "另一个问题")
+    ws2.apply_brief(BRIEF)
+    assert ws2.prefix_signature() != sig
+
+
+def test_rollup_keeps_prefix_stable():
+    ws = _ws()
+    ws.add_progress("s1", "done", "ok", "a://s1.py")
+    sig_before = ws.prefix_signature()
+    ws.rollup()
+    assert ws.prefix_signature() == sig_before
+
+
+def test_serialized_leading_is_stable():
+    ws = _ws()
+    head_before = _leading_json(ws)
+    ws.add_progress("s1", "done", "ok", "a://s1.py")
+    ws.add_issue("s1", "a://f.py#L1", "obs", "exp", "try", "ask")
+    assert _leading_json(ws) == head_before
+
+
+def _leading_json(ws: Workspace) -> str:
+    d = ws.data
+    stable = {k: d.get(k) for k in ("version", "request_id", "query", "meta", "brief")}
+    return json.dumps(stable, ensure_ascii=False, sort_keys=True)

任务拆解与执行计划.md

@@ -76,7 +76,7 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
 | T7 | 网关扩展（/chat 切 v2，/chat/legacy） | ✅ 完成 | T7 |
 | T8 | 人工检验队列 ReviewQueue | ✅ 完成 | T8 |
 | T9 | token 计量与账单 | ✅ 完成 | T9 |
-| T10 | rollup + prefix cache 调优 | ⬜ | |
+| T10 | rollup + prefix cache 调优 | ✅ 完成 | T10 |
 | T11 | 打包分发 setup_runtime.py | ⬜ | |
 | T12 | 实验脚本 bench_tokens.py + 数据集 | ⬜ | |
 | T13 | E1–E5 跑数到 research/v2_experiments/ | ⬜ | |