codebee 0.1.40 → 0.1.41

package/CHANGELOG.md

@@ -6,6 +6,12 @@ README 元数据带回，供老版本在「发现新版本」时展示新版更
 
 ## 未发布
 
+## v0.1.41（2026-09-22）
+
+### 🛠 问题修复
+
+- 断点续跑/重跑更省钱：同一任务再次生成大纲或执行计划时直接命中上次结果的精确缓存（1 小时内有效），不再重复花一次全量规划调用；大纲首试成功但格式不合规时的重试仍会重新生成（保「换一次样本」语义，不会拿同一份坏文本空转）。
+
 ## v0.1.40（2026-09-21）
 
 ### 🛠 问题修复

package/README.md

@@ -23,6 +23,9 @@ Kimi Code、MiMo Code、Grok Build、Pi、DeepSeek Harness……），提供
 <!-- relnotes:start -->
 ### 🛠 问题修复 · Fixes
 
+- 断点续跑/重跑更省钱：同任务再次生成大纲或计划直接命中精确缓存（1 小时内），不重复花全量规划调用；格式不合规的重试仍重新生成。
+- Resuming or re-running a task now hits an exact cache for outline/plan regeneration (within 1 hour), skipping a full planning call; format-failure retries still regenerate.
+
 - 修复长任务上下文撑爆后重试仍失败：压缩摘要此前只写日志、没真正替换发往模型的上下文；现在重试用「压缩摘要+最近消息」重建请求并放弃膨胀的旧会话，超长任务能真正续跑。
 - Fixed long-task overflow retries failing again: compaction summaries used to go only to logs. Retries now rebuild the request from the summary plus recent messages and drop the bloated old session.
 

package/app/core/planner.py

@@ -114,13 +114,13 @@ def _log_usage(source, role, task, res, agent=None, tool="", model="", provider=
     except Exception:
         pass
 
-CODE_PLAN_PROMPT = """你是技术负责人。请把下面的开发目标拆解为 __N__ 个以内、按顺序执行的子任务，
-并判定任务难度和后续质量步骤。只输出一个 ```json 代码块，不要输出其他内容。JSON 结构：
-{"difficulty": "easy 或 hard", "workflow": {"review_required": true, "max_repair_rounds": 0, "allow_switch": false}, "subtasks": [{"title": "简短标题", "detail": "具体要做什么，给执行工程师的直接指令", "files": ["涉及的文件路径"]}]}
-难度判定：常规增删改查/小函数/格式调整 = easy；跨模块改动/架构调整/复杂算法/安全相关 = hard。
-质量步骤判定：涉及安全、数据迁移、并发、权限、公共 API 或跨模块改动时必须评审；
-有明确自动验收的低风险小改可不做模型评审。max_repair_rounds 取 0-2，只有复杂任务才建议换将。
-子任务粒度要可独立验证；最后一个子任务必须包含整体联调/收尾。
+CODE_PLAN_PROMPT = """你是技术负责人。请把下面的开发目标拆解为 __N__ 个以内、按顺序执行的子任务，
+并判定任务难度和后续质量步骤。只输出一个 ```json 代码块，不要输出其他内容。JSON 结构：
+{"difficulty": "easy 或 hard", "workflow": {"review_required": true, "max_repair_rounds": 0, "allow_switch": false}, "subtasks": [{"title": "简短标题", "detail": "具体要做什么，给执行工程师的直接指令", "files": ["涉及的文件路径"]}]}
+难度判定：常规增删改查/小函数/格式调整 = easy；跨模块改动/架构调整/复杂算法/安全相关 = hard。
+质量步骤判定：涉及安全、数据迁移、并发、权限、公共 API 或跨模块改动时必须评审；
+有明确自动验收的低风险小改可不做模型评审。max_repair_rounds 取 0-2，只有复杂任务才建议换将。
+子任务粒度要可独立验证；最后一个子任务必须包含整体联调/收尾。
 
 ## 先探索再计划（重要，借鉴 OpenSpec explore）
 拆解之前先用你的读文件/搜索工具**实际查看工作目录**，找到目标相关的真实文件与函数，再据此拆解。
@@ -137,12 +137,12 @@ __CONTEXT__
 ## 验收命令（最终必须通过）
 __VERIFY__"""
 
-REVIEW_OUTLINE_PROMPT = """你是内容主编。请为下面的创作任务拟一份写作大纲（要点列表，3-8 条），
-并按内容风险建议后续质量步骤。
-只输出一个 ```json 代码块，不要输出其他内容。JSON 结构：
-{"outline": ["要点1", "要点2", ...], "workflow": {"reviewers": 1, "review_rounds": 1}}
-reviewers 取 1-2，review_rounds 取 1-3；研究报告、技术方案、长文和高发布门槛应增加，
-普通邮件、短翻译、工作汇报等低风险短内容保持 1。
+REVIEW_OUTLINE_PROMPT = """你是内容主编。请为下面的创作任务拟一份写作大纲（要点列表，3-8 条），
+并按内容风险建议后续质量步骤。
+只输出一个 ```json 代码块，不要输出其他内容。JSON 结构：
+{"outline": ["要点1", "要点2", ...], "workflow": {"reviewers": 1, "review_rounds": 1}}
+reviewers 取 1-2，review_rounds 取 1-3；研究报告、技术方案、长文和高发布门槛应增加，
+普通邮件、短翻译、工作汇报等低风险短内容保持 1。
 
 ## 创作任务
 __GOAL__
@@ -327,9 +327,14 @@ def make_serial_outline(task, author_agent=None, workdir=None, ev=None, log_path
             # 否则 stop_reason=max_tokens、正文为空（实测 2048 全被思考吞掉）
             # 流式回调把增量实时写进步骤日志，直连生成不再是一行标题的黑箱
             cb = _log_streamer(log_path)
-            res = modelhub.chat(prov["id"], model, prompt,
-                                max_tokens=16000, timeout=300, on_delta=cb,
-                                reasoning_effort=_reasoning_effort(task))
+            # §07 T2.2 幂等精确缓存（TTL 1h）：断点续跑/重跑同任务时大纲
+            # prompt 逐字节相同，命中省一次全量规划调用。第 2 次尝试故意
+            # 旁路缓存——首试 ok 但 JSON 不合法时，重试要的是「再抽一次样
+            # 本」而不是同一份坏文本。
+            res = modelhub.chat(prov["id"], model, prompt,
+                                max_tokens=16000, timeout=300, on_delta=cb,
+                                reasoning_effort=_reasoning_effort(task),
+                                cache_ttl=3600 if _attempt == 1 else 0)
             cb.flush()
             _log_usage("outline", "outline", task, res, model=model,
                        provider=prov.get("name", prov.get("id", "")),
@@ -409,56 +414,56 @@ def _norm_subtasks(data):
     return steps or None
 
 
-def _plan_difficulty(data):
+def _plan_difficulty(data):
     d = str((data or {}).get("difficulty") or "").lower()
-    return d if d in ("easy", "hard") else None
-
-
-def _plan_workflow(data):
-    """规范化规划器给出的后续步骤建议；最终安全下限由任务编译器裁决。"""
-    raw = (data or {}).get("workflow")
-    if not isinstance(raw, dict):
-        return None
-    out = {}
-    if isinstance(raw.get("review_required"), bool):
-        out["review_required"] = raw["review_required"]
-    try:
-        if "max_repair_rounds" in raw:
-            out["max_repair_rounds"] = max(0, min(2, int(raw["max_repair_rounds"])))
-    except (TypeError, ValueError):
-        pass
-    if isinstance(raw.get("allow_switch"), bool):
-        out["allow_switch"] = raw["allow_switch"]
-    return out or None
-
-
-def _orchestrator():
-    """编排者可用时返回 (provider, model)，否则 None。"""
-    try:
-        return modelhub.resolve_orchestrator()
-    except Exception:
-        return None
-
-
-def _reasoning_effort(task):
-    """把任务级思考程度映射为直连编排者支持的 reasoning 参数。"""
-    task = task if isinstance(task, dict) else {}
-    selected = str(task.get("thinking") or "standard").strip().lower()
-    if selected in ("low", "standard", "high"):
-        return {"low": "low", "standard": "medium", "high": "high"}[selected]
-    mode = str(task.get("mode") or "auto").lower()
-    difficulty = str(task.get("difficulty") or "auto").lower()
-    ttype = str(task.get("type") or "").lower()
-    try:
-        threshold = float(task.get("threshold") or 7.0)
-    except (TypeError, ValueError):
-        threshold = 7.0
-    if mode == "fast" or difficulty == "easy" or threshold <= 6.0:
-        return "low"
-    if (mode == "expert" or difficulty == "hard" or threshold >= 8.5
-            or ttype in ("research", "tech_proposal", "serial_novel")):
-        return "high"
-    return "medium"
+    return d if d in ("easy", "hard") else None
+
+
+def _plan_workflow(data):
+    """规范化规划器给出的后续步骤建议；最终安全下限由任务编译器裁决。"""
+    raw = (data or {}).get("workflow")
+    if not isinstance(raw, dict):
+        return None
+    out = {}
+    if isinstance(raw.get("review_required"), bool):
+        out["review_required"] = raw["review_required"]
+    try:
+        if "max_repair_rounds" in raw:
+            out["max_repair_rounds"] = max(0, min(2, int(raw["max_repair_rounds"])))
+    except (TypeError, ValueError):
+        pass
+    if isinstance(raw.get("allow_switch"), bool):
+        out["allow_switch"] = raw["allow_switch"]
+    return out or None
+
+
+def _orchestrator():
+    """编排者可用时返回 (provider, model)，否则 None。"""
+    try:
+        return modelhub.resolve_orchestrator()
+    except Exception:
+        return None
+
+
+def _reasoning_effort(task):
+    """把任务级思考程度映射为直连编排者支持的 reasoning 参数。"""
+    task = task if isinstance(task, dict) else {}
+    selected = str(task.get("thinking") or "standard").strip().lower()
+    if selected in ("low", "standard", "high"):
+        return {"low": "low", "standard": "medium", "high": "high"}[selected]
+    mode = str(task.get("mode") or "auto").lower()
+    difficulty = str(task.get("difficulty") or "auto").lower()
+    ttype = str(task.get("type") or "").lower()
+    try:
+        threshold = float(task.get("threshold") or 7.0)
+    except (TypeError, ValueError):
+        threshold = 7.0
+    if mode == "fast" or difficulty == "easy" or threshold <= 6.0:
+        return "low"
+    if (mode == "expert" or difficulty == "hard" or threshold >= 8.5
+            or ttype in ("research", "tech_proposal", "serial_novel")):
+        return "high"
+    return "medium"
 
 
 def make_code_plan(task, planner_agent, workdir, ev=None, resume=None, log_path=None):
@@ -482,10 +487,10 @@ def make_code_plan(task, planner_agent, workdir, ev=None, resume=None, log_path=
     _log_usage("plan", "plan", task, res, agent=planner_agent)
     data = runner.extract_json(res.get("text") or "")
     steps = _norm_subtasks(data)
-    if steps:
-        return {"source": "llm(%s)" % planner_agent["id"], "steps": steps,
-                "difficulty": _plan_difficulty(data),
-                "workflow": _plan_workflow(data)}
+    if steps:
+        return {"source": "llm(%s)" % planner_agent["id"], "steps": steps,
+                "difficulty": _plan_difficulty(data),
+                "workflow": _plan_workflow(data)}
     return _fallback_code_plan(task, "LLM 计划解析失败，退化为单步模板")
 
 
@@ -496,13 +501,15 @@ def _fallback_code_plan(task, note):
 
 def _orch_code_plan(task, prov, model, log_path=None):
     cb = _log_streamer(log_path)
+    # §07 T2.2：代码计划同任务同 prompt（断点续跑/重试），TTL 1h 精确缓存
     res = modelhub.chat(prov["id"], model,
                         (CODE_PLAN_PROMPT.replace("__N__", str(MAX_SUBTASKS))
                          .replace("__GOAL__", task["goal"])
                          .replace("__CONTEXT__", task.get("context") or "（无）")
                          .replace("__VERIFY__", task.get("verify_command") or "（未配置）")),
-                        max_tokens=8000, timeout=300, on_delta=cb,
-                        reasoning_effort=_reasoning_effort(task))
+                        max_tokens=8000, timeout=300, on_delta=cb,
+                        reasoning_effort=_reasoning_effort(task),
+                        cache_ttl=3600)
     cb.flush()
     _log_usage("plan", "plan", task, res, model=model,
                provider=prov.get("name", prov.get("id", "")))
@@ -515,9 +522,9 @@ def _orch_code_plan(task, prov, model, log_path=None):
     if not steps:
         _append_log(log_path, "编排者计划返回内容无法解析为子任务，回退 CLI")
         return None
-    return {"source": "编排者(%s · %s)" % (prov.get("name", prov["id"]), model),
-            "steps": steps, "difficulty": _plan_difficulty(data),
-            "workflow": _plan_workflow(data)}
+    return {"source": "编排者(%s · %s)" % (prov.get("name", prov["id"]), model),
+            "steps": steps, "difficulty": _plan_difficulty(data),
+            "workflow": _plan_workflow(data)}
 
 
 def make_review_outline(task):
@@ -530,8 +537,9 @@ def make_review_outline(task):
                         (REVIEW_OUTLINE_PROMPT
                          .replace("__GOAL__", task["goal"])
                          .replace("__CONTEXT__", task.get("context") or "（无）")),
-                        max_tokens=8000, timeout=300,
-                        reasoning_effort=_reasoning_effort(task))
+                        max_tokens=8000, timeout=300,
+                        reasoning_effort=_reasoning_effort(task),
+                        cache_ttl=3600)   # §07 T2.2：同任务重试幂等，TTL 1h
     _log_usage("outline", "outline", task, res, model=model,
                provider=prov.get("name", prov.get("id", "")))
     if not res["ok"]:
@@ -541,20 +549,20 @@ def make_review_outline(task):
     if not isinstance(outline, list):
         return None
     items = [str(x).strip()[:120] for x in outline if str(x).strip()][:8]
-    if not items:
-        return None
-    workflow = data.get("workflow") if isinstance(data.get("workflow"), dict) else {}
-    try:
-        reviewers = max(1, min(2, int(workflow.get("reviewers") or 1)))
-    except (TypeError, ValueError):
-        reviewers = 1
-    try:
-        review_rounds = max(1, min(3, int(workflow.get("review_rounds") or 1)))
-    except (TypeError, ValueError):
-        review_rounds = 1
-    return {"source": "编排者(%s · %s)" % (prov.get("name", prov["id"]), model),
-            "items": items,
-            "workflow": {"reviewers": reviewers, "review_rounds": review_rounds}}
+    if not items:
+        return None
+    workflow = data.get("workflow") if isinstance(data.get("workflow"), dict) else {}
+    try:
+        reviewers = max(1, min(2, int(workflow.get("reviewers") or 1)))
+    except (TypeError, ValueError):
+        reviewers = 1
+    try:
+        review_rounds = max(1, min(3, int(workflow.get("review_rounds") or 1)))
+    except (TypeError, ValueError):
+        review_rounds = 1
+    return {"source": "编排者(%s · %s)" % (prov.get("name", prov["id"]), model),
+            "items": items,
+            "workflow": {"reviewers": reviewers, "review_rounds": review_rounds}}
 
 
 def make_novel_plan(task, author, critics):

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codebee",
-  "version": "0.1.40",
+  "version": "0.1.41",
   "description": "CodeBee · 多智能体编排台 — 统一编排本机 AI 编码 CLI（Codex / Claude Code / QwenCode / OpenCode / Aider…）：目标→拆解→路由→执行→验证→跨厂商评审→修复→报告。纯 Python 标准库实现，本地 Web UI。",
   "license": "MIT",
   "bin": {