codebee 0.1.95 → 0.1.97

package/app/core/step_runner.py

@@ -1,133 +1,133 @@
-# -*- coding: utf-8 -*-
-"""Step 执行器：压缩触发的重试守门（复用已渲染 prompt，不重复组装）。
-
-设计稿：docs/migration/02-context-compaction.md §1D。
-参考 dsh packages/core/agent-loop/src/agent.ts:307-378：
-  失败重试只在「前进 surface generation」时返回 {kind:'retry'}，
-  复用已渲染 PromptAssembly，不重跑 pre-step / 组装 / 用户消息准入。
-
-Tutti 语境：runner.run_agent 自带 prompt（已渲染）；重试的风险不是重复渲染
-（prompt 是同一个字符串），而是「不压缩直接重试」必然再次撑爆。
-本模块的守门：仅当 maybe_compact 真正前进了 surface generation 才重试一次。
-"""
-from __future__ import annotations
-
-import logging
-
-from .compaction import maybe_compact, DEFAULT_PRESSURE_THRESHOLD
-from .error_codes import ErrorCode
-
-log = logging.getLogger(__name__)
-
-# 触发「压缩后重试」的错误码
-_OVERFLOW_CODES = {ErrorCode.CONTEXT_OVERFLOW, ErrorCode.MAX_TOKENS}
-
-# 事前预检阈值：最近一次调用的上下文锚点已占本步模型容量九成 → 第一个请求
-# 大概率被供应商拒。只拦「大概率必死」，不提前压缩平时的高压（used() 是累计
-# 口径，比真实上下文偏大；0.8 的响应式阈值语义与此不同，勿混用）。
-_PRECHECK_RATIO = 0.9
-
-
-def _v2_compaction_tuning():
-    """settings_v2 orchestrator.compaction 微调 → (threshold|None, retain|None)。
-
-    None = 未配置（走 maybe_compact 的模块默认）；读取/注册失败静默回落。
-    retain 允许 0（尾部不保留），与 None（未配置）语义不同，勿合并。"""
-    try:
-        from .settings_schema import get as ss_get, register_default_namespaces
-        register_default_namespaces()
-        c = (ss_get("orchestrator") or {}).get("compaction") or {}
-        th = c.get("pressure_threshold")
-        rt = c.get("retain_tail_tokens")
-        return (float(th) if th else None,
-                int(rt) if rt is not None else None)
-    except Exception:
-        return None, None
-
-
-def execute_step(session, run_agent_fn, prompt, *, model: str = "",
-                 llm_caller=None, retain_tail_tokens=None, **kwargs):
-    """执行一次 step；撑爆时压缩并守门重试一次。
-
-    事前预检（借鉴 freebuff base-chat 每步容量重估）：换将切到小窗口模型后，
-    响应式路径要等供应商报 CONTEXT_OVERFLOW 才压缩——报得不干净就直接卡死。
-    先用「最近一次调用上下文 vs 本步模型容量」判一次，超阈先压缩再发。
-
-    Args:
-        session: session_log.Session（记录 replace_generation）
-        run_agent_fn: runner.run_agent（或测试替身），签名 (prompt, **kwargs) -> dict
-        prompt: 已渲染的提示词（重试时原样复用，不重新渲染）
-        model: 本步将要使用的模型名（容量与压力估算都按它算）
-        llm_caller: 压缩摘要 LLM（None = 不压缩，永不重试）
-        **kwargs: 透传 run_agent_fn
-
-    Returns:
-        (result, retried: bool)
-    """
-    v2_th, v2_rt = _v2_compaction_tuning()
-    compact_kwargs = {}
-    if retain_tail_tokens is not None:
-        compact_kwargs["retain_tail_tokens"] = retain_tail_tokens
-    elif v2_rt is not None:
-        compact_kwargs["retain_tail_tokens"] = v2_rt
-    if v2_th is not None:
-        compact_kwargs["threshold"] = v2_th
-    if llm_caller is not None and model:
-        try:
-            from .token_meter import token_meter
-            ctx = token_meter.last_context(session.run_id)
-            cap = token_meter.capacity(model)
-            if ctx > 0 and cap > 0 and ctx > cap * _PRECHECK_RATIO:
-                gen_p0 = session.replace_generation()
-                if maybe_compact(session, model=model, llm_caller=llm_caller,
-                                 run_id=session.run_id, **compact_kwargs) \
-                        and session.replace_generation() > gen_p0:
-                    log.info("precheck: last ctx~%d > %.0f%% of cap(%s)=%d, "
-                             "compacted before step", ctx, _PRECHECK_RATIO * 100,
-                             model, cap)
-        except Exception:
-            log.debug("precheck skipped", exc_info=True)
-    gen0 = session.replace_generation()
-    result = run_agent_fn(prompt, **kwargs)
-    code = result.get("error_code") or ""
-    if code not in _OVERFLOW_CODES or llm_caller is None:
-        return result, False
-    # 撑爆 → 尝试压缩
-    compacted = maybe_compact(session, model=model, llm_caller=llm_caller,
-                              run_id=session.run_id, **compact_kwargs)
-    gen1 = session.replace_generation()
-    if not compacted or gen1 <= gen0:
-        # 压缩未发生或未前进 generation：重试必然再次撑爆，不重试
-        log.info("overflow but no compaction progress (gen %s->%s), no retry",
-                 gen0, gen1)
-        return result, False
-    # 压缩生效：surface 已折叠。**真正替换 CLI 请求上下文**（2026-09-21
-    # 清最老待深挖项：此前重试仍发原 prompt + 续旧 resume 会话——摘要只进
-    # 审计日志，CLI 看到的上下文一点没小，重试必然再爆）：
-    # ①重试 prompt 改用 session 派生转录（折叠摘要 + 近尾消息，体量骤降）；
-    # ②弃 resume（旧会话本体仍是膨胀态，续上就前功尽弃）。
-    msgs = []
-    try:
-        msgs = session.derive_messages()
-    except Exception:
-        log.debug("derive_messages failed on retry", exc_info=True)
-    kwargs2 = dict(kwargs)
-    if msgs:
-        transcript = "\n\n".join("[%s] %s" % (m.get("role"), m.get("content"))
-                                 for m in msgs)
-        if len(transcript) > 24_000:
-            transcript = transcript[-24_000:]
-        retry_prompt = transcript
-        # 本步指令必须在场：被折叠/截断掉就显式补在尾部（缺失指令比重复严重）
-        tail = transcript[-2000:]
-        if prompt[:400] not in tail:
-            retry_prompt = transcript + "\n\n" + prompt
-        kwargs2.pop("resume", None)
-        log.info("compaction advanced gen %s->%s, retrying with derived "
-                 "context (%d msgs, resume dropped)", gen0, gen1, len(msgs))
-        result2 = run_agent_fn(retry_prompt, **kwargs2)
-    else:
-        log.info("compaction advanced gen %s->%s but no surface derived, "
-                 "retrying same prompt", gen0, gen1)
-        result2 = run_agent_fn(prompt, **kwargs)
+# -*- coding: utf-8 -*-
+"""Step 执行器：压缩触发的重试守门（复用已渲染 prompt，不重复组装）。
+
+设计稿：docs/migration/02-context-compaction.md §1D。
+参考 dsh packages/core/agent-loop/src/agent.ts:307-378：
+  失败重试只在「前进 surface generation」时返回 {kind:'retry'}，
+  复用已渲染 PromptAssembly，不重跑 pre-step / 组装 / 用户消息准入。
+
+Tutti 语境：runner.run_agent 自带 prompt（已渲染）；重试的风险不是重复渲染
+（prompt 是同一个字符串），而是「不压缩直接重试」必然再次撑爆。
+本模块的守门：仅当 maybe_compact 真正前进了 surface generation 才重试一次。
+"""
+from __future__ import annotations
+
+import logging
+
+from .compaction import maybe_compact, DEFAULT_PRESSURE_THRESHOLD
+from .error_codes import ErrorCode
+
+log = logging.getLogger(__name__)
+
+# 触发「压缩后重试」的错误码
+_OVERFLOW_CODES = {ErrorCode.CONTEXT_OVERFLOW, ErrorCode.MAX_TOKENS}
+
+# 事前预检阈值：最近一次调用的上下文锚点已占本步模型容量九成 → 第一个请求
+# 大概率被供应商拒。只拦「大概率必死」，不提前压缩平时的高压（used() 是累计
+# 口径，比真实上下文偏大；0.8 的响应式阈值语义与此不同，勿混用）。
+_PRECHECK_RATIO = 0.9
+
+
+def _v2_compaction_tuning():
+    """settings_v2 orchestrator.compaction 微调 → (threshold|None, retain|None)。
+
+    None = 未配置（走 maybe_compact 的模块默认）；读取/注册失败静默回落。
+    retain 允许 0（尾部不保留），与 None（未配置）语义不同，勿合并。"""
+    try:
+        from .settings_schema import get as ss_get, register_default_namespaces
+        register_default_namespaces()
+        c = (ss_get("orchestrator") or {}).get("compaction") or {}
+        th = c.get("pressure_threshold")
+        rt = c.get("retain_tail_tokens")
+        return (float(th) if th else None,
+                int(rt) if rt is not None else None)
+    except Exception:
+        return None, None
+
+
+def execute_step(session, run_agent_fn, prompt, *, model: str = "",
+                 llm_caller=None, retain_tail_tokens=None, **kwargs):
+    """执行一次 step；撑爆时压缩并守门重试一次。
+
+    事前预检（借鉴 freebuff base-chat 每步容量重估）：换将切到小窗口模型后，
+    响应式路径要等供应商报 CONTEXT_OVERFLOW 才压缩——报得不干净就直接卡死。
+    先用「最近一次调用上下文 vs 本步模型容量」判一次，超阈先压缩再发。
+
+    Args:
+        session: session_log.Session（记录 replace_generation）
+        run_agent_fn: runner.run_agent（或测试替身），签名 (prompt, **kwargs) -> dict
+        prompt: 已渲染的提示词（重试时原样复用，不重新渲染）
+        model: 本步将要使用的模型名（容量与压力估算都按它算）
+        llm_caller: 压缩摘要 LLM（None = 不压缩，永不重试）
+        **kwargs: 透传 run_agent_fn
+
+    Returns:
+        (result, retried: bool)
+    """
+    v2_th, v2_rt = _v2_compaction_tuning()
+    compact_kwargs = {}
+    if retain_tail_tokens is not None:
+        compact_kwargs["retain_tail_tokens"] = retain_tail_tokens
+    elif v2_rt is not None:
+        compact_kwargs["retain_tail_tokens"] = v2_rt
+    if v2_th is not None:
+        compact_kwargs["threshold"] = v2_th
+    if llm_caller is not None and model:
+        try:
+            from .token_meter import token_meter
+            ctx = token_meter.last_context(session.run_id)
+            cap = token_meter.capacity(model)
+            if ctx > 0 and cap > 0 and ctx > cap * _PRECHECK_RATIO:
+                gen_p0 = session.replace_generation()
+                if maybe_compact(session, model=model, llm_caller=llm_caller,
+                                 run_id=session.run_id, **compact_kwargs) \
+                        and session.replace_generation() > gen_p0:
+                    log.info("precheck: last ctx~%d > %.0f%% of cap(%s)=%d, "
+                             "compacted before step", ctx, _PRECHECK_RATIO * 100,
+                             model, cap)
+        except Exception:
+            log.debug("precheck skipped", exc_info=True)
+    gen0 = session.replace_generation()
+    result = run_agent_fn(prompt, **kwargs)
+    code = result.get("error_code") or ""
+    if code not in _OVERFLOW_CODES or llm_caller is None:
+        return result, False
+    # 撑爆 → 尝试压缩
+    compacted = maybe_compact(session, model=model, llm_caller=llm_caller,
+                              run_id=session.run_id, **compact_kwargs)
+    gen1 = session.replace_generation()
+    if not compacted or gen1 <= gen0:
+        # 压缩未发生或未前进 generation：重试必然再次撑爆，不重试
+        log.info("overflow but no compaction progress (gen %s->%s), no retry",
+                 gen0, gen1)
+        return result, False
+    # 压缩生效：surface 已折叠。**真正替换 CLI 请求上下文**（2026-09-21
+    # 清最老待深挖项：此前重试仍发原 prompt + 续旧 resume 会话——摘要只进
+    # 审计日志，CLI 看到的上下文一点没小，重试必然再爆）：
+    # ①重试 prompt 改用 session 派生转录（折叠摘要 + 近尾消息，体量骤降）；
+    # ②弃 resume（旧会话本体仍是膨胀态，续上就前功尽弃）。
+    msgs = []
+    try:
+        msgs = session.derive_messages()
+    except Exception:
+        log.debug("derive_messages failed on retry", exc_info=True)
+    kwargs2 = dict(kwargs)
+    if msgs:
+        transcript = "\n\n".join("[%s] %s" % (m.get("role"), m.get("content"))
+                                 for m in msgs)
+        if len(transcript) > 24_000:
+            transcript = transcript[-24_000:]
+        retry_prompt = transcript
+        # 本步指令必须在场：被折叠/截断掉就显式补在尾部（缺失指令比重复严重）
+        tail = transcript[-2000:]
+        if prompt[:400] not in tail:
+            retry_prompt = transcript + "\n\n" + prompt
+        kwargs2.pop("resume", None)
+        log.info("compaction advanced gen %s->%s, retrying with derived "
+                 "context (%d msgs, resume dropped)", gen0, gen1, len(msgs))
+        result2 = run_agent_fn(retry_prompt, **kwargs2)
+    else:
+        log.info("compaction advanced gen %s->%s but no surface derived, "
+                 "retrying same prompt", gen0, gen1)
+        result2 = run_agent_fn(prompt, **kwargs)
     return result2, True

package/app/core/store.py

@@ -180,6 +180,13 @@ def create_task(payload):
         "created_at": time.strftime("%Y-%m-%d %H:%M:%S"),
         "status": "created",
     }
+    # Every new task must persist one verified execution policy. If policy
+    # normalization fails, do not silently replace it with a broader default.
+    try:
+        from . import policy
+        task["sandbox"] = policy.normalize_sandbox(payload.get("sandbox"), wd)
+    except Exception as exc:
+        raise ValueError("沙箱策略无法验证，任务未创建") from exc
     thinking = str(payload.get("thinking") or "standard").strip().lower()
     task["thinking"] = thinking if thinking in ("auto", "low", "standard", "high") else "auto"
     # 流程修订出处（借鉴 WorkDSH ADR-0010）：创建时钉住当时流程定义的指纹。
@@ -1738,6 +1745,23 @@ def retry_task(task_id):
                         "chapter_scores": scores,
                     }
                     break
+        else:
+            # 断点续跑（代码等非连载任务，A+B）：上一遍失败/取消/超时 run 的
+            # 实现会话（inherit_session，跨 run 续 CLI 会话）与断点简报
+            # （handoff，走到哪/改了什么/还剩什么）继承给新 run——重试从断点
+            # 继续而不是从零烧 token。done 的 run 不继承：主动重跑语义就是重来。
+            prev_runs = sorted((r for r in _RUNS.values()
+                                if r.get("task_id") == task_id and r["id"] != run["id"]),
+                               key=lambda r: r["id"], reverse=True)
+            for prev in prev_runs:
+                if prev.get("status") not in ("failed", "cancelled", "timeout"):
+                    continue
+                if prev.get("impl_session") and "inherit_session" not in run:
+                    run["inherit_session"] = dict(prev["impl_session"])
+                if prev.get("handoff") and "handoff" not in run:
+                    run["handoff"] = dict(prev["handoff"])
+                if "inherit_session" in run or "handoff" in run:
+                    break
         task["status"] = "queued"
         _save_json(paths.TASKS_DIR / (task_id + ".json"), task)
         _save_json(paths.RUNS_DIR / run["id"] / "run.json", run)