codebee 0.1.23 → 0.1.24

package/app/core/pipeline.py

@@ -19,7 +19,7 @@ import re
 import threading
 import time
 
-from . import aiflavor, catalog, history, jobs, knowledge, manager, modelhub, mocks, paihang, planner, registry, router, runner, skills, store, task_compile, usage
+from . import aiflavor, catalog, dispatch_log, history, jobs, knowledge, manager, modelhub, mocks, paihang, planner, registry, router, runner, skills, store, task_compile, usage
 from . import builtin_agent
 from . import diagnostics
 from . import paths as paths_mod
@@ -60,6 +60,20 @@ def _task_images(task, workdir):
         return []
 
 
+def _refresh_attachment_context(task, workdir):
+    """为新旧任务统一生成附件正文上下文；失败时保留原上下文。"""
+    if not task.get("attachments"):
+        return task
+    try:
+        from . import attachments as att_mod
+        out = dict(task)
+        out["context"] = att_mod.merge_context(
+            task.get("context") or "", task.get("attachments") or [], workdir=workdir)
+        return out
+    except Exception:
+        return task
+
+
 def _ms_name(raw):
     name = re.sub(r"[\\/\x00]+", "_", str(raw or "")).strip()
     name = re.sub(r"\.{2,}", "_", name).lstrip(".")
@@ -253,6 +267,7 @@ def _drain_directives(run_id, workdir, role=None, step_n=None):
         lines.append("本步为评审步骤：请把上述用户意见作为评分依据之一，"
                      "在相应维度的分数与 issues 中明确体现（引用用户原话）。")
     imgs = []
+    attachment_paths = []
     for m in msgs:
         stamp = m.get("created_at") or ""
         sender = m.get("sender") or "用户"
@@ -261,6 +276,7 @@ def _drain_directives(run_id, workdir, role=None, step_n=None):
                      else "- [%s %s]（附件指令，见下方文件）" % (stamp, sender))
         for rel in (m.get("attachments") or []):
             rel = str(rel)
+            attachment_paths.append(rel)
             low = rel.lower()
             if low.endswith((".png", ".jpg", ".jpeg", ".gif", ".webp", ".bmp")):
                 ap = os.path.join(workdir or "", rel) if workdir else rel
@@ -271,6 +287,13 @@ def _drain_directives(run_id, workdir, role=None, step_n=None):
                     lines.append("  · 图片附件：%s" % rel)
             else:
                 lines.append("  · 文件附件：%s（位于工作目录，可直接读取）" % rel)
+    if attachment_paths:
+        try:
+            from . import attachments as att_mod
+            items = att_mod.items_from_paths(attachment_paths, workdir)
+            lines.append(att_mod.context_block(items, workdir=workdir))
+        except Exception:
+            lines.append("附件预读失败：执行者必须逐个打开上述路径，无法读取时明确说明。")
     return "\n".join(lines), imgs
 
 
@@ -609,6 +632,9 @@ def _record_usage(run_id, role, agent, res, source="pipeline", step=0):
     try:
         run = store.get_run(run_id) or {}
         task = store.get_task(run.get("task_id")) if run.get("task_id") else None
+        provider = res.get("provider") or agent.get("provider") or {}
+        provider_id = provider.get("id") if isinstance(provider, dict) else ""
+        provider_name = provider.get("name") if isinstance(provider, dict) else ""
         usage.record(
             source=source, run_id=run_id, step=step,
             task_id=run.get("task_id") or "",
@@ -617,13 +643,14 @@ def _record_usage(run_id, role, agent, res, source="pipeline", step=0):
             agent_label=agent.get("label", ""),
             tool=agent.get("kind", ""),
             model=res.get("model") or "",
+            provider=res.get("provider_id") or provider_id or provider_name or "",
             ok=bool(res.get("ok")),
             duration_s=float(res.get("raw", {}).get("duration") or 0.0),
             cost_usd=float(res.get("cost_usd") or 0.0),
             usage=res.get("usage"))
         # 告警模块：CLI 调用成功/失败上报（provider 名与 usage 台账一致）
         from . import health
-        prov = agent.get("provider") or {}
+        prov = res.get("provider") or agent.get("provider") or {}
         prov_name = (prov.get("name") if isinstance(prov, dict) else "") or ""
         if prov_name:
             if res.get("ok"):
@@ -663,6 +690,9 @@ __GOAL__
 ## 未通过的原因
 __ISSUES__
 
+## 原始背景与附件
+__CONTEXT__
+
 ## 要求
 - 只针对上述问题修复；不要无关重构。
 - 完成后用 2-3 句话说明改了什么。
@@ -688,6 +718,9 @@ __GOAL__
 ## 验收命令
 __VERIFY__
 
+## 原始背景与附件要求
+__CONTEXT__
+
 ## 变更内容（git diff，若为空表示无法获取）
 __DIFF__
 """
@@ -746,6 +779,7 @@ def _run_review(run_id, task, workdir, reviewer, ev):
     prompt = (CODE_REVIEW_PROMPT
               .replace("__GOAL__", task["goal"])
               .replace("__VERIFY__", task.get("verify_command") or "（未配置）")
+              .replace("__CONTEXT__", task.get("context") or "（无）")
               .replace("__DIFF__", diff or "（无法获取 git diff，请综合任务目标谨慎评审）"))
     res = _run_step(run_id, "review", reviewer, prompt, workdir, readonly=True, ev=ev,
                     images=_task_images(task, workdir))
@@ -904,6 +938,31 @@ def _record_actual_route(run_id, task, agents, stats, implementer,
     store.update_run(run_id, route_plan={
         "task": spec, "implement": impl_plan, "review": review_plan,
     })
+    task_id = task.get("id") or (store.get_run(run_id) or {}).get("task_id") or ""
+    difficulty = (task.get("difficulty") or "auto")
+    for plan in (impl_plan, review_plan):
+        dispatch_log.record_event(
+            run_id=run_id, task_id=task_id, task_type=spec.get("type") or "",
+            difficulty=difficulty, role=plan.get("role") or "",
+            phase="selected", selected=plan.get("selected") or "",
+            participants=plan.get("participants") or (),
+            candidates=plan.get("candidates") or (),
+            fallback=plan.get("fallback") or (),
+            selection_reason=plan.get("selection_reason") or "")
+
+
+def _record_dispatch_completed(run_id, task, result, verify_pass=None, review_pass=None):
+    """记录运行终态，供调度回放与线上指标复盘使用。"""
+    try:
+        spec = task.get("_compiled_spec") or task_compile.compile_task(task)
+        dispatch_log.record_event(
+            run_id=run_id,
+            task_id=task.get("id") or (store.get_run(run_id) or {}).get("task_id") or "",
+            task_type=spec.get("type") or "", difficulty=task.get("difficulty") or "auto",
+            role="", phase="completed", result=result,
+            verify_pass=verify_pass, review_pass=review_pass)
+    except Exception:
+        pass
 
 
 def _run_code(run, task, agents, ev, stats, mode):
@@ -1080,8 +1139,9 @@ def _run_code(run, task, agents, ev, stats, mode):
         return False
 
     def review_and_score():
-        review_json = _run_review(run_id, task, workdir, modelhub.bind_agent(reviewer, difficulty), ev)
         verify_pass, verify_ran = _run_verify(run_id, task, workdir, ev)
+        # 先让确定性验证落盘，再执行模型评审；终态判断会同时使用两份证据。
+        review_json = _run_review(run_id, task, workdir, modelhub.bind_agent(reviewer, difficulty), ev)
         return review_json, verify_pass, verify_ran
 
     attempt_note = route.get("implementer", "") if mode == "auto" else ""
@@ -1099,6 +1159,7 @@ def _run_code(run, task, agents, ev, stats, mode):
             prompt = (CODE_FIX_PROMPT
                       .replace("__GOAL__", task["goal"])
                       .replace("__ISSUES__", issues_txt)
+                      .replace("__CONTEXT__", task.get("context") or "（无）")
                       .replace("__VERIFY_HINT__", _verify_hint(task)))
             res = _run_step(run_id, "fix-r%d" % round_no, modelhub.bind_agent(impl, difficulty),
                             prompt, workdir, readonly=False, ev=ev,
@@ -1218,6 +1279,7 @@ __GOAL__
 __CONTEXT__
 
 ## 要求
+- 背景与上下文里若有「附件材料」，先逐个读取再处理——附件是必要输入，不读附件就作答视为未完成。
 - 能改直接改、能写直接写（限本工作目录内），产出文件一律 UTF-8 编码（PowerShell 写文件显式 -Encoding UTF8）。
 - 正文直接交代结果与答案：不要写「本轮做了什么」这类开场总结，也不要「回复：」这类引导词。
 - 回复的最后一行单独输出一行交代结果：
@@ -1247,6 +1309,7 @@ __GOAL__
 __CONTEXT__
 
 ## 要求
+- 背景与上下文里若有「附件材料」，先逐个读取再处理——附件是必要输入，不读附件就作答视为未完成。
 - 能改直接改、能写直接写（用工具，限本工作目录内），产出文件一律 UTF-8 编码。
 - 完成后直接给用户结论与答案，需要时顺带交代产出/修改了哪些文件；不要写「本轮做了什么」这类开场白。
 __FOLLOWUPS__"""
@@ -1408,6 +1471,8 @@ def _run_direct(run, task, agents, ev, stats, mode):
                 # 扫榜选材（借鉴 oh-story 扫榜）：抓七猫排行榜公开数据注入，
                 # AI 做选题洞察；抓取失败回落普通直连提示词
                 prompt = paihang.rank_scan_prompt(task.get("goal") or "") or ""
+                if prompt and task.get("context"):
+                    prompt += "\n\n## 用户背景与附件\n" + task["context"]
             if not prompt:
                 if bi is not None:
                     prompt = (BUILTIN_DIRECT_PROMPT
@@ -1598,6 +1663,9 @@ __GOAL__
 ## 评审汇总（各维度均分与主要问题）
 __CRITIQUE__
 
+## 原始背景与附件
+__CONTEXT__
+
 ## 要求
 - 针对性改进所有 major 问题；保持既定风格与设定。
 - 完成后用 3 句话说明本轮改了什么。"""
@@ -1738,6 +1806,8 @@ def _ensure_critique_placeholders(tpl):
         tpl += "\n\n## 待评审稿件\n---\n__MANUSCRIPT__\n---"
     if "__DIMKEYS__" not in tpl:
         tpl = ("请按维度打分（1-10 分）。\n\n" + tpl)
+    if "__CONTEXT__" not in tpl:
+        tpl += "\n\n## 原始任务背景与附件参考\n__CONTEXT__"
     return tpl
 
 
@@ -1774,6 +1844,9 @@ __GOAL__
 ## 本章评审意见
 __CRITIQUE__
 
+## 原始背景与附件
+__CONTEXT__
+
 ## 要求
 - 针对性解决所有 major 问题，保持与前后的剧情衔接；字数仍约 __WORDS__ 字。"""
 
@@ -1789,6 +1862,9 @@ SERIAL_GLOBAL_PROMPT = """你是网文主编（不要修改任何文件）。全
 ## 全书目标
 __GOAL__
 
+## 原始背景与附件
+__CONTEXT__
+
 ## 全文
 ---
 __MANUSCRIPT__
@@ -1976,8 +2052,9 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
         kb = knowledge.block_for(task)
         if kb:
             tpl = tpl.replace("## 待评审稿件", "%s\n\n## 待评审稿件" % kb, 1)
-        return tpl.replace("__DIMKEYS__", dimkey).replace(
-            "__MANUSCRIPT__", text or "（稿件为空！）")
+        return (tpl.replace("__DIMKEYS__", dimkey)
+                .replace("__CONTEXT__", task.get("context") or "（无）")
+                .replace("__MANUSCRIPT__", text or "（稿件为空！）"))
 
     # ---- 1) 大纲（断点续跑时直接继承上一遍，保证全书结构一致）
     inherit = run.get("inherit") or {}
@@ -2175,8 +2252,11 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
         else:
             # stable_order：同一任务 8 个章节的技能块必须字节级一致（§07 T1.2' 前缀缓存）
             sk_block, _ = skills.block_for(task, stable_order=True)
-            if bible:
-                sk_block = (sk_block + "\n\n" + bible) if sk_block else bible
+        if bible:
+            sk_block = (sk_block + "\n\n" + bible) if sk_block else bible
+        if task.get("context"):
+            task_ctx = "## 任务背景与附件\n" + task["context"]
+            sk_block = (sk_block + "\n\n" + task_ctx) if sk_block else task_ctx
             kb_block = knowledge.block_for(task)
             if kb_block:
                 sk_block = (sk_block + "\n\n" + kb_block) if sk_block else kb_block
@@ -2497,6 +2577,7 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
                           .replace("__I__", str(i)).replace("__FILE__", ch_file)
                           .replace("__GOAL__", task["goal"])
                           .replace("__CRITIQUE__", "\n".join(crit_lines))
+                          .replace("__CONTEXT__", task.get("context") or "（无）")
                           .replace("__WORDS__", str(wpc)))
                 _run_step(run_id, "revise-c%d" % i, modelhub.bind_agent(impl, difficulty), prompt,
                           step_wd, readonly=False, ev=ev, timeout=2400,
@@ -2553,6 +2634,7 @@ def _run_serial_review(run, task, agents, ev, stats, mode, critics, impl, route,
                 res = _run_step(run_id, "global-critique", modelhub.bind_agent(agent, difficulty),
                                 (gtpl.replace("__DIMKEYS__", dimkey)
                                  .replace("__GOAL__", task["goal"])
+                                 .replace("__CONTEXT__", task.get("context") or "（无）")
                                  .replace("__MANUSCRIPT__", full_text[:60000])),
                                 workdir, readonly=True, ev=ev, timeout=2400)
                 gj = _critique_json(res, dims)
@@ -3009,6 +3091,7 @@ def _run_content_review(run, task, agents, ev, stats, mode):
         crit_prompt = (_ensure_critique_placeholders(
             _tpl(task, "critique_prompt", NOVEL_CRITIQUE_PROMPT))
             .replace("__DIMKEYS__", dimkey)
+            .replace("__CONTEXT__", task.get("context") or "（无）")
             .replace("__MANUSCRIPT__", manuscript or "（稿件为空！）"))
         # AI 味确定性检测（借鉴 oh-story 去AI味）：客观参考线随评审下发，
         # 命中才追加——评审官结合上下文判断是否真问题，脚本不直接扣分
@@ -3076,6 +3159,7 @@ def _run_content_review(run, task, agents, ev, stats, mode):
             prompt = (NOVEL_REVISE_PROMPT.replace("__FILE__", ms_name)
                       .replace("__ROLE__", _content_role(task))
                       .replace("__GOAL__", task["goal"])
+                      .replace("__CONTEXT__", task.get("context") or "（无）")
                       .replace("__CRITIQUE__", "\n".join(crit_lines)))
             prompt += _content_contract(task)
             _run_step(run_id, "revise-r%d" % r, modelhub.bind_agent(impl, difficulty), prompt,
@@ -3405,6 +3489,7 @@ def execute_run(run_id):
                      task_spec_summary=task_compile.summary(task_spec),
                      difficulty=task_spec["difficulty"])
     task = dict(task)
+    task = _refresh_attachment_context(task, task.get("workdir") or "")
     task["_compiled_spec"] = task_spec
     # 运行内统一使用编译后的难度；store 中历史任务常带 difficulty=auto，
     # 不能让这个兼容值覆盖 easy/default/hard 的模型调度决策。
@@ -3515,6 +3600,33 @@ def execute_run(run_id):
         except Exception:
             pass
     finally:
+        # 全类型统一写调度终态与质量反馈。调用成功只代表传输可靠；真正用于
+        # 在线推荐的成功率以 verify/review/publishable 等验收结果为准。
+        try:
+            final_run = store.get_run(run_id) or {}
+            final_status = final_run.get("status") or ""
+            if final_status in ("done", "failed"):
+                final_verdict = final_run.get("verdict") or {}
+                if "pass" in final_verdict:
+                    quality_ok = bool(final_verdict.get("pass"))
+                elif "publishable" in final_verdict:
+                    quality_ok = bool(final_verdict.get("publishable"))
+                else:
+                    quality_ok = final_status == "done"
+                _record_dispatch_completed(
+                    run_id, task, "passed" if quality_ok else "failed",
+                    verify_pass=final_verdict.get("verify_pass"),
+                    review_pass=final_verdict.get(
+                        "review_pass", final_verdict.get("publishable")))
+                final_agent = (((final_run.get("route_plan") or {})
+                                .get("implement") or {}).get("selected") or "")
+                if final_agent.startswith("builtin:"):
+                    final_agent = "builtin"
+                usage.record_quality_for_run(run_id, quality_ok, agent=final_agent)
+            elif final_status == "cancelled":
+                _record_dispatch_completed(run_id, task, "cancelled")
+        except Exception:
+            pass
         # 任务分支收尾（git_rev 隔离链的第二半）：先只读快照本 run 的全部变更
         # 落 run 记录供人审，再把产物提交到 tutti/<task-id> 并切回原分支。
         # 放 finally：done/failed/cancelled/异常一律保存现场；收尾自身绝不抛错，

package/app/core/portguard.py

@@ -0,0 +1,111 @@
+# -*- coding: utf-8 -*-
+"""启动端口清场（用户拍板 2026-09-21）：端口被占时先清场再启动。
+
+- **自家旧实例**（命令行含本包 main.py 完整路径，或 npm 包内 app/main.py）
+  → 杀树（TerminateProcess 直杀）——升级/重启最常见的占用者就是没退干净
+  的 CodeBee 自己；控制台进程对温和信号（WM_CLOSE）无反应，温和关不掉
+  正是用户「旧进程太难杀」的痛点，所以自家实例直接强杀。
+- **别人的进程** → 只报告占用者，不发送任何信号；启动流程不得替用户
+  关闭可能正在开发中的服务。
+- 系统/自身进程拒关（portscan 内置护栏）。
+"""
+from __future__ import annotations
+
+import logging
+import os
+import shlex
+import subprocess
+
+from . import portscan, runner
+
+log = logging.getLogger(__name__)
+
+PS_EXE = "C:\\Windows\\System32\\WindowsPowerShell\\v1.0\\powershell.exe"
+
+
+def proc_cmdline(pid):
+    """读进程命令行；失败返回空串。仅用于确认旧 CodeBee 实例身份。"""
+    try:
+        pid = int(pid)
+    except (TypeError, ValueError):
+        return ""
+    if pid <= 0:
+        return ""
+    if os.name != "nt":
+        try:
+            with open("/proc/%d/cmdline" % pid, "rb") as fh:
+                return fh.read().decode("utf-8", "replace").replace("\x00", " ").strip()
+        except Exception:
+            return ""
+    try:
+        r = subprocess.run(
+            [PS_EXE, "-NoProfile", "-Command",
+             "(Get-CimInstance Win32_Process -Filter 'ProcessId = %d').CommandLine" % pid],
+            capture_output=True, timeout=10)
+        return r.stdout.decode("utf-8", "replace").strip() if r.returncode == 0 else ""
+    except Exception:
+        return ""
+
+
+def _command_tokens(cmdline):
+    try:
+        return [token.strip().strip('"').strip("'")
+                for token in shlex.split(str(cmdline or ""), posix=False)]
+    except (TypeError, ValueError):
+        return []
+
+
+def is_own_instance(cmdline, main_script, port=None):
+    """按启动器、脚本独立参数和端口精确确认 CodeBee 实例。"""
+    tokens = _command_tokens(cmdline)
+    if len(tokens) < 2:
+        return False
+    launcher = os.path.splitext(os.path.basename(tokens[0]))[0].lower()
+    if launcher not in ("python", "python3", "py"):
+        return False
+    script = ""
+    for token in tokens[1:]:
+        if token.startswith("-"):
+            continue
+        script = token
+        break
+    if not script or not os.path.isabs(script):
+        return False
+    actual = os.path.normcase(os.path.normpath(os.path.abspath(script)))
+    expected = os.path.normcase(os.path.normpath(os.path.abspath(str(main_script))))
+    packaged = actual.replace("\\", "/").lower().endswith(
+        "/node_modules/codebee/app/main.py")
+    if actual != expected and not packaged:
+        return False
+    if port is None:
+        return True
+    try:
+        wanted = str(int(port))
+    except (TypeError, ValueError):
+        return False
+    explicit_port = None
+    for i, token in enumerate(tokens):
+        if token == "--port" and i + 1 < len(tokens):
+            explicit_port = tokens[i + 1]
+        elif token.startswith("--port="):
+            explicit_port = token.split("=", 1)[1]
+    # argparse 默认端口可不出现在旧实例参数中；非默认端口必须显式相符。
+    return ((explicit_port == wanted) if explicit_port is not None
+            else wanted == "8765")
+
+
+def clear_stale_port(port, main_script):
+    """清掉占用端口的进程。返回 (是否清掉, 人话说明)。"""
+    holders = [p for p in portscan.listening_ports() if p.get("port") == int(port)]
+    if not holders:
+        return True, ""
+    pid = int(holders[0].get("pid") or 0)
+    if pid <= 4 or pid == os.getpid():
+        return False, "占用者是系统进程/自身，拒绝清理"
+    if is_own_instance(proc_cmdline(pid), main_script, port=port):
+        try:
+            runner._kill_tree(pid)
+            return True, "已结束旧实例 PID %d" % pid
+        except Exception as e:
+            return False, "旧实例 PID %d 清理失败 %s" % (pid, str(e)[:80])
+    return False, "占用者非 CodeBee（PID %d），未自动关闭" % pid