feat: 文档/图片工具进任务镜像 + 结果截断从截断 JSON 改为提取 text 上限 20000 字
This commit is contained in:
1 parent
9ac0199eba
commit
c1d25d11b1
4 files changed
+93
-13
No files matched your search
+14
-2
@@ -24,9 +24,21 @@
|
||||
# workspace-write 的控制面约束,见 cloud/README.md。
|
||||
ARG PYTHON_VERSION=3.12-slim
|
||||
FROM python:${PYTHON_VERSION}
|
||||
# 文档 / 图片工具(document-tools、image-workflow 技能的核查项):
|
||||
# - pandoc:通用文档转换;poppler-utils:pdftotext(PDF 读回)。
|
||||
# - latexmk + texlive-latex-base:pdflatex 编译(英文 LaTeX;
|
||||
# 中文 LaTeX 仍需 xelatex + CJK 字体,见下,本镜像故意不装:
|
||||
# texlive-xetex + fonts-noto-cjk 会再增数百 MB,主机盘已 89%,
|
||||
# 中文 LaTeX 缺 xelatex 时技能按规则明说,不伪造)。
|
||||
# - pip:python-docx / openpyxl / python-pptx / PyMuPDF(fitz)/
|
||||
# Pillow——技能核查的 Python 库与图片处理(Pillow 有其一即可
|
||||
# 做裁剪/缩放/转格式)。图片“生成”仍未配置:本镜像不装、
|
||||
# 不接任何生成端点,生成需求技能仍按规则明说无法生成。
|
||||
# - --no-cache-dir:pip 不留缓存层,镜像不虚胖。
|
||||
RUN apt-get update \
|
||||
&& apt-get install -y --no-install-recommends git \
|
||||
&& rm -rf /var/lib/apt/lists/*
|
||||
&& apt-get install -y --no-install-recommends git pandoc poppler-utils latexmk texlive-latex-base \
|
||||
&& rm -rf /var/lib/apt/lists/* \
|
||||
&& pip install --no-cache-dir python-docx openpyxl python-pptx PyMuPDF Pillow
|
||||
ARG CODEX_BIN_DIR=bin
|
||||
COPY ${CODEX_BIN_DIR}/codex /usr/local/bin/codex
|
||||
COPY ${CODEX_BIN_DIR}/codex-code-mode-host /usr/local/bin/codex-code-mode-host
|
||||
|
||||
+38
-5
@@ -30,6 +30,15 @@ from cloud.marker_tools import extract_wait_threads as _extract_wait_threads
|
||||
from cloud.peer_messages import extract_peer_messages as _extract_peer_messages
|
||||
|
||||
MAX_STDOUT_CHARS = 2000
|
||||
# 任务结果文本上限:2000 字截断的是 host.py 打印的整段 JSON,
|
||||
# text 超过 2000 字时 JSON 本身被截断,worker 再解析 stdout
|
||||
# 取 text 必然失败,存库的就变成截断的 JSON(2026-10-30
|
||||
# 生产审计任务 dbe0c39f… 正是这样在 "PI" 处断掉)。修法:
|
||||
# 解析出的 text(或非 JSON 的原文本)单独截到 20000 字,
|
||||
# 超限时带截断标记,绝不把截断的 JSON 当结果存库。
|
||||
# 状态 / usage / 标记块仍从完整 stdout 解析,不受此截断影响。
|
||||
MAX_RESULT_CHARS = 20000
|
||||
_RESULT_TRUNCATED_MARKER = "\n\n[结果已截断:原文超 20000 字,仅显示前 20000 字]"
|
||||
|
||||
# 启动前分段日志(2026-10-29 任务 A):docker run 是阻塞调用,
|
||||
# 容器输出要等返回才回传,中途卡死时 worker 侧没有任何标记;
|
||||
@@ -223,6 +232,21 @@ def _truncate_stdout(text):
|
||||
return str(text)[:MAX_STDOUT_CHARS]
|
||||
|
||||
|
||||
def _result_text(parsed, stdout_full):
|
||||
"""执行结果存库用的文本:优先取 host.py JSON 的 text 字段,
|
||||
取不到时用原文本;截到 MAX_RESULT_CHARS,超限带截断标记。
|
||||
调用方必须在解析出 parsed 后再调本函数,绝不能先截断
|
||||
stdout 再解析 JSON(截断的 JSON 解析失败,text 全丢)。
|
||||
"""
|
||||
if isinstance(parsed, dict) and parsed.get("text"):
|
||||
text = str(parsed["text"])
|
||||
else:
|
||||
text = str(stdout_full or "")
|
||||
if len(text) > MAX_RESULT_CHARS:
|
||||
return text[:MAX_RESULT_CHARS] + _RESULT_TRUNCATED_MARKER
|
||||
return text
|
||||
|
||||
|
||||
def _default_runner(cmd, env=None):
|
||||
"""subprocess.run 的薄封装:捕获输出、文本模式。
|
||||
|
||||
@@ -353,7 +377,6 @@ class LocalExecutor(Executor):
|
||||
"""
|
||||
returncode = getattr(result, "returncode", 1)
|
||||
stdout_full = getattr(result, "stdout", "") or ""
|
||||
stdout = _truncate_stdout(stdout_full)
|
||||
status = "failed"
|
||||
usage = None
|
||||
task_messages = []
|
||||
@@ -364,12 +387,14 @@ class LocalExecutor(Executor):
|
||||
task_retries = []
|
||||
turn_id = None
|
||||
root_turn_id = None
|
||||
parsed = None
|
||||
if returncode == 0:
|
||||
status = "completed"
|
||||
# JSON 解析必须用完整 stdout:截头后的前 2000 字在
|
||||
# text 很长时根本不是合法 JSON,旧代码会把 host 的
|
||||
# failed 误判成 completed;usage 在 JSON 末尾,截头
|
||||
# 后同样丢。存库的 stdout 仍用截头版,不变。
|
||||
# 后同样丢。存库的 stdout 是解析出的 text(见
|
||||
# _result_text),不再是截断的 JSON。
|
||||
try:
|
||||
parsed = json.loads(stdout_full)
|
||||
except (TypeError, ValueError):
|
||||
@@ -404,6 +429,7 @@ class LocalExecutor(Executor):
|
||||
host_status = None
|
||||
else:
|
||||
host_status = None
|
||||
stdout = _result_text(parsed, stdout_full)
|
||||
return {
|
||||
"executor": "local",
|
||||
"task_id": task_id,
|
||||
@@ -796,7 +822,12 @@ class LocalExecutor(Executor):
|
||||
raise ValueError("任务缺少 id")
|
||||
result = self.runner(self._command(task_id))
|
||||
returncode = getattr(result, "returncode", 1)
|
||||
stdout = _truncate_stdout(getattr(result, "stdout", ""))
|
||||
stdout_full = getattr(result, "stdout", "") or ""
|
||||
try:
|
||||
parsed = json.loads(stdout_full)
|
||||
except (TypeError, ValueError):
|
||||
parsed = None
|
||||
stdout = _result_text(parsed, stdout_full)
|
||||
return {
|
||||
"executor": "local",
|
||||
"task_id": task_id,
|
||||
@@ -1758,7 +1789,6 @@ def _parse_host_result(task_id, fake):
|
||||
"""
|
||||
returncode = getattr(fake, "returncode", 1)
|
||||
stdout_full = getattr(fake, "stdout", "") or ""
|
||||
stdout = _truncate_stdout(stdout_full)
|
||||
status = "failed"
|
||||
host_status = None
|
||||
usage = None
|
||||
@@ -1770,12 +1800,14 @@ def _parse_host_result(task_id, fake):
|
||||
task_retries = []
|
||||
turn_id = None
|
||||
root_turn_id = None
|
||||
parsed = None
|
||||
if returncode == 0:
|
||||
status = "completed"
|
||||
# 与 _result_from_runner 同理:JSON 解析用完整 stdout,
|
||||
# 否则长 text 会把 host 的 failed 误判成 completed,
|
||||
# 也会丢掉 JSON 末尾的 usage。标记块同样从完整解析
|
||||
# 出的 text 里提取。
|
||||
# 出的 text 里提取。存库的 stdout 是解析出的 text,
|
||||
# 不再是截断的 JSON(见 _result_text)。
|
||||
try:
|
||||
parsed = json.loads(stdout_full)
|
||||
except (TypeError, ValueError):
|
||||
@@ -1797,6 +1829,7 @@ def _parse_host_result(task_id, fake):
|
||||
# 时保持 None,worker 写回时不覆盖旧值。
|
||||
turn_id = parsed.get("turn_id")
|
||||
root_turn_id = parsed.get("root_turn_id")
|
||||
stdout = _result_text(parsed, stdout_full)
|
||||
return {
|
||||
"executor": "docker",
|
||||
"task_id": task_id,
|
||||
|
||||
@@ -20,7 +20,8 @@ sys.path.insert(0, os.path.dirname(os.path.dirname(
|
||||
|
||||
from queue import TaskQueue, budget_exceeded, order_tasks, ready_tasks
|
||||
from sandbox import (CloudExecutor, DockerExecutor, Executor,
|
||||
LocalExecutor, _parse_goal_output)
|
||||
LocalExecutor, _parse_goal_output,
|
||||
_parse_host_result, _result_text)
|
||||
|
||||
|
||||
def _task(task_id, depends_on=None, priority=0, status="queued"):
|
||||
@@ -177,12 +178,38 @@ class TestExecutors(unittest.TestCase):
|
||||
self.assertEqual(result["stdout"], "boom")
|
||||
|
||||
def test_local_executor_truncates_stdout(self):
|
||||
# 20000 字上限:5000 字不再截断,25000 字截断并带标记。
|
||||
def fake_runner(cmd):
|
||||
return SimpleNamespace(returncode=0, stdout="x" * 5000)
|
||||
|
||||
result = LocalExecutor("/fake/host.py",
|
||||
runner=fake_runner).execute({"id": "t3"})
|
||||
self.assertEqual(len(result["stdout"]), 2000)
|
||||
self.assertEqual(len(result["stdout"]), 5000)
|
||||
|
||||
def long_runner(cmd):
|
||||
return SimpleNamespace(returncode=0, stdout="y" * 25000)
|
||||
|
||||
result = LocalExecutor("/fake/host.py",
|
||||
runner=long_runner).execute({"id": "t4"})
|
||||
self.assertIn("结果已截断", result["stdout"])
|
||||
self.assertTrue(result["stdout"].startswith("y" * 20000))
|
||||
|
||||
def test_result_text_extracts_long_json_text(self):
|
||||
# 2026-10-30 生产审计任务的复发:host.py JSON 的 text
|
||||
# 超过旧 2000 字截断时,存库的不能是截断的 JSON,
|
||||
# 必须是完整提取出的 text(本例 5000 字,全部保留)。
|
||||
text = "z" * 5000
|
||||
payload = json.dumps({"status": "completed", "text": text})
|
||||
result = _parse_host_result(
|
||||
"t5", SimpleNamespace(returncode=0, stdout=payload,
|
||||
stderr=""))
|
||||
self.assertEqual(result["stdout"], text)
|
||||
self.assertEqual(result["status"], "completed")
|
||||
|
||||
def test_result_text_marks_over_limit(self):
|
||||
text = "w" * 21000
|
||||
self.assertIn("结果已截断", _result_text(
|
||||
{"text": text}, "unused"))
|
||||
|
||||
def test_docker_executor_no_docker_raises_not_local(self):
|
||||
with mock.patch("sandbox.shutil.which", return_value=None):
|
||||
|
||||
+12
-4
@@ -714,7 +714,8 @@ class Worker:
|
||||
"""pick_next 的输入:只取 queued/completed/merged 三类任务。
|
||||
|
||||
旧实现是 SELECT * 全表:prompt/result_text/error_text
|
||||
(后两者各可达 4000 字)全量进内存,且每轮 poll 都来一次;
|
||||
(result_text 可达 21000 字、error_text 可达 4000 字)
|
||||
全量进内存,且每轮 poll 都来一次;
|
||||
任务表随时间变大后这是纯浪费。pick_next 只需要 queued
|
||||
任务和已完成 id(依赖判断),running/failed/interrupted
|
||||
的任务对它既不是就绪、也不是已完成,排除它们不影响结果。
|
||||
@@ -1467,7 +1468,7 @@ class Worker:
|
||||
" turn_id = COALESCE(?, turn_id),"
|
||||
" root_turn_id = COALESCE(?, root_turn_id),"
|
||||
" updated_at = ? WHERE id = ?" + status_clause,
|
||||
((result_text or "")[:4000] if result_text else None,
|
||||
((result_text or "")[:21000] if result_text else None,
|
||||
(error_text or "")[:4000] if error_text else None,
|
||||
actual_model, started_at, finished_at,
|
||||
turn_id, root_turn_id,
|
||||
@@ -1481,7 +1482,7 @@ class Worker:
|
||||
" turn_id = COALESCE(?, turn_id),"
|
||||
" root_turn_id = COALESCE(?, root_turn_id),"
|
||||
" updated_at = ? WHERE id = ?" + status_clause,
|
||||
((result_text or "")[:4000] if result_text else None,
|
||||
((result_text or "")[:21000] if result_text else None,
|
||||
(error_text or "")[:4000] if error_text else None,
|
||||
actual_model, started_at, finished_at, child_pid,
|
||||
turn_id, root_turn_id,
|
||||
@@ -3893,7 +3894,8 @@ class Worker:
|
||||
|
||||
超时用统一的超时文本;runner 异常用异常文本;同步路径的
|
||||
stdout 是 host.py 的 JSON 时取它的 text 字段(比整段 JSON
|
||||
可读);最后退到 host_status token 或通用文本。绝不返回
|
||||
可读),执行器已提取出 text 时直接用该文本;最后退到
|
||||
host_status token 或通用文本。绝不返回
|
||||
任何密钥(stdout 里本就不应含密钥,text 字段同样不含)。
|
||||
"""
|
||||
if result is not None and result.get("status") == "timeout":
|
||||
@@ -3922,6 +3924,12 @@ class Worker:
|
||||
stderr = (result or {}).get("stderr", "") or ""
|
||||
if not stdout.strip() and stderr.strip():
|
||||
return stderr.strip()[-1000:]
|
||||
# 执行器现在返回的是解析出的 text(不再是整段 JSON):
|
||||
# JSON 解析失败且 stdout 非空时,stdout 本身就是报错
|
||||
# 文本,绝不能丢掉它退到 host_status=404 这种代号
|
||||
# (2026-10-30 全量因此丢了 "primary 404 boom")。
|
||||
if stdout.strip():
|
||||
return stdout.strip()[:1000]
|
||||
if token is not None:
|
||||
return "host_status=%s" % token
|
||||
return stdout.strip() or "host execution failed"
|
||||
|
||||
Reference in new issue
Block a user