v2.3.4

Patch Generation Robustness Overhaul

This release focuses on hardening the LLM-driven patch generation pipeline — fixing path resolution, diff application, JSON parsing, and observability.

Bug Fixes

• Use git apply --3way for LLM-generated patches (better conflict resolution)
• Fuzzy-match diff file paths when LLM fabricates wrong directories
• Include full file contents in patch generator prompt for better context
• Bug-specific source search instead of reusing diagnose-stage locations
• Replace hardcoded --include extensions with --exclude-dir approach
• Keyword search prioritizes specific identifiers over generic terms
• Handle LLM illegal JSON escapes (`' and bare newlines)
• Patch generation uses repo-relative paths instead of absolute paths
• Fix severity type comparison in _deploy_patch
• Fix for-loop variable _ shadowing i18n.t import (TypeError)
• Restore ANSI colors by using sys.__stdout__
• _display_width strips ANSI escape codes before calculating width
• Treat East Asian Ambiguous width chars as 2 columns in banner
• Exclude .bak and .orig files from grep search
• Fix _sanitize_diff orphaned call (reverted but still referenced)

Features

• Log prompt length and LLM empty response details (finish_reason)
• Log LLM raw response in PatchGenerator before parse
• Log diagnose LLM raw response for JSON parse failure debugging
• INFO-level logging for patch generation and application
• Add patch info to trace output

Docs

• Add Docker container patrol requirements to observe.md

v2.3.3

v2.3.3 Bug Fixes

• fix: patch generation uses repo-relative paths instead of absolute paths — source_locator.render() now outputs repo-relative paths (e.g. backend/... instead of root/project/backend/...), fixing git apply "No such file or directory" errors
• fix: patch applier pre-checks diff file references exist before running git apply
• fix: patch generator prompt explicitly requires repo-relative paths in diff headers
• fix: _display_width strips ANSI escape codes before calculating width — banner border │ no longer sticks to content
• fix: console output uses sys.__stdout__ via _raw_print — ANSI colors restored without prompt_toolkit garbage
• fix: for _ in range() variable no longer shadows i18n t import — TypeError: 'int' object is not callable resolved

v2.3.2

v2.3.2 Bug Fixes

• fix: for-loop variable _ shadows i18n t import, causing TypeError: 'int' object is not callable in incident loop (#635)
• fix: 修复 _deploy_patch 中 severity 类型比较错误，改用 status==NORMAL 判断
• fix: remove raw=True from patch_stdout to prevent ANSI escape garbage output
• docs: add Docker container patrol requirements to observe.md

v2.3.1

v2.3.1

Bug Fixes

• Fix banner hardcoded Chinese not going through i18n t()
• Fix dynamic banner width for zh/en alignment (CJK double-width character support)

Documentation

• Add English translations: README.en.md, USER_GUIDE.en.md
• Add language switcher links on all documentation
• Add SCENARIOS.md + SCENARIOS.en.md — 8 usage scenario guides
• Add --lang usage guide to README and USER_GUIDE

Misc

• Close GitHub Issue #2 (lessons/ and audit/ not generating content)

---

项目代码统计

类别	行数	文件数
Python 源码	14,965	56
Python 测试	6,243	17
Prompt 模板	1,561	25
i18n 字典	866	2
文档	5,448	6
配置/脚本	1,593	—
总计	30,676	—

模块	行数
src/agent	3,982
src/infra	3,294
src/safety	1,411
src/repair	1,583
src/reliability	844
src/i18n	119

v2.3.0 — i18n Support

v2.3.0 — 中英文双语支持

Features

• i18n 基础设施：src/i18n/ 模块，YAML 字典 + prompt 目录双轨架构
• 语言切换：--lang zh|en 参数、OPS_AGENT_LANG 环境变量、配置文件，默认中文
• 9 个 prompt 双语版本：7 个核心 prompt + reporter + agents_md
• 327 个 i18n key：覆盖所有用户可见文本
• UI 文本国际化：human.py / pipeline.py / core.py / chat.py / check.py
• 通知国际化：notifier.py / reporter.py / production_watcher.py
• 边角文件：pr_workflow.py / agents_md.py
• 向后兼容：默认 zh，现有行为不变

Bug Fixes

• 修复变量名遮蔽 i18n t() 函数（loop var t → tgt）
• 修复 pr_body_default 占位符与 _build_pr_body() 不匹配

Usage

# 中文（默认）
python main.py

# 英文
python main.py --lang en

# 环境变量
OPS_AGENT_LANG=en python main.py

v2.2.5

Bug Fixes

• Fix false rollback when grep/find returns exit code 1 — 信息收集类命令（grep/find/test）无匹配时返回退出码 1，此前被误判为步骤失败并触发回滚

Changes

• CommandResult.is_tolerable() — 支持步骤级容忍退出码（tools.py）
• _execute 执行前检查 tolerate_exit_codes 字段（pipeline.py）
• plan prompt 新增 tolerate_exit_codes 字段文档（plan.md）

v2.2.4

v2.2.4

New Features

• LLM Response Validator — 解析前校验 LLM 输出，校验失败写入 notebook/llm_errors/YYYY-MM-DD.jsonl
  • 校验阶段：DIAGNOSE / PLAN / VERIFY (JSON) + ASSESS (键值对)
  • 错误类型：empty_response / json_parse_failed / missing_required_field / invalid_field_value
  • 每条日志含：timestamp, phase, provider, model, error_type, validation checks, request/response snippets, duration
• trace-view analyzer — 支持 JSON 结构化 verify response 解析

Bug Fixes

• verify prompt 误判修复 — LLM 在 evidence 中提及 FAILED 不再导致 SUCCESS 被误判（改为 JSON 结构化输出）

Changes

• max\_plan\_rounds 默认值从 3 调整为 8
• 81 个 LLM validator 测试用例，95% 覆盖率

v2.2.3

Bug Fix

verify prompt 误判修复：LLM 在 evidence 中提到 FAILED 时不再导致 SUCCESS 结果被误判为 FAILED。

Changes

• verify prompt 改为 JSON 结构化输出，与 diagnose/plan 保持一致
• _parse_verify_response 改为 JSON 优先解析，JSON 解析失败时 fallback 到关键词匹配
• 版本号更新至 2.2.3

v2.2.2

What's Changed

🔧 Prompt 优化

• 移除 request 中与 system prompt 重复的字段：system_map、permissions、mode 不再在每个 request prompt 中重复注入，每次 LLM 调用节省 200-500 token

🔍 智能检索重排序（依赖 smart-notebook v2+）

• Hybrid 排序（默认）：关键词命中(0.5) + 图距离(0.3) + 热点惩罚(0.2)
• LLM 排序：可选，LLM 对候选结果按相关性排序，异常自动回退 hybrid
• 配置项位于 notebook/config/notebook.yaml → smart.knowledge.ranking

🐛 Bug 修复

• 修复 system prompt 构建时 UnboundLocalError（变量未定义）
• 修复 reflect 中 playbook 内容被截断（改用 code fence 包裹）
• 修复 UTF-8 解码错误导致内部错误（添加 errors="replace"）
• 修复 incident 关闭时跳过 reflect
• 修复 plan READY 但无有效 steps 时未降级为 COLLECT_MORE
• 修复 COLLECT_MORE 超轮次限制时展示空方案

✨ 新功能

• 新增 list trace / lt 命令查看 trace 文件列表
• 新增 trace-view 子命令分析 trace 日志
• max_plan_rounds 可通过 LimitsConfig 配置

Full Changelog: v2.2.1...v2.2.2

v2.2.1 — 修复 UnboundLocalError

修复

• 修复 _ask_llm 中 system prompt trace 引用未赋值变量导致 UnboundLocalError 的问题
• 新增 list trace (lt) 命令，列出 notebook/trace 下的 trace 文件

v2.2.0 新功能回顾

• ops-agent trace-view 子命令 — trace 日志分析器
• Context Budget Breakdown + 截断检测
• 模板感知 section 切分
• 交互式浏览模式