v1.5.0 - Salience Enforcement & Official Contract Alignment / 注意力加固与官方契约对齐

English Release Notes

🚀 Highlights & Features

• RFC 2119 Restrictive Modal Capitalization & Attention Salience:
  • Upgraded periodic convergence reminder to NEVER omit verification. to strictly counter premature conclusion under step budget pressure.
  • Standardized all-caps formatting for hard boundaries (ONLY, NEVER, STRICTLY) across convergence guidance, default templates, and runtime intercept reasons.
• Precision Tool Alignment with Official DSH Contracts:
  • Synchronized edit_batch schema and parameter descriptions verbatim with the official edit tool contract (Provide file_path before old_string and new_string in the arguments).
  • Decoupled tool homonyms in guidance (produce deliverables instead of write deliverables) to eliminate semantic collision with native write.
• Refined Command Chaining Policy:
  • Fine-tuned CHAINED_NETWORK_COMMAND runtime defense to unblock non-blocking compound workflows (such as git push in chained scripts) while continuing to strictly guard against Windows handle hangs on credential-bound commands (npm publish, gh PR creation).
• 27/27 Tests Passing: Verified against official @deepseek-ai/dsh-agent-loop with zero regressions.

---

中文发布说明

🚀 核心亮点与特性

• RFC 2119 强约束语气与注意力锚定加固：
  • 收敛软提醒全面升级为 NEVER omit verification.，杜绝步数压力下的偷懒与提前交卷。
  • 在常驻收敛规范、推荐模板及拦截提示中全局统一大写强约束词（ONLY、NEVER、STRICTLY），强化大模型注意力边界。
• 工具定义与官方 DSH 契约像素级对齐：
  • edit_batch 参数描述精准对齐官方 edit 工具约定（显式要求提供 file_path 优先顺序）。
  • 彻底消除动词混淆（统一为 produce deliverables），确保模型复用对原生工具的先验理解。
• 细粒度复合网络命令拦截放行：
  • 精准优化 CHAINED_NETWORK_COMMAND 防护规则，在彻底杜绝 npm publish 等 Windows 凭据与发布操作悬挂的同时，完全放行本地与协同的复合 git push 流水线。
• 27 项测试全绿：完全基于官方 @deepseek-ai/dsh-agent-loop 跑通全量回归测试。

v1.1.0 - Standalone Refined System Prompt & 80% Token Reduction / 纯粹独立系统提示词与 80% Token 极致瘦身

English Release Notes

🚀 Highlights & Features

• Pure Standalone Mode (placement: 'replace-all'):
  • Completely stripped the legacy, ineffective replace-persona-suffix mode.
  • Standardized on replace-all as the default and primary placement mechanism.
  • Flushes all verbose official DSH system sections (deliverable:file-references, web:surface, duplicate tool instructions, etc.) while preserving optimizer-native enhancements.
  • Slashes system prompt overhead from ~2.5K tokens to ~300-400 tokens (~80% token reduction).
• Ultra-Lean Core Directives:
  • Global ~/.dsh/SYSTEM.md reduced to 21 lines / 1963 characters of pure, high-density engineering principles.
  • Retains 100% of the critical operational tenets: action over speculation, zero pre-action narration/filler, dual-batching (read_batch / edit_batch), prohibited shell file ops, AST/LSP intelligence, two-failure pivoting, and high-density reporting.
• 21/21 Tests Passing: Verified against official @deepseek-ai/dsh-agent-loop with zero regressions.

---

中文发布说明

🚀 核心亮点与特性

• 纯粹独立精简模式 (placement: 'replace-all')：
  • 彻底剔除了替换不彻底且存在冗余的过渡模式 replace-persona-suffix。
  • 全面标准化以 replace-all 为核心与默认置换模式。
  • 彻底清空并剥离官方冗长庞大的内置分段（deliverable:file-references、web:surface、重复的各工具长篇说明等），系统提示词开销从 ~2.5K 骤降至 ~300-400 tokens（直降 ~80% 系统开销）。
• 极致紧凑的高密度工程师心智模型：
  • 全局 ~/.dsh/SYSTEM.md 精炼至 21 行、1963 字符，彻底剔除叙事赘语。
  • 100% 捍卫行动优先、零预告废话、双批处理工具优先（read_batch / edit_batch）、严禁 Shell 管道读写文件与大内联脚本、两步止损与精准验证等黄金准则。
• 21 项测试全绿：完全基于官方 @deepseek-ai/dsh-agent-loop 跑通全量回归测试。

v1.0.0 - Reactive File-Driven System Prompt Engine & Stable Architecture / 响应式系统提示词引擎与稳定架构

English

Milestone v1.0.0 - Production-Ready Reactive Architecture

This milestone release establishes the stable v1.0.0 architecture of dsh-awesome-optimizer, introducing a file-driven reactive system prompt customization engine alongside the dual-batching and convergence foundation.

1. Reactive File-Driven System Prompt Engine (src/custom-prompt.js)

• Dual-Layer SYSTEM.md Convention: Automatically discovers global ~/.dsh/SYSTEM.md (or $DSH_HOME/SYSTEM.md) and workspace-level .dsh/SYSTEM.md (or SYSTEM.md), cleanly separating developer system defaults from repo-specific instructions and distinct from role-oriented AGENTS.md.
• Millisecond Hot-Reload: Powered by debounced fs.watch watchers with atomic rename recognition; editor saves take effect immediately on subsequent turns without restarting the DSH server.
• In-Memory Zero-IO Cache: Prompt rendering is instantaneous from memory without adding disk I/O overhead to model turns.
• Cascading & Placement Modes:
  • mergeStrategy: Supports concat (cascading global + workspace) and workspace-override (project-level completely overrides global).
  • placement: Supports suffix (order 10300, highest constraint authority at prompt tail), replace-persona-suffix (replaces official persona suffix), and prefix.

2. Fully Verified 21-Test Quality Suite

• Complete end-to-end integration coverage across runtime policy, composite batching (read_batch, edit_batch), soft convergence, large-write recovery, hot-reloading prompt engine, and A/B telemetry. All 21 tests passing.

---

中文

里程碑 v1.0.0 - 生产级文件响应式架构

本里程碑版本标志着 dsh-awesome-optimizer 正式步入 v1.0.0 稳定生产级架构，在双向批处理（read_batch + edit_batch）与软收敛的基础上，全面融入极客级文件驱动的响应式系统提示词定制引擎。

1. 文件驱动响应式系统提示词定制 (src/custom-prompt.js)

• 统一的双层 SYSTEM.md 规约：自动寻找全局配置 ~/.dsh/SYSTEM.md（或 $DSH_HOME/SYSTEM.md）与项目级工作区配置 .dsh/SYSTEM.md（或 SYSTEM.md），与面向角色的 AGENTS.md 权责清晰分离。
• 毫秒级热加载（Hot-Reload）：基于 fs.watch + 50ms 窗口防抖重载，支持原子重命名感知；直接修改 Markdown 文件保存即刻生效，无需重启 DSH 服务。
• 内存零 I/O 缓存：运行期直接通过内存缓存渲染提示词，不增加每次模型轮次的磁盘 I/O 延迟。
• 智能级联与挂载：
  • mergeStrategy：支持 concat（全局+工作区级联合并）与 workspace-override（项目级完全覆盖全局）等策略；
  • placement：支持 suffix（默认尾部分段，order 10300，具备最高约束力）、replace-persona-suffix（替换官方 Persona 后缀）与 prefix 模式。

2. 完备的 21 项测试工程验证

• 全面覆盖运行期硬防护、对偶复合批处理、微指令软收敛、大文件写入恢复与响应式提示词热重载，21 个测试全部绿灯通过。

v0.9.0 - Dual Composite Batching with edit_batch

English

What's New in v0.9.0

• edit_batch Composite Batch Tool:
  • Introduced edit_batch alongside read_batch to collapse multi-file or multi-location literal text replacements into a single model round-trip.
  • Sub-dispatches sequentially via official ToolExecutionPipeline (PTC), strictly preserving sandbox modes, file mutation policies, and observation tracking.
  • Completely eliminates fragile Unified Diff line-number calculation errors and indentation mismatches typical of apply_patch.
  • Supports partial success with structured per-edit status reporting without losing valid changes.
• Config & Guidance Updates:
  • Added compositeTools.editBatch (default: true) and compositeTools.maxBatchEdits (default: 10).
  • Updated convergence guidance to steer models toward edit_batch when making multi-file modifications.
• Verification:
  • Expanded test suite to 17 automated tests, all passing.

---

简体中文

v0.9.0 更新日志

• 新增 edit_batch 复合批量修改工具：
  • 与 read_batch 形成对称闭环，支持单次调用顺序执行最多 10 处跨单文件或多文件的字面量文本替换，将多文件重构轮次压缩至 1 轮。
  • 通过官方 ToolExecutionPipeline（PTC）顺序分发每一个子操作，完整继承沙箱权限、文件观察策略与审查记录。
  • 彻底规避官方 apply_patch 因 Unified Diff 行号少算/换行符微差导致的拒收错误，成功率与稳定性大幅提升。
  • 支持局部错误容忍与结构化状态返回，单点失误不破坏同批次其他有效修改。
• 配置与收敛指导升级：
  • 新增 compositeTools.editBatch: true 与 compositeTools.maxBatchEdits: 10 配置项。
  • 收敛提示词自动引导模型在多文件/多位置修改时优先采用 edit_batch。
• 测试用例扩充：
  • 自动化测试用例扩充至 17 项，全部通过。

v0.8.0 - Interval Convergence Guidance, Error Suppression & Backend Stat Probing

English

What's New in v0.8.0

• Interval Convergence Guidance:
  • Upgraded soft step budget to periodic interval reminders (stepInterval: 8, default cap maxRemindersPerTurn: 4).
  • Switched reminder payload to a concise, high-signal micro-directive:

    Accelerate execution: batch remaining actions and conclude if criteria are met; do not omit verification.

  • Saves ~40% token overhead per turn while maintaining steady long-range convergence control.
• Dynamic Tool-Error Suppression:
  • Automatically suspends reminder injection when a preceding tool execution encountered an error (result.isError or result.error).
  • Allows model full uninterrupted latitude to diagnose and resolve errors without prompt pressure.
• Virtual Backend Probe via ctx.fs:
  • fileMutationPolicy checks existence via fsService.resolve and fsService.stat before falling back to host filesystem, preventing false overwrites across virtual/sandboxed environments.
• Windows Path Case-Folding Resiliency:
  • Normalizes path casing on Windows across invocation keys and failure fingerprints.
• Multi-Session Telemetry Stream Attribution:
  • Correctly routes interleaved model streams to their individual session IDs.

---

简体中文

v0.8.0 更新日志

• 周期性收敛软提醒升级：
  • 将单次软预算升级为周期性步数微提醒（默认 stepInterval: 8 步，每轮上限 maxRemindersPerTurn: 4 次）。
  • 全新精炼微指令型提示词（单次仅 ~15 Token，节约 40% Token 消耗，提供长程收敛牵引力）：

    Accelerate execution: batch remaining actions and conclude if criteria are met; do not omit verification.

• 错误感知与智能避让机制：
  • 当模型上一步工具调用发生错误（result.isError 或抛错）时，自动静默避让提醒注入，优先保证排错排障上下文连贯，避免催促干扰。
• 支持后端沙箱 ctx.fs 探测：
  • 文件策略优先通过 ctx.fs.resolve + ctx.fs.stat 进行后端虚拟文件探测，有效避免沙箱环境下的误判与强制覆盖。
• Windows 大小写折叠容错增强：
  • 针对 Windows 文件路径不区分大小写特性进行规范化处理，防止相同报错工具调用漏计。
• 多会话并发遥测流隔离：
  • 完善 telemetry 在多会话并发流交替输出时的隔离归属，会话销毁时安全释放追踪状态。

v0.6.0

What's Changed in v0.6.0

🚀 Enhancements & Bug Fixes

• Multi-Session Telemetry Stream Attribution: Accurately attribute LLM stream tokens, latency, and requests to concurrent sessions using request sessionId, and clean up session records on session/disposed.
• Windows Path Casing Normalization: Ensure case-insensitive path normalization and invocation identity hashing on Windows, reliably intercepting repeated identical failures (NO_PROGRESS) regardless of case variations.
• Unified ctx.fs Backend Existence Probe: Refactor fileMutationPolicy to probe file existence through the shared ctx.fs backend (resolve and stat), preventing false negatives on virtual, sandbox, or non-host filesystems even without prior reads in the current turn.

🧪 Testing & Verification

• Added mock ctx.fs test fixture to the test runtime mapping virtualFs.
• Added regression tests for concurrent stream attribution, Windows path case-folding, and blind large write interception over backend files.
• All 13 integration tests passing with zero TypeScript / linter errors.

---

🇨🇳 中文更新日志 (Release Notes in Chinese)

🚀 功能增强与问题修复

• 多会话遥测流精准归属：支持按请求附带的 sessionId 精确归属 LLM 流式 Token、请求数及耗时，彻底解决并发交错流的遥测统计串线问题；在 session/disposed 生命周期事件触发时安全释放映射，防止内存泄漏。
• Windows 路径大小写归一化：在 Windows 环境下对盘符与完整路径进行不区分大小写的标准归一化与调用指纹哈希，可靠拦截因路径大小写差异导致的无效重试死循环（NO_PROGRESS）。
• 统一接入 ctx.fs 后端状态探针：fileMutationPolicy 优先调用底层 ctx.fs.resolve 与 ctx.fs.stat 进行文件存在性探测，彻底消除虚拟文件系统、沙箱环境与宿主物理磁盘的视图不一致；即使模型在本轮中未先执行 read 探测，也能精准拦截对已有文件的大量全写覆盖。

🧪 测试与质量验证

• 在集成测试运行时提供规范的 ctx.fs Fixture，无缝映射虚拟文件系统（virtualFs）。
• 新增针对并发交错流遥测、Windows 路径大小写折叠以及未预读后端文件的全写阻断 3 项回归测试。
• 全套 13 项官方 agent loop 集成测试全部通过，TypeScript 声明严格校验零错误。

v0.5.0

Initial release of dsh-plugin-agent-optimizer: non-invasive runtime policy guards, PTC composite batch tools, soft convergence guidance, and A/B turn telemetry for DeepSeek Harness.