@mengyuly/dsh-ponytail 0.3.3 → 0.3.4

package/CHANGELOG.md

@@ -5,6 +5,19 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## Unreleased
 
+## [0.3.4] - 2026-10-07
+
+### Added
+
+- English README and bilingual navigation (PR #1), included in the npm package.
+
+### Fixed
+
+- Mode/default status notices now use non-waking injection instead of steering
+  an idle agent into an extra model turn (issue #2). Mode changes apply
+  immediately; notices reach the model on its next real step. Explicit
+  review/audit/debt/gain/help invocations still request an ordinary turn.
+
 ## [0.3.3] - 2026-09-09
 
 ### Changed

package/README.md

@@ -1,5 +1,7 @@
 # dsh-ponytail
 
+[English](README_EN.md) | 简体中文
+
 ![CI](https://github.com/MengYuil/dsh-ponytail/actions/workflows/ci.yml/badge.svg)
 ![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)
 [![npm](https://img.shields.io/npm/v/@mengyuly/dsh-ponytail)](https://www.npmjs.com/package/@mengyuly/dsh-ponytail)
@@ -13,15 +15,15 @@
 
 - **最新版（latest）**：
   `https://github.com/MengYuil/dsh-ponytail/releases/latest/download/mengyuly-dsh-ponytail.tgz`
-- **固定版本 v0.3.2**（按 Tag 不可变）：
-  `https://github.com/MengYuil/dsh-ponytail/releases/download/v0.3.2/mengyuly-dsh-ponytail.tgz`
+- **固定版本 v0.3.4**（按 Tag 不可变）：
+  `https://github.com/MengYuil/dsh-ponytail/releases/download/v0.3.4/mengyuly-dsh-ponytail.tgz`
 
 说明：
 
 - **latest**：适合快速安装体验，会随最新 Release 更新；固定资产名
   `mengyuly-dsh-ponytail.tgz` 在每个 Release 中保持不变，因此该 URL
   不会因版本号变化而失效，**不适合作为不可变依赖**。
-- **固定版本**：适合可复现安装，URL 中固定 Tag（如 `v0.3.2`），按 Tag
+- **固定版本**：适合可复现安装，URL 中固定 Tag（如 `v0.3.4`），按 Tag
   不可变；资产名同样为 `mengyuly-dsh-ponytail.tgz`。
 - npm 安装仍走 npm Registry 或 `dsh plugin` 命令。
 - 固定资产名由 `scripts/release-assets.mjs` 生成并验证（`node scripts/release-assets.mjs`，仅仓库维护者）。
@@ -54,25 +56,26 @@ dsh plugin --profile web add @mengyuly/dsh-ponytail
 ## 功能
 
 - **核心模式** `/ponytail` — 每轮注入结构化的懒惰开发者规则集，**三个档位是真实不同的 Prompt 片段**（不只是换一行）：
-  - **Common（所有非 off 档共享）**：先把请求转成可观察的完成条件；沿真实调用流取证，但决策阶梯是快速反射、不是研究项目；按“检查 → 修改 → 最窄有效验证 → 检查最终 diff”闭环执行；非平凡逻辑只留一个最小可执行检查，不额外搭测试框架或 fixtures；只汇报已验证结果。
+  - **Common（所有非 off 档共享）**：先把请求转成可观察的完成条件；沿真实调用流取证，但决策阶梯是快速反射、不是研究项目；按“检查 → 修改 → 最窄有效验证 → 检查最终 diff”闭环执行；非平凡逻辑只留一个最小可执行检查，不额外搭测试框架或 fixtures；只汇报已验证结果。
   - **Safety（任何档位都不可删）**：输入校验、防数据丢失的错误处理、安全措施、无障碍、明确验收项、先理解问题、「最小 diff ≠ 正确修复」。
-  - **`lite`**：直接执行、减少仪式；完整交付明确要求；可以一句话指出更简方案，但**不挑战明确需求**。
+  - **`lite`**：直接执行、减少仪式；完整交付明确要求；可以一句话指出更简方案，但**不挑战明确需求**。
   - **`full`（默认）**：完整七级阶梯（YAGNI → 复用 → 标准库 → 原生 → 已装依赖 → 一行 → 最小实现），默认选最短正确实现，修根因而非症状。
-  - **`ultra`**：新增代码前先要证据，优先删除或复用；主动质疑投机性功能/缓存/抽象/配置/新依赖；复杂需求先给最小正确版并说明扩大条件；**不是无脑拒绝**。
+  - **`ultra`**：新增代码前先要证据，优先删除或复用；主动质疑投机性功能/缓存/抽象/配置/新依赖；复杂需求先给最小正确版并说明扩大条件；**不是无脑拒绝**。
   - `off`：完全不注入。
   - 档位**会话级**（会话 A 不影响会话 B，会话结束自动释放）。
   - 裸 `/ponytail`：已启用时只报告；`off` 时恢复到有效默认档（默认也是 `off` 则回 `full`）。
-  - `/ponytail status`：只查询、永不修改，并显示当前模式来自会话覆盖还是配置默认值。
-  - `/ponytail reset`：清除当前会话覆盖，重新跟随有效配置默认值。
+  - `/ponytail status`：只查询、永不修改，并显示当前模式来自会话覆盖还是配置默认值。
+  - `/ponytail reset`：清除当前会话覆盖，重新跟随有效配置默认值。
   - `/ponytail lite|full|ultra|off`：显式切换。
   - `/ponytail default <mode>`：持久化默认值到**用户级配置文件**（env/Profile 仍优先，命令分别提示 saved 与 effective）。
+  - 状态操作立即生效，但不唤醒空闲 Agent；模型通知在下一次真实请求时接收，避免切档位额外触发模型调用。
 - **一次性技能**（用哪个载哪个，不进常驻 prompt）：
-  - `/ponytail-review` — 针对最近改动找过度工程；每条包含位置、替代方案和实际调用证据，不猜测精确收益。
-  - `/ponytail-audit` — 全仓库过度工程审计；区分可安全删除与需要先验证的候选，最多返回 10 条高价值发现。
+  - `/ponytail-review` — 针对最近改动找过度工程；每条包含位置、替代方案和实际调用证据，不猜测精确收益。
+  - `/ponytail-audit` — 全仓库过度工程审计；区分可安全删除与需要先验证的候选，最多返回 10 条高价值发现。
   - `/ponytail-debt` — 收割所有 `ponytail:` 注释成债务账本。
   - `/ponytail-gain` — 上游 Benchmark 参考计分板（代码减少；Token/成本/延迟效果取决于模型与任务，**非本适配版保证**）。
   - `/ponytail-help` — 参考卡。
-- **停用**：说 `stop ponytail`、`normal mode`、`停止 ponytail`、`关闭 ponytail`、`普通模式` 或 `正常模式`（兼容中英文句末标点）；随时 `/ponytail` 恢复。
+- **停用**：说 `stop ponytail`、`normal mode`、`停止 ponytail`、`关闭 ponytail`、`普通模式` 或 `正常模式`（兼容中英文句末标点）；随时 `/ponytail` 恢复。
 - **默认值优先级**（代码/测试/文档一致）：
   ```
   会话 override > PONYTAIL_DEFAULT_MODE > Profile config.defaultMode > 用户 config.json > full
@@ -104,9 +107,9 @@ prompt 与推理开销变得更贵。
 
 | 档位 | 字符数 | UTF-8 字节 | 说明 |
 |------|--------|-----------|------|
-| lite | 1915 | 1917 | 实测生成 |
-| full | 3022 | 3038 | 实测生成 |
-| ultra | 2797 | 2813 | 实测生成 |
+| lite | 1915 | 1917 | 实测生成 |
+| full | 3022 | 3038 | 实测生成 |
+| ultra | 2797 | 2813 | 实测生成 |
 | off | 0 | 0 | 不注入 |
 
 这些是 **Prompt 体积测量，不是账单金额，也不是对所有模型成立的节省
@@ -162,8 +165,8 @@ Smoke Benchmark 只提供方向性证据（见 `docs/dsh-smoke-summary.md`）。
   node scripts/verify-dist.mjs         # 静态一致性：src/d.ts 导出一致、关键签名、运行时导出、provenance
   node scripts/verify-pack.mjs         # tarball 边界（含无 scripts/ 暴露回归检查）、版本、安装后 smoke
   node scripts/test-consumer.mjs       # NodeNext + skipLibCheck:false 声明消费测试（对打包产物）
-  node scripts/test-regressions.mjs    # 验证工具自身的回归测试
-  node scripts/test-core.mjs           # 核心 Prompt 字节、安全边界、模式与 Skill 表面
+  node scripts/test-regressions.mjs    # 验证工具自身的回归测试
+  node scripts/test-core.mjs           # 核心 Prompt 字节、安全边界、模式与 Skill 表面
   node scripts/measure-prompt.mjs      # 各模式 Prompt 段体积（依赖未发布的 src/）
   node scripts/check-release-links.mjs # README/CHANGELOG/docs 无版本化 latest 资产链接
   node scripts/check-release-consistency.mjs --version <v>  # 四方发布一致性（git tag/npm/GitHub/provenance）
@@ -180,4 +183,4 @@ Smoke Benchmark 只提供方向性证据（见 `docs/dsh-smoke-summary.md`）。
 
 ## 许可
 
-MIT，© 2026 DietrichGebert（上游）+ MengYuil（移植）。详见 [LICENSE](LICENSE)。
+MIT，© 2026 DietrichGebert（上游）+ MengYuil（移植）。详见 [LICENSE](LICENSE)。

package/README_EN.md

@@ -0,0 +1,311 @@
+# dsh-ponytail
+
+English | [简体中文](README.md)
+
+![CI](https://github.com/MengYuil/dsh-ponytail/actions/workflows/ci.yml/badge.svg)
+![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)
+[![npm](https://img.shields.io/npm/v/@mengyuly/dsh-ponytail)](https://www.npmjs.com/package/@mengyuly/dsh-ponytail)
+[![dsh.so security](https://www.dsh.so/badge/dsh-ponytail-4.svg)](https://www.dsh.so/artifact/dsh-ponytail-4/)
+
+Adapts the minimal-coding principles and skills of
+[DietrichGebert/ponytail](https://github.com/DietrichGebert/ponytail) to
+[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness): a YAGNI
+decision ladder, session-level Lite / Full / Ultra / Off modes, and a set of
+skills for code slimming, review, auditing, and technical-debt tracking.
+
+This project aligns with the upstream core ideas and main workflows, but DSH's
+model loop, prompt assembly, skill mechanism, and tool calls differ. **The
+upstream benchmark is a reference only and does not imply that this port has
+the same token, cost, or latency gains** (see “Efficiency (conditional gains,
+not guarantees)”).
+
+## Download from GitHub Releases
+
+- **Latest release**:
+  `https://github.com/MengYuil/dsh-ponytail/releases/latest/download/mengyuly-dsh-ponytail.tgz`
+- **Pinned version v0.3.4** (immutable per tag):
+  `https://github.com/MengYuil/dsh-ponytail/releases/download/v0.3.4/mengyuly-dsh-ponytail.tgz`
+
+Notes:
+
+- **latest**: good for a quick install; it tracks the newest release. The fixed
+  asset name `mengyuly-dsh-ponytail.tgz` stays the same in every release, so
+  the URL never breaks on a version bump — which also means it is **not
+  suitable as an immutable dependency**.
+- **Pinned version**: good for reproducible installs; the URL pins a tag
+  (e.g. `v0.3.4`) that is immutable per tag; the asset name is likewise
+  `mengyuly-dsh-ponytail.tgz`.
+- npm installs still go through the npm Registry or the `dsh plugin` command.
+- The fixed asset name is produced and verified by
+  `scripts/release-assets.mjs` (`node scripts/release-assets.mjs`, maintainers
+  only).
+
+## Installation
+
+Install into a profile (replace `web` with `tui` or a custom name):
+
+```bash
+# Option 1: local link (current dsh cores >= 0.1.x)
+dsh plugin --profile web add link:$(pwd)
+
+# Option 2: install straight from GitHub
+dsh plugin --profile web add github:MengYuil/dsh-ponytail
+
+# Option 3: Release tarball (download the tgz first — the latest fixed asset
+# name never changes)
+#   https://github.com/MengYuil/dsh-ponytail/releases/latest/download/mengyuly-dsh-ponytail.tgz
+dsh plugin --profile web add file:./mengyuly-dsh-ponytail.tgz
+
+# Option 4: npm
+dsh plugin --profile web add @mengyuly/dsh-ponytail
+```
+
+Restart the profile after installing (`dsh web` / `dsh tui`) for it to take
+effect. Once loaded, the session skill catalog shows 6 `ponytail*` skills; send
+`/ponytail-help` to verify immediately.
+
+> `lib/index.js` is a self-contained bundle (`dsh-llm` / `dsh-skill` are
+> already inlined — npm has no compatible versions) with two published
+> runtime peers: `@deepseek-ai/cordis` (4.0.1) and
+> `@deepseek-ai/schemastery` (3.18.x). `schemastery` is deliberately kept
+> external rather than inlined: its schema DSL compiles `callback` strings
+> with `new Function`, so keeping it out means **the shipped artifact contains
+> no dynamic code execution** (CI has a dedicated check). None of the three
+> install paths — GitHub, tgz, npm — needs a dsh source tree.
+
+> Note: `src/` is the source, `lib/` is the prebuilt artifact (loadable
+> out of the box, no build step). The authoritative source lives in the
+> deepseek-harness monorepo at `packages/community/ponytail`; after changing
+> source, rebuild and sync the full `lib/` with
+> `DSH_CHECKOUT=/path/to/deepseek-harness node scripts/sync-dist.mjs`
+> (see “Release maintenance” below).
+
+## Features
+
+- **Core mode** `/ponytail` — injects a structured lazy-developer ruleset
+  every turn; **the three levels are genuinely different prompt fragments**
+  (not just a swapped line):
+  - **Common (shared by every non-off level)**: turn the request into an
+    observable completion condition first; gather evidence along the real call
+    flow, but the decision ladder is a quick reflex, not a research project;
+    follow the “inspect → change → narrowest effective verification → inspect
+    the final diff” loop; non-trivial logic keeps exactly one minimal runnable
+    check, without standing up a test framework or fixtures; report only
+    verified results.
+  - **Safety (never removable in any level)**: input validation, error
+    handling that prevents data loss, security measures, accessibility,
+    explicit acceptance criteria, understanding the problem first, and
+    “minimal diff ≠ correct fix”.
+  - **`lite`**: execute directly, less ceremony; still complete explicit
+    deliverables; you may point out a simpler approach in one sentence, but
+    **do not challenge explicit requirements**.
+  - **`full`** (default): the full seven-rung ladder (YAGNI → reuse → stdlib
+    → native → installed dependencies → one line → minimal implementation);
+    pick the shortest correct implementation by default; fix root causes, not
+    symptoms.
+  - **`ultra`**: demand evidence before adding code; prefer deleting or
+    reusing; proactively question speculative features/caching/abstraction/
+    configuration/new dependencies; for complex requests deliver the minimal
+    correct version first and state the conditions for expanding — **not a
+    blanket refusal**.
+  - `off`: no injection at all.
+  - Levels are **session-scoped** (session A does not affect session B;
+    released automatically when the session ends).
+  - Bare `/ponytail`: when enabled it only reports; when `off` it restores the
+    effective default level (falling back to `full` if that default is also
+    `off`).
+  - `/ponytail status`: query only, never modifies, and shows whether the
+    current mode comes from a session override or the configured default.
+  - `/ponytail reset`: clears the current session override and follows the
+    effective configured default again.
+  - `/ponytail lite|full|ultra|off`: explicit switch.
+  - `/ponytail default <mode>`: persists the default to the **user-level
+    config file** (env/profile still take priority; the command reports both
+    `saved` and `effective`).
+  - Status changes take effect immediately without waking an idle agent.
+    Notices reach the model on its next real step, avoiding an extra model
+    turn just to switch levels.
+- **One-shot skills** (load on demand, never part of the standing prompt):
+  - `/ponytail-review` — find over-engineering in recent changes; every
+    finding includes location, replacement, and actual call evidence; no
+    guessed savings numbers.
+  - `/ponytail-audit` — whole-repo over-engineering audit; distinguishes
+    safe-to-delete from verify-first candidates; at most 10 high-value
+    findings.
+  - `/ponytail-debt` — harvests all `ponytail:` comments into a debt ledger.
+  - `/ponytail-gain` — upstream benchmark reference scoreboard (less code;
+    token/cost/latency effects depend on model and task, **not guaranteed by
+    this port**).
+  - `/ponytail-help` — reference card.
+- **Deactivation**: say `stop ponytail`, `normal mode`, `停止 ponytail`,
+  `关闭 ponytail`, `普通模式`, or `正常模式` (trailing Chinese/English
+  punctuation tolerated); `/ponytail` re-enables at any time.
+- **Default priority** (consistent across code/tests/docs):
+  ```
+  session override > PONYTAIL_DEFAULT_MODE > Profile config.defaultMode > user config.json > full
+  ```
+  - **Profile-level config** (official Cordis plugin config API; each profile
+    can differ):
+    ```yaml
+    # add config to the ponytail row in ~/.dsh/profiles/tui/cordis.patch.yml
+    - insert:
+        - id: ponytail
+          name: '@mengyuly/dsh-ponytail'
+          config:
+            defaultMode: lite
+    ```
+    Example: `web → full`, `tui → lite`, `automation → off`. Profile config is
+    read when the plugin initializes (Cordis has no public config-change
+    event), so **restart that profile after changing it**; an invalid value
+    warns once and falls back without breaking startup. The user `config.json`
+    stays hot-reloaded.
+  - **User config.json** (`~/.config/ponytail/config.json`, Windows
+    `%APPDATA%\ponytail\config.json`): `{"defaultMode": "lite"}`,
+    hot-reloaded (~1s polling); invalid content keeps the last valid value.
+- **Subagents (honest boundaries)**: DSH's built-in `subagent` tool is an
+  **isolated fork**, but the global system-prompt section participates in each
+  subagent's own assembly by default; this is not parent-prompt or session
+  state inheritance. `PONYTAIL_SUBAGENT_MATCHER` (a regex matched against the
+  subagent's `agentPreset`) only filters which subagents can enter this prompt
+  pipeline — it is not an inheritance switch; without a preset nothing is
+  excluded by the matcher. DSH currently has no public parent→child prompt
+  inheritance API, so **no parent-child prompt inheritance is claimed** (a
+  read-only mode-snapshot propagation may follow once an official API
+  exists). An invalid regex warns once and fails open.
+- **Config errors**: invalid JSON / invalid `defaultMode` / read failures /
+  invalid regex warn exactly once (no log spam); a missing config file is
+  normal and never warns.
+
+## Efficiency (conditional gains, not guarantees)
+
+Ponytail adds a small fixed ruleset to every model request. Its benefit is
+**conditional**: when the agent tends to over-engineer, the reduced code,
+tool calls, and rework may offset or exceed that overhead; when the task is
+already simple, the gain may be near zero — or the extra input overhead may
+lose outright. It is not a “save tokens switch” and does not guarantee savings
+across models — some reasoning models may get more expensive due to prompt
+and reasoning overhead.
+
+Measured prompt-section sizes for this DSH port (`node
+scripts/measure-prompt.mjs`, generated from the real
+`getPonytailInstructions()`):
+
+| Level | Characters | UTF-8 bytes | Notes |
+|------|--------|-----------|------|
+| lite | 1915 | 1917 | measured output |
+| full | 3022 | 3038 | measured output |
+| ultra | 2797 | 2813 | measured output |
+| off | 0 | 0 | not injected |
+
+These are **prompt-size measurements, not billing amounts, and not a savings
+ratio that holds for every model** (there is no universal tokenizer;
+`estimated_tokens` in `measure:prompt` output is null; characters/4 is only a
+rough estimate). Same-mode output is byte-stable, so KV-cache prefixes hit.
+
+**Upstream numbers are not a guarantee for this DSH port**: the upstream
+Ponytail single-shot results (code −80–94%, cost −42–75%, latency 3.1–5.8×)
+and agentic results (LOC −54% etc.) are references only; this DSH port has
+**not** established stable token/cost/latency savings rates. The DSH smoke
+benchmark provides directional evidence only (see
+`docs/dsh-smoke-summary.md`).
+
+## Known limitations
+
+- The levels differ in **rule semantics** (see above); the three prompt
+  sizes are similar (measured in the table above).
+- The upstream Claude-only statusline badge has no DSH counterpart; the MCP
+  server was dropped because DSH has a first-class system-prompt injection
+  point.
+- User `config.json` hot-reloads; `PONYTAIL_DEFAULT_MODE` and profile config
+  need a restart.
+- The shipped `lib/` is a prebuilt artifact; to change behavior, rebuild in
+  the main repo and re-sync.
+
+## Compatibility matrix (measured, not fabricated)
+
+| Component | Verified environment | Notes |
+|---|---|---|
+| Node.js | 22.x / 24.x | CI matrix, 4 combinations green |
+| OS | ubuntu-latest / windows-latest | CI matrix |
+| DSH | commit `b150a551` (build checkout) | exact mapping to an official release **TBD** |
+| Cordis | 4.0.1 (build vendor) | same |
+| web profile | verified | long-running on a real local profile + three isolated install-path tests (npm / GitHub / tgz) |
+| tui profile | not verified | not started inside a tui profile |
+| headless profile | not verified | not fully started; plugin unit tests run in a UI-less environment |
+| npm tarball | verified | contents/version/post-install smoke/NodeNext consumer |
+
+- `dist-provenance.json` records the actual build sources (checkout commit +
+  node/typescript/tsdown/cordis versions).
+- Do not hide failures with `continue-on-error` — the matrix is green only if
+  every cell is green.
+
+## Test environments and source of truth
+
+- Verified locally (Linux, Node.js **v24.16.0**, deepseek-harness checkout
+  build) and in the CI matrix (**ubuntu-latest + windows-latest × Node
+  22/24**). The exact published DSH/Cordis release they correspond to is
+  **TBD** — the checkout is a prerelease working tree, not a release tag.
+- The authoritative source is `packages/community/ponytail`
+  (`@deepseek-ai/dsh-ponytail`) in the deepseek-harness monorepo; this
+  repository (`@mengyuly/dsh-ponytail`) is a **distribution mirror**: it
+  ships build artifacts with the package and is not an independent source of
+  truth.
+
+## Release maintenance
+
+> **The commands below are for source-repository maintainers only.** The
+> `scripts/` directory is deliberately **excluded from the npm tarball**, so
+> these commands are unavailable after installing the published package — npm
+> users never need to run maintenance checks; they exist for maintainers and
+> CI before a release. The published `package.json` exposes no `scripts/`
+> commands (no maintenance entry points, no install lifecycle hooks),
+> enforced by a regression check in `node scripts/verify-pack.mjs`.
+
+- **Authoritative source**: `packages/community/ponytail` in the
+  deepseek-harness monorepo (this repository is a distribution mirror that
+  publishes build artifacts only).
+- **Maintainer commands** (run `node scripts/<script>.mjs` inside the source
+  repository; see `package.dev.json` for the shortcut list):
+  ```bash
+  node scripts/check-bundle.mjs        # bundle external-dependency allowlist + no new Function/eval
+  node scripts/verify-dist.mjs         # static consistency: src/d.ts export parity, key signatures, runtime exports, provenance
+  node scripts/verify-pack.mjs         # tarball boundary (incl. no scripts/ exposure regression), versions, post-install smoke
+  node scripts/test-consumer.mjs       # NodeNext + skipLibCheck:false declaration consumer test (against the packed artifact)
+  node scripts/test-regressions.mjs    # regression tests for the verification tooling itself
+  node scripts/test-core.mjs           # core prompt bytes, safety boundaries, modes and skill surface
+  node scripts/measure-prompt.mjs      # prompt-section size per mode (depends on the unpublished src/)
+  node scripts/check-release-links.mjs # README/CHANGELOG/docs contain no versioned latest asset links
+  node scripts/check-release-consistency.mjs --version <v>  # four-way release consistency (git tag/npm/GitHub/provenance)
+  ```
+- **Fully regenerate and sync `lib/`** (JS and declarations must be synced as
+  one artifact; copying a single JS file is forbidden):
+  ```bash
+  DSH_CHECKOUT=/path/to/deepseek-harness node scripts/sync-dist.mjs
+  ```
+  This rebuilds inside the authoritative checkout (`tsc` for declarations +
+  `tsdown` for the runtime bundle), syncs `lib/index.js`, `lib/invariant.js`,
+  `lib/types/*.d.ts`, generates `dist-provenance.json` (recording the
+  authoritative checkout's real commit SHA and toolchain versions), and runs
+  the consistency checks automatically; it prompts you to commit when
+  artifacts changed. **Full build consistency is produced by this command in
+  the release process — the mirror repository's CI never rebuilds the
+  authoritative monorepo.**
+- **CI capability boundary (honest)**: CI (ubuntu + windows matrix) runs the
+  static verification and pack/consumer tests above, but **does not rebuild
+  the authoritative monorepo**; `verify:dist` is an export-surface/signature/
+  runtime-export consistency check, **not** a byte-level equivalence proof
+  against the authoritative build — that is guaranteed by `sync:dist` in the
+  release process.
+- `dist-provenance.json` ships with the npm package for build-source audits.
+- When verifying locally, if `npm_execpath` points at another package manager
+  (e.g. a pnpm/yarn shim), the scripts fall back to `npm` on PATH; set
+  `PONYTAIL_VERIFY_KEEP_TEMP=1` to keep temp directories on failure.
+- **Security**: `scripts/**` exists only for development/build/release
+  verification — **not in the npm tarball**, no install lifecycle hooks, not
+  referenced by the runtime entry point, and not exposed in the published
+  `package.json`; `child_process` warnings there are accepted development-tool
+  risk. See [SECURITY.md](SECURITY.md).
+
+## License
+
+MIT, © 2026 DietrichGebert (upstream) + MengYuil (port). See [LICENSE](LICENSE).

package/cordis.patch.yml

@@ -1,4 +1,4 @@
 # dsh bundle patch: inserts the ponytail persona plugin into a profile's layer stack.
 - insert:
     - id: ponytail
-      name: '@mengyuly/dsh-ponytail'
+      name: '@mengyuly/dsh-ponytail'

package/lib/index.js

@@ -1696,7 +1696,7 @@ function registerCommands(deps, commandCtx) {
 				deps.setDefault(effective);
 				const reason = defaultOverrideReason(process.env, deps.profileMode);
 				if (reason !== null) {
-					agent.steer(createUserMessage({
+					agent.inject(createUserMessage({
 						content: [{
 							type: "text",
 							text: `PONYTAIL DEFAULT SET — saved ${written}, effective ${effective} (${reason}).`
@@ -1711,7 +1711,7 @@ function registerCommands(deps, commandCtx) {
 						text: `Saved default: ${written}. Effective default: ${effective}, overridden by ${reason}.`
 					};
 				}
-				agent.steer(createUserMessage({
+				agent.inject(createUserMessage({
 					content: [{
 						type: "text",
 						text: `PONYTAIL DEFAULT SET — new sessions start in ${written}.`
@@ -1739,7 +1739,7 @@ function registerCommands(deps, commandCtx) {
 				const changed = deps.store.has(key);
 				deps.store.clear(key);
 				const current = deps.defaultMode();
-				if (changed) agent.steer(createUserMessage({
+				if (changed) agent.inject(createUserMessage({
 					content: [{
 						type: "text",
 						text: modeNotice(current)
@@ -1760,7 +1760,7 @@ function registerCommands(deps, commandCtx) {
 				if (current === "off") {
 					if (effectiveDefault === "off") {
 						deps.store.set(sessionKey(agent), "full");
-						agent.steer(createUserMessage({
+						agent.inject(createUserMessage({
 							content: [{
 								type: "text",
 								text: "PONYTAIL MODE CHANGED — level: full"
@@ -1776,7 +1776,7 @@ function registerCommands(deps, commandCtx) {
 						};
 					}
 					deps.store.clear(sessionKey(agent));
-					agent.steer(createUserMessage({
+					agent.inject(createUserMessage({
 						content: [{
 							type: "text",
 							text: `PONYTAIL MODE ACTIVE — level: ${effectiveDefault}`
@@ -1791,7 +1791,7 @@ function registerCommands(deps, commandCtx) {
 						text: `Ponytail re-enabled. Effective default: ${effectiveDefault}.`
 					};
 				}
-				agent.steer(createUserMessage({
+				agent.inject(createUserMessage({
 					content: [{
 						type: "text",
 						text: `PONYTAIL MODE ACTIVE — level: ${current}`
@@ -1812,7 +1812,7 @@ function registerCommands(deps, commandCtx) {
 				text: "Usage: /ponytail [status|reset|default <mode>|lite|full|ultra|off]"
 			};
 			deps.store.set(sessionKey(agent), mode);
-			agent.steer(createUserMessage({
+			agent.inject(createUserMessage({
 				content: [{
 					type: "text",
 					text: modeNotice(mode)

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@mengyuly/dsh-ponytail",
   "description": "Lazy senior dev mode for DeepSeek Harness: always-on minimal-code ruleset, intensity switching, and short review/audit/debt/gain/help skills",
-  "version": "0.3.3",
+  "version": "0.3.4",
   "publishConfig": {
     "access": "public"
   },
@@ -25,6 +25,7 @@
     "./package.json": "./package.json"
   },
   "files": [
+    "README_EN.md",
     "dist-provenance.json",
     "CHANGELOG.md",
     "lib/index.js",