@fuxi-ailabs/dsh-ecat 0.0.0-stage → 0.1.0-alpha.2

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Fuxi AI Labs
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -1,3 +1,169 @@
-# Temporary Holding Version
+# dsh-ecat
+
+Two DSH commands over the ECAT refinement loop.
+
+| Command | What it does |
+|---|---|
+| `/ecat-setup` | Hands a prompt to the agent, which runs the `ecat-setup` skill: interview, profile, scoring plugins |
+| `/ecat-refine` | Runs the `ecat-refine` scripts directly: `status`, `open`, `preflight`, `start`, `stop` |
+
+`/ecat-setup` never writes files itself — every write goes through the agent, so the normal
+approval and sandbox rules apply. `/ecat-refine` only executes fixed subcommands with an
+argv array; the raw input is never passed to a shell.
+
+## Install
+
+```
+dsh plugin --profile <profile> add dsh-ecat
+```
+
+Or add the market index source and install from the discovers page. The package is MIT
+licensed; its full contents are the two skills, the ECAT wheel, and the command code.
+
+## Requirements
+
+- DSH `^0.2.1-rc.2`
+- Python 3.11+ somewhere on the host (`python3`, `python`, or the `py` launcher)
+
+ECAT itself is not a prerequisite: the package ships `assets/wheels/ecat-*.whl` and
+`/ecat-setup` creates `~/.dsh/ecat-venv`, installs the wheel into it with `uv pip`
+(falling back to `python -m venv` + `pip`), and points every later command at that
+interpreter. Dependencies are resolved from the network on that first run, which is why
+the installation lives in setup rather than in the loop commands: configuration is the
+point where a missing environment should surface, not the first status check.
+
+If the environment cannot be prepared, `/ecat-setup` fails before asking any questions
+and reports the uv and pip errors; `/ecat-refine` has no install step of its own.
+
+If you already installed ECAT globally (`uv tool install --editable <repo>`), the plugin
+uses that interpreter instead and never creates a venv.
+
+## Configuration
+
+| Key | Default | Meaning |
+|---|---|---|
+| `pythonPath` | auto | Interpreter to use. When set it always wins, broken or not. |
+| `ecatPath` | `ecat` | The CLI, used only for preflight's availability check. |
+| `workspaceRoot` | empty | Fixed ECAT workspace; otherwise the hand-off file or session cwd. |
+| `uvPath` | `uv` | Used by `/ecat-setup` to build the venv. Falls back to `python -m venv`. |
+
+`ECAT_VENV_DIR` relocates the managed environment; tests and multi-user machines use it.
+
+## Usage
+
+```
+/ecat-setup
+/ecat-setup help
+/ecat-refine status
+/ecat-refine preflight
+/ecat-refine start
+/ecat-refine open
+/ecat-refine stop
+```
+
+Both commands advertise their arguments in the composer hint as soon as you type the
+name: `/ecat-setup [help | <task description>]` and
+`/ecat-refine [help | status | open | start | stop | preflight] [workspace]`.
+
+`/ecat-refine` with no subcommand prints the same list as `/ecat-refine help`.
+`/ecat-refine run` starts the runtime-selection conversation explicitly.
+
+Help output is echoed back through one agent turn. DSH's chat view treats a session
+whose only nodes are commands as blank and shows the welcome screen, so a command
+that merely returns text stays invisible until something creates a conversation turn.
+
+### Worked example — `/ecat-setup help`
+
+The environment check lives here, not on `/ecat-refine`, because that is where you go
+when you want to know whether the Python side is usable and where a short wait is
+expected. Real output, DSH `0.2.1-rc.2`, plugin `0.1.0-alpha.1`, Windows:
+
+```
+Usage: /ecat-setup [help | <task description>]
+  no argument   interview me about the task, then write profile and plugins
+  <text>        same, reusing what this message already says about the task
+  help          this text, plus whether every ECAT dependency is usable
+
+interpreter: C:\Users\me\.dsh\ecat-venv\Scripts\python.exe
+python: 3.11.13
+environment: ready — all 13 runtime dependencies usable
+  anthropic 2.1.0
+  typer 0.27.3
+  ...
+ecat: C:\Users\me\.dsh\ecat-venv\Lib\site-packages\ecat\__init__.py
+```
+
+The list is not hard-coded: `src/probe_deps.py` reads the installed wheel's own
+`Requires-Dist` entries through `importlib.metadata`, so a dependency added to
+`pyproject.toml` is covered the next time the wheel is built. Each entry is
+checked for presence, for the version range the wheel asks for, and for a
+resolvable top-level module. A failure names the package and how to fix it:
+
+```
+environment: NOT ready — 2 of 13 dependencies failed
+  openpyxl is not installed
+  pydantic has version 2.6.1, needs pydantic>=2.7
+  fix with: <interpreter> -m pip install openpyxl
+```
+
+This is deliberately not `import ecat`: `ecat/__init__.py` is a version string,
+so importing it succeeds with every third-party package missing. `resolvePython`
+uses the same probe when it chooses an interpreter, so a half-installed
+environment is never picked for `status`, `start`, or anything else.
+
+Before the first setup it reads `environment: NOT ready — <venv> does not exist.` and
+tells you to run `/ecat-setup` with no argument. Nothing is installed and no interview
+starts from `help`.
+
+### Worked example — `/ecat-refine help`
+
+```
+Usage: /ecat-refine [status|open|start|stop|preflight] [workspace]
+  status     latest run: H, iterations, worst dimension (default)
+  ...
+No subcommand runs the loop: the agent picks the runtime with you first.
+Check the environment with /ecat-setup help; /ecat-refine has no install step.
+```
+
+This is synchronous — it does not spawn Python, so it answers immediately even when
+no interpreter can import ECAT.
+
+### Worked example — a common failure
+
+`/ecat-refine status` with no workspace configured and no hand-off file:
+
+```
+kind: error
+No ECAT workspace found. Run /ecat-setup first, pass it explicitly
+(/ecat-refine status <workspace>), or set workspaceRoot in the plugin config.
+```
+
+This is the path a first-time user hits; it names all three ways to resolve it rather
+than silently reading the session's cwd, which could start the loop on the wrong tree.
+
+## Failure modes
+
+| Symptom | Cause |
+|---|---|
+| `No ECAT workspace found` | `/ecat-setup` has not run and no workspace was passed |
+| `spawn failed: ENOENT` | `pythonPath` does not exist on the DSH host |
+| preflight `exit 15` | `ecat` is not on the host PATH |
+| preflight `exit 13` | the variable named by `api_key_env` has no value in `<ws>/.env` |
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+## Releasing the npm package
+
+Everything needed to build and publish ships on the `dsh-plugin` branch: `src/`,
+the two skills under `assets/skills/`, and the ECAT wheel under `assets/wheels/`.
+The `ECAT_core` branch is not required for a release (refresh the wheel only when
+core behaviour that the plugin relies on changed).
+
+1. `git checkout dsh-plugin` and make sure the tree is clean.
+2. `node integrations/dsh/scripts/sync-skills.mjs` — keeps `assets/skills` identical to `./skills`.
+3. If ECAT core changed: build a wheel from that revision (`uv build`) and replace `assets/wheels/ecat-*.whl`.
+4. Bump the version: `npm version prerelease --preid=alpha --no-git-tag-version`.
+5. `cd integrations/dsh && npm pack --dry-run` — expect the same 37 files, including `src/client.js`.
+6. `npm publish --tag alpha`.
+7. `git push origin dsh-plugin`.
+
+Keep `name` in `package.json` and the `insert` entry in `cordis.patch.yml`
+identical (`@fuxi-ailabs/dsh-ecat`); a mismatch makes the plugin fail to import.

package/assets/skills/ecat-refine/SKILL.md

@@ -0,0 +1,126 @@
+---
+name: ecat-refine
+description: 启动、监控、停止 ECAT 的 Android→HarmonyOS 对抗式迁移循环（ecat refine）。当用户说"跑一下 ECAT""继续优化鸿蒙仓""跑迁移修复循环""ECAT 跑到哪了""看看熵降到多少了""停掉 ECAT"时触发。也用于跑之前检查 ECAT 配置是否符合预期。这是小时级后台任务，本 skill 只负责启停与查询，不会同步等待。
+---
+
+# ecat-refine
+
+驱动 ECAT：生成器改鸿蒙代码、判别器打分算熵（H），逐轮对抗直到收敛。
+
+## 这是长任务，绝不要同步等
+
+一轮的耗时由 target 规模、维度数量和模型速度决定，从几分钟到几十分钟不等；轮次上限默认 300。启动脚本后台跑并立即返回；之后每次想知道进展就调一次 `watch.py`，它秒回。
+
+**永远不要**用 `ecat refine` 直接前台跑，也不要在等待循环里反复 sleep。
+
+## 第零步：先问用户跑哪个 runtime，打印 config 路径让用户确认
+
+**先分清两份配置**：`ecat-setup` 写的是 `<WS>/ecat.config.yaml`（你的 profile、插件集、
+权重、路径），四份 `config.*.yaml` 只是 **runtime preset**（决定用哪个 CLI、endpoint 怎么走）。
+`preflight.py` / `start.py` **默认读 workspace 那份**；只有第一次在某台机器上试跑、或
+workspace 里还没有 `ecat.config.yaml` 时才用 preset。
+
+如果 workspace 里已有 `ecat.config.yaml` 而你又显式传了 preset，preflight 会 WARN 并
+列出被忽略的键（`profile` / `entropy.plugin_set` / `paths.target`）——看到这条就别继续。
+
+1. **问用户用哪个 runtime**（四选一，列出来让用户选）：
+
+| 选项 | preset | 说明 |
+|---|---|---|
+| Claude Code | `config.claudecode.yaml` | 走 `claude` CLI + DeepSeek 的 Anthropic 兼容层。可观测性完整（stream-json 事件流、会话续跑、多目录访问） |
+| opencode | `config.opencode.yaml` | 走 `opencode` CLI + OpenAI 兼容端点。**已知降级**：`generator_compact_interval` 无效；附加目录靠公共祖先合并成单个 `--dir` |
+| deveco | `config.deveco.yaml` | 走 `deveco` CLI（npm `@deveco/deveco-code`，opencode 的 fork）+ OpenAI 兼容端点。配置 schema / 事件流 / 429 软中止与 opencode 同构；**已知降级**：同 opencode（`generator_compact_interval` 无效） |
+| DeepSeek Harness | `config.dsh.yaml` | 走 `dsh --profile headless` 的一次性模式（npm `@deepseek-ai/dsh`）。**两处特殊**：`base_url` 必须留空（端点由 dsh 的 profile 决定，配了会被 ECAT 拒绝）；模型靠每次调用的临时 patch 指定。**已知降级**：`generator_compact_interval` 无效（无会话内压缩），改由 dsh 自带的 compaction-basic 接管；`.ecat/agents/*/agents/*.md` 不会被加载，judge 的 rubric 走派发提示词 |
+
+2. **打印实际会用的 config 绝对路径**，念给用户确认：workspace 有 `ecat.config.yaml` 时
+   就是它；没有时才念所选 preset 的路径。
+
+3. **用户确认后**才执行 preflight（体检）+ start（启动）。未确认/用户选别的，不要擅自跑。
+
+四份 preset 的差异只在 runtime / base_url / api_key_env / model（deveco 与 opencode
+同构，仅 runtime 字段和命令名不同），其余调参完全一致。**`ecat-setup` 已把这些键写进
+`<WS>/ecat.config.yaml`，不要用 preset 覆盖它。**
+
+> 选 `dsh` 时额外确认两件事：`dsh` 在 PATH 上（`npm i -g @deepseek-ai/dsh`），以及 `.env` 里有 `DEEPSEEK_API_KEY`。首次运行会在 `.ecat/agents/<role>/` 下自动生成 DSH 的 profile 和 `AGENTS.md`——preflight 会提示"尚未生成"，那是正常的。
+
+> 确认依据：跑 `preflight.py` 会打印「── 确认 ──」段（runtime + config 绝对路径），**把它念给用户确认后再执行 start.py**。preflight 不是确认，它只是体检；发车前的确认以用户点头为准。
+
+## 四步
+
+全部是 Python 脚本（Windows / Linux / macOS 通用，不需要 bash）。
+`S` = 本 skill 目录，`WS` = ECAT 工作区根目录。
+
+```
+# CFG = config.claudecode.yaml / config.opencode.yaml / config.deveco.yaml / config.dsh.yaml（第零步问出来的）
+
+# 1. 体检（必做）—— 摊开"实际会怎么跑"，逐项对现实核验
+python <S>/scripts/preflight.py --config <S>/<CFG> --workspace <WS>
+
+# 2. 启动（体检 exit 0 之后）
+python <S>/scripts/start.py --config <S>/<CFG> --workspace <WS>
+
+# 3. 查进度（随时，秒回）
+python <S>/scripts/watch.py --workspace <WS>
+#    想让用户看曲线：加 --open 用默认浏览器打开 charts/entropy.html
+python <S>/scripts/watch.py --workspace <WS> --open
+
+# 4. 停止
+python <S>/scripts/stop.py --workspace <WS>
+```
+
+解释器名按平台取 `python` 或 `python3`。脚本只用标准库 + `pyyaml`，
+进程操作可选依赖 `psutil`（ECAT 本身已依赖；缺失时自动退化到 tasklist/kill）。
+
+`<WS>` = ECAT 工作区根目录（含 `.ecat/`、鸿蒙仓、安卓源码那个目录）。必须由用户给出，不要猜。
+
+不传 `<WS>` 时，脚本和 `/ecat-refine` 命令按 `~/.ecat/active-workspace.json` 解析——那是**最近一次 `ecat-setup` 的 `check.py` 通过时**写入的指针，所以换项目只要重跑 setup，裸命令会自动跟随。命令行显式给的 `<WS>` 永远优先；指针指向的目录被删掉时会跳过它，退回插件配置/会话 cwd。
+
+## 配置就在这个 skill 里
+
+两份 config 都随 skill 分发。**每次启动前先跑 preflight**，它会打印这份配置实际指向什么，末尾单列一块「项目相关配置」。
+
+配置分两类，处理方式不同：
+
+| | 字段 | 怎么做 |
+|---|---|---|
+| **项目相关**（每次调用都可能变） | `project`、`paths.android`、`paths.harmony`、`paths.spec`、`paths.feature_checklist` | 把 preflight 末尾那一块念给用户确认；不对就改所选的 config 重跑 |
+| **策略/调参**（ECAT 的经验值） | `loop.*` 阈值、`entropy.weights`、`judge_max_issues`、`timeout_s`、模型与 endpoint | **不要动**。用户明确要求才改，改完说明改了什么 |
+
+`paths.spec` 与 `paths.feature_checklist` 是选填：spec 缺失时生成器退化为按文件名匹配安卓源码；feature_checklist 未配置时该熵维度自动跳过（`skipped: no_feature_checklist_configured`），都不会报错。但**配了却指向不存在的文件**是 FAIL——那是写错了，不是没配。
+
+## 退出码
+
+`preflight.py`
+
+| 码 | 含义 | 怎么办 |
+|---|---|---|
+| 0 | 就绪 | 可以启动 |
+| 10 | 配置本身有问题 | 改所选的 config；若报 endpoint 不匹配，多半是两份 config 用混了 |
+| 11 | 路径对不上现实 | 改 `config.yaml` 的 `paths.*`；harmony 不是 git 仓要先 `git init` |
+| 12 | 运行时缺失 | 装/加 PATH：`claude` 或 `opencode`（按所选 runtime） |
+| 13 | 凭据缺失 | 往 `<WS>/.env` 里补 `api_key_env` 指名的变量 |
+| 14 | workspace 未初始化 | 在 `<WS>` 跑 `ecat init` |
+| 15 | `ecat` 不在 PATH | `uv tool install --editable <ECAT 仓>` |
+
+`start.py`：0=已启动 / 1=参数或路径错 / 2=已有 run 在跑 / 3=启动失败（看打印的 boot log 末尾）
+
+`watch.py`：0=正在跑 / 1=已结束 / 2=异常终止 / 3=没有任何 run
+
+`stop.py`：0=已停 / 3=没有在跑的 run
+
+## 怎么读 watch 的输出
+
+- **H 曲线**是主信号，逐轮下降说明生成器在真的修东西；长期不动说明卡住了
+- **维度分数**指出当前最差的一维（`worst:`），生成器下一轮主要修它
+- **清单行**（配了功能清单时）显示 `OK/ISSUE/TODO` 与上轮推进量——推进量远低于目标说明判别器在敷衍
+- 异常终止（exit 2）时看 `.ecat/skill-refine.boot.log` 和 `runs_*/run.log` 末尾
+
+`--open` 打开的是 `runs_*/charts/entropy.html`：每轮判别完成后重写、页面每 30 秒自刷新，
+所以停在浏览器里即可，不必反复 `watch`。首轮判别完成前这个文件还不存在，脚本会直接说明。
+只想确认路径、不想弹浏览器时用 `--open --dry-run`。
+
+## 边界
+
+- 一个 workspace 同时只允许一个 refine（两个会互相 revert 同一个 git 仓）。`start.py` 会拦。
+- 本 skill 不改鸿蒙代码、不改熵插件、不动 `.ecat/agents/`——那些是 ECAT 自己的事。
+- 跑起来之后 ECAT 会自行 commit 每轮改动到鸿蒙仓；`stop.py` 不回滚已提交的内容。

package/assets/skills/ecat-refine/config.claudecode.yaml

@@ -0,0 +1,101 @@
+# ─── Identity ────────────────────────────────────────────────────────────────
+project: my-android-app
+work_dir: .
+
+# ─── Source / Target paths ───────────────────────────────────────────────────
+# `target` is required; `source` is optional (single-tree tasks omit it).
+paths:
+  source: ./Android_code/AIPPT
+  target: ./aippt_cc_glm_0705
+
+
+# ─── Agents ──────────────────────────────────────────────────────────────────
+# Each agent can use a different runtime, provider, model, and base_url.
+# base_url: set to your provider's OpenAI-compatible endpoint, or null for default.
+# api_key_env: the NAME of the env var in .env (not the key itself).
+
+generator:
+  runtime: claude_code       # claude_code | opencode | deveco_code | dsh
+  model: deepseek-v4-flash[1m]  # any model your provider supports
+  provider: deepseek
+  base_url: https://api.deepseek.com/anthropic             # e.g. https://api.deepseek.com/v1
+  api_key_env: ANTHROPIC_AUTH_TOKEN
+  config_dir: .ecat/agents/generator
+  timeout_s: 5400
+  extra_args: []
+  resume_session: true
+
+discriminator:
+  runtime: claude_code
+  model: deepseek-v4-flash[1m]
+  provider: deepseek
+  base_url: https://api.deepseek.com/anthropic
+  api_key_env: ANTHROPIC_AUTH_TOKEN
+  config_dir: .ecat/agents/discriminator
+  timeout_s: 3600
+  extra_args: []
+  evolution_write_paths:
+    - .ecat/agents/generator/skills
+    - .ecat/agents/generator/memory
+
+# ─── Adversarial loop ────────────────────────────────────────────────────────
+loop:
+  max_iterations: 300
+  entropy_threshold: 0.01
+  # stable_iters_to_stop: H 需连续 N 轮低于 entropy_threshold 才提前退出（防单轮"安静"过早收敛）。默认 2。
+  stable_iters_to_stop: 2
+  # regression_tolerance: 相邻两轮之间的回归容忍度。
+  # 如果本轮 H 比上轮 H 增加超过此值，判定为"回归"（代码质量变差）。
+  # revert_on_regression=true 时会 git checkout 撤销本轮 Generator 的改动。
+  # 作用：防止单次迭代引入破坏性变更。
+  regression_tolerance: 0.1
+  revert_on_regression: true
+
+  # oscillation_guard: 检测循环是否陷入"震荡"（反复修改同一问题无法收敛）。
+  # window: 回看最近 N 轮的 H 值序列。
+  # min_delta: 如果最近 window 轮中相邻两轮的 H 变化绝对值都 < min_delta，
+  #   说明 Generator 在原地打转，触发提前终止。
+  # 作用：避免浪费 token 在无效迭代上。
+  oscillation_guard:
+    enabled: true
+    window: 5
+    min_delta: 0.02
+
+  generator_compact_interval: 15
+  # max_items_per_iter: 每轮传给生成器的 work-list 最大条目数。
+  # 调小可缩短每轮耗时但需更多轮次收敛；调大每轮改动更多但速度变慢。默认 25。
+  max_items_per_iter: 50
+
+# ─── Entropy plugin set ───────────────────────────────────────────────────────
+# Every dimension listed in weights must be registered by a plugin in the
+# selected set; a weight with no plugin is silently ignored, so the dimension
+# never scores and the Generator never gets work for it.
+entropy:
+  plugin_set: example
+  # judge_max_issues: 每个 agentic judge 子代理最多上报的问题数。
+  # 调小让 judge 更早停止探索文件，加快判别器；调大获取更多细节。默认 10。
+  judge_max_issues: 20
+  weights:
+    compile:          0.30
+    feature_parity:   0.35
+
+# ─── Retry ────────────────────────────────────────────────────────────────────
+retry:
+  max_attempts: 3
+  base_delay_s: 2
+  max_delay_s: 30
+  jitter: true
+
+# ─── Phase gates ─────────────────────────────────────────────────────────────
+features:
+  memory_evolution: true
+  skill_evolution:  false
+  multi_discriminator: false
+  in_loop_intervention: false
+
+# ─── Reporting ───────────────────────────────────────────────────────────────
+reporting:
+  chart_format: png
+  chart_on_each_iter: true
+  keep_run_artifacts: -1
+  log_level: info

package/assets/skills/ecat-refine/config.deveco.yaml

@@ -0,0 +1,109 @@
+# ─── deveco_code 变体 ────────────────────────────────────────────────────────
+# 与 config.opencode.yaml 的差异只在 runtime：deveco 是 opencode 的 fork
+# （npm @deveco/deveco-code），配置 schema / 事件流 / 429 软中止与 opencode 同构，
+# 只是命令名为 `deveco`、自动批准用 `--auto`、配置目录环境变量是 DEVECO_CONFIG_DIR、
+# 配置文件名是 deveco.jsonc。ECAT 引擎由 DevecoCodeRunner（继承 OpencodeRunner）驱动。
+#
+# runtime  claude_code → deveco_code
+# 其余（base_url / api_key_env / model / provider）与 opencode 变体一致：
+#   * 走 @ai-sdk/openai-compatible 裸 OpenAI 端点（不要 /anthropic 后缀）
+#   * api_key_env 用 DEEPSEEK_API_KEY 之类标准 AI SDK 槽位（同 opencode）
+# ─── Identity ────────────────────────────────────────────────────────────────
+project: my-android-app
+work_dir: .
+
+# ─── Source / Target paths ───────────────────────────────────────────────────
+# `target` is required; `source` is optional (single-tree tasks omit it).
+paths:
+  source: ./Android_code/AIPPT
+  target: ./ai_ppt_ecat_deveco_skill0910
+
+
+# ─── Agents ──────────────────────────────────────────────────────────────────
+# Each agent can use a different runtime, provider, model, and base_url.
+# base_url: set to your provider's OpenAI-compatible endpoint, or null for default.
+# api_key_env: the NAME of the env var in .env (not the key itself).
+
+generator:
+  runtime: deveco_code       # claude_code | opencode | deveco_code | dsh
+  model: glm-5.3-flash  # 干净模型名（不带 provider 前缀；deveco runner 会拼成 zhipu/glm-5.3-flash）
+  provider: zhipu
+  base_url: https://open.bigmodel.cn/api/coding/paas/v4
+  api_key_env: GLM_API_KEY_YQ
+  config_dir: .ecat/agents/generator
+  timeout_s: 5400
+  extra_args: []
+  resume_session: true
+
+discriminator:
+  runtime: deveco_code
+  model: glm-5.3-flash
+  provider: zhipu
+  base_url: https://open.bigmodel.cn/api/coding/paas/v4
+  api_key_env: GLM_API_KEY_YQ
+  config_dir: .ecat/agents/discriminator
+  timeout_s: 3600
+  extra_args: []
+  evolution_write_paths:
+    - .ecat/agents/generator/skills
+    - .ecat/agents/generator/memory
+
+# ─── Adversarial loop ────────────────────────────────────────────────────────
+loop:
+  max_iterations: 100
+  entropy_threshold: 0.01
+  stable_iters_to_stop: 2
+  regression_tolerance: 0.1
+  revert_on_regression: true
+  oscillation_guard:
+    enabled: true
+    window: 5
+    min_delta: 0.02
+  generator_compact_interval: 15
+  max_items_per_iter: 50
+
+# ─── Entropy plugin set ───────────────────────────────────────────────────────
+# Every dimension listed in weights must be registered by a plugin in the
+# selected set; a weight with no plugin is silently ignored, so the dimension
+# never scores and the Generator never gets work for it.
+entropy:
+  plugin_set: example
+  # judge_max_issues: 每个 agentic judge 子代理最多上报的问题数。
+  # 调小让 judge 更早停止探索文件，加快判别器；调大获取更多细节。默认 10。
+  judge_max_issues: 20
+  weights:
+    compile:          0.30
+    feature_parity:   0.35
+
+    # entry_nav:      0.10
+    # feature_check:  0.30
+    # page_audit:     0.10
+    # confidence:     0.10
+    # spec_coverage:  0.10
+    # source_parity:  0.10
+    # hallucination:      0.10
+
+    # node_align:        0.0
+    # ui_align:          0.0
+    # runtime_liveness:  0.0
+
+# ─── Retry ────────────────────────────────────────────────────────────────────
+retry:
+  max_attempts: 3
+  base_delay_s: 2
+  max_delay_s: 30
+  jitter: true
+
+# ─── Phase gates ─────────────────────────────────────────────────────────────
+features:
+  memory_evolution: false
+  skill_evolution:  false
+  multi_discriminator: false
+  in_loop_intervention: false
+
+# ─── Reporting ───────────────────────────────────────────────────────────────
+reporting:
+  chart_format: png
+  chart_on_each_iter: true
+  keep_run_artifacts: -1
+  log_level: info

package/assets/skills/ecat-refine/config.dsh.yaml

@@ -0,0 +1,104 @@
+# ECAT — DeepSeek Harness (dsh) runtime, headless one-shot mode.
+#
+# Differences from the claude_code config:
+#   * no base_url — dsh owns its endpoint (official route or DEEPSEEK_BASE_URL);
+#     ECAT rejects base_url for this runtime rather than silently ignoring it.
+#   * model is written into a per-call patch, because headless has no --model.
+#   * api_key_env defaults to DEEPSEEK_API_KEY; the runner exports the resolved
+#     value under that name whatever variable the config points at.
+#   * each role gets its own DSH_HOME under .ecat/agents/<role>/ (profile +
+#     AGENTS.md), created automatically on the first run.
+#
+# Verify with: python <skill>/scripts/preflight.py --config <this file> --workspace <WS>
+
+# ─── Identity ────────────────────────────────────────────────────────────────
+project: my-project
+work_dir: .
+
+# ─── Prompt profile ──────────────────────────────────────────────────────────
+# Domain fragment under ecat/prompts/profiles/<name>/. "a2h" ships with the
+# repo; create your own with the ecat-setup skill.
+profile: a2h
+
+# ─── Source / Target paths ───────────────────────────────────────────────────
+paths:
+  # source: ./reference      # optional: only for a task with a ground-truth tree
+  target: ./target           # required, must exist
+
+# ─── Agents ──────────────────────────────────────────────────────────────────
+generator:
+  runtime: dsh
+  model: deepseek-flash
+  provider: deepseek
+  base_url: null             # must stay null for dsh
+  api_key_env: DEEPSEEK_API_KEY
+  config_dir: .ecat/agents/generator
+  timeout_s: 5400
+  extra_args: []
+  # Continue the same dsh session across iterations (--session-id). Saves
+  # cache reads; context grows, so pair it with the loop's compaction setting.
+  resume_session: true
+
+discriminator:
+  runtime: dsh
+  model: deepseek-flash
+  provider: deepseek
+  base_url: null
+  api_key_env: DEEPSEEK_API_KEY
+  config_dir: .ecat/agents/discriminator
+  timeout_s: 3600
+  extra_args: []
+  evolution_write_paths:
+    - .ecat/agents/generator/skills
+    - .ecat/agents/generator/memory
+
+# ─── Adversarial loop ────────────────────────────────────────────────────────
+loop:
+  max_iterations: 100
+  entropy_threshold: 0.01
+  stable_iters_to_stop: 2
+  regression_tolerance: 0.1
+  revert_on_regression: true
+  oscillation_guard:
+    enabled: true
+    window: 5
+    min_delta: 0.02
+  # dsh runs one process per call and has no in-session /compact, so this knob
+  # is a no-op here. Context growth is handled by dsh's own compaction-basic
+  # (threshold 0.8 of the model window) inside each resumed session.
+  generator_compact_interval: 0
+  max_items_per_iter: 25
+
+# ─── Entropy plugin set ──────────────────────────────────────────────────────
+# `example` ships with the repo. Replace with the set you build; every
+# dimension listed in weights must be registered by some plugin.
+entropy:
+  plugin_set: example
+  judge_max_issues: 10
+  weights:
+    compile: 0.30
+    feature_parity: 0.35
+
+# ─── Retry ────────────────────────────────────────────────────────────────────
+# dsh retries transient provider failures inside the turn (dsh-llm-retry).
+# Keep this modest so an upstream outage surfaces as a failed iteration rather
+# than hours of silent retrying.
+retry:
+  max_attempts: 2
+  base_delay_s: 2
+  max_delay_s: 30
+  jitter: true
+
+# ─── Phase gates ─────────────────────────────────────────────────────────────
+features:
+  memory_evolution: false
+  skill_evolution: false
+  multi_discriminator: false
+  in_loop_intervention: false
+
+# ─── Reporting ───────────────────────────────────────────────────────────────
+reporting:
+  chart_format: png
+  chart_on_each_iter: true
+  keep_run_artifacts: -1
+  log_level: info