@crazyhappyone/dsh-tui 0.1.0-alpha.48 → 0.1.0-alpha.49

package/README.md

@@ -62,6 +62,10 @@ dsh-tui
 
 In this mode, the launcher automatically detects the installed `dsh` executable and mounts the bundled prebuilt TUI extension without cloning the 2GB monorepo.
 
+Detection checks an absolute `DSH_BIN` override, PATH, npm and pnpm global packages, then packages resolvable from the working directory and launcher directory. Each probe subprocess has a three-second timeout. A broken explicit `DSH_BIN` is reported without selecting another installation. Package bin entries run through the launcher's Node executable; probing establishes executability, not compatibility with older Harness releases.
+
+Lightweight launch uses `dsh --profile tui --patch <bundled-patch>` with a custom `tui` profile containing only `@deepseek-ai/dsh-base` and `patchReload: startup`. Existing profile manifests are preserved. The patch supplies the terminal runtime; the current `headless` profile adds its own runner and is not used. This release composition follows Harness `0.1.6-alpha.2`.
+
 ### Mode B: Source Mode (Developer / Debugging)
 To debug or hack on TUI or DSH source directly:
 
@@ -86,7 +90,7 @@ dsh-tui
 
 Run `dsh-tui update` when you want to refresh the dedicated runtime before launching it again.
 
-Ordinary arguments pass to the `deepseek-tui` profile unchanged:
+Ordinary arguments pass unchanged to `tui` in lightweight mode or `deepseek-tui` in source mode:
 
 ```text
 dsh-tui "review this repository"
@@ -179,6 +183,8 @@ The proxy listens on `127.0.0.1:4097` only; remove the `zen-proxy` row from
 <a id="development-workflow"></a>
 ## Development workflow
 
+With the matching `../deepseek-harness` checkout and its build dependencies installed, run `npm run build`. This invokes its supported `scripts/build-dsh-tui-bundle.ts`, mirrors the TUI sources, validates bundle imports, and generates both release patches from `packages/tui/tui/cordis.patch.yml`, changing only bundled plugin paths. Run `npm test` and `npm run pack:check` before release.
+
 Implement TUI behavior in the complete DSH checkout, run the checks selected by the changed behavior, and export the confirmed `packages/tui/` tree here. Develop the launcher in this repository with `npm test`; test installation through a packed tarball rather than a workspace link.
 
 The launcher runtime must remain separate from the development checkout. To test against another source safely, point `DSH_TUI_RUNTIME_DIR` at a temporary absolute directory and configure `DSH_TUI_SOURCE_URL` and `DSH_TUI_SOURCE_REF`.

package/README.zh.md

@@ -178,6 +178,12 @@ llm-pi-ai:
 <a id="development-workflow"></a>
 ## 开发工作流
 
+在相邻的 `../deepseek-harness` checkout 安装构建依赖后，运行 `npm run build`。该命令调用现有 `scripts/build-dsh-tui-bundle.ts`，镜像 TUI 源码并检查 bundle 导入，再从 `packages/tui/tui/cordis.patch.yml` 生成两份发布 patch，仅替换预构建插件的路径。发布前运行 `npm test` 和 `npm run pack:check`。
+
+轻量启动按绝对路径 `DSH_BIN`、PATH、npm/pnpm 全局包、工作目录和 launcher 目录可解析的本地包依次探测，每个探测子进程限时三秒；显式 `DSH_BIN` 失败时不会选择其他安装。包内 bin 使用当前 Node 执行，探测成功仅表示可执行，不保证旧版 Harness 兼容。
+
+轻量模式执行 `dsh --profile tui --patch <bundled-patch>`；首次创建的 `tui` profile 仅含 `@deepseek-ai/dsh-base`，设置 `patchReload: startup`，已有 manifest 保持不变。发布 patch 提供终端运行时，并与 Harness `0.1.6-alpha.2` 对齐。当前 `headless` profile 会添加独立 runner，因此不用于轻量启动；源码模式继续使用 `deepseek-tui`。
+
 在完整 DSH checkout 中实现 TUI 行为，运行受变更行为影响的检查，并将确认后的 `packages/tui/` tree 导出到这里。在本仓库中使用 `npm test` 开发 launcher；通过 packed tarball 测试安装，不使用 workspace link。
 
 Launcher runtime 必须与开发 checkout 分离。要安全测试其他源码，请将 `DSH_TUI_RUNTIME_DIR` 指向临时绝对目录，并配置 `DSH_TUI_SOURCE_URL` 与 `DSH_TUI_SOURCE_REF`。

package/bin/dsh-tui.js

@@ -5,6 +5,8 @@ import { lstatSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'
 import { constants, homedir } from 'node:os'
 import { fileURLToPath } from 'node:url'
 import { spawnSync } from 'node:child_process'
+import { createRequire } from 'node:module'
+import { join } from 'node:path'
 import { parseLauncherInvocation, resolveLauncherSettings, runLauncher } from '../src/launcher.js'
 
 function inspectPath(path) {
@@ -28,6 +30,8 @@ function run(command, args, options = {}) {
     env: process.env,
     maxBuffer: capture ? 16 * 1024 * 1024 : undefined,
     shell: false,
+    timeout: options.timeoutMs,
+    killSignal: 'SIGKILL',
     stdio: options.stdio === 'inherit' ? 'inherit' : capture ? ['ignore', 'pipe', 'pipe'] : 'inherit',
   })
   if (result.error !== undefined) {
@@ -61,6 +65,16 @@ try {
     packageVersion: manifest.version,
     env: process.env,
     adapters: {
+      nodeExecutable: process.execPath,
+      currentDirectory: process.cwd(),
+      resolveDshPackage(directory) {
+        try {
+          return createRequire(join(directory, 'package.json')).resolve('@deepseek-ai/dsh/package.json')
+        } catch (error) {
+          if (error.code === 'MODULE_NOT_FOUND') return undefined
+          throw error
+        }
+      },
       inspectPath,
       makeDirectory: path => mkdirSync(path, { recursive: true, mode: 0o700 }),
       readText: path => readFileSync(path, 'utf8'),

package/cordis.patch.yml

@@ -1,30 +1,40 @@
+# yaml-language-server: $schema=none
 # The deepseek-tui bundle patch: terminal-surface mode directly over dsh-base.
 # It mounts no Host, HTTP server, Web runtime, or browser plugin. An ordinary
 # provider plugin parses --resume/--cwd and the task positional, then the
 # terminal driver creates an Agent through the core registry and prints its
 # durable result. Unlike the web surface, the agent plane stays process-wide:
 # dsh-base keeps tool-bash/tool-fs/tool-todo and friends because the TUI is
-# single-session and composes its agent process-wide.
+# single-session and composes its agent process-wide. Insert and config rows
+# resolve through the 0.1.6-alpha.2 runtime Loader; this file is not a
+# compile-time plugin graph.
 
 - id: system-prompt
   config:
     persona: >-
-      You are a coding agent powered by the {{model}} model. Your working directory is {{cwd}}.
-      You are interacting with the user through the deepseek-tui terminal interface.
+      You are a coding agent powered by the {{model}} model.
+      Your working directory is {{cwd}}.
+      You are interacting with the user through the deepseek-tui
+      terminal interface.
 
-# The shared module-reload HMR row stays off; the launcher's watch-only
-# fallback still keeps the user patch layers live until the run exits.
+# HMR stays disabled. Profile and home patch edits apply on the next
+# process start; /reload restarts the terminal with the current session.
 - id: hmr
   disabled: true
 
 - id: tools
   config:
     # Keep the same temporary process-wide Code Mode opt-in as the Web surface.
+    # pi-lens-ignore: YAML:0
     mode: !!js process.env.DSH_TOOLS_MODE
 
 - id: approval
+  # pi-lens-ignore: YAML:0
   config:
-    policy: !!js "(process.env.DSH_PERMISSION_MODE ?? 'workspace-write') === 'danger-full-access' ? 'never' : 'ask'"
+    policy: !!js >-
+      (process.env.DSH_PERMISSION_MODE ?? 'workspace-write')
+      === 'danger-full-access'
+      ? 'never' : 'ask'
 
 # Full-text session search stays disabled until the first Ctrl+K search (the
 # base row's `openAt: never`). The TUI has no browser-latency concern, so the
@@ -34,36 +44,60 @@
     path: ':memory:'
     openAt: first-search
 
+# Inherit the 0.1.6-alpha.2 llm-deepseek DEFAULT_MODELS catalog
+# (`deepseek-flash`, `deepseek-v4-pro`). Do not re-list removed V4 Flash
+# / V4 Flash Vision Exp ids here. Protocol stays Messages unless a PTY
+# snapshot sets DSH_LLM_PROTOCOL=chat-completions for the mock server.
 - id: llm-deepseek
   config:
-    models:
-      - id: deepseek-v4-flash
-        name: DeepSeek-V4-Flash
-        description: Fast, efficient, and economical; suited to focused, routine, or parallel tasks.
-        contextWindow: 1000000
-      - id: deepseek-v4-pro
-        name: DeepSeek-V4-Pro
-        description: Stronger agentic coding, knowledge, and difficult reasoning; suited to complex or quality-critical tasks at higher cost.
-        contextWindow: 1000000
-      - id: deepseek-v4-flash-vision-exp
-        name: DeepSeek-V4-Flash-Vision-Exp
-        contextWindow: 1000000
-        inputModalities: [text, image]
-      - id: deepseek-v4.1-flash-expires-on-0910
-        name: DeepSeek-V4.1-Flash
-        description: DeepSeek V4.1 Flash intermediate beta with native multimodal support, higher capability, faster speed, and lower cost.
-        contextWindow: 1000000
-        inputModalities: [text, image]
+    # pi-lens-ignore: YAML:0
+    protocol: !!js "process.env.DSH_LLM_PROTOCOL ?? 'messages'"
 
 - id: agent-default-model
   config:
+    # pi-lens-ignore: YAML:0
     provider: deepseek-official
-    model: !!js "process.env.DSH_MODEL ?? 'deepseek-v4-flash'"
+    # pi-lens-ignore: YAML:0
+    model: !!js "process.env.DSH_MODEL ?? 'deepseek-flash'"
+
+# pi-lens-ignore: YAML:0
+- id: session-title-llm
+  name: "@deepseek-ai/dsh-session-title-first-prompt-llm"
+  config:
+    targetWords: 5
+    targetCjkCharacters: 10
+    maxInputBytes: 16384
+    maxOutputTokens: 64
+    timeoutMs: 60000
 
 - insert:
+    # Per-round token/duration stats for the terminal status line.
+    - id: session-stats
+      name: "@deepseek-ai/dsh-session-stats"
+
+    # Base owns the shared storage/domain/projection-cache stack. This layer
+    # contributes only the TUI-facing feedback consumer over that stack.
+    - id: message-feedback
+      name: "@deepseek-ai/dsh-message-feedback"
+      config:
+        maxNoteBytes: 8192
+
     - id: tool-ask-user
       name: "@deepseek-ai/dsh-tool-ask-user"
 
+    - id: tui-startup
+      name: "./dist/startup.js"
+
+    # Zero-config official OpenAI/Anthropic routes. They register
+    # openai-official / anthropic-official so they do not collide with the
+    # dormant llm-pi-ai catalog keys openai / anthropic that the Web Models
+    # page already owns.
+    - id: llm-openai
+      name: "@deepseek-ai/dsh-llm-openai"
+
+    - id: llm-anthropic
+      name: "@deepseek-ai/dsh-llm-anthropic"
+
     # In-process OpenCode Zen header proxy (vendored from
     # zhangyang-crazy-one/dsh-zen-proxy, MIT). Listens on localhost
     # only; point an opencode provider baseURL at
@@ -75,20 +109,33 @@
         host: 127.0.0.1
         port: 4097
 
-    - id: tui-startup
-      name: "./dist/startup.js"
+    # Records each top-level turn's changed files from git working-tree
+    # snapshots; the TUI projects a bounded transcript card from Host
+    # summaries. Bounds match the plugin Config defaults.
+    - id: workspace-changes
+      name: '@deepseek-ai/dsh-workspace-changes'
+      config:
+        timeoutMs: 30000
+        outputMaxBytes: 8388608
+        maxFiles: 500
+        maxFileBytes: 2097152
+        diffTimeoutMs: 100
 
     # Reads its task from the ordinary tuiStartup provider.
     - id: tui-runtime
+      # pi-lens-ignore: YAML:0
       name: "./dist/index.js"
       inject: [tuiStartup]
+      # pi-lens-ignore: YAML:0
       config:
+        # pi-lens-ignore: YAML:0
         task: !!js ctx.tuiStartup.task
         resume: !!js ctx.tuiStartup.resume
+        # pi-lens-ignore: YAML:0
         cwd: !!js ctx.tuiStartup.cwd
         model: !!js ctx.tuiStartup.model
         frameStats: !!js ctx.tuiStartup.frameStats
-        # Render policy (overscan + stream pacing + scroll pacing + cache budgets).
+        # Render policy (overscan, stream pacing, scroll pacing, cache).
         # Loaded by the runtime Config schema; the values below mirror the
         # schema defaults so the deployment layer owns the tunables. Override
         # here to retune without touching the plugin.