dsh-exec-extension
其他 活跃维护

dsh-exec-extension

LvDAO/dsh-exec-extension

为无头模式dsh运行器的exec模块扩展额外执行参数支持,可直接灵活配置执行规则,无需调整原有运行逻辑,满足不同自动化脚本调度的参数需求。

0
Stars 标星
0
Forks 分支
0
Watchers 关注
0
Open Issues
JavaScript
主要语言
MIT
开源协议
201 KB
仓库大小
1 个月前
最后推送
一键安装扩展 / 插件指令
dsh plugin --profile web add github:LvDAO/dsh-exec-extension
git clone https://github.com/LvDAO/dsh-exec-extension.git
git clone git@github.com:LvDAO/dsh-exec-extension.git
README.md main

dsh-exec-extension

npm · 中文文档

A DeepSeek Harness bundle for one-shot exec, in the same shape as official surface plugins (dsh.bundle.patch + parseCmdline + commander). Stock headless-startup only declares [task...]; this bundle disables that row and inserts a replacement that still provides headlessStartup.task for unmodified headless-runner.

The CLI follows OpenCode run and Pi -p: piped stdin is merged into the prompt, @file / -f attaches files, --format json streams session events, --dir sets cwd, --yolo skips permission prompts (it does not disable the sandbox). Model overlay is in-process (Rust WASM) and never writes $DSH_HOME/settings.yaml.

Install

dsh plugin --profile exec add @deepseek-ai/dsh-headless@$(npm view @deepseek-ai/dsh version)
dsh plugin --profile exec add dsh-exec-extension
dsh --profile exec --help

Follow the same npm latest channel as the dsh CLI (npx @deepseek-ai/dsh / npm install --global @deepseek-ai/dsh). Do not use @next unless you installed dsh@next. @deepseek-ai/dsh-headless has a stale latest tag, so the first line takes the CLI's latest version instead of headless's own. dsh-exec-extension uses its own latest. Freeze with @<version> only when you need a snapshot. npm install does not need a prepare allowlist.

Git fallback: dsh plugin --profile exec add github:LvDAO/dsh-exec-extension

Dedicated profile only. Stock dsh --profile headless --model x "t" must still fail.

CLI

dsh --profile exec [options] [--] [task|-]
cat README.md | dsh --profile exec "summarize this"
dsh --profile exec @notes.md --full-auto "use the notes"
Flag Meaning
-m, --model This-process model (never part of the task)
--effort, --reasoning-effort off\|high\|max
--thinking Pi thinking tiers; xhigh/maxmax, minimal/low/medium/highhigh
--provider This-process provider
-C, --cwd / --cd / --dir Working directory (published on headlessStartup.cwd)
--timeout <seconds> appExit(1) after N seconds
-s, --sandbox / --permission-mode read-only | workspace-write | danger-full-access
--approval ask\|never\|allow ask fail-closes without a UI; never auto-denies; allow auto-grants
--full-auto workspace-write + auto-allow (CI)
--yolo / --dangerously-skip-permissions Skip permission prompts (auto-allow). Does not set danger-full-access; default sandbox stays workspace-write. Unrestricted access is --sandbox danger-full-access
--tools-mode native\|code\|both Tools presentation
-f, --file / @path Attach file text into the task
--output-schema <path> Prompt-level JSON Schema constraint (not constrained decode)
-o, --output-last-message <path> Also write the final assistant text
--format text\|json / --mode json json is session-event JSONL
-c, --config key=value Repeatable this-process overlay
--env KEY=VALUE / --api-key Process env only (not written to disk)
--print-selection Print overlaid selection JSON and exit

Unknown flags error. Overlay never calls saveSelection().

--output-schema is a prompt constraint, not constrained decoding. --approval allow and --yolo install a process-local auto-grant because dsh never auto-denies and headless has no UI. --yolo does not set danger-full-access; use --sandbox danger-full-access for unrestricted tools.

headlessStartup is { task, cwd, permissionMode, approvalPolicy, autoApprove, toolsMode?, format }. sandbox-policy, approval, and tools inject that service (!!js ctx.headlessStartup.*). Stock headless-runner still only reads task.

Non-goals

MCP, Host, HTTP, resume, --image, forking headless-runner. Put MCP on the profile patch + process env, not on this CLI.

Develop

Node ≥ 22.19. WASM is committed; prepare skips rustc.

npm test
DSH_BIN=/path/to/dsh npm run acceptance