dsh-subagent-agy

Delegate tasks to Google Antigravity CLI (agy) — a separate Gemini-powered coding agent on your machine — straight from DeepSeek Harness. No Gemini API key is involved: agy authenticates through your Antigravity login (system keyring / Google Sign-In).

中文版见 README.zh.md · Source: github.com/ZEM17/dsh-subagent-agy

Installation

Requirements

1. agy installed: irm https://antigravity.google/cli/install.ps1 | iex (Windows). The plugin locates it automatically (PATH → AGY_BIN → %LOCALAPPDATA%\agy\bin → ~/.local/bin → ~/.gemini/bin → ~/go/bin).
2. agy logged in once: run agy in a terminal and complete the sign-in. If a run hits "authentication required", the plugin pops the interactive login window automatically.
3. DeepSeek Harness with the standard web profile (it ships dsh-subagent, dsh-subprocess, dsh-tools, dsh-jobs).

Install the plugin

From GitHub (recommended):

dsh plugin --profile web add github:ZEM17/dsh-subagent-agy

From a tarball: download dsh-external-dsh-subagent-antigravity-0.2.0.tgz from the releases page, then:

dsh plugin --profile web add /path/to/dsh-external-dsh-subagent-antigravity-0.1.0.tgz

That's it — no restart needed. The model in any session can now use the antigravity tools.

Usage

The model calls the tools automatically when you ask for it; you can also prompt it explicitly:

Example prompts:

"Use agy to rewrite the README of D:\projects\my-project and list the changes."

"用 antigravity 优化这个项目的页面" (in a Chinese session)

Image analysis

agy's model is multimodal — put an image path in the prompt and it will describe or analyze the image:

"Use agy to describe this image: C:\projects\design\mockup.png"

"用 antigravity 识别这张图片：C:\projects\screenshots\ui.png"

Images inside the session workspace work directly; images elsewhere need the directory in the tool's dirs parameter (or the addDirs config).

• Tasks targeting files outside the session workspace need the target directory: the tool takes a dirs parameter (the model passes it when you give it the path).
• Long tasks: the model uses background: true for long jobs — it keeps you posted on live progress and reports the final result when done.
• Continue a task: tell the model to continue using the task key from a previous call.
• Find a task again: antigravity_tasks lists running and recent finished tasks — task keys are recorded in ~/.dsh/agy-conversations.json and survive DSH restarts, so a key can always be recovered for follow-up.
• Stop a task: antigravity_cancel with the running task's key stops it (also possible via job_kill for background jobs).

Configuration

All fields are optional; sane defaults are shown. Configure through the plugin's settings section (e.g. $DSH_HOME/settings.yaml):

dsh-subagent-antigravity:
  model: gemini-3.1-pro-high   # agy models; empty = agy default
  effort: high                 # low | medium | high
  printTimeoutMs: 300000       # foreground ceiling (5 min)
  backgroundTimeoutMs: 3600000 # background ceiling (1 h)
  skipPermissions: true        # auto-approve agy tool calls (headless needs it)
  addDirs: []                  # absolute dirs added to the agy workspace (--add-dir)
  proxy: ""                    # proxy URL; falls back to AGY_PROXY, then ~/.dsh/agy-proxy.txt

How it works (short version)

Each task spawns a fresh agy -p <task> process in the session workspace and returns the final answer. Runs use --output-format json, whose reply carries agy's conversation_id; the plugin records it in a durable registry (~/.dsh/agy-conversations.json), and antigravity_followup resumes the same conversation via --conversation <id>. Background runs use stream-json for live progress and register with DSH's jobs system (job_output / job_kill). A local watchdog terminates hung runs, and all timeouts interrupt: nothing ever waits forever.

antigravity_tasks surfaces the durable registry (restart-safe) plus live in-flight runs, and antigravity_cancel stops a running one. The registry self-heals: a corrupted file is backed up once (*.corrupt) and reset instead of being silently overwritten, and all registry writes are serialized, so concurrent runs never lose each other's records.

Troubleshooting

Development

bash scripts/build.sh            # junction-link @deepseek-ai deps from the DSH runtime, then tsc src/ → lib/
npm run selftest                 # offline checks: agy presence, flag probe, registry round trip
npm run selftest:e2e             # + live PONG round trip and follow-up (needs an agy login)

Dev workflow on a local DSH (super-injector): dev_build_plugin (build + pack) → dev_inject_plugin (runtime injection, no restart); iterate with dev_build_plugin → dev_reload_package.