dsh-pet-in-frame
DeepSeek Harness 的代理桌面宠物叠加层：通过静态姿势或逐帧动画姿势响应代理状态和工具调用。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-pet-in-frame@0.6.0Pet In Frame 以 dsh-pet-in-frame 发布，当前版本为 0.6.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A single-package DeepSeek Harness plugin: a small desktop pet that floats in the corner of the Web GUI and reacts to what the agent is doing — thinking, searching the web, reading code, running commands, editing files, hitting an error, or idling.
One package = one bundle = one loader row. The host half listens to the agent lifecycle (agent/status, tools/execute, tools/result, agent/error), watches the assets/ directory for image and manifest.json changes, and serves five HTTP routes (/dsh-pet-in-frame/state, /texts, /frames/<action>, /assets/<file>, /wake). The browser half registers the pet into shell.overlay, polls state, and swaps static poses or cycles frame animations.
The differentiator: the assets directory is the config. Drop a bash.png into assets/ and the pet shows it whenever a command runs; add a frame list plus one manifest.json line and you get an animation. No rebuild, no code change — changes hot-reload within a few seconds.
Explore screen — the pet idles in the bottom-right corner:
In a conversation — the pet switches poses with the agent's activity and shows a status bubble on click:
Poses switch with agent state and tool calls:
| Action | Triggered by | Image |
|---|---|---|
think | running, no active tool | think.png / {imgs:[...],delay} |
search | web_fetch | search.png |
learn | web_search, skill | learn.png → default.png |
read | read, glob, grep | read.png → search.png → default.png |
bash | bash, pwsh | command.png / command.webp, plus a pinned bash_comp / pwsh_comp badge |
edit | edit, write | edit.png → default.png |
plan | todo_*, goal tools, workflow, and goal/changed events (brief flash) | plan.png / plan.webp |
ask | ask_user_question | ask.png → default.png |
subagent | subagent/start, subagent/end events; subagent / subagent_fork / send_message / interrupt_agent / report tool dispatch | subagent.png → default.png |
cordis | cordis_define, cordis_run, cordis_stop, cordis_undefine, cordis_inspect_* | cordis.png / cordis.webp |
permission | while an approval request is pending (sandbox escalation / sandbox_permissions) | permission.png / permission.webp |
error | ; tool results with (failed operations, a failed subagent run); / non-zero exit (foreground result or background job); cancelled calls are excluded |
Subagent working indicator: while any subagent is running, a badge pinned inside the pet image (top-left at ~4.63% / 31.66% of the pet, sized ~26.63% × 36.92% — matching a 334×463 badge on a 1254×1254 pet sprite) cycles the subagent_comp1~3 frames (500 ms interval). Because it lives inside the pet's own container it follows drag and resize. At start/end the pet flashes the subagent pose for ~2 s, and the badge lingers ~2 s after the last subagent settles so both disappear in sync. The pose flash and the running indicator are independent. If the main agent calls a tool during the flash, the pet switches to that tool's pose immediately (the flash is cancelled); pure thinking keeps the subagent pose until the window ends, then switches to think.
Subagent isolation: tool calls, status and error events originating from live subagents are ignored — only the main agent's activity drives the pet pose. So a foreground subagent keeps the pet in the subagent pose for its whole run, and a background subagent never steals the pose from the main agent's ongoing work.
Command badge: while the main agent runs a foreground bash / pwsh tool, the pet switches to the command pose and pins the matching badge inside the pet image — bash → bash_comp1~2 frames, pwsh → pwsh_comp1~2 frames (500 ms interval; badge top-left at 74.24% / 45.61% of the pet, sized 15.95% × 15.95%). The badge and the pose stay in sync: both hold for ~1.5 s after a quick command settles (MIN_TOOL_SHOW_MS), so even a sub-second command shows its badge. Shell calls from inside subagents never trigger it (subagent isolation).
Background commands: when the main agent dispatches bash / pwsh with run_in_background: true, the pet mirrors the subagent pattern — it flashes the command pose for ~2 s and pins the matching comp badge (bash_comp / pwsh_comp) while the job runs; when the job settles (done or killed) it flashes command again for ~2 s, and a job that failed (explicit failed status, or a completed job with a non-zero exit code) shows the error pose for ~6 s instead. The badge lingers ~2 s after the last job so both disappear in sync. Job lifecycle comes from ctx.jobs.onJobDone (deployments without the jobs service skip background tracking; the foreground badge is unaffected).
Fast state polling: the client polls /state every 250 ms, so pose switches land within ~0.25 s (average ~0.125 s). Each tool pose (and its comp badge) is additionally held for ~1.5 s (MIN_TOOL_SHOW_MS) as a readable display duration, so even a sub-second tool call is visible.
Idle sleep: after 30 s of continuous idle the pet falls asleep (sleep frames, 500 ms interval) and stays asleep; click the pet to wake it back to default and restart the countdown — 30 s later it sleeps again. Any tool call, thinking, or error also interrupts sleep and restarts the timer.
Two ways to configure each action: "action": "image.png" (static) or "action": { "imgs": [...], "delay": 1000 } (frame animation, delay in ms, default 500).
Hot reload: edit an image or manifest.json and the pet updates within ~3–4 s. No restart.
Draggable with viewport clamping, hover −/+ size control (100–320 px), click for a status bubble, × to hide / 🐾 to bring back.
No build step: plain JavaScript (Node ESM host + window.__ModuleLoader__ client).
Prereqs: a running dsh web profile, pnpm available to the dsh CLI.
From npm (recommended):
dsh plugin --profile web add dsh-pet-in-frame
From GitHub (pin a commit for auditability):
dsh plugin --profile web add github:HarmlessFunny/dsh-pet-in-frame#<commit-sha>
From a local checkout (development):
dsh plugin --profile web add file:E:/path/to/dsh-pet-in-frame
Then restart the dsh web process — new plugin entries load at boot. The package's cordis.patch.yml inserts its own loader row, so no manual cordis.yml edits are needed.
Images live in the package's assets/ directory. You can point elsewhere with the DSH_PET_ASSETS environment variable or a config.assetsDir on the loader row:
- id: dsh-pet-in-frame
name: dsh-pet-in-frame
config:
assetsDir: C:/path/to/your/assets
Convention mode (zero config) — assets/<action>.<ext>, extensions png/jpg/jpeg/gif/webp/svg:
assets/
├── default.png # fallback
├── think.png # thinking
├── search.png # web search / code reading
└── bash.png # optional; falls back to default.png
manifest.json — explicit control, including animations:
{
"default": "default.png",
"think": { "imgs": ["think1.png", "think2.png", "think3.png"], "delay": 1000 },
"sleep": { "imgs": ["sleep1.png", "sleep2.png", "sleep3.png"], "delay": 500 },
"bash": "bash.png",
"read": "search.png"
}
{imgs, delay} = frame animation.Asset optimization: originals are usually 1254px+ while the pet displays at most 320px — shipping them raw slows first load noticeably. Run scripts/optimize-assets.cjs to downscale to 512px WebP (~1/10 the size) in place, then update manifest.json to the .webp names.
The host half exposes five loopback routes:
| Route | Returns |
|---|---|
/dsh-pet-in-frame/state | { status, action, tool, error, idleMs, rev } |
/dsh-pet-in-frame/texts | bubble copy for each action |
/dsh-pet-in-frame/frames/<action> | { frames: [url...], delay } (URLs carry the current rev) |
/dsh-pet-in-frame/assets/<file> | image bytes |
/dsh-pet-in-frame/wake | POST; wake the pet and reset the idle countdown on click |
The client polls state every 250 ms and refetches frames whenever rev changes, so pose switches are near-instant and image/manifest edits propagate without a reload.
[\w.-]+, so no path traversal.JSON.parse and treated as data only; a bad manifest degrades to the convention mode.POST /wake, which only resets a timer (no request body read, nothing written).~/Desktop/assets, copy those files into the package's assets/ (or set DSH_PET_ASSETS).MIT. Copyright (c) 2025 dsh-pet-in-frame contributors.
agent/errorisErrorbashpwshoops.webp |
idle / default | nothing active | default.png |
sleep | after 30 s of continuous idle | sleep.png / {imgs:[...],delay} |