@picgo/dsh-plugin

Upload images and files to your image host from DeepSeek Harness, powered by PicGo.

Harness can show your agent a screenshot, but it has no way to turn a local file into a link. So when the agent writes a README, renders a chart, or captures a screenshot, the image stays on disk and ![](./out.png) becomes a dead link the moment you push. This plugin closes that gap.

It uploads through whatever image host you already configured in PicGo — PicGo Cloud, GitHub, S3, Tencent COS, Qiniu, or any third-party uploader plugin you installed. Nothing to re-configure. If you've never used PicGo, it walks you into PicGo Cloud's free tier.

If the PicGo desktop app is running, uploads go through it, reusing the image host you set up in its window. Otherwise they run in-process. See Upload routes.

Install

dsh plugin --profile web add @picgo/dsh-plugin

Then boot as usual:

dsh --profile web

What you get

picgo_upload — a tool the model calls on its own when a local file needs to become a link. Returns structured results, so in Code Mode you can use it directly:

const { uploaded } = await tools.picgo_upload({ paths: ['/tmp/chart.png'] })
console.log(uploaded[0].imgUrl)

/picgo — a command that uploads without spending a model turn:

Command	What it does
/picgo	Upload the clipboard image
/picgo <path>...	Upload one or more files
/picgo status	Show the active host and sign-in state
/picgo login [token]	Sign in to PicGo Cloud
/picgo logout	Sign out

A bundled skill that teaches the model when to upload — inserting a screenshot into docs is the primary case — and when not to (you named a specific destination, you want a local copy).

Upload routes

There are two ways this plugin can reach your image host, and it picks one per upload:

1. The PicGo desktop app, when it is running. It exposes a local upload server, and using it means your uploads honour the host you configured in the app's window.
2. In-process PicGo, otherwise — the picgo library reading ~/.picgo/config.json.

This matters because the two read different config files. The desktop app stores its settings in the system application-data directory (on macOS ~/Library/Application Support/picgo/data.json), while the library reads ~/.picgo/config.json. If you have only ever used the GUI, the library route sees a config you never touched — most likely empty and defaulting to PicGo Cloud. Preferring the app is what makes "it just uses my existing setup" true.

Upgrading from 0.1.x? If the desktop app is running and the two configs point at different hosts, your uploads will now land on the app's host rather than the CLI's. Set gui.mode: off to keep the old behaviour.

/picgo status tells you which route is active. Two things worth knowing about the desktop-app route:

• The app copies each URL to your clipboard and shows a notification — that is the app's own behaviour and this plugin cannot turn it off.
• If an upload there fails with a login error, sign in from the app's own window. /picgo login writes the PicGo CLI config, which the desktop app does not read.

First run

If you've never configured PicGo, uploads default to PicGo Cloud, which needs a one-time sign-in. The free tier covers casual use.

/picgo login

That opens your browser and reports back when it completes. If you already have a token from the PicGo Cloud dashboard, /picgo login <token> is instant.

The model will never run this for you: with no token the sign-in blocks waiting on a browser callback, which would hang the session. It relays the instruction and waits.

Already using GitHub, S3, or another host in PicGo? None of this applies — your existing config is used as-is and no sign-in is involved.

Configuration

Every field has a working default. Override them from your profile's cordis.patch.yml:

- id: picgo
  name: '@picgo/dsh-plugin'
  config:
    silent: true
    timeoutMs: 120000

Field	Default	Meaning
configPath	''	PicGo config file; empty uses PicGo's own default (~/.picgo/config.json)
silent	true	Suppress PicGo's console output and its picgo.log writes
timeoutMs	120000	How long to wait for one upload
registerSkill	true	Register the bundled picgo-upload skill
registerCommand	true	Register the /picgo command
announceSignIn	true	On startup, point a signed-out PicGo Cloud user at /picgo login
gui	see below	How to reach a running PicGo desktop app

gui controls the desktop-app route described in Upload routes:

Field	Default	Meaning
gui.mode	auto	auto uses the app when it answers; off never tries; only requires it rather than uploading to a different host
gui.host	127.0.0.1	Where the app's upload server listens
gui.port	36677	"
gui.secret	''	Auth secret, if you enabled one in the app. Empty falls back to $PICGO_SERVER_SECRET
gui.probeTimeoutMs	1500	How long to wait for the app to answer a heartbeat
gui.probeTtlMs	5000	How long a heartbeat result is reused, so a multi-file upload does not re-probe per file
gui.timeoutMs	0	Upload deadline for this route; 0 inherits timeoutMs

A patch replaces a row's entire config rather than merging keys, so restate every field you want to keep. Within gui, unset keys still fall back to the defaults above — overriding gui.mode alone is fine.

Notes

Uploaded links are public. Anyone with the URL can open it, and a deleted file may stay cached. Fine for screenshots and doc images; think before uploading a contract PDF or an internal archive. The bundled skill tells the model to confirm first for anything that looks sensitive.

Your PicGo config is treated as read-only, with one exception outside this plugin's control: when PicGo Cloud rejects a stored token, PicGo itself clears it from ~/.picgo/config.json. Signing in and out through /picgo login / /picgo logout also writes the token, as you'd expect.

Clipboard uploads need a desktop session and are only reachable through /picgo — the model is never given a way to upload your clipboard, since it can't know what's on it.

Falling back to the in-process route is deliberately narrow. If the desktop app stops between the check and the upload, the upload is retried in-process. But a rejected upload, a missing auth secret, or a timeout is reported rather than retried elsewhere: the app may have already accepted the file, and silently re-uploading would put a second copy on a host you did not pick.

If an upload through the app hangs, check whether "rename before upload" is enabled in its settings — that opens a dialog and waits for a human, which nothing is going to answer in an agent session. The timeout message says so too.

Development

pnpm install
pnpm build
pnpm test

pnpm test:live additionally exercises the desktop-app route against a running PicGo app. It is kept out of pnpm test because it uploads real files to your real image host.

To run it against a dsh source checkout without packaging, write a cordis.dev.yml (gitignored — the path is specific to your machine):

- insert:
    - id: picgo
      name: '/absolute/path/to/dsh-plugin/lib/index.js'

Then, from the dsh checkout:

pnpm dsh web --patch /absolute/path/to/dsh-plugin/cordis.dev.yml

The path must be absolute: a patch adds config but does not move the loader's resolution root.

Releasing

@picgo/bump-version bumps the version, writes the changelog, commits, and tags in one step:

pnpm release          # patch: 0.1.0 -> 0.1.1
pnpm release:minor    # 0.1.0 -> 0.2.0
pnpm release:major    # 0.1.0 -> 1.0.0
pnpm release:beta     # 0.1.0 -> 0.1.1-beta.0
pnpm release:dry      # print what would happen, change nothing

Then push the tag — that is what triggers publishing:

pnpm push-release

The release workflow runs typecheck, tests, and build before publishing, and refuses to publish if the tag does not match package.json. Prerelease tags pick their own dist-tag (-beta.x → beta, -alpha.x → alpha, anything else prerelease → next), so npm install @picgo/dsh-plugin never resolves to a prerelease.

npm authentication

npm cannot configure a trusted publisher for a package that does not exist yet, so the first release and every later one authenticate differently.

First release — needs an NPM_TOKEN repository secret (a granular token with publish rights to the @picgo scope):

gh secret set NPM_TOKEN --repo PicGo/dsh-plugin

After that first release lands, switch to trusted publishing so no long-lived token is involved. On npmjs.com, open the package → Settings → Trusted Publisher, and register:

Field	Value
Publisher	GitHub Actions
Organization or user	PicGo
Repository	dsh-plugin
Workflow filename	release.yml (filename only, not a path)
Environment name	leave empty
Allowed actions	npm publish

The workflow already sets id-token: write, so nothing changes on this side — npm picks OIDC over the token automatically. Once a trusted-publish release succeeds, delete the NPM_TOKEN secret and revoke the token, then set Settings → Publishing access to "Require two-factor authentication and disallow tokens".

Trusted publishing needs npm ≥ 11.5.1, so the release workflow runs on Node 24 (which ships npm 11.x). Node 22 ships npm 10.x and fails with a misleading 404. That choice affects only the machine doing the publishing — the package itself still supports Node ^22.19.0 || >=24.0.0, and CI tests against 22.

Compatibility

Tested against DeepSeek Harness 0.1.1-rc.2 (2026-08-22) and PicGo Core 3.0.1. Requires Node ^22.19.0 || >=24.0.0.

The desktop-app route needs a PicGo app new enough to expose the local upload server. If yours is older, gui.mode: auto simply never finds it and everything runs in-process as before.

Harness is a developer preview and its APIs change often. If a release breaks this plugin, please open an issue.

License

MIT