dsh-searxng-search
Keyless web_search for DeepSeek Harness via a self-hosted SearXNG instance. A thin ctx.web provider (searxng-local) plus on-demand podman start — no API key, no MCP hop. The start is detached into a transient systemd scope, so restarting dsh-web cannot kill the container's port-forward (see why).
Install
One-liner for podman (system dependency, outside pnpm):
sudo apt update && sudo apt install -y podman
Install the bundle into your web profile (pnpm only — no profile hand-editing):
dsh plugin --profile web add "github:anthony1x6000/dsh-searxng-search#main"
Verify the layer without booting:
dsh --profile web --dump-config | grep -B 1 -A 4 searxng
Then restart dsh web for the new bundle layer to take effect. Build artifacts (lib/) are committed, so installing from git runs no build and needs no pnpm allowBuilds approval. The first search creates the container if missing (~1–2 min for the image pull), later ones start it on demand after reboots.
Why the container start is scoped
podman run / podman start inherit the cgroup of whoever invokes them. Installed as a bundle that caller is dsh-web.service, so the container's conmon and pasta — the processes that actually own the 127.0.0.1:8888 forward — end up inside that service's cgroup. A systemctl --user restart dsh-web then tears the cgroup down with the default KillMode=control-group: SearXNG keeps running, but nothing listens on the port any more, so web_search reports the provider down while podman ps still shows the container Up.
Every container-mutating CLI call is therefore wrapped in a transient per-container scope:
systemd-run --user --scope --collect --quiet --unit dsh-searxng-.scope podman …
A scope is a sibling of dsh-web.service under user@1000.service/app.slice, so restarting DSH never touches it. Inspect it with systemctl --user status dsh-searxng-searxng.scope. Where no user systemd manager is reachable (macOS, a bare CI container, systemd-run missing) the plugin calls the CLI directly and logs a one-line warning; SEARXNG_DETACH=0 forces that behaviour. searxng-podman.sh uses the same wrapper.
The bundle pins searchProvider: searxng-local in its own layer because the base bundle defaults to deepseek-official (hosted API, needs a key) — without the pin, searches never reach your local instance. Your profile patch can still override it. To go back to hosted search, dsh plugin --profile web remove dsh-searxng-search (then restart) and the layer — pin included — is gone.
Local development: pnpm install && pnpm build, then dsh plugin --profile web add link:/path/to/dsh-searxng-search.
Files
| File | Role |
|---|---|
index.ts |
The plugin source: WebSearchProvider over GET {base}/search?format=json, on-demand podman start scoped out of the caller's cgroup |
lib/index.js |
Built artifact (pnpm build), committed — what installed profiles actually load |
package.json |
Bundle manifest (dsh.bundle.patch); build-only devDeps, zero runtime deps |
cordis.patch.yml |
Bundle layer: package-name entry + searxng-local provider pin |
searxng-podman.sh |
One-shot SearXNG setup: settings.yml (secret + json format), container run, health check |
selfcheck.ts |
Mapping + URL-gating + packaging asserts: pnpm selfcheck |
Config
| Env / row field | Default | Meaning |
|---|---|---|
SEARXNG_BASE_URL / baseURL |
http://127.0.0.1:8888 |
SearXNG base; /search is appended |
SEARXNG_CONTAINER / containerName |
searxng |
Container started on demand |
SEARXNG_CLI / containerCLI |
podman |
podman, docker tried as fallback |
SEARXNG_START_TIMEOUT_MS / startTimeoutMs |
60000 |
Start + readiness budget per search |
SEARXNG_AUTO_CREATE / autoCreate |
true |
Create the container when missing (set 0/false to start-only) |
SEARXNG_IMAGE / image |
docker.io/searxng/searxng:latest |
Image pulled on first creation |
SEARXNG_PORT / port |
8888 |
Host port the container binds |
SEARXNG_DETACH / detach |
true |
Start the container in a transient systemd --user scope so it leaves the caller's cgroup (0 calls the CLI directly) |
Autostart only manages loopback bases (localhost, 127.0.0.1, ::1); remote bases fail fast. Redirects are rejected (redirect: 'error'); SearXNG returns no generated answer, so results carry sources only.
