dsh-zed-provider

Run your Zed AI subscription (Zed Pro / Student — the monthly prompt credits)
inside DeepSeek Harness. The bundle registers one LLM provider route,
zed-agent, whose turns are minted and streamed through Zed Cloud
(https://cloud.zed.dev) exactly as the Zed editor does it.

Install

dsh plugin --profile web add dsh-zed-provider

Link your Zed account

Type /zed-login in the Harness composer. The command opens Zed's native-app
sign-in page, waits for the loopback callback, and stores the resulting grant in
the Harness credential store (llm-zed/zed-agent). Later requests reuse it, and
the short-lived gateway token is minted automatically and refreshed on demand.

Credentials are resolved in this order, so any one of them is enough:

1. config.credentials — a literal " " (or JSON) value.
2. config.apiKeyEnv (default ZED_CREDENTIALS) — resolved through the
   credential seam. The Models settings page shows this as the Zed card's key
   field, so a credential pasted there works without /zed-login.
3. The llm-zed/zed-agent grant record written by /zed-login.
4. The Zed editor's own local keyring entry, on macOS (zed.dev) and Linux
   (secret-tool), imported once into the grant record — sign in inside Zed and
   this plugin picks it up by itself.

apiKeyEnv is spelled that way because the Models page reads exactly that field
name; its value is a credential reference, and what it holds is the
" " pair Zed issues.

Settings UI

Settings → Models → the Zed card carries this bundle's own editor
(lib/client.js, a client-module half registered on the page's
settings.models.provider-card seat for the llm-zed namespace):

• credential status, from the card's credential join plus a live
  credentials/describe;
• a credential field that stores " " under the reference
  apiKeyEnv names (read live from the settings document, default
  ZED_CREDENTIALS), with Remove to unset it;
• a refusal instead of a silent no-op when that reference is supplied by a
  read-only source such as an environment variable.

It also carries the model chooser: the Host registers
llm.registerModelDiscovery for this namespace, the page reads it through
llm/discoverModels, and the tick boxes write visibleModels through
settings/mutate. That field is volatile, so a tick changes the picker on the
next catalog read without a reload; an empty or complete selection unsets it,
meaning "offer every model this plugin can run".

baseUrl stays a composition-only field in cordis.patch.yml.

The client fiber hard-injects only slots, locale and remote, and picks the
credential remote up through ctx.inject(['remote.credentials'], …). That is
deliberate: the web boot asserts that every client entry activated, so a root
client plugin that waits on a service the settings UI keeps private (such as
settingsScope) takes the whole GUI boot down with
: pending (waiting for service: …).

Models

listModels reads GET /models with your credential and advertises the models
Zed reports for the backends this plugin implements — Claude (Anthropic
Messages) and GPT (OpenAI Responses). Zed's Gemini and Grok entries are
filtered out rather than offered broken; selecting one anyway fails fast with
UNSUPPORTED_CONTENT.

Each model carries Zed's own context window, output cap, and — when
supports_thinking — the effort levels off / low / high / max, mapped
onto Anthropic extended thinking budgets.

Configuration

Quota

Zed owns the accounting: no plan or no credits left answers HTTP 402, which
surfaces as a QUOTA failure naming https://zed.dev/account. The plugin does
not invent percentages or limits.

Development

pnpm install (which runs the prepare hook), then pnpm check —
typecheck, build, tests. Sources are TypeScript under src/; lib/ is build
output and is git-ignored, because the prepare/prepack hooks rebuild it
wherever the package is consumed and the published tarball carries the built
copy through the files allowlist.

A profile can install this bundle as a link: to a working copy, in which case
that copy's lib/ is what the Host imports: rebuild after changing src/, or
the Host keeps running the previous build.

node --test test/*.test.mjs covers the pieces that fail silently in
production: the NDJSON → StreamChunk translation for both backends
(src/frames.ts framing, src/parse.ts Anthropic events, src/responses.ts
Responses events), the RSA sign-in handshake (src/signin.ts), the request
projections (src/anthropic.ts, src/responses.ts), and the card's slot
registration, credential and model-selection wiring (src/client.ts, loaded
under a stubbed module loader in test/client.test.mjs). The tests import the
built lib/, because that is what the Harness actually runs. Only
src/index.ts touches the Harness runtime.

Two install-time facts worth knowing: the Host half is re-imported only by a
profile restart (ESM caches by URL), and the client half is composed from the
Loader's entries at Host boot, so adding or removing exports["./client"] takes
a restart too — after that it is served per request under a content hash, so
client-side edits reach the browser on the next load.