DeepSeek Harness Provider
DeepSeek Harness as a session provider: DeepSeek's OWN agent harness CLI (dsh) driven over the Agent Client Protocol, how it differs from the existing DeepSeek model provider, how to install the CLI and what gates a spawn, and what it does and does not do in v1.
On this page
Two different DeepSeek things, and only one of them is new.
DeepSeekruns DeepSeek models inside the Claude Code harness.DeepSeek Harnessruns DeepSeek's own agent harness — a separate product with its own tools, skills and runtime modes. If you are looking for cheaper DeepSeek models, you want the first one.
What it is
DeepSeek Harness is DeepSeek's own agent harness, published as the dsh CLI. Omniscio runs
it as a session provider the same way it runs Gemini and Kimi Code: a kept-alive child process
speaking the Agent Client Protocol (ACP) over stdio.
- The engine spawns
dsh --profile acp. - It is a persistent-external engine — one long-lived process per session, not one process per turn.
- Sessions stream as they work, and a second message continues the same conversation.
Verified live against dsh 0.1.5-rc.2: the handshake, creating a session, sending a prompt,
streaming the answer back, and a clean end_turn finish.
Where to find it
Enabling it
Two things must be in place, and Omniscio reports the first gap it finds:
- Settings → Accounts → Allow DeepSeek Harness sessions — the opt-in toggle, off by default. Nothing spawns until you turn it on.
- The
dshCLI on your PATH. The account card links to the install page when it is missing. - A DeepSeek API key in Settings → Accounts. This is the same key the existing DeepSeek provider uses — you do not add a second one.
DeepSeek Harness is a developer preview. DeepSeek ships it as
dsh0.1.x and says so. It is offered here as a normal pickable engine rather than hidden behind a flag, so treat it as preview-quality; the opt-in toggle above is always your off switch.
Which version to install
Use dsh 0.2.1-rc.2. Measured on a fresh machine (2026-10-21): 0.1.5-rc.2 cannot boot its
ACP profile at all — it stops with plugin tree failed to load: @deepseek-ai/dsh-sandbox-local —
while 0.2.1-rc.2 initialises, creates a session and completes a turn. That is the version the
app's install recipe pins (ENGINE_TOOLS.dsh in src/shared/providers/engine-tool-recipes.ts,
@deepseek-ai/dsh@0.2.1-rc.2), which is also the recipe the conformance real-engine lane installs
through — so the suite and the product run the same build.
How it behaves
Choosing a model
Pick a model and a reasoning effort on the session, like any other engine. dsh selects both through
the protocol session rather than a command-line flag, so Omniscio asks the engine what it offers
on the session it just created and sets the choice there — the picker only ever sends a value the
running dsh advertised. The four models offered are the same four the DeepSeek provider offers
(V4.1 Flash is the default), and the effort offers Low, High and Max.
The choice takes effect when the session starts. Changing it on an running session is saved for
the next one: the dsh process is kept alive for the session's whole life, and no Omniscio engine
re-configures a running child.
If a choice cannot be applied — the engine no longer offers that model, or it refuses the value — the session still starts on the engine's default and says so in the transcript.
Cost
Turns on this engine report no cost figure. The protocol sends a usage frame, but it carries
context-window occupancy — how full the conversation is — not a billable per-turn input/output
split, so there is nothing to price. Omniscio surfaces this as "cost not reported" rather than a
misleading $0.
A slow first start is normal
The first time a dsh session starts, the CLI resolves its plugin profile — a few hundred
modules. On a busy machine that measured over two minutes. Omniscio deliberately allows a
long first start rather than declaring the engine broken, and later starts are fast.
Running out of room
If a turn is rejected because dsh's own working memory outgrew the model's window (usually a
command that printed a huge amount of output), the message says it ran out of room and that the
conversation is saved. Omniscio has already replaced the overflowed dsh process, so your next
message starts fresh on a compact copy of the conversation — usually no new session is needed.
If it still runs out of room, the conversation may be too long for the model, so start a new session.
What v1 does not do
Stated plainly, because each is a real limit rather than an oversight:
- No image input. dsh itself reports it does not accept images.
- A server whose command cannot be resolved to a real, spawnable path. The project's MCP servers do ride the session, but this engine refuses the WHOLE session over one server it cannot start, so Omniscio resolves each command first and holds back only what it cannot — naming those servers in the transcript rather than losing the session, or the other servers with it.
- No SSH / remote sessions.
- No automatic resume after a crash. The protocol offers resume, but it has not been proven here, so dsh is deliberately left out of the crash-recovery path rather than promising a resume that might not arrive.
- No Claude Code guardrail hooks. Permission requests follow your permission level through the
shared ACP gate, so a call is auto-approved, put to you as an approval card, or denied because it
touches a protected path — that last one at every level, Full trust included. None of that is a
guardrail, which is why a DeepSeek Harness session in a git project is isolated in its own
worktree by default. On a Windows computer that changes code without worktrees (the patch road),
it runs inside a Windows write sandbox instead: Windows lets it change files only in its own
session folder (the session shows a line naming that folder), it starts with a fresh copy of its
profile rather than your
~/.dsh, and it cannot land work itself — its diff is landed from that folder. Off Windows it is refused on the patch road.
Related
Kimi Code and Gemini are the other engines that speak the Agent Client Protocol the same way, so their pages describe the same session behaviour. DeepSeek covers the DeepSeek models provider — the different thing this page opens by warning about — and Who pays & who serves covers who pays for those model sessions and which company serves them. The harness is not on that list: it always runs on your own DeepSeek key.
Last verified 2026-10-06