@raullenchai/dsh-provider

A native Rapid-MLX provider for
DeepSeek Harness — so dsh
gets its model facts from the server instead of from whatever you typed into
settings.yaml.

Status: published to npm as
@raullenchai/dsh-provider.
The end-to-end dsh run in Verified was on an M3 Ultra against
dsh 0.1.0-rc.7; dsh 0.1.0-rc.8 is API-compatible — the LlmAdapter
contract is byte-identical and the only changes are additive — and the
adapter is re-verified against rc.8 at the protocol and unit-test level.
DSH is still a developer preview that moves fast, so treat this as tracking
a moving target, not a frozen compatibility promise.

What it does for you

DSH can already talk to a local Rapid-MLX server through its generic
openai-completions provider. That route works — but it knows nothing about
your model beyond what you hand-wrote:

# what the generic route makes you maintain, by hand, per model
llm-pi-ai:
  providers:
    rapid-mlx:
      baseURL: http://localhost:8000/v1
      defaultContextWindow: 262144      # you looked this up. is it still right?
      models:
        - id: qwen3.6-35b-8bit
          contextWindow: 262144
          reasoningEfforts: {off: none, low: low, medium: medium, high: high}

Rapid-MLX's /v1/models already publishes all of that and more. This adapter
reads it, so:

1. Nothing to hand-write, and nothing to re-write when you switch models.
Swap what rapid-mlx serve is running and dsh follows. No re-running setup,
no stale numbers.

2. The reasoning control tells the truth. Rapid-MLX reports whether a model
actually has a reasoning parser. A model that can't reason no longer shows an
off/low/medium/high selector that does nothing.

3. Compaction is timed with the capacity that actually fits this Mac, not a
number that drifted. This is the one that quietly costs you.
dsh-compaction-basic asks the provider for the route's capacity and compacts
at thresholdRatio × capacity (0.8 by default). The provider prefers the
server's max_model_len — Rapid-MLX's memory-fitted ceiling (what fits in
unified memory: weights + KV cache), in the vLLM/SGLang-standard field — over
the native context_window, and falls back to context_window on an older
server that doesn't report it. So compaction is timed to what the machine can
actually hold, not the model's advertised window (which it may not have room
for) and not a hand-written number copied from another model.

Install

Needs Node ≥ 22.15 (dsh imports Node's Zstd stream API without declaring it)
and a running Rapid-MLX server.

# From npm:
dsh plugin --profile web add @raullenchai/dsh-provider

# …or straight from source — the package ships plain JS with no build step:
dsh plugin --profile web add github:raullenchai/rapid-mlx-dsh-provider

export RAPID_MLX_BASE_URL=http://localhost:8000/v1     # optional; this is the default
dsh web

Then point the agent at the route:

# $DSH_HOME/settings.yaml
agent-default-model:
  provider: rapid-mlx
  model: qwen3.6-35b-8bit

Verified: that command installs and activates as a profile layer against
dsh 0.1.0-rc.7. To hack on it locally instead, see
Local development.

Model management (v0.2.0)

Beyond the provider route, the plugin registers five tools and a /rapid-mlx
command so the agent can see and manage models without leaving the session.
The split follows which surface actually owns each fact: served-model facts
come from the structured HTTP /v1/models; the download cache and pull/remove
are CLI-only, so those — and only those — shell out to rapid-mlx through the
harness subprocess seam.

Tool	Source	What it does
rapid_mlx_serving	HTTP /v1/models	The model(s) served right now, deduped, with context window, reasoning/tool parsers, MoE/hybrid, and modalities.
rapid_mlx_cached	rapid-mlx models --cached	Downloaded models and their on-disk size.
rapid_mlx_pull	rapid-mlx pull <name>	Download a model (alias or HF repo id). Cancellable; no fixed deadline.
rapid_mlx_remove	rapid-mlx rm -y <name>	Delete a cached model to free disk.
rapid_mlx_health	HTTP + rapid-mlx --version	API up? CLI reachable? Reported as two independent facts.

/rapid-mlx prints a one-shot overview: health, the served model and its facts,
and total cache disk usage.

The CLI is resolved from the cliCommand config (default rapid-mlx on
PATH, or $RAPID_MLX_CLI); set it to an absolute path if the binary is not on
the harness's PATH. rapid_mlx_pull/rapid_mlx_remove are the only tools
that change anything on disk, and they run non-interactively (rm is forced
with -y because the subprocess seam ignores stdin).

Verified

Against dsh 0.1.0-rc.7 on an M3 Ultra:

• Installs and activates as a profile layer (no "declares no dsh.bundle"
  warning; the entry shows up in dsh --profile headless --dump-config).
• Registers the rapid-mlx route with ctx.llm and serves real queries.
• Plain chat, a single tool call, and the multi-step bug-fix task that gates
  Rapid-MLX releases — the last one fixed the bug and made the target repo's own
  test pass, verified independently, in 36 s on qwen3.6-35b-8bit.

Not done yet

Being explicit, because the point of the adapter is to use what the server
says and some of it is still only read:

• recommended_sampling — should be applied automatically per model.
• tool_call_parser — should let dsh fail fast on a model that cannot emit
  tool_calls, instead of looping.
• is_hybrid / is_moe / capabilities — read, not yet acted on.
• Memory-aware capacity. Today resolveModel() reports the model's
  advertised context window. On a Mac the real ceiling is unified memory, and
  reporting that instead is the biggest remaining win — it needs Rapid-MLX to
  expose a usable-capacity figure first.
• Images are not carried through stream() — but they now refuse with
  LlmError(..., 'UNSUPPORTED') rather than being dropped, per the cookbook.
  Text, reasoning and tool calls are carried.
• The route is registered as rapid-mlx. If your settings.yaml also declares
  a rapid-mlx provider under llm-pi-ai, the two compete for one route name
  (registerAdapter owns provider exclusivity). Use one or rename ours.

Conformance with the official adapter contract

Built against
docs/cookbook/adding-an-llm-adapter.md
and its "protocol obligations" section. Each item has a test:

Obligation	How it is met
usage before finish, nothing after finish	usage is buffered and flushed at end-of-stream, so a trailing usage-only chunk cannot reorder it
Tool-call arguments are raw JSON strings end to end	fragments stream as argumentsDelta and reassemble unparsed
Block indexes in first-seen order, reused per block	verified across a reasoning-then-text response
Errors take exactly two sanctioned paths	transport/protocol failures throw LlmError with a stable code; nothing ends the stream quietly
Honor options.signal	passed to fetch and to the SSE reader; an AbortError is re-thrown unchanged, not reclassified
A field the provider cannot honor throws UNSUPPORTED	image content refuses instead of being narrowed away
Config is a schemastery schema with env fallback	export const Config, fed from cordis.patch.yml via !!js process.env.RAPID_MLX_BASE_URL

finish.replayState is not emitted: Rapid-MLX needs no native response ids
or signatures on follow-up calls, so there is nothing lossless to project.

Three things worth knowing before you edit this

Each of these cost real debugging time:

1. dsh.bundle in package.json is what makes this a plugin. Without it
   the package installs as an inert dependency and dsh only warns. It is
   also what gets it appended to the profile's dsh.profile.bundles. CI fails
   if it goes missing.
2. LlmReasoningEffortInfo.name is required. Returning {id} alone fails
   the whole model with INVALID_MODEL_REASONING — an error that names the
   model, not the missing field.
3. DSH has no tool role. Message.role is only system|user|assistant; a
   tool result is a user-role message whose source.kind === 'tool'
   carries the callId and whose content holds a ToolResultBlock. Flatten
   those into plain user text and the model reissues the same call forever —
   the symptom is an empty answer and a non-zero exit, with nothing on
   stderr.

Local development

pnpm links a local path outside the profile tree, so Node's parent-walk
never reaches $DSH_HOME/profiles/node_modules and the peer deps fail to
resolve. Symlink them in — dev only, node_modules is gitignored and excluded
from the published files:

mkdir -p node_modules/@deepseek-ai
ln -sfn <dsh-install>/node_modules/@deepseek-ai/dsh-llm node_modules/@deepseek-ai/dsh-llm
ln -sfn <dsh-install>/node_modules/@deepseek-ai/cordis  node_modules/@deepseek-ai/cordis

export DSH_HOME=/tmp/dsh-dev          # never your real ~/.dsh
dsh plugin --profile headless add "$PWD"
export RAPID_MLX_BASE_URL=http://127.0.0.1:8000/v1
dsh --profile headless "say hello"

A real npm install needs none of this: the package lands inside the profile
tree, where the flat fallback resolves bare names normally.

When testing agent behaviour, use a strong 8-bit model. A multi-step task here
failed on qwen3.5-9b-4bit and passed on qwen3.6-35b-8bit — 4-bit confounds
"weak model" with "broken integration".

The engine side guards these fields

Living in its own repo means a rename in Rapid-MLX would break this package
silently — nothing there imports it and this CI does not run there. So the
fields are pinned on the side that owns them, by
tests/test_model_card_client_contract.py in
Rapid-MLX, which names this package
as its reason. It pins the wire shape: field names, nullability, and the fact
that ModelInfo does not set exclude_none — which is what makes
"reasoning_parser": null distinguishable from an older server that omits the
key entirely.

If you start reading a new /v1/models field here, add it there too.
Otherwise the guard silently stops covering what this package actually uses.

License

Apache-2.0, matching Rapid-MLX.