Skip to content

dsh-mistral-vibe

Verified

dsh-mistral-vibe Β· v1.1.1 Β· MIT Β· Web UI

Mistral Vibe in DeepSeek Harness: a self-contained Mistral La Plateforme provider with the current Vibe model catalogue, per-model thinking levels mapped to reasoning_effort, streaming reasoning blocks, and a settings page that stores the whole model list

Install

dsh plugin add dsh-mistral-vibe

Confirm the layer applied with dsh --profile default --dump-config β€” see the install guide.

Source

Tags

Readme

dsh-mistral-vibe

Mistral Vibe in DeepSeek Harness: a provider route for Mistral La Plateforme that ships the current Vibe model catalogue, maps each model's thinking levels onto reasoning_effort, streams reasoning back as a thinking block, and gives the route a real settings page.

Self-contained by design: no imports from @deepseek-ai/* or pi-ai, no install-time dependencies. The two schema helpers it needs are vendored under vendor/.

Features

  • One provider route, mistral-vibe β€” chat and streaming through POST /v1/chat/completions, with tool calls, usage, and providerRetryPolicy.
  • Thinking levels per model, taken from what the catalogue actually publishes: off β†’ reasoning_effort: "none", low/medium/high/max sent verbatim. A model with no ladder gets no level control and never receives an unsupported effort.
  • Reasoning streamed as a block, not as text: delta.content arrays of {type: "thinking"} become reasoning-delta, so the harness renders them as thinking.
  • A model list that is data, not code. One fetch writes the whole list (shipped rows + everything the API reported) into the row's modelsJson; the chat picker, the settings page and id resolution all read that store, so nothing has to be fetched twice.
  • One row per model. /v1/models lists every alias as its own row (64 rows for 13 models, type: "base" on all of them); discovery unions each row with its aliases array and emits one entry per model.
  • models.dev supplies the ladder, read field by field rather than guessed from ids: type for chattability, modalities.input for vision (attachment: true is also set on pdf/audio-only models), limit.* for the windows, reasoning_options for the levels.
  • A settings page at Settings β†’ Mistral Vibe: both credentials with an eye toggle, the endpoint, and the model list with its ladder, modality and window on every row.
  • Secrets stay out of the row file. A saved key goes to the credential store; the YAML keeps only the reference.
  • No quota fiction. Mistral exposes no readable allowance (/v1/billing/subscription returns a plan name and nothing else), so the page shows no meter. The route still captures rate-limit headers and refuses a call locally when a header already reported remaining: 0.

Install

Requires DeepSeek Harness with the dsh CLI.

# from a directory containing this repo (or publish it, then use the bare name);
# <name> is the profile you boot, e.g. `web`
dsh plugin --profile <name> add ./dsh-mistral-vibe
dsh --profile <name>

Then open Settings β†’ Models, pick Mistral Vibe, and paste a Vibe Code API key.

Other install methods

From a packed tarball (npm run pack produces dsh-mistral-vibe-<version>.tgz):

dsh plugin --profile <name> add ./dsh-mistral-vibe-<version>.tgz

This package is not built: lib/ is the shipped source, so a profile loads exactly what is in the repo.

Configuration

Settings β†’ Mistral Vibe owns every field. The same row is editable in cordis.patch.yml if you prefer text.

field default meaning
displayName Mistral Vibe the label the picker shows for this route
baseURL https://api.mistral.ai/v1 chat and model-list endpoint
apiKey β€” Vibe Code key, sent as Authorization: Bearer
studioApiKey β€” Studio/API key for account reads; falls back to the Vibe Code key
apiKeyEnv MISTRAL_API_KEY credential-store name consulted when the row is empty
disabledModels [] ids hidden from the picker, still listed on the page
modelsJson "" the stored model list, written by Fetch available models
bypassRewrite true send x-bili-passthrough: 1 so an in-process rewriter relays the body verbatim
allowRemote true answer the settings routes for a non-loopback request (same-origin or with the token)
accessToken β€” token a cross-origin request must present; falls back to MISTRAL_VIBE_ACCESS_TOKEN

Both keys are role("secret") and volatile(). The page never receives either value β€” it gets hasKey / hasStudioKey booleans.

Fetching the model list

Fetch available models calls GET /v1/models with the draft key, merges the answer into the stored list, and saves it. What lands in the store:

  • rows the shipped catalogue already serves, with their real limits and ladder;
  • rows Mistral serves that the bundle does not ship, marked new;
  • nothing that models.dev calls an embedding, speech or transcription record.

If the row cannot be written (read-only profile), the page says so instead of silently dropping the result.

Exposed deployments (HTTP/S, not just loopback)

The settings routes are served by the same web server as the page, so exposing the web UI over HTTP/S does not need any extra setup: the page's own requests are same-origin and are accepted.

request answer
loopback (127.0.0.1, localhost, [::1]) always accepted
same origin (Origin/Referer host equals Host, or Sec-Fetch-Site: same-origin) accepted
different origin, with the token in x-mistral-vibe-token or Authorization: Bearer … accepted
different origin, no token configured 401 {"code":"access_token_required"}
different origin, wrong token 401 {"code":"access_token_rejected"}
allowRemote: false 403 {"code":"loopback_only"}

The reason the token matters: these routes can move baseURL and therefore where the key is sent. Remote access on the page stores the token in this browser's localStorage and sends it on every request; the value never enters the bundle.

Thinking levels

A level is sent only when the model publishes one:

model levels wire
deepseek-v4-1-flash-1017 off, low, high, max none, low, high, max
mistral-large-4 low, medium, high verbatim
mistral-medium-2604 off, high none, high
codestral-2508, ministral-*, voxtral-* none no reasoning_effort

Development

npm test                 # host + client suites, offline
npm run test:host        # adapter, discovery, settings, wire encoding
npm run test:client      # the browser half, rendered against a React stub
npm run test:contract    # the emitted stream, through the harness's own assembler
npm run check            # node --check both halves, then the suites and the audit
npm run audit:harness    # run this adapter through the installed harness's own validators
npm run sync:models      # rewrite MODEL_CATALOGUE from models.dev
npm run probe            # live /v1/models read with the stored key
npm run stamp            # rewrite the BUILD stamp in lib/index.js

Layout:

lib/index.js       host half: adapter, catalogue, discovery, HTTP surface
lib/client.js      browser half: the settings page and its nav glyph
vendor/            schemastery + cosmokit, vendored as ESM
locale/            en/zh titles for the settings entry
scripts/           sync-models, live-check, client-half-toggle
test/              host.test.mjs, contract.test.mjs, client.test.mjs

The suites need no network and no key: the host suite runs the real adapter against a local SSE server, and the models.dev read is pointed at a local stub through MISTRAL_MODELS_DEV_URL.

Session-safety invariants

A stream that is wrong here does not fail a turn β€” it writes a session log that later refuses to load. The rules the host actually enforces, and what this route guarantees against each:

rule where it bites this route
a tool-call id must be unique for the whole session assistant/message repeats advertised tool call <id> ids are a 9-character hash of the provider id plus a serial that never resets, so a provider that numbers calls call_1, call_2 per response cannot repeat across turns
tool/call must echo the advertised name and arguments byte for byte tool/call <id> does not match one advertised tool call the block-end carries the same accumulated name/arguments the deltas streamed, and {} when nothing arrived, so the value always parses
only text, reasoning and tool-call blocks cannot assemble incomplete block of type … no other block type is ever started
chunk types are a closed set; argumentsDelta is a string; every index is a non-negative safe integer AssistantStreamAccumulator.push throws enforced by the contract test, chunk by chunk
usage precedes finish, nothing follows finish a stream that ends early loses usage the generator yields usage then finish, last
each call gets its own block two calls merging concatenates their arguments a missing index, or an index reused by a different id, allocates a fresh slot

npm run test:contract runs the emitted stream through the harness's real AssistantStreamAccumulator and BlockAssembler (imported from the installed DeepSeek Harness; the file prints SKIP when that install is absent) and then applies the tool-lifecycle rules above. It is the only check that would have caught the chatcmplt corruption before it hit a log.

Which build is running

A host keeps an applied plugin module in memory: re-applying the entry writes applied to the operation log but does not re-import the file, so a fix can be on disk and still not run. BUILD makes that observable β€” it is the hash of lib/index.js without the stamp line, logged on mount, returned by GET /mistral-vibe/config, and printed under the model list on the settings page. "host build unknown" means the host is still serving a module loaded before this field existed; restart DeepSeek Harness to pick the file up.

Route surface

route effect
GET /mistral-vibe/config the row as the page needs it, plus build β€” no key values
POST /mistral-vibe/config save apiKey, studioApiKey, baseURL, displayName, disabledModels
POST /mistral-vibe/discover merge GET /v1/models into the stored list; answers {source, added, models, persisted}

All three accept loopback and same-origin requests, and answer anything else with 401 {"code":"access_token_required"} or 401 {"code":"access_token_rejected"} unless allowRemote: false, which restores 403 {"code":"loopback_only"}.

Auditing against the installed harness

npm run audit:harness imports the harness's real LlmRuntime and BlockAssembler from %LOCALAPPDATA%\Programs\DSH NEXT\resources\app\node_modules\@deepseek-ai\dsh-llm (or DSH_LLM_PATH) and feeds this adapter's output through them, so the checks are the harness's own code rather than a restatement of it:

check the failure it prevents
catalog passes listModels validation duplicate or unnamed id β†’ INVALID_CATALOG
every advertised id passes normalizeModelInfo bad name / window / maxTokens β†’ INVALID_MODEL_INFO, INVALID_MODEL_CONTEXT
a retired alias echoes the requested id resolved.id !== model β†’ INVALID_MODEL_INFO on every turn
an unknown id still returns valid metadata a session that stored a dropped id keeps working
ladders have unique efforts and a known default INVALID_MODEL_REASONING
a streamed reply builds through BlockAssembler an unknown chunk, a bad index or a non-string argumentsDelta

It exits 0 with SKIP when no harness is installed, so it is safe in CI.

Notes / limitations

  • No quota reading. /v1/billing/subscription returns {"plan_name": …} and no money fields; roughly three dozen other candidates (/v1/billing/usage, /v1/credits, /v1/vibe-code/usage, …) answer 404 no Route matched. The console's API and Vibe Code allowances are not readable through the API, so this plugin does not pretend to show them.
  • A discovered model's ladder depends on models.dev. An id models.dev does not know keeps "levels unknown": no level control, and no reasoning_effort on the wire.
  • Replay is sanitised. Every request is rebuilt from the fields Mistral declares: role, content, and tool_calls with id/type/function. Reasoning replays the way Vibe replays it β€” as {"type":"thinking","thinking":[{"type":"text",…}]} chunks inside content, gated on the model declaring an effort ladder β€” never as a sibling reasoning_content field, which Mistral answers with a body-wide 422 extra_forbidden.
  • An in-process request rewriter is asked to stay out of the way. billion-context ("bili") injects a blank reasoning_content into any assistant turn carrying tool_calls when the model id matches /deepseek/i β€” which is every deepseek-v4-1-flash-1017 turn that used a tool β€” and Mistral rejects the field. Every request here therefore carries x-bili-passthrough: 1, bili's own verbatim-relay marker; Mistral ignores unknown headers. Set bypassRewrite: false to stop sending it behind a gateway that rejects unknown headers. The trade-off is that bili does no folding on this route while the header is on.
  • Tool-call ids are hashed, never truncated. Harness ids share a long prefix (chatcmpltool…), so the first nine alphanumerics were the same string for every call in a message β€” chatcmplt β€” and the session log then refused to load: "assistant/message repeats advertised tool call chatcmplt". Ids are now a 9-character hash of the whole id, with a per-turn collision guard.
  • Alias ids still resolve. mistral-medium-latest is not a row of its own; a session that selected it resolves to mistral-medium-2604 with that entry's limits.

License

MIT Β© Meltartica β€” see LICENSE.