Khuraie/dsh-mistral-vibe-provider

Mistral Vibe is an unofficial DSH provider route for Mistral's Vibe Code subscription.
PKCE browser sign-in · the real Vibe and Studio/API meters · prices on every model · a live Session cost.

Features · Install · Images · Routes · Notes & limits

It registers the LLM route mistral-vibe against https://api.mistral.ai/v1, signs you in with the console's own PKCE flow, and — unlike the bundle it grew out of — shows the real allowance meters for the Vibe Code budget and the Studio/API budget.

Community integration, not affiliated with Mistral. Your own Mistral account's plan and terms apply; usage made through this route is billed by Mistral to the same allowance the Vibe CLI draws on.

Features

	Feature	What it does
🔐	PKCE browser sign-in	Settings → Mistral Vibe → Sign in with Mistral runs console.mistral.ai/api/vibe/sign-in, opens the approval page, polls the process and, once approved, stores the minted key directly in the credential store. No copy-paste.
📊	Two real meters	GET /v1/billing/budget returns the Vibe Code allowance and the Studio/API allowance separately; the account card draws both, with percent, used-of-total, currency and reset date. GET /v1/billing/subscription supplies the plan name.
📈	The meters in the sidebar	The same two allowances sit in the sidebar foot, directly above the account row: dollars used of the allowance, its percentage, and a bar, with the reset date in the tooltip. Click to re-read; it also refreshes itself every ten minutes.
🗂️	A live catalogue	/v1/models fills the picker; rows the key cannot use are kept visible but marked, so a tier_not_allowed 403 is a badge, not a mid-turn surprise.
💵	Prices on the model rows	Every catalogue model shows its own price line — $0.30 in · $0.007 cached · $1.20 out / 1M tokens for DeepSeek V4.1 Flash 1017, taken from Mistral's model card — so the cost of a model is visible before you pick it. A model discovered at runtime has no published price and shows no line.
🧮	A price badge and a session cost run	The picker badges each model with its input/output price ($0.3/$1.2), and the composer's token-usage pill — the one reading 27.3M tok · Cache hit 99% — gains · ≈$1.51 at the end of its own text run for the Session you are in. Folded from that Session's assistant/message usage at list prices, so a model switch mid-session prices each message on the model that actually ran it.
🧠	Per-model thinking levels	Mapped onto reasoning_effort, with reasoning streamed as a thinking block rather than as text.
🖼️	Screenshots that do not wreck the cache	Downscaled to 1024px / 768 KB, capped at Mistral's 20-image ceiling through the harness's durable offload protocol, and kept visible by default so a visual workflow can compare one render against the next. See Images.
🔑	Secrets stay out of YAML	Keys live in the credential record named by apiKeyEnv (MISTRAL_API_KEY by default); the row keeps only a reference.

Install

The package is not published on npm yet, so install it from the repository. dsh plugin forwards its arguments to pnpm inside the profile directory, which means any spec pnpm accepts works — a path, file:, github:owner/repo, or a published version.

From the repository:

dsh plugin --profile <profile> add github:Khuraie/dsh-mistral-vibe-provider

From a clone — what a development setup wants, because the profile links the checkout and an edit is live after a restart:

git clone https://github.com/Khuraie/dsh-mistral-vibe-provider
dsh plugin --profile <profile> add ./dsh-mistral-vibe-provider

dsh plugin only installs the dependency. The launcher stacks the patch layers listed in dsh.profile.bundles, so add the package name there too — in ~/.dsh/profiles/<profile>/package.json:

"dsh": {
  "profile": {
    "bundles": ["…", "@khuraie/dsh-mistral-vibe-provider"]
  }
}

The in-app plugin manager does both steps in one go. Restart DSH afterwards.

Then Settings → Mistral Vibe and either sign in with Mistral or paste a Vibe Code API key. Hit Fetch available models to persist the live catalogue, and Refresh on the account card to re-read the meters. The two meters also appear in the sidebar foot above the account row, so the remaining allowance is visible without opening Settings.

Once the package is published, the shortest form becomes dsh plugin --profile <profile> add @khuraie/dsh-mistral-vibe-provider.

Images

Screenshots are the one input that can wreck Mistral's prompt cache, so this route manages them rather than forwarding whatever the harness hands it. The cache is a longest-common-prefix cache: it only serves a request that repeats the previous one byte for byte, and an image payload is the biggest thing that can change.

Downscaling

Every image is re-encoded before it is sent: 1024px on the longest edge, 768 KB encoded. In a measured Session, steps carrying a 1–4 MB screenshot fell to 1–26% cache hit while text-only steps in the same Session stayed at 86–100%. The harness default for other providers is 1568px / 4 MB; this route trades that detail for cache.

setting	default	effect
imageMaxEdge	1024	longest edge, in pixels, of the image sent
imageMaxBytesKb	768	encoded byte budget for one request image

The 20-image ceiling

Mistral refuses a request carrying more than 20 images — 400 invalid_request_invalid_args, "Total number of images exceeds the maximum allowed of 20", code 3051. The route does not drop images itself: that would rewrite the transcript on every request and break the prefix cache. Instead it fails the step with IMAGE_OFFLOAD_REQUIRED and the count it needs, and the harness's dsh-compaction-image-offload plugin records one durable image/offload selection before retrying. The count is retired four at a time, so the retained window stays between 17 and 20 images instead of collapsing to one.

An offloaded image becomes a named placeholder rather than a silent gap:

[image omitted from this request to stay inside the image limit (not a failed
 read); "render-04.png" (sha256:…)]

The "not a failed read" wording is load-bearing: without it a model finds the placeholder where its own read_image result used to be and reports the read as broken, then re-reads the file.

Keeping images visible

By default images stay on the wire up to the ceiling, because a visual workflow — render, look, edit, re-render — needs the previous render to compare against. offloadSeenImages is the opt-in cache mode, off by default the way Command Code ships its equivalent: once the model has answered on an image, that image becomes a placeholder, so a new screenshot no longer rolls the cache back. The cost is that seeing an old image again means reading it again.

setting	default	effect
offloadSeenImages	false	opt-in: retire images the model has already answered on
imageKeepBatches	3	in that mode, how many recent image batches keep their pixels

Measured

Session	images	offloads	cache efficiency
52-image judging, batch-cap design	52	4	78.8%
192-step visual iteration, current default	40	17 (2 after the fix)	95.4%

Keeping images and letting the budget retire four at a time measured better than the aggressive offloading it replaced: fewer transcript rewrites means more of the prefix survives.

Routes

route	effect
GET/POST /mistral-vibe/config	the row as the page needs it (no key values)
POST /mistral-vibe/discover	merge a live /v1/models read into the stored list
GET /mistral-vibe/budget	plan name plus the Vibe and Studio/API meters
GET /mistral-vibe/session-cost?sessionId=…	what one Session's Mistral Vibe usage cost at list prices
GET /mistral-vibe/usage	token, cache and cost totals across the most recent Sessions
POST /mistral-vibe/sign-in	start a PKCE attempt, answer the console URL
GET /mistral-vibe/sign-in/status?processId=…	poll an attempt; on approval stores the key

All routes are loopback/same-origin by default and accept a configured access token for an exposed web UI.

Notes and limits

• Thinking is on by default. Mistral's own default for a request that carries no reasoning_effort is off, and a plugin that simply forwards "no level chosen" ends up running DeepSeek V4.1 Flash with no reasoning trace at all. This route declares high as the default level for every model that can think (the official DeepSeek route does the same), so a fresh Session thinks unless you pick off in the model picker.
• Images are managed, not forwarded. Downscaling, the 20-image ceiling, the durable offload protocol and the opt-in cache mode are all in Images above.
• Sticky routing for the prefix cache. Mistral's prompt cache lives on the backend instance that served a prefix, so a request routed to another instance misses even when the prefix is byte-identical — the whole-prefix drops that show up a few times per Session. Every request now carries x-affinity: <session id>, the same header Mistral's own Vibe CLI sends, so a Session stays on one instance. It is only sent to *.mistral.ai hosts and can be turned off with sendAffinity: false.
• Tier gating is server-side per model. On a free plan the live catalogue simply omits paid models such as deepseek-v4-1-flash-1017, and calling one answers 403 {"type":"tier_not_allowed","code":"1910"}. The picker marks rows that never appeared in a live read; nothing is hidden automatically.
• Thinking models answer with content as an array of typed blocks ({"type":"thinking",…}), which this adapter normalises into reasoning deltas.
• max_tokens includes reasoning tokens, so a small cap can end a turn before any answer text arrives.
• The bundle is derived from dsh-mistral-vibe by Meltartica (MIT): the wire adapter, catalogue, thinking-level contract and client shell are theirs. This bundle replaces its settings half with the account card and adds the sign-in and billing routes. See NOTICE.

License

MIT © 2026 Meltartica and Mars-Sea — see LICENSE and NOTICE.