Skip to main content
L

dsh-mistral-pro

lordraiden/dsh-mistral-pro

Maintained fork of dsh-mistral-pro: Mistral subscription (Vibe Pro/Team) model route, multi-account failover, prompt-cache tuning, and a sidebar usage card with a statistics panel.

Install

dsh plugin --profile web add github:lordraiden/dsh-mistral-pro

README

dsh-mistral-pro

Mistral subscription (Vibe Pro / Team) provider for DeepSeek Harness (DSH). Adds the mistral-pro model route, multi-account failover, prompt-cache tuning, and a sidebar usage card with a statistics panel.

version DSH license EN · 中 · 日

English · 中文 · 日本語

The Mistral Pro statistics panel: Vibe and API budgets with reset countdowns, prompt-cache statistics and configured accounts

Contents

At a glance

Model routemistral-pro
Model catalog19 built-in models (11 visible by default), refreshed from /v1/models
FailoverSeveral Mistral accounts, switched before the first streamed output
Prompt cachingauto / shared / off cache keys, compaction-aware rotation, hit-rate route
UISidebar usage card, statistics panel, Settings → Mistral Pro
TransportNative node:https, with an opt-in fetch path

Why this plugin

  • Mistral models inside DSH. Registers the mistral-pro route with a 19-model catalog (11 visible by default), a live /v1/models refresh, and per-model context-window, output and reasoning-effort overrides.
  • Multi-account failover. Several Mistral keys are tried in order; an account that fails before the first streamed token is skipped and cooled down for 300 seconds.
  • Prompt caching that can actually hit. Stable prefixes, auto / shared / off cache keys, compaction-aware key rotation, and a route that reports the real cached_tokens.
  • Subscription usage at a glance. Quota windows, reset countdown, prepaid credit and current-month spend in the composer toolbar.
  • Tested and observable. 11 regression suites, CI on Node 22 and 24, a 0600 wire log, and read-only diagnostic routes.

This is a fork. lordraiden/dsh-mistral-pro merges the two independent fix lineages that existed on top of upstream 0.1.2: santera-pixel/dsh-mistral-pro (0.2.0-0.2.4) and the upstream issue #1 patch series (0.1.3-0.1.8). See CHANGELOG.md for the full reconciliation and Fork lineage for the version policy.

Requirements

  • DSH >= 0.2.1-rc.5
  • Node ^22.19.0 || >=24.0.0
  • A Mistral account — Free, Vibe Pro / Team, or pay-as-you-go

Installation

Install the fork pinned to a release tag:

dsh plugin --profile web add github:lordraiden/dsh-mistral-pro#v0.4.0

Or add it to the profile's package.json and install:

{
  "dependencies": {
    "dsh-mistral-pro": "github:lordraiden/dsh-mistral-pro#v0.4.0"
  }
}
pnpm install

From a local checkout, pack a tarball (recommended) or link the directory:

npm pack
dsh plugin --profile web add /absolute/path/to/dsh-mistral-pro-0.4.0.tgz

Linking a directory creates a symlink. Because pnpm does not resolve peerDependencies for symlinked targets, the repo includes a node_modules/@deepseek-ai/* symlink directory pointing to the active profile's shared dependencies (excluded when packing the tarball).

Once installed, enable dsh-mistral-pro on the Plugins page and reload. A host restart is required after changing the installed version.

Adding accounts

Open Settings → Mistral Pro → Accounts:

  1. Click Add account and enter a display name (used only for display and ordering).
  2. Set Credential reference to an environment variable name, such as MISTRAL_API_KEY.
  3. Paste your API key into the password field and click Set key. Keys are stored in the DSH credential store and are never sent back to the browser.
  4. Click Test to run a check against /v1/models.

You can generate an API key under Code › Vibe CLI. The same key works across Free, subscription plans, and pay-as-you-go; requests consume your monthly plan allowance first before falling back to pay-as-you-go billing.

Accounts are tried in list order (use ↑ / ↓ to adjust). The first enabled account acts as the primary. When a failover succeeds, the working account is promoted in memory for the rest of the process lifetime without modifying your saved config.

Two keys: Vibe and Studio

For the common setup — a Vibe key with a Studio key behind it — Settings → Mistral Pro → Keys has one row per key:

RowCredential referenceAccountOrder
Vibe API keyMISTRAL_VIBE_KEYVibeprimary
Studio API keyMISTRAL_API_KEYStudiofallback
  • Vibe API key — the key the Vibe CLI uses. After vibe --setup it is in ~/.vibe/.env; you can also create one under Code › Extensions → Advanced → Vibe API Key.
  • Studio API key — a pay-as-you-go key from the Mistral admin: API › API Keys.

Each row has a password field, Save and Test, and shows whether the key is configured. Save stores the key in the DSH credential store and creates or updates its account (Vibe or Studio), keeping Vibe first and Studio second even when other accounts exist. Test runs a /v1/models check against that account. The key value is never shown again and never logged.

Requests go to Vibe first. When Vibe answers 402 (budget exhausted) or 429 (rate limit), the adapter fails over to Studio before the first streamed token, and the failed account cools down for 300 seconds. The advanced account list stays available for more than two accounts or custom endpoints.

Terms of service. Using the Vibe key outside the official Vibe CLI may not be permitted by Mistral. Whether to do it is your decision; the plugin never ships or bundles keys.

Routing

Settings → Mistral Pro → Routing decides which key is tried first and whether a failure falls back:

  • Vibe first / Studio first — the named key is always tried first; a fallback is never promoted.
  • Automatic — the account that last worked is promoted for the rest of the process.
  • Automatic account failover — when off, only the first key is used and a failure surfaces instead of switching keys.

Changing the priority also reorders the Vibe and Studio accounts, so the account list below never contradicts it; their reorder arrows are disabled there.

Models

Settings → Mistral Pro → Models combines the built-in catalog with models returned by /v1/models. Uncheck Visible to hide a model from the picker. Context window, max output tokens, and Default effort (reasoning_effort) can be overridden per model.

Built-in catalog (2026-11, 19 models total, 11 visible by default). With a key, the live /v1/models listing wins and aliases collapse into one entry:

ModelContextReasoning
mistral-large-41Mlow / medium / high
mistral-small-2603256Knone / low / medium / high
mistral-medium-3-5128Knone / low / medium / high
zai-glm-5-31Mlow / high / max
zai-glm-5-3-flash1Mlow / high / max
deepseek-v4-1-flash-10171Mnone / low / high / max
ministral-3-{14b,8b,3b}-25-12256K—
codestral-25-08 / codestral-latest256K—

mistral-large-2512, devstral-2512, and mistral-medium-2508 are hidden by default and can be enabled in the model table.

Sidebar usage card and statistics panel

The plugin adds a card to the sidebar footer, beside Settings. It shows one combined figure for the two Mistral wallets — the Vibe budget and the API budget — with a two-colour bar: grey for Vibe, the primary text colour for API. The percentage is the shared consumption (used / total across both budgets), so it no longer reads 0% while the API wallet is being spent.

The sidebar usage card with the combined Vibe and API spend

Clicking the card opens the statistics panel:

  • each wallet with its used / total, percentage and reset countdown;
  • prepaid credit balance and current-month API spend;
  • the combined total;
  • prompt-cache statistics: steady-state hit rate, a stacked mix bar of the per-request hit-rate ranges, a sparkline of the last 20 requests, a one-line summary, failovers and the last request;
  • configured accounts and the model count;
  • manual refresh and last-updated timestamp.

The cache card replaces the six per-range distribution rows with a stacked mix bar (one segment per range, with a compact counter legend), a sparkline with one thin bar per request over the last 20 (height is the hit rate, colour is the range, the tooltip shows model and percentage), and a one-line summary such as 18 of 20 requests ≥ 76% · 2 cold starts. The remaining cache figures are unchanged.

Visibility follows Settings → Mistral Pro → Usage → Sidebar card visibility: auto shows the card when an account is configured, always pins it, and off hides it. Background tabs pause polling and refresh immediately when brought back into focus.

Usage comes from api.mistral.ai/v1/billing/budget, read with the API key you saved. No extra setup is needed. The endpoint returns the Vibe and API monthly budgets: percentage used, budget amount, and reset time.

If the key lacks the usage permission (the billing endpoint answers 403), you can use an admin-site cookie instead:

  1. Sign in to Mistral Admin.
  2. Open DevTools → Network, select any request to admin.mistral.ai, and copy the full Cookie request header. It must include ory_session_* (if csrftoken is present, the plugin sends X-CSRFTOKEN automatically).
  3. Paste the header under Settings → Mistral Pro → Advanced and click Set cookie.

The cookie path queries GET /subscription, GET /api/billing/v2/usage?month=&year=, and GET /api/billing/credits independently, so a failure on one endpoint won't block the others. If the session expires and redirects to the login page, the panel prompts you to update the cookie.

Request flow and failover

The llm service passes session messages to the adapter, which formats them into OpenAI-compatible payloads, POSTs to https://api.mistral.ai/v1/chat/completions, and parses the SSE response into DSH StreamChunks. Assistant reasoning history is replayed back as thinking chunks as required by Mistral's API.

Account failover only triggers before the stream emits content:

  • On 401 / 403, 402 (API budget exhausted), 429, quota exhaustion, or a missing/invalid key, the adapter switches to the next account and places the failed one on a 300-second in-memory cooldown.
  • Once text, reasoning, or a tool call has been emitted, errors are surfaced directly without switching accounts so partial responses from different accounts are never mixed.

Stream contract

The adapter follows DSH's stream contract on the way back:

  • Streamed tool calls are accumulated from continuation chunks, and only a non-empty function.name is accepted — Mistral repeats function.name: "" on every chunk of a streamed tool call.
  • Exactly one block-start is emitted per block, so the live thinking panel does not collapse while streaming.
  • id is always a string on tool-call-delta.
  • A truncated stream raises TRANSPORT instead of ending the turn "successfully" with partial output.
  • The retry policy is the resolved, flattened, JSON-safe shape.
  • No max_tokens cap is injected by default (advanced.maxTokensCap is 0): reasoning shares the completion budget, so a cap that looks generous for text can be consumed by thinking alone.

Prompt caching

Mistral's prompt caching reuses previously computed prompt tokens when requests share the same prefix. Cached tokens are billed at 10% of the standard input price and are reported in usage.prompt_tokens_details.cached_tokens.

Two facts from the documentation drive every decision here:

  • prompt_cache_key is a routing hint. It "increases the chance of a cache hit, but doesn't guarantee one" — a hit happens when the API finds a compatible cached prefix. The key does not create the cache; prefix stability does.
  • Cache blocks are 64 tokens. cached_tokens is always a multiple of 64, and prompts under 64 tokens never hit. Longer shared prefixes reuse more.

What the plugin does to keep the prefix stable:

  • It never injects volatile content (timestamps, session ids, counters) into the request.
  • It preserves message order and replays assistant reasoning as thinking blocks, so the prefix stays byte-identical between turns.
  • It drops non-wire keys per role (sanitizeMessage), deterministically, so a foreign field can never change the prefix on one turn and not the next.
  • It sends stream_options.include_usage, without which cached_tokens would never be reported at all.

advanced.promptCacheKey selects the key:

ValueKey derivationBest for
auto (default)model + system prompt + first user messageOne long conversation: every appended turn keeps the same key.
sharedmodel + system prompt only; falls back to auto when the session has no system/developer messageMany short conversations that reuse the same (large) system prompt: the first turn of a new conversation can still hit the cached system-prompt prefix.
offfield omittedProviders or proxies that reject the field.
any other stringsent as-isManaging keys externally (one per workflow, team, or subscription).

auto and shared are a trade-off, not a ranking. auto isolates conversations, which is what Mistral recommends ("use a stable application-level identifier, such as a conversation ID"). shared gives up that isolation so separate conversations can reuse the system prompt. The provider still verifies prefix compatibility, so a shared key can only miss — it can never return another conversation's content.

With no leading system or developer message there is no shared prefix to reuse, so shared falls back to the per-conversation auto key instead of emitting a model-only key that every conversation on the model would share.

A compaction rewrites the conversation prefix, so the derived key (auto and shared) rotates once per compaction: the newest compactionId from DSH's compaction checkpoint is mixed into the hash, giving the rewritten prefix a clean cache entry instead of filing it under a key that pointed at the old prefix. Sessions that never compact are unaffected — their key derivation is byte-identical.

The plugin also sends an x-affinity header with the same value as the cache key. The header is undocumented: Mistral's prompt-caching documentation does not mention it, but Mistral has stated that its API "uses it for cache routing ahead of prompt_cache_key" (mistral-vibe#1481), and Vibe derives it from its session family's root id. An A/B run against the live API measured no difference in cache hits (7/8 with the header vs 8/8 without; the only miss was the first, cold call), and Mistral separately acknowledged serving-side cache misses that no client change can fix (#1512). Treat it as a hint that may help, not a guarantee that requests land on the same server. cache.affinity and cache.reportUsage can be turned off independently; turning off reportUsage also costs you the cache-hit numbers.

To see what actually happened, GET /plugins/dsh-mistral-pro/wire?limit=20 returns the request and usage entries of the recent requests. The usage entry carries promptTokens, cachedTokens, the serving account, a conversation digest and an affinity digest. The affinity digest is taken from the x-affinity header that was actually placed on the request, not from the setting that was expected to produce it, and it is omitted when no header was sent. While the adapter sends the cache key as the affinity, the two digests are equal; they diverge the moment the two identities do.

The request entry carries prefixRelation and prefixReused, the only client-side signal that can tell normal append-only growth from a rewritten prefix. prefixRelation is one of:

ValueMeaning
firstNo earlier request of this session was observed in this process (first turn, plugin reload, or the session was evicted from the bounded cache).
appendThe whole previous message list matched the head of a longer current list. Verified even when the current list exceeds the comparison cap, because every previous message was retained and compared.
repeatBoth message lists were fully retained and are identical.
rewriteA divergence was found inside the compared range, or the whole current list matched the head of a longer previous list (a truncation).
unknownThe compared range matched, but at least one list extends beyond the comparison cap, so the unexamined suffixes could still differ. Equality is never assumed for messages that were not compared.

prefixReused is how many leading messages are byte-identical to the previous request, so rewrite with prefixReused: 1 means only the system message survived. The comparison is bounded and in memory: it covers at most the first 512 messages of a session and the 32 most recently used sessions, it keeps only one digest per message, and it never writes message content anywhere. It is also process-local: after a restart the first observed request is first.

The identity is DSH's session id only. A prompt-cache key is never used as a substitute: shared mode and custom keys are designed to be shared across independent conversations, so comparing them would invent relationships between unrelated sessions. A request without a session id carries no relation, and auxiliary calls (session title, compaction, auto-review) are not part of the conversation's message sequence, so they carry none either and do not disturb the next comparison.

The observer is only updated for a request that actually reaches the outbound transport: a call that fails first (missing credentials, an already-aborted request, a transport that cannot be prepared) is not observed, while a request that reaches the transport counts as attempted even if the transport then fails. One logical request is observed once, even when it fails over between accounts.

This compares the wire messages list only. It does not cover tools, tool_choice, temperature, max_tokens or any other request field, and it says nothing about tokenization, the provider's cache placement or routing. append therefore means "the messages this client sent are an append-only extension of the previous message list", not "the complete effective prompt and every cache-relevant input are unchanged". A hit rate near 0% with prefixRelation: append usually means the prompt is under 64 tokens, the cache expired between turns, the provider missed on its side, or a non-message request field changed.

GET /plugins/dsh-mistral-pro/cache reports the whole-log requests, promptTokens, cachedTokens and hitRate, plus the views that answer "why is it low":

  • steadyState — the same numbers without the first observed request of each prompt-cache-key digest in the retained window. That group is a cache-key family, not necessarily one conversation: shared and custom keys can span independent conversations. The excluded request is not necessarily a real cold start, and the figure is neither a lifetime metric nor a guaranteed warm-conversation hit rate;
  • steadyStateByRoute — the same numbers without the first request of each (account, model, affinity, prompt-key) route. A hit also needs the same serving account, the same model and the same effective affinity, so an account failover, a model change, a key rotation after compaction or a new affinity is a cold start even when the conversation itself is warm. The model is part of the identity because a custom prompt_cache_key does not carry it. This is the honest number when you switch accounts; steadyState keeps its original prompt-key-only meaning;
  • window — the retained log window the figures were computed over (maxLines, entries). Every rate here is relative to that window: the request a steady-state view drops is the first one still visible, not necessarily a real cold start, and a route may have been warm long before its earlier entries rotated out;
  • distribution — how many requests landed in each per-request hit-rate bucket (0, 1-25, 26-50, 51-75, 76-99, 100);
  • byAccount, byModel, byAffinity — the same totals per serving account, per model and per affinity digest, busiest first, with a null group for requests that carried no label;
  • failovers and coldAfterFailover — account switches and the cold request that follows one: prompt caches are per account, so that miss is expected;
  • recent — the 20 newest requests with their account, affinity and conversation digests.

Only completed responses whose supplier reported usage are recorded, so an interrupted stream is absent from every figure instead of being counted as a zero-cache request.

Reading the numbers, in order:

  1. A low whole-log hitRate with a high steadyState usually means cold starts, not a broken cache.
  2. A high steadyState but a low steadyStateByRoute is account failovers or key rotations — look at coldAfterFailover and at byAccount.
  3. One low byAccount entry with the others healthy means that account's traffic is missing; the plugin is not rewriting anything.
  4. A low rate in every account and model, with prefixRelation: append (or unknown once the conversation exceeds the comparison cap) on the missing turns in /wire, points at provider-side eviction — the case Mistral acknowledged in #1512. No client setting fixes it. A rewrite instead means this client changed the prefix, and that is actionable.

Usage accounting

DSH counts token usage in disjoint buckets: inputTokens is uncached input, cacheReadTokens is the cached prefix, and the billed prompt is their sum. Mistral folds the cached prefix into prompt_tokens, so the adapter subtracts it before reporting. The /cache route keeps reporting the total prompt count, so its hitRate stays cachedTokens / promptTokens.

Settings

  • Keys: the Vibe and Studio key rows with Save and Test (see Two keys).
  • Accounts: add, remove, reorder, enable/disable, store a key (save & enable), test connectivity.
  • Models: toggle visibility, override context window, max output, and default reasoning effort.
  • Usage: sidebar card visibility (auto / always / off), refresh interval, credit balance, prompt-cache section, and the optional admin cookie.
  • Advanced: automatic failover, cooldown, timeouts, catalog refresh, output cap, prompt-cache key, cache affinity, stream usage reporting, transport, proxy, button display mode, refresh interval, and an optional cookie fallback.
    • Send x-affinity (cache.affinity, on by default): sends the cache key as x-affinity; an undocumented, unverified cache-routing hint (see Prompt caching).
    • Report stream usage (cache.reportUsage, on by default): sends stream_options.include_usage; turning it off also hides the cachedTokens counts.
    • Show prompt-cache hit rate (usage.showCache, on by default): the statistics-panel section that reads /cache and shows the rate, the mix bar, the sparkline, the summary line and the last request.
    • Transport (advanced.transport, default https): https posts the serialized body over node:https; fetch routes through globalThis.fetch so in-process wrappers can see and rewrite the request.

The Mistral Pro settings page in DeepSeek Harness, with account, model and advanced sections

All settings take effect immediately without restarting the plugin.

Diagnostic routes

Same-origin read-only routes registered for troubleshooting:

RouteDescription
/plugins/dsh-mistral-pro/healthaccount status, registered providers, model count
/plugins/dsh-mistral-pro/metaactive usage / advanced / accounts configuration
/plugins/dsh-mistral-pro/modelsmodel catalog, visibility, limits, reasoning options
/plugins/dsh-mistral-pro/usage?account=&force=1aggregated usage snapshot
/plugins/dsh-mistral-pro/test?account=run a /v1/models check against a specific account
/plugins/dsh-mistral-pro/refreshforce a live model catalog refresh
/plugins/dsh-mistral-pro/settingssettings namespace and revision
/plugins/dsh-mistral-pro/wire?limit=20newest wire fingerprints, per-message key layout, cache usage, conversation id, prefix digest and serving account
/plugins/dsh-mistral-pro/cacheaggregate, steady-state and route-aware hit rates over the retained log window, per-account / per-model / per-affinity breakdowns, per-request distribution, failover counts, and the 20 newest usage entries

Transport and the wire log

The chat request is posted over node:https with the body the adapter already serialized, not through globalThis.fetch.

Proxy plugins replace globalThis.fetch and re-serialize request bodies on the way out. billion-context's ACP loop, for one, adds reasoning_content: "" to every assistant tool-call message whenever the model id matches /deepseek/i — and deepseek-v4-1-flash-1017 does. Mistral forbids that field on assistant messages, so the rewrite turns every turn after the first tool call into 422 extra_forbidden. node:https cannot be patched by another plugin in the same process, so the bytes the adapter builds are the bytes that leave the machine.

  • DSH_MISTRAL_TRANSPORT=fetch restores the old path; use it only when no wrapper rewrites bodies.
  • A configured advanced.proxy always uses fetch + undici ProxyAgent, because that path needs the dispatcher.
  • Trade-off: traffic sent this way skips a proxy plugin's compression and rewrite. Correctness comes first.

Every request appends one line to %TEMP%\dsh-mistral-wire.log (override with DSH_MISTRAL_WIRE_LOG): model id, message count, each message's role and keys, how many messages carry reasoning_content, the serving account, a 16-hex digest of the exact x-affinity value placed on the request (omitted when no header was sent), prefixRelation and prefixReused (omitted when there is no conversation identity to compare), and a 16-hex SHA-256 of the serialized body. A second usage line records promptTokens, cachedTokens, the serving account, the conversation digest, the affinity digest, messages and sha256; an account switch appends a failover line with from, to and the failure code. The log is trimmed to 400 lines and a logging failure never interrupts a call. Only a completed response whose supplier reported usage produces a usage line, so an interrupted stream leaves no zero-cache record behind. Every aggregate on /cache is computed over that retained window, which window reports.

The log is created with mode 0600: %TEMP% is usually /tmp (1777), and the entries are prompt metadata. If an older version left the file world-readable, the next write tightens it to 0600, and rotation keeps that mode.

FAQ

Does it work with a free Mistral account?

Yes. The same API key works across Free, subscription plans, and pay-as-you-go. Requests consume your monthly plan allowance first and fall back to pay-as-you-go billing after that.

Where do I get a Mistral API key?

The Vibe key is in ~/.vibe/.env after vibe --setup, or under Code › Extensions → Advanced → Vibe API Key. The Studio key comes from the Mistral admin under API › API Keys. Paste either one into Settings → Mistral Pro → Keys — see Two keys.

Does it support tool calls, reasoning and images?

Yes. Streamed tool calls are accumulated across continuation chunks, assistant reasoning is replayed as thinking blocks, and image inputs use DSH's attachments service (falling back to placeholder text when that service is unavailable).

Why is my prompt-cache hit rate 0%?

Usually one of three things: the prompt is under 64 tokens, the cache expired between turns, or something upstream is rewriting the first messages. Check /plugins/dsh-mistral-pro/wire?limit=20 for the per-request cachedTokens and prefixRelation, and /plugins/dsh-mistral-pro/cache for the steady-state rate, the route-aware rate and the per-account breakdown — a low whole-log rate with a high steady-state rate means cold starts, not a broken cache; a healthy steadyState with a low steadyStateByRoute means account failovers, model changes or key rotations; a low rate in every account with prefixRelation: append is provider-side eviction, which no plugin setting can fix. A prefixRelation: rewrite is the one case this client can act on: something changed the messages it sends.

Is my API key sent anywhere besides Mistral?

No. Chat requests go only to the configured Mistral API endpoint, catalog refreshes send only the API key and User-Agent, and usage queries go to admin.mistral.ai with the cookie you provide. Keys and cookies stay in the host credential store; the browser only sees whether a credential is set.

How is this different from 01NGVI-QINGYI/dsh-mistral-pro?

Upstream stopped at 0.1.2. This fork merges the two independent fix lineages built on top of it, adds CI and regression suites, and ships the native node:https transport, prompt-cache key modes, the sidebar usage card, the statistics panel and the diagnostic routes. Version numbers are not comparable between the two repositories — see Fork lineage.

Data and privacy

  • Chat requests are sent only to the configured Mistral API endpoint.
  • Model catalog refreshes send only the API key and User-Agent.
  • Usage queries are sent only to admin.mistral.ai (configurable) using the cookie you provide.
  • API keys and cookies are stored in the host credential store; the browser only sees whether a credential is set.

Known limitations

  • The /v1/billing/budget and admin usage endpoints are undocumented. The plugin reads them by field name, so upstream schema changes may hide individual windows.
  • prompt_cache_key and x-affinity are hints, not guarantees: a stable prefix and a stable affinity can still miss after the provider's cache expires, prompts under 64 tokens never hit, and Mistral has acknowledged serving-side cache misses (mistral-vibe#1512) that no client change can fix.
  • prefixRelation is a process-local comparison: it covers the first 512 messages of a session and the 32 most recently used sessions, reports unknown when either list extends past that cap, and reports first for the first request after a plugin restart. It describes the messages this client sent, not the provider's tokenized prefix or cache state.
  • steadyState and steadyStateByRoute are calculations over the retained 400-line log window, not lifetime counters: the request they drop is the first one still visible, which may already have been warm.
  • Failover account promotion is kept in memory only and resets to the configured order when the process restarts.
  • Image inputs require DSH's attachments service and fall back to placeholder text if unavailable.
  • Proxy support requires undici to be resolvable in the host environment; otherwise proxied requests fail with an explicit error.

Development

npm test                  # every test/*.mjs suite, no network, no keys
npm run check             # node --check over every shipped JS file
node test/tool-name.mjs   # or run a single suite directly

Tests run against a local mock HTTP server simulating Mistral SSE streams, with no external network calls and no API keys.

Fork lineage and upstream sync

01NGVI-QINGYI 0.1.2 (263a4f0)
├── santera-pixel 0.2.0 … 0.2.4
└── upstream issue #1 patch series 0.1.3 … 0.1.8
        └── this fork 0.3.0 → 0.4.0

This fork tracks upstream 01NGVI-QINGYI/dsh-mistral-pro by cherry-picking concrete fixes — never by merging upstream main, which would revert the 0.2.x/0.3.0 work. Add the remote once and inspect what is new:

git remote add upstream https://github.com/01NGVI-QINGYI/dsh-mistral-pro
git fetch upstream --depth=50
git log --oneline upstream/main

The two 0.x lines are independent: a higher number in one repository does not mean a newer tree in the other. See CONTRIBUTING.md for the merge policy, the version policy and how to record a sync in CHANGELOG.md.

License

MIT · originally by 01NGVI-QINGYI, fork maintained by lordraiden.

Related plugins