@project-ava-supernova/core 0.2.71

package/dist/providers/provider-registry.js

@@ -0,0 +1,197 @@
+import { DeepSeekProvider } from './deepseek/index.js';
+import { KimiProvider } from './kimi/index.js';
+import { QwenProvider } from './qwen/index.js';
+import { ZhipuProvider } from './zhipu/index.js';
+import { MistralProvider } from './mistral/index.js';
+import { MiniMaxProvider } from './minimax/index.js';
+import { XiaomiProvider } from './xiaomi/index.js';
+import { TencentProvider } from './tencent/index.js';
+import { NvidiaProvider } from './nvidia/index.js';
+// Single source of truth for the model catalogue — also exported browser-safe
+// via `@ava/core/models` so the IDE renderer shares the exact same list.
+import { ALL_MODELS } from './catalog.js';
+const BUILT_IN_PROVIDERS = {
+    deepseek: (config) => new DeepSeekProvider(config),
+    kimi: (config) => new KimiProvider(config),
+    qwen: (config) => new QwenProvider(config),
+    zhipu: (config) => new ZhipuProvider(config),
+    mistral: (config) => new MistralProvider(config),
+    minimax: (config) => new MiniMaxProvider(config),
+    xiaomi: (config) => new XiaomiProvider(config),
+    tencent: (config) => new TencentProvider(config),
+    nvidia: (config) => new NvidiaProvider(config),
+};
+/**
+ * Providers that Ava used to support and deliberately no longer does.
+ *
+ * Named rather than forgotten, because the two are different messages: a
+ * forgotten provider reads as a typo or a bug, and this was a decision. The
+ * key stays on the user's disk untouched — withdrawing support is ours to do,
+ * and deleting somebody's credential is not.
+ */
+const RETIRED_PROVIDERS = {
+    anthropic: 'Anthropic models are no longer supported in Ava. Pick another model — ' +
+        'the open-weight providers (Qwen, DeepSeek, Kimi, GLM, Mistral) cover the ' +
+        'same ground, and your Anthropic key has been left alone.',
+};
+export class ProviderRegistry {
+    providers = new Map();
+    constructor() {
+        // No built-in free provider — free users get Qwen through the platform provider
+    }
+    register(name, config) {
+        const factory = BUILT_IN_PROVIDERS[name];
+        if (!factory) {
+            // Retired providers get their own sentence. Somebody whose key worked
+            // yesterday deserves to be told it was withdrawn on purpose, not handed
+            // "Unknown provider" as though they had typed it wrong.
+            const retired = RETIRED_PROVIDERS[name];
+            throw new Error(retired ?? `Unknown provider: ${name}`);
+        }
+        this.providers.set(name, factory(config));
+    }
+    registerCustom(name, provider) {
+        this.providers.set(name, provider);
+    }
+    get(name) {
+        return this.providers.get(name);
+    }
+    resolveModel(qualifiedId) {
+        const [providerName, modelId] = qualifiedId.includes(':')
+            ? qualifiedId.split(':', 2)
+            : [undefined, qualifiedId];
+        if (providerName) {
+            const provider = this.providers.get(providerName);
+            if (!provider)
+                return undefined;
+            const model = this.findModel(provider, modelId);
+            if (!model)
+                return undefined;
+            return { provider, model };
+        }
+        for (const [, provider] of this.providers) {
+            const model = this.findModel(provider, modelId);
+            if (model)
+                return { provider, model };
+        }
+        return undefined;
+    }
+    /**
+     * Find a model on a provider, bridging the platform/native id-suffix duality.
+     *
+     * The same logical model is listed under different ids depending on surface:
+     * the managed platform serves DeepSeek/Mistral under `<id>-platform` rows
+     * (the server `models` table key), while a BYOK provider lists the native
+     * `<id>`. A routing table written against either form must resolve on both
+     * surfaces. So after an exact match, also try the id with `-platform`
+     * stripped and appended. Qwen uses one native id on both surfaces and is
+     * unaffected. Returns the concrete model so callers use its real `.id`.
+     */
+    findModel(provider, modelId) {
+        // Disabled models are invisible to resolution — auto-routing skips them and
+        // a stale saved selection falls back instead of failing. See ModelDefinition.disabled.
+        const models = provider.listModels().filter((m) => !m.disabled);
+        const exact = models.find((m) => m.id === modelId);
+        if (exact)
+            return exact;
+        const alt = modelId.endsWith('-platform')
+            ? modelId.slice(0, -'-platform'.length) // platform route → native BYOK id
+            : `${modelId}-platform`; // BYOK/native route → platform row
+        return models.find((m) => m.id === alt);
+    }
+    listAllModels() {
+        const models = [];
+        for (const [, provider] of this.providers) {
+            models.push(...provider.listModels().filter((m) => !m.disabled));
+        }
+        return models;
+    }
+    /**
+     * Build a fallback chain for a given model. Returns the primary entry plus
+     * compatible fallbacks from other registered providers, sorted by capability
+     * match, context window, and pricing. Capped at 3 fallbacks.
+     *
+     * Returns undefined if the primary model cannot be resolved.
+     */
+    buildFallbackChain(qualifiedId) {
+        const primary = this.resolveModel(qualifiedId);
+        if (!primary)
+            return undefined;
+        const candidates = [];
+        const primaryModel = primary.model;
+        for (const [, provider] of this.providers) {
+            if (provider === primary.provider)
+                continue; // skip primary's own provider
+            for (const model of provider.listModels()) {
+                // Never fall back onto a disabled model
+                if (model.disabled)
+                    continue;
+                // Must support tool calls if primary does
+                if (primaryModel.supportsToolCalls && !model.supportsToolCalls)
+                    continue;
+                // Must support streaming if primary does
+                if (primaryModel.supportsStreaming && !model.supportsStreaming)
+                    continue;
+                // Score: higher is better
+                let score = 0;
+                // Prefer matching thinking capability
+                if (primaryModel.supportsThinking && model.supportsThinking)
+                    score += 10;
+                // Prefer matching vision capability
+                if (primaryModel.supportsVision && model.supportsVision)
+                    score += 5;
+                // Larger context window is better
+                score += Math.min(model.contextWindow / 10_000, 20);
+                // Cheaper is better (if pricing available on both)
+                if (primaryModel.pricing && model.pricing) {
+                    if (model.pricing.inputPerMillion < primaryModel.pricing.inputPerMillion)
+                        score += 3;
+                }
+                candidates.push({ provider, model, score });
+            }
+        }
+        // Sort by score descending, take top 3
+        candidates.sort((a, b) => b.score - a.score);
+        const fallbacks = candidates.slice(0, 3).map(({ provider, model }) => ({ provider, model }));
+        return [{ provider: primary.provider, model: primary.model }, ...fallbacks];
+    }
+    /**
+     * List every model Ava supports, regardless of whether the provider is configured.
+     * Each entry includes `available: true` if the provider is registered (user has API key),
+     * or `available: false` if not.
+     */
+    listAllPossibleModels() {
+        const results = [];
+        for (const [providerName, models] of Object.entries(ALL_MODELS)) {
+            const isAvailable = this.providers.has(providerName);
+            for (const m of models) {
+                if (m.disabled)
+                    continue; // hidden from the picker while disabled
+                // Superseded models stay in the catalogue so routing and fallback can
+                // still resolve them, but they are not things to OFFER. The flag has
+                // existed on qwen3.5-plus and kimi-k2.6 the whole time and nothing
+                // read it, so both kept appearing in every picker — the operator saw
+                // them listed as choices on 2026-08-18.
+                //
+                // Filtered HERE and not in resolveModel: a fallback chain must still
+                // be able to reach them by id. "Do not offer this" and "do not use
+                // this" are different claims, and only the first one is being made.
+                if (m.hiddenFromPicker)
+                    continue;
+                results.push({ ...m, available: isAvailable });
+            }
+        }
+        // Include any custom/generic providers that aren't in ALL_MODELS
+        for (const [name, provider] of this.providers) {
+            if (!(name in ALL_MODELS)) {
+                for (const m of provider.listModels()) {
+                    if (m.disabled)
+                        continue;
+                    results.push({ ...m, available: true });
+                }
+            }
+        }
+        return results;
+    }
+}
+//# sourceMappingURL=provider-registry.js.map

package/dist/providers/provider-registry.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-registry.js","sourceRoot":"","sources":["../../src/providers/provider-registry.ts"],"names":[],"mappings":"AAGA,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACvD,OAAO,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AAC/C,OAAO,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AAC/C,OAAO,EAAE,aAAa,EAAE,MAAM,kBAAkB,CAAC;AACjD,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AACrD,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AACrD,OAAO,EAAE,cAAc,EAAE,MAAM,mBAAmB,CAAC;AACnD,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AACrD,OAAO,EAAE,cAAc,EAAE,MAAM,mBAAmB,CAAC;AACnD,8EAA8E;AAC9E,yEAAyE;AACzE,OAAO,EAAE,UAAU,EAAE,MAAM,cAAc,CAAC;AAI1C,MAAM,kBAAkB,GAAoC;IAC1D,QAAQ,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,gBAAgB,CAAC,MAAM,CAAC;IAClD,IAAI,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,YAAY,CAAC,MAAM,CAAC;IAC1C,IAAI,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,YAAY,CAAC,MAAM,CAAC;IAC1C,KAAK,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,aAAa,CAAC,MAAM,CAAC;IAC5C,OAAO,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,eAAe,CAAC,MAAM,CAAC;IAChD,OAAO,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,eAAe,CAAC,MAAM,CAAC;IAChD,MAAM,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,cAAc,CAAC,MAAM,CAAC;IAC9C,OAAO,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,eAAe,CAAC,MAAM,CAAC;IAChD,MAAM,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,IAAI,cAAc,CAAC,MAAM,CAAC;CAC/C,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,iBAAiB,GAA2B;IAChD,SAAS,EACP,wEAAwE;QACxE,2EAA2E;QAC3E,0DAA0D;CAC7D,CAAC;AAEF,MAAM,OAAO,gBAAgB;IACnB,SAAS,GAAG,IAAI,GAAG,EAAoB,CAAC;IAEhD;QACE,gFAAgF;IAClF,CAAC;IAED,QAAQ,CAAC,IAAY,EAAE,MAAsB;QAC3C,MAAM,OAAO,GAAG,kBAAkB,CAAC,IAAI,CAAC,CAAC;QACzC,IAAI,CAAC,OAAO,EAAE,CAAC;YACb,sEAAsE;YACtE,wEAAwE;YACxE,wDAAwD;YACxD,MAAM,OAAO,GAAG,iBAAiB,CAAC,IAAI,CAAC,CAAC;YACxC,MAAM,IAAI,KAAK,CAAC,OAAO,IAAI,qBAAqB,IAAI,EAAE,CAAC,CAAC;QAC1D,CAAC;QACD,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC;IAC5C,CAAC;IAED,cAAc,CAAC,IAAY,EAAE,QAAkB;QAC7C,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;IACrC,CAAC;IAED,GAAG,CAAC,IAAY;QACd,OAAO,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAClC,CAAC;IAED,YAAY,CAAC,WAAmB;QAC9B,MAAM,CAAC,YAAY,EAAE,OAAO,CAAC,GAAG,WAAW,CAAC,QAAQ,CAAC,GAAG,CAAC;YACvD,CAAC,CAAE,WAAW,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,CAAsB;YACjD,CAAC,CAAC,CAAC,SAAS,EAAE,WAAW,CAAC,CAAC;QAE7B,IAAI,YAAY,EAAE,CAAC;YACjB,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;YAClD,IAAI,CAAC,QAAQ;gBAAE,OAAO,SAAS,CAAC;YAChC,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;YAChD,IAAI,CAAC,KAAK;gBAAE,OAAO,SAAS,CAAC;YAC7B,OAAO,EAAE,QAAQ,EAAE,KAAK,EAAE,CAAC;QAC7B,CAAC;QAED,KAAK,MAAM,CAAC,EAAE,QAAQ,CAAC,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YAC1C,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;YAChD,IAAI,KAAK;gBAAE,OAAO,EAAE,QAAQ,EAAE,KAAK,EAAE,CAAC;QACxC,CAAC;QAED,OAAO,SAAS,CAAC;IACnB,CAAC;IAED;;;;;;;;;;OAUG;IACK,SAAS,CAAC,QAAkB,EAAE,OAAe;QACnD,4EAA4E;QAC5E,uFAAuF;QACvF,MAAM,MAAM,GAAG,QAAQ,CAAC,UAAU,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;QAChE,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,OAAO,CAAC,CAAC;QACnD,IAAI,KAAK;YAAE,OAAO,KAAK,CAAC;QAExB,MAAM,GAAG,GAAG,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAC;YACvC,CAAC,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC,kCAAkC;YAC1E,CAAC,CAAC,GAAG,OAAO,WAAW,CAAC,CAAgB,mCAAmC;QAC7E,OAAO,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,GAAG,CAAC,CAAC;IAC1C,CAAC;IAED,aAAa;QACX,MAAM,MAAM,GAAsB,EAAE,CAAC;QACrC,KAAK,MAAM,CAAC,EAAE,QAAQ,CAAC,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YAC1C,MAAM,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC,UAAU,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC;QACnE,CAAC;QACD,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;;;;;OAMG;IACH,kBAAkB,CAAC,WAAmB;QACpC,MAAM,OAAO,GAAG,IAAI,CAAC,YAAY,CAAC,WAAW,CAAC,CAAC;QAC/C,IAAI,CAAC,OAAO;YAAE,OAAO,SAAS,CAAC;QAE/B,MAAM,UAAU,GAA6C,EAAE,CAAC;QAChE,MAAM,YAAY,GAAG,OAAO,CAAC,KAAK,CAAC;QAEnC,KAAK,MAAM,CAAC,EAAE,QAAQ,CAAC,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YAC1C,IAAI,QAAQ,KAAK,OAAO,CAAC,QAAQ;gBAAE,SAAS,CAAC,8BAA8B;YAE3E,KAAK,MAAM,KAAK,IAAI,QAAQ,CAAC,UAAU,EAAE,EAAE,CAAC;gBAC1C,wCAAwC;gBACxC,IAAI,KAAK,CAAC,QAAQ;oBAAE,SAAS;gBAC7B,0CAA0C;gBAC1C,IAAI,YAAY,CAAC,iBAAiB,IAAI,CAAC,KAAK,CAAC,iBAAiB;oBAAE,SAAS;gBACzE,yCAAyC;gBACzC,IAAI,YAAY,CAAC,iBAAiB,IAAI,CAAC,KAAK,CAAC,iBAAiB;oBAAE,SAAS;gBAEzE,0BAA0B;gBAC1B,IAAI,KAAK,GAAG,CAAC,CAAC;gBAEd,sCAAsC;gBACtC,IAAI,YAAY,CAAC,gBAAgB,IAAI,KAAK,CAAC,gBAAgB;oBAAE,KAAK,IAAI,EAAE,CAAC;gBACzE,oCAAoC;gBACpC,IAAI,YAAY,CAAC,cAAc,IAAI,KAAK,CAAC,cAAc;oBAAE,KAAK,IAAI,CAAC,CAAC;gBACpE,kCAAkC;gBAClC,KAAK,IAAI,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,aAAa,GAAG,MAAM,EAAE,EAAE,CAAC,CAAC;gBACpD,mDAAmD;gBACnD,IAAI,YAAY,CAAC,OAAO,IAAI,KAAK,CAAC,OAAO,EAAE,CAAC;oBAC1C,IAAI,KAAK,CAAC,OAAO,CAAC,eAAe,GAAG,YAAY,CAAC,OAAO,CAAC,eAAe;wBAAE,KAAK,IAAI,CAAC,CAAC;gBACvF,CAAC;gBAED,UAAU,CAAC,IAAI,CAAC,EAAE,QAAQ,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,CAAC;YAC9C,CAAC;QACH,CAAC;QAED,uCAAuC;QACvC,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;QAC7C,MAAM,SAAS,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,QAAQ,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,QAAQ,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC;QAE7F,OAAO,CAAC,EAAE,QAAQ,EAAE,OAAO,CAAC,QAAQ,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,EAAE,EAAE,GAAG,SAAS,CAAC,CAAC;IAC9E,CAAC;IAED;;;;OAIG;IACH,qBAAqB;QACnB,MAAM,OAAO,GAAoD,EAAE,CAAC;QACpE,KAAK,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,UAAU,CAAC,EAAE,CAAC;YAChE,MAAM,WAAW,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;YACrD,KAAK,MAAM,CAAC,IAAI,MAAM,EAAE,CAAC;gBACvB,IAAI,CAAC,CAAC,QAAQ;oBAAE,SAAS,CAAC,wCAAwC;gBAClE,sEAAsE;gBACtE,qEAAqE;gBACrE,mEAAmE;gBACnE,qEAAqE;gBACrE,wCAAwC;gBACxC,EAAE;gBACF,qEAAqE;gBACrE,mEAAmE;gBACnE,oEAAoE;gBACpE,IAAI,CAAC,CAAC,gBAAgB;oBAAE,SAAS;gBACjC,OAAO,CAAC,IAAI,CAAC,EAAE,GAAG,CAAC,EAAE,SAAS,EAAE,WAAW,EAAE,CAAC,CAAC;YACjD,CAAC;QACH,CAAC;QACD,iEAAiE;QACjE,KAAK,MAAM,CAAC,IAAI,EAAE,QAAQ,CAAC,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YAC9C,IAAI,CAAC,CAAC,IAAI,IAAI,UAAU,CAAC,EAAE,CAAC;gBAC1B,KAAK,MAAM,CAAC,IAAI,QAAQ,CAAC,UAAU,EAAE,EAAE,CAAC;oBACtC,IAAI,CAAC,CAAC,QAAQ;wBAAE,SAAS;oBACzB,OAAO,CAAC,IAAI,CAAC,EAAE,GAAG,CAAC,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;gBAC1C,CAAC;YACH,CAAC;QACH,CAAC;QACD,OAAO,OAAO,CAAC;IACjB,CAAC;CACF"}

package/dist/providers/qwen/index.d.ts

@@ -0,0 +1,9 @@
+import { BaseProvider } from '../base-provider.js';
+import type { ModelDefinition } from '../../core/types.js';
+export declare class QwenProvider extends BaseProvider {
+    readonly name = "qwen";
+    readonly displayName = "Qwen (Alibaba Cloud)";
+    protected getDefaultBaseUrl(): string;
+    listModels(): ModelDefinition[];
+}
+//# sourceMappingURL=index.d.ts.map

package/dist/providers/qwen/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/providers/qwen/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AACnD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAG3D,qBAAa,YAAa,SAAQ,YAAY;IAC5C,QAAQ,CAAC,IAAI,UAAU;IACvB,QAAQ,CAAC,WAAW,0BAA0B;IAE9C,SAAS,CAAC,iBAAiB,IAAI,MAAM;IAIrC,UAAU,IAAI,eAAe,EAAE;CAShC"}

package/dist/providers/qwen/index.js

@@ -0,0 +1,13 @@
+import { BaseProvider } from '../base-provider.js';
+import { QWEN_MODELS } from './models.js';
+export class QwenProvider extends BaseProvider {
+    name = 'qwen';
+    displayName = 'Qwen (Alibaba Cloud)';
+    getDefaultBaseUrl() {
+        return 'https://dashscope-intl.aliyuncs.com/compatible-mode/v1';
+    }
+    listModels() {
+        return QWEN_MODELS;
+    }
+}
+//# sourceMappingURL=index.js.map

package/dist/providers/qwen/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/providers/qwen/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AAEnD,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAE1C,MAAM,OAAO,YAAa,SAAQ,YAAY;IACnC,IAAI,GAAG,MAAM,CAAC;IACd,WAAW,GAAG,sBAAsB,CAAC;IAEpC,iBAAiB;QACzB,OAAO,wDAAwD,CAAC;IAClE,CAAC;IAED,UAAU;QACR,OAAO,WAAW,CAAC;IACrB,CAAC;CAOF"}

package/dist/providers/qwen/models.d.ts

@@ -0,0 +1,3 @@
+import type { ModelDefinition } from '../../core/types.js';
+export declare const QWEN_MODELS: ModelDefinition[];
+//# sourceMappingURL=models.d.ts.map

package/dist/providers/qwen/models.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../../../src/providers/qwen/models.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAE3D,eAAO,MAAM,WAAW,EAAE,eAAe,EAkKxC,CAAC"}

package/dist/providers/qwen/models.js

@@ -0,0 +1,164 @@
+export const QWEN_MODELS = [
+    // Qwen 3.8 Max (3 August 2026) — Alibaba's flagship, and a straight upgrade
+    // on 3.7 Max rather than a trade-off: cheaper ($2/$6 vs $2.50/$7.50), double
+    // the output ceiling (131,072 vs 65,536), 2.4T MoE / ~95B active, and video
+    // in as well as images. Id verified live against DashScope intl — the docs
+    // page had not caught up yet, the API had.
+    //
+    // Still BYOK-first for single-pick, same reasoning as 3.7 Max: $2/$6 does not
+    // fit the credit-plan margin as a default. Qwen 3.7 Plus stays the Maestro
+    // conductor.
+    {
+        id: 'qwen3.8-max',
+        name: 'Qwen 3.8 Max',
+        provider: 'qwen',
+        contextWindow: 1000000,
+        maxOutputTokens: 131072,
+        supportsToolCalls: true,
+        supportsStreaming: true,
+        supportsThinking: true,
+        supportsVision: true,
+        desktopCapable: true,
+        pricing: { inputPerMillion: 2.00, outputPerMillion: 6.00 },
+    },
+    // Qwen 3.7 Max was here until 2026-08-09, retired in favour of 3.8 Max.
+    // Even on your own key there is no reason to send it: same provider, lower
+    // price, higher output ceiling, more capable. The id rolls forward in
+    // model-ids.ts so a saved selection keeps working.
+    // Qwen 3.7 Plus (June 2026) — the Maestro conductor. 1M context, vision +
+    // video, reasoning, native function calling. Supersedes Qwen 3.6 Plus and the
+    // 3.5 Omni tier: better agentic coding, multimodal, and cheaper ($0.40/$1.60).
+    {
+        id: 'qwen3.7-plus',
+        name: 'Qwen 3.7 Plus',
+        provider: 'qwen',
+        contextWindow: 1000000,
+        maxOutputTokens: 65536,
+        supportsToolCalls: true,
+        supportsStreaming: true,
+        supportsThinking: true,
+        supportsVision: true,
+        desktopCapable: true,
+        pricing: { inputPerMillion: 0.40, outputPerMillion: 1.60 },
+    },
+    {
+        id: 'qwen3.5-plus',
+        name: 'Qwen 3.5 Plus',
+        provider: 'qwen',
+        contextWindow: 1000000,
+        maxOutputTokens: 128000,
+        supportsToolCalls: true,
+        supportsStreaming: true,
+        supportsThinking: true,
+        supportsVision: true,
+        desktopCapable: true,
+        hiddenFromPicker: true, // superseded by Qwen 3.7 Plus — kept for routing/fallback only
+        pricing: { inputPerMillion: 0.20, outputPerMillion: 1.20 },
+    },
+    {
+        id: 'qwen3.5-flash',
+        name: 'Qwen 3.5 Flash',
+        provider: 'qwen',
+        contextWindow: 256000,
+        maxOutputTokens: 8192,
+        supportsToolCalls: true,
+        supportsStreaming: true,
+        // FALSE IS DELIBERATE, AND THE NAME IS MISLEADING. This model DOES reason
+        // by default — measured 2026-08-16, 142 reasoning tokens on "reply with
+        // one word". So read as a capability claim, the flag is simply wrong.
+        //
+        // But the only thing it controls is whether agent.ts REPLAYS
+        // reasoning_content back to the provider on the next turn, and Qwen
+        // rejects that with a 422 (the companion route carries the same note:
+        // "reasoning_content is output-only — replaying it as input 422s on Qwen
+        // and Mistral and 400s on DeepSeek"). So false produces the correct
+        // behaviour — strip it — and flipping it to true to "fix" the label would
+        // start 422-ing every multi-turn Qwen conversation.
+        //
+        // If you want to stop this model reasoning, that is a different lever:
+        // enable_thinking: false on the request (see ChatCompletionRequest). The
+        // intent gate does exactly that.
+        supportsThinking: false,
+        supportsVision: false,
+        pricing: { inputPerMillion: 0.05, outputPerMillion: 0.40 },
+    },
+    {
+        // Qwen 3.7 Flash (Jul 2026) — 1M context, multimodal, function calling.
+        // Verified against the live DashScope model list, not inferred from a
+        // version number.
+        //
+        // OFFERED, BUT NOT USED AS THE INTENT GATE, and that is deliberate. It is
+        // dramatically cheaper than 3.5 Flash on short prompts and dearer on long
+        // ones, which is fine for a model somebody chooses. The gate is different:
+        // it runs on EVERY request and its whole job is returning a reliable tool
+        // call quickly. Reported figures are an ~8.9% tool-call error rate and a
+        // P99 above 90 seconds, so it stays out of that seat until we have measured
+        // it ourselves.
+        //
+        // IT REASONS BY DEFAULT, AND YOU PAY FOR IT. A live call asking for the
+        // single word "ok" came back with 167 completion tokens, 162 of them
+        // reasoning. That is roughly ten times the output a trivial answer needs,
+        // billed at the output rate — so the real cost of a short exchange is far
+        // above what the per-token price suggests, and it is a second reason the
+        // intent gate stays where it is.
+        //
+        // PRICING IS TIERED BY PROMPT LENGTH and this field is flat, so it cannot
+        // be told the whole truth: $0.03/$0.13 under 32K, $0.10/$0.40 to 256K,
+        // $0.20/$0.80 to 1M. The middle tier is quoted here because an agentic turn
+        // carries a system prompt and context and clears 32K almost immediately —
+        // quoting the $0.03 headline would understate nearly every real call, and a
+        // cost estimate that flatters us is worse than no estimate.
+        id: 'qwen3.7-flash',
+        name: 'Qwen 3.7 Flash',
+        provider: 'qwen',
+        contextWindow: 1000000,
+        maxOutputTokens: 32768,
+        supportsToolCalls: true,
+        supportsStreaming: true,
+        supportsThinking: true,
+        supportsVision: true,
+        pricing: { inputPerMillion: 0.10, outputPerMillion: 0.40 },
+    },
+    {
+        // Qwen3 Coder Next — the coder tier we did not have. Ava is an agentic
+        // coding tool that shipped exactly ONE coder-specialised model, while these
+        // sat on an endpoint we already call with a key we already hold.
+        //
+        // 80B total, 3B active, so it prices like a small model and scores like a
+        // large one: 70.6% SWE-bench Verified at $0.12/$0.80 — against kimi-k2.7-code
+        // at $0.95/$4.00 for the same class of work. Native 256K.
+        //
+        // Verified live before shipping: it answers, and it does NOT burn reasoning
+        // tokens on trivial replies (2 completion tokens for "ok", where Qwen 3.7
+        // Flash spent 167).
+        id: 'qwen3-coder-next',
+        name: 'Qwen3 Coder Next',
+        provider: 'qwen',
+        contextWindow: 256000,
+        maxOutputTokens: 65536,
+        supportsToolCalls: true,
+        supportsStreaming: true,
+        supportsThinking: false,
+        supportsVision: false,
+        pricing: { inputPerMillion: 0.12, outputPerMillion: 0.80 },
+    },
+    {
+        // Qwen3 Coder Flash — 1M context, tool-calling coder.
+        //
+        // The name is misleading and it is worth saying so here rather than letting
+        // somebody assume: this is DEARER than Coder Next ($0.195/$0.975 against
+        // $0.12/$0.80), not cheaper. What it buys is the 1M context window, which
+        // Coder Next does not have. Pick it for context length, never for price.
+        id: 'qwen3-coder-flash',
+        name: 'Qwen3 Coder Flash',
+        provider: 'qwen',
+        contextWindow: 1000000,
+        maxOutputTokens: 65536,
+        supportsToolCalls: true,
+        supportsStreaming: true,
+        supportsThinking: false,
+        supportsVision: false,
+        pricing: { inputPerMillion: 0.195, outputPerMillion: 0.975 },
+    },
+];
+//# sourceMappingURL=models.js.map

package/dist/providers/qwen/models.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"models.js","sourceRoot":"","sources":["../../../src/providers/qwen/models.ts"],"names":[],"mappings":"AAEA,MAAM,CAAC,MAAM,WAAW,GAAsB;IAC5C,4EAA4E;IAC5E,6EAA6E;IAC7E,4EAA4E;IAC5E,2EAA2E;IAC3E,2CAA2C;IAC3C,EAAE;IACF,8EAA8E;IAC9E,2EAA2E;IAC3E,aAAa;IACb;QACE,EAAE,EAAE,aAAa;QACjB,IAAI,EAAE,cAAc;QACpB,QAAQ,EAAE,MAAM;QAChB,aAAa,EAAE,OAAO;QACtB,eAAe,EAAE,MAAM;QACvB,iBAAiB,EAAE,IAAI;QACvB,iBAAiB,EAAE,IAAI;QACvB,gBAAgB,EAAE,IAAI;QACtB,cAAc,EAAE,IAAI;QACpB,cAAc,EAAE,IAAI;QACpB,OAAO,EAAE,EAAE,eAAe,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE;KAC3D;IACD,wEAAwE;IACxE,2EAA2E;IAC3E,sEAAsE;IACtE,mDAAmD;IACnD,0EAA0E;IAC1E,8EAA8E;IAC9E,+EAA+E;IAC/E;QACE,EAAE,EAAE,cAAc;QAClB,IAAI,EAAE,eAAe;QACrB,QAAQ,EAAE,MAAM;QAChB,aAAa,EAAE,OAAO;QACtB,eAAe,EAAE,KAAK;QACtB,iBAAiB,EAAE,IAAI;QACvB,iBAAiB,EAAE,IAAI;QACvB,gBAAgB,EAAE,IAAI;QACtB,cAAc,EAAE,IAAI;QACpB,cAAc,EAAE,IAAI;QACpB,OAAO,EAAE,EAAE,eAAe,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE;KAC3D;IACD;QACE,EAAE,EAAE,cAAc;QAClB,IAAI,EAAE,eAAe;QACrB,QAAQ,EAAE,MAAM;QAChB,aAAa,EAAE,OAAO;QACtB,eAAe,EAAE,MAAM;QACvB,iBAAiB,EAAE,IAAI;QACvB,iBAAiB,EAAE,IAAI;QACvB,gBAAgB,EAAE,IAAI;QACtB,cAAc,EAAE,IAAI;QACpB,cAAc,EAAE,IAAI;QACpB,gBAAgB,EAAE,IAAI,EAAE,+DAA+D;QACvF,OAAO,EAAE,EAAE,eAAe,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE;KAC3D;IACD;QACE,EAAE,EAAE,eAAe;QACnB,IAAI,EAAE,gBAAgB;QACtB,QAAQ,EAAE,MAAM;QAChB,aAAa,EAAE,MAAM;QACrB,eAAe,EAAE,IAAI;QACrB,iBAAiB,EAAE,IAAI;QACvB,iBAAiB,EAAE,IAAI;QACvB,0EAA0E;QAC1E,wEAAwE;QACxE,sEAAsE;QACtE,EAAE;QACF,6DAA6D;QAC7D,oEAAoE;QACpE,sEAAsE;QACtE,yEAAyE;QACzE,oEAAoE;QACpE,0EAA0E;QAC1E,oDAAoD;QACpD,EAAE;QACF,uEAAuE;QACvE,yEAAyE;QACzE,iCAAiC;QACjC,gBAAgB,EAAE,KAAK;QACvB,cAAc,EAAE,KAAK;QACrB,OAAO,EAAE,EAAE,eAAe,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE;KAC3D;IACD;QACE,wEAAwE;QACxE,sEAAsE;QACtE,kBAAkB;QAClB,EAAE;QACF,0EAA0E;QAC1E,0EAA0E;QAC1E,2EAA2E;QAC3E,0EAA0E;QAC1E,yEAAyE;QACzE,4EAA4E;QAC5E,gBAAgB;QAChB,EAAE;QACF,wEAAwE;QACxE,qEAAqE;QACrE,0EAA0E;QAC1E,0EAA0E;QAC1E,yEAAyE;QACzE,iCAAiC;QACjC,EAAE;QACF,0EAA0E;QAC1E,uEAAuE;QACvE,4EAA4E;QAC5E,0EAA0E;QAC1E,4EAA4E;QAC5E,4DAA4D;QAC5D,EAAE,EAAE,eAAe;QACnB,IAAI,EAAE,gBAAgB;QACtB,QAAQ,EAAE,MAAM;QAChB,aAAa,EAAE,OAAO;QACtB,eAAe,EAAE,KAAK;QACtB,iBAAiB,EAAE,IAAI;QACvB,iBAAiB,EAAE,IAAI;QACvB,gBAAgB,EAAE,IAAI;QACtB,cAAc,EAAE,IAAI;QACpB,OAAO,EAAE,EAAE,eAAe,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE;KAC3D;IACD;QACE,uEAAuE;QACvE,4EAA4E;QAC5E,iEAAiE;QACjE,EAAE;QACF,0EAA0E;QAC1E,8EAA8E;QAC9E,0DAA0D;QAC1D,EAAE;QACF,4EAA4E;QAC5E,0EAA0E;QAC1E,oBAAoB;QACpB,EAAE,EAAE,kBAAkB;QACtB,IAAI,EAAE,kBAAkB;QACxB,QAAQ,EAAE,MAAM;QAChB,aAAa,EAAE,MAAM;QACrB,eAAe,EAAE,KAAK;QACtB,iBAAiB,EAAE,IAAI;QACvB,iBAAiB,EAAE,IAAI;QACvB,gBAAgB,EAAE,KAAK;QACvB,cAAc,EAAE,KAAK;QACrB,OAAO,EAAE,EAAE,eAAe,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE;KAC3D;IACD;QACE,sDAAsD;QACtD,EAAE;QACF,4EAA4E;QAC5E,yEAAyE;QACzE,0EAA0E;QAC1E,yEAAyE;QACzE,EAAE,EAAE,mBAAmB;QACvB,IAAI,EAAE,mBAAmB;QACzB,QAAQ,EAAE,MAAM;QAChB,aAAa,EAAE,OAAO;QACtB,eAAe,EAAE,KAAK;QACtB,iBAAiB,EAAE,IAAI;QACvB,iBAAiB,EAAE,IAAI;QACvB,gBAAgB,EAAE,KAAK;QACvB,cAAc,EAAE,KAAK;QACrB,OAAO,EAAE,EAAE,eAAe,EAAE,KAAK,EAAE,gBAAgB,EAAE,KAAK,EAAE;KAC7D;CACF,CAAC"}

package/dist/providers/request-shaping/index.d.ts

@@ -0,0 +1,29 @@
+/**
+ * Shared LLM request-shaping — the single source of truth both the PLATFORM
+ * path (packages/web routes) and the BYOK path (core providers) import so the
+ * outbound request body is identical regardless of who holds the key.
+ *
+ * Transport-free by design: callers do their own fetch/auth; this only shapes
+ * the body. OpenAI-compatible providers use `shapeOpenAICompatBody` (or the
+ * granular pieces).
+ */
+export { MODEL_API_NAMES, VISION_REROUTE, resolveApiModel, messagesHaveImages, } from './model-ids.js';
+export { stripReasoningContent, reorderSystemForQwen, markTrailingAssistantPrefix, shapeMessages, } from './messages.js';
+export { isZhipuFlashModel, shapeParams, } from './params.js';
+export type { ShapeableParams } from './params.js';
+import { type ShapeableParams } from './params.js';
+export interface ShapeOpenAICompatInput extends ShapeableParams {
+    /** Provider key: 'qwen' | 'deepseek' | 'mistral' | 'zhipu' | 'minimax' | ... */
+    provider: string;
+    /** Friendly model id (pre-translation). */
+    model: string;
+    messages: ReadonlyArray<Record<string, unknown>>;
+}
+/**
+ * Build a complete OpenAI-compatible chat-completion request body: translate
+ * the model id (with vision reroute when images are present), shape the
+ * messages array, and apply per-provider parameter quirks. Returns the body
+ * object ready to JSON.stringify — minus auth, which the caller supplies.
+ */
+export declare function shapeOpenAICompatBody(input: ShapeOpenAICompatInput): Record<string, unknown>;
+//# sourceMappingURL=index.d.ts.map

package/dist/providers/request-shaping/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/providers/request-shaping/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,EACL,eAAe,EACf,cAAc,EACd,eAAe,EACf,kBAAkB,GACnB,MAAM,gBAAgB,CAAC;AAExB,OAAO,EACL,qBAAqB,EACrB,oBAAoB,EACpB,2BAA2B,EAC3B,aAAa,GACd,MAAM,eAAe,CAAC;AAEvB,OAAO,EACL,iBAAiB,EACjB,WAAW,GACZ,MAAM,aAAa,CAAC;AACrB,YAAY,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAInD,OAAO,EAAe,KAAK,eAAe,EAAE,MAAM,aAAa,CAAC;AAEhE,MAAM,WAAW,sBAAuB,SAAQ,eAAe;IAC7D,gFAAgF;IAChF,QAAQ,EAAE,MAAM,CAAC;IACjB,2CAA2C;IAC3C,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,aAAa,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC;CAClD;AAED;;;;;GAKG;AACH,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,sBAAsB,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAQ5F"}

package/dist/providers/request-shaping/index.js

@@ -0,0 +1,31 @@
+/**
+ * Shared LLM request-shaping — the single source of truth both the PLATFORM
+ * path (packages/web routes) and the BYOK path (core providers) import so the
+ * outbound request body is identical regardless of who holds the key.
+ *
+ * Transport-free by design: callers do their own fetch/auth; this only shapes
+ * the body. OpenAI-compatible providers use `shapeOpenAICompatBody` (or the
+ * granular pieces).
+ */
+export { MODEL_API_NAMES, VISION_REROUTE, resolveApiModel, messagesHaveImages, } from './model-ids.js';
+export { stripReasoningContent, reorderSystemForQwen, markTrailingAssistantPrefix, shapeMessages, } from './messages.js';
+export { isZhipuFlashModel, shapeParams, } from './params.js';
+import { resolveApiModel, messagesHaveImages } from './model-ids.js';
+import { shapeMessages } from './messages.js';
+import { shapeParams } from './params.js';
+/**
+ * Build a complete OpenAI-compatible chat-completion request body: translate
+ * the model id (with vision reroute when images are present), shape the
+ * messages array, and apply per-provider parameter quirks. Returns the body
+ * object ready to JSON.stringify — minus auth, which the caller supplies.
+ */
+export function shapeOpenAICompatBody(input) {
+    const { provider, model, messages, ...params } = input;
+    const hasImages = messagesHaveImages(messages);
+    return {
+        model: resolveApiModel(model, hasImages),
+        messages: shapeMessages(provider, messages),
+        ...shapeParams(provider, model, params),
+    };
+}
+//# sourceMappingURL=index.js.map

package/dist/providers/request-shaping/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/providers/request-shaping/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,EACL,eAAe,EACf,cAAc,EACd,eAAe,EACf,kBAAkB,GACnB,MAAM,gBAAgB,CAAC;AAExB,OAAO,EACL,qBAAqB,EACrB,oBAAoB,EACpB,2BAA2B,EAC3B,aAAa,GACd,MAAM,eAAe,CAAC;AAEvB,OAAO,EACL,iBAAiB,EACjB,WAAW,GACZ,MAAM,aAAa,CAAC;AAGrB,OAAO,EAAE,eAAe,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAC;AACrE,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AAC9C,OAAO,EAAE,WAAW,EAAwB,MAAM,aAAa,CAAC;AAUhE;;;;;GAKG;AACH,MAAM,UAAU,qBAAqB,CAAC,KAA6B;IACjE,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,QAAQ,EAAE,GAAG,MAAM,EAAE,GAAG,KAAK,CAAC;IACvD,MAAM,SAAS,GAAG,kBAAkB,CAAC,QAAQ,CAAC,CAAC;IAC/C,OAAO;QACL,KAAK,EAAE,eAAe,CAAC,KAAK,EAAE,SAAS,CAAC;QACxC,QAAQ,EAAE,aAAa,CAAC,QAAQ,EAAE,QAAQ,CAAC;QAC3C,GAAG,WAAW,CAAC,QAAQ,EAAE,KAAK,EAAE,MAAM,CAAC;KACxC,CAAC;AACJ,CAAC"}

package/dist/providers/request-shaping/messages.d.ts

@@ -0,0 +1,61 @@
+/**
+ * Shared message-array shaping for OpenAI-compatible chat requests.
+ *
+ * Same rationale as model-ids: the platform path (web routes) and the BYOK
+ * path (core providers) both have to massage the messages array into a shape
+ * the upstream provider accepts, and the copies drifted (the `reasoning_content`
+ * 422 on Mistral was exactly this — web stripped it for Qwen only, core kept it
+ * for thinking models). One implementation, both paths import it.
+ *
+ * Loosely typed (Record<string, unknown>) so both a web route's plain JSON and
+ * a core provider's typed messages can pass through without a type dance.
+ *
+ * Scope note: ANSI/control-char + null-content cleaning is deliberately NOT here
+ * yet — it's a core-only concern (raw shell output reaching the model) and the
+ * web path doesn't do it today. Core keeps its existing inline cleaning; folding
+ * it in is a later convergence step, kept out now to preserve exact web parity.
+ */
+type LooseMessage = Record<string, unknown>;
+/**
+ * reasoning_content is an OUTPUT-only field that reasoning models EMIT — they
+ * reject it as INPUT (Qwen + Mistral 422 `extra_forbidden`, DeepSeek 400). No
+ * provider needs its own prior thinking replayed. Strip it for every provider.
+ */
+export declare function stripReasoningContent<T extends LooseMessage>(messages: ReadonlyArray<T>): T[];
+/**
+ * Qwen (DashScope) requires system messages at the very start of the array.
+ * Merge multiple system messages into one leading block; if a single system
+ * message exists but isn't first, move it to the front. No-op when already valid.
+ */
+export declare function reorderSystemForQwen<T extends LooseMessage>(messages: ReadonlyArray<T>): T[];
+/**
+ * Mistral refuses a conversation whose last message is an assistant turn:
+ *
+ *   400 invalid_request_message_order — "Expected last role User or Tool
+ *   (or Assistant with prefix True) for serving but got assistant"
+ *
+ * That state is reachable. The agent appends a PARTIAL assistant reply and
+ * loops when a stream is interrupted mid-answer, relying on the interjection
+ * drain to add a user message behind it — and the queue can be emptied in
+ * between (a stop clears it). Interrupting a stream went from impossible to a
+ * button people press on 2026-08-17, so this stopped being theoretical.
+ *
+ * Mistral's own error names the fix: `prefix: true` marks a trailing assistant
+ * message as one to CONTINUE rather than one to reply to — which is exactly
+ * what a half-finished reply is. So the message is kept and completed rather
+ * than thrown away, which is the honest handling: the user already watched
+ * that text stream onto their screen.
+ *
+ * Only Mistral is strict about this today. The others accept it, so nothing
+ * else is touched — a rule applied where it isn't needed is how "why is this
+ * here?" comments start.
+ */
+export declare function markTrailingAssistantPrefix<T extends LooseMessage>(provider: string, messages: ReadonlyArray<T>): T[];
+/**
+ * The shared OpenAI-compatible message pipeline: always strip reasoning_content,
+ * reorder system messages for Qwen, and mark a trailing assistant as a Mistral
+ * prefix continuation. Provider-aware.
+ */
+export declare function shapeMessages<T extends LooseMessage>(provider: string, messages: ReadonlyArray<T>): T[];
+export {};
+//# sourceMappingURL=messages.d.ts.map

package/dist/providers/request-shaping/messages.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"messages.d.ts","sourceRoot":"","sources":["../../../src/providers/request-shaping/messages.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAIH,KAAK,YAAY,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;AAE5C;;;;GAIG;AACH,wBAAgB,qBAAqB,CAAC,CAAC,SAAS,YAAY,EAAE,QAAQ,EAAE,aAAa,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAS7F;AAED;;;;GAIG;AACH,wBAAgB,oBAAoB,CAAC,CAAC,SAAS,YAAY,EAAE,QAAQ,EAAE,aAAa,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAW5F;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAgB,2BAA2B,CAAC,CAAC,SAAS,YAAY,EAChE,QAAQ,EAAE,MAAM,EAChB,QAAQ,EAAE,aAAa,CAAC,CAAC,CAAC,GACzB,CAAC,EAAE,CA8BL;AAED;;;;GAIG;AACH,wBAAgB,aAAa,CAAC,CAAC,SAAS,YAAY,EAClD,QAAQ,EAAE,MAAM,EAChB,QAAQ,EAAE,aAAa,CAAC,CAAC,CAAC,GACzB,CAAC,EAAE,CAIL"}