@project-ava-supernova/core 0.2.71

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

@@ -0,0 +1,114 @@
+/**
+ * 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.
+ */
+import { logger } from '../../core/logger.js';
+/**
+ * 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 function stripReasoningContent(messages) {
+    return messages.map((m) => {
+        if ('reasoning_content' in m) {
+            const { reasoning_content: _rc, ...rest } = m;
+            void _rc;
+            return rest;
+        }
+        return m;
+    });
+}
+/**
+ * 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 function reorderSystemForQwen(messages) {
+    const sys = messages.filter((m) => m.role === 'system');
+    const nonSys = messages.filter((m) => m.role !== 'system');
+    if (sys.length > 1) {
+        const merged = sys.map((m) => (typeof m.content === 'string' ? m.content : '')).join('\n\n');
+        return [{ role: 'system', content: merged }, ...nonSys];
+    }
+    if (sys.length === 1 && messages[0]?.role !== 'system') {
+        return [sys[0], ...nonSys];
+    }
+    return messages;
+}
+/**
+ * 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 function markTrailingAssistantPrefix(provider, messages) {
+    if (provider !== 'mistral' || messages.length === 0)
+        return messages;
+    const last = messages[messages.length - 1];
+    if (last?.role !== 'assistant' || 'prefix' in last)
+        return messages;
+    // A trailing assistant WITH tool_calls used to be skipped here, on the
+    // reasoning that it is a different fault — dangling calls with no results —
+    // which prefix would not rescue, so better to fail loudly.
+    //
+    // It failed loudly and rescued nothing. The operator hit it on 2026-08-19
+    // after a present_plan turn:
+    //
+    //   Provider mistral returned 400: Expected last role User or Tool
+    //   (or Assistant with prefix True) for serving but got assistant
+    //
+    // Mistral's own message names prefix:true as an accepted shape — it does not
+    // exclude assistant turns carrying tool_calls. So the skip turned a
+    // recoverable request into a hard error, and the user saw a raw provider 400
+    // at the end of a turn that had otherwise worked.
+    //
+    // Marked rather than skipped now. If the underlying state really is a
+    // dangling call, a continuation is still a better outcome than a 400 the user
+    // has to read — and it is logged so the state is visible rather than silent.
+    const trailingCall = Array.isArray(last.tool_calls) && last.tool_calls.length > 0;
+    if (trailingCall) {
+        logger.warn('[mistral] Trailing assistant carries tool_calls with no results — marking as a ' +
+            'prefix continuation. Provider rejects it outright otherwise.');
+    }
+    return [...messages.slice(0, -1), { ...last, prefix: true }];
+}
+/**
+ * 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 function shapeMessages(provider, messages) {
+    const stripped = stripReasoningContent(messages);
+    const ordered = provider === 'qwen' ? reorderSystemForQwen(stripped) : stripped;
+    return markTrailingAssistantPrefix(provider, ordered);
+}
+//# sourceMappingURL=messages.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"messages.js","sourceRoot":"","sources":["../../../src/providers/request-shaping/messages.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAE,MAAM,EAAE,MAAM,sBAAsB,CAAC;AAI9C;;;;GAIG;AACH,MAAM,UAAU,qBAAqB,CAAyB,QAA0B;IACtF,OAAO,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACxB,IAAI,mBAAmB,IAAI,CAAC,EAAE,CAAC;YAC7B,MAAM,EAAE,iBAAiB,EAAE,GAAG,EAAE,GAAG,IAAI,EAAE,GAAG,CAAC,CAAC;YAC9C,KAAK,GAAG,CAAC;YACT,OAAO,IAAS,CAAC;QACnB,CAAC;QACD,OAAO,CAAC,CAAC;IACX,CAAC,CAAC,CAAC;AACL,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,oBAAoB,CAAyB,QAA0B;IACrF,MAAM,GAAG,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IACxD,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IAC3D,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACnB,MAAM,MAAM,GAAG,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAC7F,OAAO,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAkB,EAAE,GAAG,MAAM,CAAC,CAAC;IAC1E,CAAC;IACD,IAAI,GAAG,CAAC,MAAM,KAAK,CAAC,IAAI,QAAQ,CAAC,CAAC,CAAC,EAAE,IAAI,KAAK,QAAQ,EAAE,CAAC;QACvD,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,CAAC,CAAC;IAC7B,CAAC;IACD,OAAO,QAAe,CAAC;AACzB,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,MAAM,UAAU,2BAA2B,CACzC,QAAgB,EAChB,QAA0B;IAE1B,IAAI,QAAQ,KAAK,SAAS,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,QAAe,CAAC;IAC5E,MAAM,IAAI,GAAG,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC3C,IAAI,IAAI,EAAE,IAAI,KAAK,WAAW,IAAI,QAAQ,IAAI,IAAI;QAAE,OAAO,QAAe,CAAC;IAC3E,uEAAuE;IACvE,4EAA4E;IAC5E,2DAA2D;IAC3D,EAAE;IACF,0EAA0E;IAC1E,6BAA6B;IAC7B,EAAE;IACF,mEAAmE;IACnE,kEAAkE;IAClE,EAAE;IACF,6EAA6E;IAC7E,oEAAoE;IACpE,6EAA6E;IAC7E,kDAAkD;IAClD,EAAE;IACF,sEAAsE;IACtE,8EAA8E;IAC9E,6EAA6E;IAC7E,MAAM,YAAY,GAAG,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,UAAU,CAAC,IAAI,IAAI,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;IAClF,IAAI,YAAY,EAAE,CAAC;QACjB,MAAM,CAAC,IAAI,CACT,iFAAiF;YACjF,8DAA8D,CAC/D,CAAC;IACJ,CAAC;IACD,OAAO,CAAC,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,GAAG,IAAI,EAAE,MAAM,EAAE,IAAI,EAAO,CAAC,CAAC;AACpE,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,aAAa,CAC3B,QAAgB,EAChB,QAA0B;IAE1B,MAAM,QAAQ,GAAG,qBAAqB,CAAC,QAAQ,CAAC,CAAC;IACjD,MAAM,OAAO,GAAG,QAAQ,KAAK,MAAM,CAAC,CAAC,CAAC,oBAAoB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC;IAChF,OAAO,2BAA2B,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;AACxD,CAAC"}

package/dist/providers/request-shaping/model-ids.d.ts

@@ -0,0 +1,44 @@
+/**
+ * Single source of truth for upstream model-id translation + vision rerouting.
+ *
+ * Both code paths that talk to upstream LLM providers must agree on the API
+ * model id: the PLATFORM path (server-side, packages/web) and the BYOK path
+ * (client-side, the core providers here). They used to keep separate copies of
+ * these maps, and the copies drifted — that's how Aurora shipped calling the
+ * non-existent `mistral-medium-3-5` (a 429 ghost) while the platform map said
+ * one thing and the BYOK map another. This module is the one map both import.
+ *
+ * Pure + transport-free on purpose: web calls it before its own fetch (with
+ * Ava's keys), core providers call it inside transformRequest (with the user's
+ * key). Neither imports the other's transport.
+ */
+/**
+ * Friendly internal id -> the exact id the upstream API accepts.
+ *
+ * Mistral uses date-stamped snapshot ids (`mistral-<tier>-YYMM`), not the
+ * friendly product names. Managed `-platform` entries strip the suffix to the
+ * canonical id the upstream API knows. Qwen legacy aliases roll forward to the
+ * current generation. Anything not listed passes through unchanged.
+ */
+export declare const MODEL_API_NAMES: Record<string, string>;
+/**
+ * When a request carries images, reroute text-only models to a vision-capable
+ * equivalent. Keys/values are POST-translation ids (run after MODEL_API_NAMES).
+ *
+ * qwen3.5-flash is text-only. DeepSeek V4 (Pro + Flash) is text-only at the API
+ * level (verified against api.deepseek.com 2026-04-25) so image requests reroute
+ * to Qwen 3.7 Plus (native vision + video). Aurora coordinates on Medium 3.5,
+ * whose vision encoder was trained from scratch, so images stay in-fleet and
+ * the EU-stack guarantee holds instead of falling to Qwen. (Large 3 — text-only
+ * — was the coordinator before the fleet was re-tiered; it's the heavy reserve
+ * now.)
+ */
+export declare const VISION_REROUTE: Record<string, string>;
+/**
+ * Translate a friendly model id to the upstream API id, applying the vision
+ * reroute when images are present. Unknown ids pass through unchanged.
+ */
+export declare function resolveApiModel(model: string, hasImages?: boolean): string;
+/** True when any message carries an image_url content part. */
+export declare function messagesHaveImages(messages: ReadonlyArray<Record<string, unknown>>): boolean;
+//# sourceMappingURL=model-ids.d.ts.map

package/dist/providers/request-shaping/model-ids.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-ids.d.ts","sourceRoot":"","sources":["../../../src/providers/request-shaping/model-ids.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH;;;;;;;GAOG;AACH,eAAO,MAAM,eAAe,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAoClD,CAAC;AAEF;;;;;;;;;;;GAWG;AACH,eAAO,MAAM,cAAc,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAOjD,CAAC;AAEF;;;GAGG;AACH,wBAAgB,eAAe,CAAC,KAAK,EAAE,MAAM,EAAE,SAAS,UAAQ,GAAG,MAAM,CAMxE;AAED,+DAA+D;AAC/D,wBAAgB,kBAAkB,CAAC,QAAQ,EAAE,aAAa,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,GAAG,OAAO,CAM5F"}

package/dist/providers/request-shaping/model-ids.js

@@ -0,0 +1,96 @@
+/**
+ * Single source of truth for upstream model-id translation + vision rerouting.
+ *
+ * Both code paths that talk to upstream LLM providers must agree on the API
+ * model id: the PLATFORM path (server-side, packages/web) and the BYOK path
+ * (client-side, the core providers here). They used to keep separate copies of
+ * these maps, and the copies drifted — that's how Aurora shipped calling the
+ * non-existent `mistral-medium-3-5` (a 429 ghost) while the platform map said
+ * one thing and the BYOK map another. This module is the one map both import.
+ *
+ * Pure + transport-free on purpose: web calls it before its own fetch (with
+ * Ava's keys), core providers call it inside transformRequest (with the user's
+ * key). Neither imports the other's transport.
+ */
+/**
+ * Friendly internal id -> the exact id the upstream API accepts.
+ *
+ * Mistral uses date-stamped snapshot ids (`mistral-<tier>-YYMM`), not the
+ * friendly product names. Managed `-platform` entries strip the suffix to the
+ * canonical id the upstream API knows. Qwen legacy aliases roll forward to the
+ * current generation. Anything not listed passes through unchanged.
+ */
+export const MODEL_API_NAMES = {
+    // Qwen legacy rolling aliases -> current generation
+    'qwen-flash': 'qwen3.5-flash',
+    'qwen-turbo': 'qwen3.5-flash',
+    // 3.7 Max retired 2026-08-09 in favour of 3.8 Max, which is better on every
+    // axis AND cheaper ($2/$6 against $2.50/$7.50, 131K output against 65K) and
+    // therefore costs FEWER credits (2.58x against 3.22x). Rolled forward rather
+    // than removed: a saved selection keeps working, and the person holding it
+    // gets a better model for less money without being asked to do anything.
+    'qwen3.7-max': 'qwen3.8-max',
+    // Managed DeepSeek V4 — strip the `-platform` disambiguator
+    'deepseek-v4-pro-platform': 'deepseek-v4-pro',
+    'deepseek-v4-flash-platform': 'deepseek-v4-flash',
+    // Managed Kimi K3 (Longxiang's lead seat) — strip the disambiguator.
+    // Moonshot only knows `kimi-k3`; sending the suffixed id returns
+    // 404 "Not found the model kimi-k3-platform or Permission denied".
+    'kimi-k3-platform': 'kimi-k3',
+    // Same for K2.7 Code. Migration 373 created the `kimi-k2.7-code-platform`
+    // row so it could be picked as a credit single, but no strip was added
+    // alongside it — so every surface reading that row (extension, IDE, hub)
+    // would have sent the suffixed id to Moonshot and taken the same 404 the
+    // line above exists to prevent.
+    'kimi-k2.7-code-platform': 'kimi-k2.7-code',
+    // Managed Mistral (Aurora's fleet) — friendly name -> the id Mistral accepts.
+    // Large 3 and Small 4 use date-stamped snapshots. Medium 3.5 is pinned to the
+    // SEMANTIC id `mistral-medium-3-5` (what Mistral's own rate-limit console
+    // lists). Mistral ALSO accepts the dated snapshot `mistral-medium-2604` (both
+    // return 200); we keep the semantic id. The dotted `mistral-medium-3.5` is rejected.
+    // NB: the 429s during the Aurora launch were a free-tier rate limit, NOT the
+    // id — fixed by upgrading the Mistral account tier, not by changing this value.
+    'mistral-small-4-platform': 'mistral-small-2603',
+    'mistral-large-3-platform': 'mistral-large-2512',
+    'mistral-medium-3.5-platform': 'mistral-medium-3-5',
+    'mistral-small-4': 'mistral-small-2603',
+    'mistral-large-3': 'mistral-large-2512',
+    'mistral-medium-3.5': 'mistral-medium-3-5',
+};
+/**
+ * When a request carries images, reroute text-only models to a vision-capable
+ * equivalent. Keys/values are POST-translation ids (run after MODEL_API_NAMES).
+ *
+ * qwen3.5-flash is text-only. DeepSeek V4 (Pro + Flash) is text-only at the API
+ * level (verified against api.deepseek.com 2026-04-25) so image requests reroute
+ * to Qwen 3.7 Plus (native vision + video). Aurora coordinates on Medium 3.5,
+ * whose vision encoder was trained from scratch, so images stay in-fleet and
+ * the EU-stack guarantee holds instead of falling to Qwen. (Large 3 — text-only
+ * — was the coordinator before the fleet was re-tiered; it's the heavy reserve
+ * now.)
+ */
+export const VISION_REROUTE = {
+    'qwen3.5-flash': 'qwen3.7-plus',
+    'qwen-flash': 'qwen3.7-plus',
+    'qwen-turbo': 'qwen3.7-plus',
+    'deepseek-v4-pro': 'qwen3.7-plus',
+    'deepseek-v4-flash': 'qwen3.7-plus',
+    'mistral-large-2512': 'mistral-medium-3-5',
+};
+/**
+ * Translate a friendly model id to the upstream API id, applying the vision
+ * reroute when images are present. Unknown ids pass through unchanged.
+ */
+export function resolveApiModel(model, hasImages = false) {
+    const mapped = MODEL_API_NAMES[model] || model;
+    if (hasImages && VISION_REROUTE[mapped]) {
+        return VISION_REROUTE[mapped];
+    }
+    return mapped;
+}
+/** True when any message carries an image_url content part. */
+export function messagesHaveImages(messages) {
+    return messages.some((m) => Array.isArray(m.content) &&
+        m.content.some((p) => p.type === 'image_url'));
+}
+//# sourceMappingURL=model-ids.js.map

package/dist/providers/request-shaping/model-ids.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-ids.js","sourceRoot":"","sources":["../../../src/providers/request-shaping/model-ids.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH;;;;;;;GAOG;AACH,MAAM,CAAC,MAAM,eAAe,GAA2B;IACrD,oDAAoD;IACpD,YAAY,EAAS,eAAe;IACpC,YAAY,EAAS,eAAe;IACpC,4EAA4E;IAC5E,4EAA4E;IAC5E,6EAA6E;IAC7E,2EAA2E;IAC3E,yEAAyE;IACzE,aAAa,EAAQ,aAAa;IAClC,4DAA4D;IAC5D,0BAA0B,EAAI,iBAAiB;IAC/C,4BAA4B,EAAE,mBAAmB;IACjD,qEAAqE;IACrE,iEAAiE;IACjE,mEAAmE;IACnE,kBAAkB,EAAY,SAAS;IACvC,0EAA0E;IAC1E,uEAAuE;IACvE,yEAAyE;IACzE,yEAAyE;IACzE,gCAAgC;IAChC,yBAAyB,EAAK,gBAAgB;IAC9C,8EAA8E;IAC9E,8EAA8E;IAC9E,0EAA0E;IAC1E,8EAA8E;IAC9E,qFAAqF;IACrF,6EAA6E;IAC7E,gFAAgF;IAChF,0BAA0B,EAAI,oBAAoB;IAClD,0BAA0B,EAAI,oBAAoB;IAClD,6BAA6B,EAAC,oBAAoB;IAClD,iBAAiB,EAAa,oBAAoB;IAClD,iBAAiB,EAAa,oBAAoB;IAClD,oBAAoB,EAAU,oBAAoB;CACnD,CAAC;AAEF;;;;;;;;;;;GAWG;AACH,MAAM,CAAC,MAAM,cAAc,GAA2B;IACpD,eAAe,EAAQ,cAAc;IACrC,YAAY,EAAW,cAAc;IACrC,YAAY,EAAW,cAAc;IACrC,iBAAiB,EAAM,cAAc;IACrC,mBAAmB,EAAI,cAAc;IACrC,oBAAoB,EAAG,oBAAoB;CAC5C,CAAC;AAEF;;;GAGG;AACH,MAAM,UAAU,eAAe,CAAC,KAAa,EAAE,SAAS,GAAG,KAAK;IAC9D,MAAM,MAAM,GAAG,eAAe,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC;IAC/C,IAAI,SAAS,IAAI,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC;QACxC,OAAO,cAAc,CAAC,MAAM,CAAC,CAAC;IAChC,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,+DAA+D;AAC/D,MAAM,UAAU,kBAAkB,CAAC,QAAgD;IACjF,OAAO,QAAQ,CAAC,IAAI,CAClB,CAAC,CAAC,EAAE,EAAE,CACJ,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC;QACvB,CAAC,CAAC,OAAoC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,WAAW,CAAC,CAC9E,CAAC;AACJ,CAAC"}

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

@@ -0,0 +1,38 @@
+/**
+ * Shared per-provider parameter quirks for OpenAI-compatible chat requests.
+ *
+ * The platform route assembles these inline; the BYOK providers each re-derive
+ * a subset. Centralised here so both produce an identical request body and a
+ * gap like "core MiniMax never remaps max_tokens" can't survive.
+ */
+export interface ShapeableParams {
+    temperature?: number;
+    max_tokens?: number;
+    top_p?: number;
+    stop?: unknown;
+    frequency_penalty?: number;
+    presence_penalty?: number;
+    tools?: unknown;
+    tool_choice?: unknown;
+    stream?: boolean;
+    stream_options?: Record<string, unknown>;
+    /** Explicit opt-out of a hybrid model's reasoning pass. Undefined leaves the
+     *  provider default alone; false suppresses it. See ChatCompletionRequest. */
+    enable_thinking?: boolean;
+    /** NIM/vLLM chat-template arguments. The only lever that stops a Nemotron
+     *  reasoning — enable_thinking does nothing on that stack. See
+     *  ChatCompletionRequest for what it costs when it is missing. */
+    chat_template_kwargs?: Record<string, unknown>;
+}
+export declare function isZhipuFlashModel(provider: string, model: string): boolean;
+export declare function leaksReasoningIntoContent(provider: string, model: string): boolean;
+/**
+ * Produce the request-body fragment (everything except `model` and `messages`)
+ * with each provider's quirks applied:
+ *  - MiniMax wants `max_completion_tokens`, not `max_tokens`.
+ *  - Qwen (DashScope) rejects `frequency_penalty`.
+ *  - Zhipu Flash: force `enable_thinking: false`.
+ *  - Streaming requests opt into usage accounting via `stream_options`.
+ */
+export declare function shapeParams(provider: string, model: string, p: ShapeableParams): Record<string, unknown>;
+//# sourceMappingURL=params.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"params.d.ts","sourceRoot":"","sources":["../../../src/providers/request-shaping/params.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,MAAM,WAAW,eAAe;IAC9B,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,IAAI,CAAC,EAAE,OAAO,CAAC;IACf,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACzC;kFAC8E;IAC9E,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B;;sEAEkE;IAClE,oBAAoB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CAChD;AAOD,wBAAgB,iBAAiB,CAAC,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAG1E;AAoBD,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAElF;AAED;;;;;;;GAOG;AACH,wBAAgB,WAAW,CACzB,QAAQ,EAAE,MAAM,EAChB,KAAK,EAAE,MAAM,EACb,CAAC,EAAE,eAAe,GACjB,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAgCzB"}

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

@@ -0,0 +1,79 @@
+/**
+ * Shared per-provider parameter quirks for OpenAI-compatible chat requests.
+ *
+ * The platform route assembles these inline; the BYOK providers each re-derive
+ * a subset. Centralised here so both produce an identical request body and a
+ * gap like "core MiniMax never remaps max_tokens" can't survive.
+ */
+/** Zhipu fast models that ship with thinking ON by default but where we want
+ *  it OFF (30–60s latency hit otherwise). Reconciles two previously-divergent
+ *  checks: the platform route used a `flash` substring; core used an explicit
+ *  set `{glm-4.5-air}` (no "flash" in the name). The union preserves both. */
+const ZHIPU_FAST_MODELS = new Set(['glm-4.5-air']);
+export function isZhipuFlashModel(provider, model) {
+    if (provider !== 'zhipu')
+        return false;
+    return ZHIPU_FAST_MODELS.has(model) || model.includes('flash') || model.includes('Flash');
+}
+/** Models that write their reasoning into the CONTENT rather than into
+ *  reasoning_content, which makes them unusable at their own defaults.
+ *
+ *  Measured 2026-08-16 on "In one sentence, what is a race condition?":
+ *  Nemotron 3.5 Lightning spent 300 tokens and 14.3 SECONDS emitting
+ *  "Here's a thinking process: 1. **Analyze User Input:**" and never reached
+ *  an answer. With chat_template_kwargs {thinking:false}: 54 tokens, 0.67s,
+ *  correct. There is no stripping our side because it is not in a separate
+ *  field — it IS the answer as far as any consumer can tell.
+ *
+ *  Nemotron 3 Ultra is deliberately NOT here. It keeps its reasoning in
+ *  reasoning_content and answers cleanly at its default, so forcing it off
+ *  would remove capability to fix a problem it does not have.
+ *
+ *  An explicit set, not a name pattern: "lightning" and "nemotron" would both
+ *  over-match the next model NVIDIA ships, and the failure mode of guessing
+ *  wrong here is silently degrading a model that was fine. */
+const CONTENT_LEAKING_REASONERS = new Set(['nvidia/nemotron-3.5-lightning-30b-a3b']);
+export function leaksReasoningIntoContent(provider, model) {
+    return provider === 'nvidia' && CONTENT_LEAKING_REASONERS.has(model);
+}
+/**
+ * Produce the request-body fragment (everything except `model` and `messages`)
+ * with each provider's quirks applied:
+ *  - MiniMax wants `max_completion_tokens`, not `max_tokens`.
+ *  - Qwen (DashScope) rejects `frequency_penalty`.
+ *  - Zhipu Flash: force `enable_thinking: false`.
+ *  - Streaming requests opt into usage accounting via `stream_options`.
+ */
+export function shapeParams(provider, model, p) {
+    const isMiniMax = provider === 'minimax';
+    const isQwen = provider === 'qwen';
+    const isZhipuFlash = isZhipuFlashModel(provider, model);
+    const leaksReasoning = leaksReasoningIntoContent(provider, model);
+    return {
+        stream: p.stream ?? false,
+        ...(p.temperature !== undefined && { temperature: p.temperature }),
+        ...(p.max_tokens !== undefined &&
+            (isMiniMax ? { max_completion_tokens: p.max_tokens } : { max_tokens: p.max_tokens })),
+        ...(p.top_p !== undefined && { top_p: p.top_p }),
+        ...(p.stop !== undefined && { stop: p.stop }),
+        ...(!isQwen && p.frequency_penalty !== undefined && { frequency_penalty: p.frequency_penalty }),
+        ...(p.presence_penalty !== undefined && { presence_penalty: p.presence_penalty }),
+        ...(p.tools !== undefined && { tools: p.tools }),
+        ...(p.tool_choice !== undefined && { tool_choice: p.tool_choice }),
+        ...(p.stream && { stream_options: { include_usage: true, ...(p.stream_options || {}) } }),
+        // Zhipu Flash is forced off unconditionally (see above). Any other caller
+        // can opt out explicitly — classifiers and gates want the answer, not the
+        // reasoning, and the reasoning pass is not bounded by max_tokens.
+        ...(isZhipuFlash
+            ? { enable_thinking: false }
+            : p.enable_thinking !== undefined && { enable_thinking: p.enable_thinking }),
+        // An explicit value always wins; otherwise a content-leaking reasoner gets
+        // thinking turned off by default, because its default is not usable. This
+        // is applied for EVERY caller, not just the intent gate — the model ships
+        // to BYOK users and a plain question is exactly where it fails worst.
+        ...(p.chat_template_kwargs !== undefined
+            ? { chat_template_kwargs: p.chat_template_kwargs }
+            : leaksReasoning ? { chat_template_kwargs: { thinking: false } } : {}),
+    };
+}
+//# sourceMappingURL=params.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"params.js","sourceRoot":"","sources":["../../../src/providers/request-shaping/params.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAsBH;;;8EAG8E;AAC9E,MAAM,iBAAiB,GAAG,IAAI,GAAG,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC;AACnD,MAAM,UAAU,iBAAiB,CAAC,QAAgB,EAAE,KAAa;IAC/D,IAAI,QAAQ,KAAK,OAAO;QAAE,OAAO,KAAK,CAAC;IACvC,OAAO,iBAAiB,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,KAAK,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;AAC5F,CAAC;AAED;;;;;;;;;;;;;;;;8DAgB8D;AAC9D,MAAM,yBAAyB,GAAG,IAAI,GAAG,CAAC,CAAC,uCAAuC,CAAC,CAAC,CAAC;AACrF,MAAM,UAAU,yBAAyB,CAAC,QAAgB,EAAE,KAAa;IACvE,OAAO,QAAQ,KAAK,QAAQ,IAAI,yBAAyB,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;AACvE,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,WAAW,CACzB,QAAgB,EAChB,KAAa,EACb,CAAkB;IAElB,MAAM,SAAS,GAAG,QAAQ,KAAK,SAAS,CAAC;IACzC,MAAM,MAAM,GAAG,QAAQ,KAAK,MAAM,CAAC;IACnC,MAAM,YAAY,GAAG,iBAAiB,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;IACxD,MAAM,cAAc,GAAG,yBAAyB,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;IAElE,OAAO;QACL,MAAM,EAAE,CAAC,CAAC,MAAM,IAAI,KAAK;QACzB,GAAG,CAAC,CAAC,CAAC,WAAW,KAAK,SAAS,IAAI,EAAE,WAAW,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;QAClE,GAAG,CAAC,CAAC,CAAC,UAAU,KAAK,SAAS;YAC5B,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,qBAAqB,EAAE,CAAC,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,CAAC,CAAC,UAAU,EAAE,CAAC,CAAC;QACvF,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,SAAS,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC;QAChD,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,SAAS,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;QAC7C,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,iBAAiB,KAAK,SAAS,IAAI,EAAE,iBAAiB,EAAE,CAAC,CAAC,iBAAiB,EAAE,CAAC;QAC/F,GAAG,CAAC,CAAC,CAAC,gBAAgB,KAAK,SAAS,IAAI,EAAE,gBAAgB,EAAE,CAAC,CAAC,gBAAgB,EAAE,CAAC;QACjF,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,SAAS,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC;QAChD,GAAG,CAAC,CAAC,CAAC,WAAW,KAAK,SAAS,IAAI,EAAE,WAAW,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;QAClE,GAAG,CAAC,CAAC,CAAC,MAAM,IAAI,EAAE,cAAc,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,GAAG,CAAC,CAAC,CAAC,cAAc,IAAI,EAAE,CAAC,EAAE,EAAE,CAAC;QACzF,0EAA0E;QAC1E,0EAA0E;QAC1E,kEAAkE;QAClE,GAAG,CAAC,YAAY;YACd,CAAC,CAAC,EAAE,eAAe,EAAE,KAAK,EAAE;YAC5B,CAAC,CAAC,CAAC,CAAC,eAAe,KAAK,SAAS,IAAI,EAAE,eAAe,EAAE,CAAC,CAAC,eAAe,EAAE,CAAC;QAC9E,2EAA2E;QAC3E,0EAA0E;QAC1E,0EAA0E;QAC1E,sEAAsE;QACtE,GAAG,CAAC,CAAC,CAAC,oBAAoB,KAAK,SAAS;YACtC,CAAC,CAAC,EAAE,oBAAoB,EAAE,CAAC,CAAC,oBAAoB,EAAE;YAClD,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,oBAAoB,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACzE,CAAC;AACJ,CAAC"}

package/dist/providers/resilient-provider.d.ts

@@ -0,0 +1,44 @@
+/**
+ * Resilient provider wrapper — automatic failover across multiple providers.
+ *
+ * Implements the Provider interface and wraps multiple {provider, model} pairs.
+ * On transient failures (429, 500, 502, 503, network errors) it automatically
+ * falls back to the next healthy provider in the chain. Non-retryable errors
+ * (401, 403, 404) are thrown immediately as they indicate configuration problems.
+ *
+ * The Agent sees this as a single Provider — failover is transparent.
+ */
+import type { ChatCompletionRequest, Provider } from './types.js';
+import type { ModelDefinition, CompletionResponse, StreamChunk } from '../core/types.js';
+import type { ProviderHealthTracker } from './health-tracker.js';
+export interface FallbackEntry {
+    provider: Provider;
+    model: ModelDefinition;
+}
+export interface ResilientProviderOptions {
+    primary: FallbackEntry;
+    fallbacks: FallbackEntry[];
+    healthTracker: ProviderHealthTracker;
+    onFallback?: (from: FallbackEntry, to: FallbackEntry, error: Error) => void;
+}
+export declare class ResilientProvider implements Provider {
+    private readonly primary;
+    private readonly fallbacks;
+    private readonly healthTracker;
+    private readonly onFallback?;
+    private _activeModel;
+    constructor(opts: ResilientProviderOptions);
+    get name(): string;
+    get displayName(): string;
+    /** The model currently in use (may differ from primary during fallback). */
+    get activeModel(): ModelDefinition;
+    listModels(): ModelDefinition[];
+    createCompletion(request: ChatCompletionRequest, signal?: AbortSignal): Promise<CompletionResponse>;
+    createStreamingCompletion(request: ChatCompletionRequest, signal?: AbortSignal): AsyncGenerator<StreamChunk, void, unknown>;
+    /**
+     * Build the ordered chain of entries to try.
+     * Healthy entries first, but always include at least the primary.
+     */
+    private buildChain;
+}
+//# sourceMappingURL=resilient-provider.d.ts.map

package/dist/providers/resilient-provider.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"resilient-provider.d.ts","sourceRoot":"","sources":["../../src/providers/resilient-provider.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,KAAK,EAAE,qBAAqB,EAAE,QAAQ,EAAE,MAAM,YAAY,CAAC;AAClE,OAAO,KAAK,EAAE,eAAe,EAAE,kBAAkB,EAAE,WAAW,EAAE,MAAM,kBAAkB,CAAC;AAEzF,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,qBAAqB,CAAC;AAIjE,MAAM,WAAW,aAAa;IAC5B,QAAQ,EAAE,QAAQ,CAAC;IACnB,KAAK,EAAE,eAAe,CAAC;CACxB;AAED,MAAM,WAAW,wBAAwB;IACvC,OAAO,EAAE,aAAa,CAAC;IACvB,SAAS,EAAE,aAAa,EAAE,CAAC;IAC3B,aAAa,EAAE,qBAAqB,CAAC;IACrC,UAAU,CAAC,EAAE,CAAC,IAAI,EAAE,aAAa,EAAE,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,KAAK,KAAK,IAAI,CAAC;CAC7E;AAID,qBAAa,iBAAkB,YAAW,QAAQ;IAChD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAgB;IACxC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAkB;IAC5C,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAwB;IACtD,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAyC;IACrE,OAAO,CAAC,YAAY,CAAkB;gBAE1B,IAAI,EAAE,wBAAwB;IAQ1C,IAAI,IAAI,IAAI,MAAM,CAEjB;IAED,IAAI,WAAW,IAAI,MAAM,CAExB;IAED,4EAA4E;IAC5E,IAAI,WAAW,IAAI,eAAe,CAEjC;IAED,UAAU,IAAI,eAAe,EAAE;IAMzB,gBAAgB,CACpB,OAAO,EAAE,qBAAqB,EAC9B,MAAM,CAAC,EAAE,WAAW,GACnB,OAAO,CAAC,kBAAkB,CAAC;IAsCvB,yBAAyB,CAC9B,OAAO,EAAE,qBAAqB,EAC9B,MAAM,CAAC,EAAE,WAAW,GACnB,cAAc,CAAC,WAAW,EAAE,IAAI,EAAE,OAAO,CAAC;IA0E7C;;;OAGG;IACH,OAAO,CAAC,UAAU;CAQnB"}

package/dist/providers/resilient-provider.js

@@ -0,0 +1,141 @@
+/**
+ * Resilient provider wrapper — automatic failover across multiple providers.
+ *
+ * Implements the Provider interface and wraps multiple {provider, model} pairs.
+ * On transient failures (429, 500, 502, 503, network errors) it automatically
+ * falls back to the next healthy provider in the chain. Non-retryable errors
+ * (401, 403, 404) are thrown immediately as they indicate configuration problems.
+ *
+ * The Agent sees this as a single Provider — failover is transparent.
+ */
+import { ProviderError } from '../core/errors.js';
+// ── Implementation ───────────────────────────────────────────────────────────
+export class ResilientProvider {
+    primary;
+    fallbacks;
+    healthTracker;
+    onFallback;
+    _activeModel;
+    constructor(opts) {
+        this.primary = opts.primary;
+        this.fallbacks = opts.fallbacks;
+        this.healthTracker = opts.healthTracker;
+        this.onFallback = opts.onFallback;
+        this._activeModel = opts.primary.model;
+    }
+    get name() {
+        return this.primary.provider.name;
+    }
+    get displayName() {
+        return this.primary.provider.displayName;
+    }
+    /** The model currently in use (may differ from primary during fallback). */
+    get activeModel() {
+        return this._activeModel;
+    }
+    listModels() {
+        return this.primary.provider.listModels();
+    }
+    // ── Completion with failover ─────────────────────────────────────────────
+    async createCompletion(request, signal) {
+        const chain = this.buildChain();
+        let lastError;
+        for (let i = 0; i < chain.length; i++) {
+            const entry = chain[i];
+            const rewrittenRequest = { ...request, model: entry.model.id };
+            const start = Date.now();
+            try {
+                const result = await entry.provider.createCompletion(rewrittenRequest, signal);
+                this.healthTracker.recordSuccess(entry.provider.name, Date.now() - start);
+                this._activeModel = entry.model;
+                return result;
+            }
+            catch (err) {
+                const error = err;
+                this.healthTracker.recordFailure(entry.provider.name, error);
+                // Non-retryable errors: throw immediately (config problem)
+                if (err instanceof ProviderError && !err.retryable) {
+                    throw err;
+                }
+                lastError = error;
+                // Notify about failover (if there's a next entry to try)
+                if (i < chain.length - 1 && this.onFallback) {
+                    this.onFallback(entry, chain[i + 1], error);
+                }
+            }
+        }
+        // All entries exhausted
+        throw lastError ?? new ProviderError('All providers failed', this.name);
+    }
+    // ── Streaming with failover ──────────────────────────────────────────────
+    async *createStreamingCompletion(request, signal) {
+        const chain = this.buildChain();
+        let lastError;
+        for (let i = 0; i < chain.length; i++) {
+            const entry = chain[i];
+            const rewrittenRequest = { ...request, model: entry.model.id };
+            const start = Date.now();
+            try {
+                const gen = entry.provider.createStreamingCompletion(rewrittenRequest, signal);
+                // Try to get the first chunk — this is where connection failures surface
+                const first = await gen.next();
+                this.healthTracker.recordSuccess(entry.provider.name, Date.now() - start);
+                this._activeModel = entry.model;
+                if (!first.done) {
+                    yield first.value;
+                    // Wrap mid-stream yield to catch failures after connection established
+                    try {
+                        yield* gen;
+                    }
+                    catch (midStreamErr) {
+                        // Mid-stream failure — record and attempt fallback with remaining chain
+                        const msError = midStreamErr;
+                        this.healthTracker.recordFailure(entry.provider.name, msError);
+                        if (midStreamErr instanceof ProviderError && !midStreamErr.retryable) {
+                            throw midStreamErr;
+                        }
+                        lastError = msError;
+                        // Signal the mid-stream failure so the caller can decide
+                        // We can't restart the stream (already yielded data), so throw
+                        // with a clear message that partial data was received
+                        if (i >= chain.length - 1) {
+                            throw new ProviderError(`Stream interrupted after partial response: ${msError.message}`, entry.provider.name, 502);
+                        }
+                        // Notify about failover
+                        if (this.onFallback) {
+                            this.onFallback(entry, chain[i + 1], msError);
+                        }
+                        continue; // Try next provider (caller gets a fresh stream)
+                    }
+                }
+                return;
+            }
+            catch (err) {
+                const error = err;
+                this.healthTracker.recordFailure(entry.provider.name, error);
+                if (err instanceof ProviderError && !err.retryable) {
+                    throw err;
+                }
+                lastError = error;
+                if (i < chain.length - 1 && this.onFallback) {
+                    this.onFallback(entry, chain[i + 1], error);
+                }
+            }
+        }
+        throw lastError ?? new ProviderError('All providers failed', this.name);
+    }
+    // ── Private ──────────────────────────────────────────────────────────────
+    /**
+     * Build the ordered chain of entries to try.
+     * Healthy entries first, but always include at least the primary.
+     */
+    buildChain() {
+        const all = [this.primary, ...this.fallbacks];
+        const healthy = all.filter(e => this.healthTracker.isHealthy(e.provider.name));
+        // If no healthy entries (all tripped), still try the primary
+        if (healthy.length === 0)
+            return [this.primary];
+        return healthy;
+    }
+}
+//# sourceMappingURL=resilient-provider.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"resilient-provider.js","sourceRoot":"","sources":["../../src/providers/resilient-provider.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAIH,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAiBlD,gFAAgF;AAEhF,MAAM,OAAO,iBAAiB;IACX,OAAO,CAAgB;IACvB,SAAS,CAAkB;IAC3B,aAAa,CAAwB;IACrC,UAAU,CAA0C;IAC7D,YAAY,CAAkB;IAEtC,YAAY,IAA8B;QACxC,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC;QAC5B,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC;QAChC,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC,aAAa,CAAC;QACxC,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,UAAU,CAAC;QAClC,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC;IACzC,CAAC;IAED,IAAI,IAAI;QACN,OAAO,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAC;IACpC,CAAC;IAED,IAAI,WAAW;QACb,OAAO,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAC;IAC3C,CAAC;IAED,4EAA4E;IAC5E,IAAI,WAAW;QACb,OAAO,IAAI,CAAC,YAAY,CAAC;IAC3B,CAAC;IAED,UAAU;QACR,OAAO,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,UAAU,EAAE,CAAC;IAC5C,CAAC;IAED,4EAA4E;IAE5E,KAAK,CAAC,gBAAgB,CACpB,OAA8B,EAC9B,MAAoB;QAEpB,MAAM,KAAK,GAAG,IAAI,CAAC,UAAU,EAAE,CAAC;QAChC,IAAI,SAA4B,CAAC;QAEjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACtC,MAAM,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC;YACvB,MAAM,gBAAgB,GAAG,EAAE,GAAG,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,EAAE,CAAC;YAC/D,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;YAEzB,IAAI,CAAC;gBACH,MAAM,MAAM,GAAG,MAAM,KAAK,CAAC,QAAQ,CAAC,gBAAgB,CAAC,gBAAgB,EAAE,MAAM,CAAC,CAAC;gBAC/E,IAAI,CAAC,aAAa,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,IAAI,CAAC,GAAG,EAAE,GAAG,KAAK,CAAC,CAAC;gBAC1E,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC,KAAK,CAAC;gBAChC,OAAO,MAAM,CAAC;YAChB,CAAC;YAAC,OAAO,GAAG,EAAE,CAAC;gBACb,MAAM,KAAK,GAAG,GAAY,CAAC;gBAC3B,IAAI,CAAC,aAAa,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;gBAE7D,2DAA2D;gBAC3D,IAAI,GAAG,YAAY,aAAa,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,CAAC;oBACnD,MAAM,GAAG,CAAC;gBACZ,CAAC;gBAED,SAAS,GAAG,KAAK,CAAC;gBAElB,yDAAyD;gBACzD,IAAI,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;oBAC5C,IAAI,CAAC,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC;gBAC9C,CAAC;YACH,CAAC;QACH,CAAC;QAED,wBAAwB;QACxB,MAAM,SAAS,IAAI,IAAI,aAAa,CAAC,sBAAsB,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;IAC1E,CAAC;IAED,4EAA4E;IAE5E,KAAK,CAAC,CAAC,yBAAyB,CAC9B,OAA8B,EAC9B,MAAoB;QAEpB,MAAM,KAAK,GAAG,IAAI,CAAC,UAAU,EAAE,CAAC;QAChC,IAAI,SAA4B,CAAC;QAEjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACtC,MAAM,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC;YACvB,MAAM,gBAAgB,GAAG,EAAE,GAAG,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,EAAE,CAAC;YAC/D,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;YAEzB,IAAI,CAAC;gBACH,MAAM,GAAG,GAAG,KAAK,CAAC,QAAQ,CAAC,yBAAyB,CAAC,gBAAgB,EAAE,MAAM,CAAC,CAAC;gBAE/E,yEAAyE;gBACzE,MAAM,KAAK,GAAG,MAAM,GAAG,CAAC,IAAI,EAAE,CAAC;gBAC/B,IAAI,CAAC,aAAa,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,IAAI,CAAC,GAAG,EAAE,GAAG,KAAK,CAAC,CAAC;gBAC1E,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC,KAAK,CAAC;gBAEhC,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;oBAChB,MAAM,KAAK,CAAC,KAAK,CAAC;oBAElB,uEAAuE;oBACvE,IAAI,CAAC;wBACH,KAAK,CAAC,CAAC,GAAG,CAAC;oBACb,CAAC;oBAAC,OAAO,YAAY,EAAE,CAAC;wBACtB,wEAAwE;wBACxE,MAAM,OAAO,GAAG,YAAqB,CAAC;wBACtC,IAAI,CAAC,aAAa,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,OAAO,CAAC,CAAC;wBAE/D,IAAI,YAAY,YAAY,aAAa,IAAI,CAAC,YAAY,CAAC,SAAS,EAAE,CAAC;4BACrE,MAAM,YAAY,CAAC;wBACrB,CAAC;wBAED,SAAS,GAAG,OAAO,CAAC;wBAEpB,yDAAyD;wBACzD,+DAA+D;wBAC/D,sDAAsD;wBACtD,IAAI,CAAC,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;4BAC1B,MAAM,IAAI,aAAa,CACrB,8CAA8C,OAAO,CAAC,OAAO,EAAE,EAC/D,KAAK,CAAC,QAAQ,CAAC,IAAI,EACnB,GAAG,CACJ,CAAC;wBACJ,CAAC;wBAED,wBAAwB;wBACxB,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;4BACpB,IAAI,CAAC,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,CAAC,CAAC;wBAChD,CAAC;wBACD,SAAS,CAAC,iDAAiD;oBAC7D,CAAC;gBACH,CAAC;gBACD,OAAO;YACT,CAAC;YAAC,OAAO,GAAG,EAAE,CAAC;gBACb,MAAM,KAAK,GAAG,GAAY,CAAC;gBAC3B,IAAI,CAAC,aAAa,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;gBAE7D,IAAI,GAAG,YAAY,aAAa,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,CAAC;oBACnD,MAAM,GAAG,CAAC;gBACZ,CAAC;gBAED,SAAS,GAAG,KAAK,CAAC;gBAElB,IAAI,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;oBAC5C,IAAI,CAAC,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC;gBAC9C,CAAC;YACH,CAAC;QACH,CAAC;QAED,MAAM,SAAS,IAAI,IAAI,aAAa,CAAC,sBAAsB,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;IAC1E,CAAC;IAED,4EAA4E;IAE5E;;;OAGG;IACK,UAAU;QAChB,MAAM,GAAG,GAAG,CAAC,IAAI,CAAC,OAAO,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC;QAC9C,MAAM,OAAO,GAAG,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,IAAI,CAAC,aAAa,CAAC,SAAS,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC;QAE/E,6DAA6D;QAC7D,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;QAChD,OAAO,OAAO,CAAC;IACjB,CAAC;CACF"}

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

@@ -0,0 +1,19 @@
+import { BaseProvider } from '../base-provider.js';
+import type { CompletionResponse, ModelDefinition, StreamChunk } from '../../core/types.js';
+/**
+ * Tencent Hunyuan — BYOK provider (OpenAI-compatible).
+ *
+ * Talks to Tencent's TokenHub gateway (tokenhub-intl.tencentmaas.com — the
+ * international endpoint, which is the one we use everywhere), mirroring
+ * the OpenAI chat-completions shape including tool calling + reasoning_effort.
+ * Bring your own Tencent Hunyuan key.
+ */
+export declare class TencentProvider extends BaseProvider {
+    readonly name = "tencent";
+    readonly displayName = "Tencent Hunyuan";
+    protected getDefaultBaseUrl(): string;
+    listModels(): ModelDefinition[];
+    protected normalizeResponse(raw: unknown): CompletionResponse;
+    protected normalizeStreamChunk(raw: unknown): StreamChunk;
+}
+//# sourceMappingURL=index.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/providers/tencent/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AACnD,OAAO,KAAK,EAAE,kBAAkB,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,qBAAqB,CAAC;AAG5F;;;;;;;GAOG;AACH,qBAAa,eAAgB,SAAQ,YAAY;IAC/C,QAAQ,CAAC,IAAI,aAAa;IAC1B,QAAQ,CAAC,WAAW,qBAAqB;IAEzC,SAAS,CAAC,iBAAiB,IAAI,MAAM;IAIrC,UAAU,IAAI,eAAe,EAAE;IAO/B,SAAS,CAAC,iBAAiB,CAAC,GAAG,EAAE,OAAO,GAAG,kBAAkB;IAc7D,SAAS,CAAC,oBAAoB,CAAC,GAAG,EAAE,OAAO,GAAG,WAAW;CAa1D"}