mindweave 0.0.1 → 2.1.2

package/dist/drivers/glm/client.js

@@ -0,0 +1,90 @@
+import { compatStreamTurn, compatToolTurn } from "../openaiCompat/wire.js";
+import { BUFFERED_OUTPUT_TOKENS, DEFAULT_MODEL, takesEffort } from "./manifest.js";
+/**
+ * Z.ai's international OpenAI-compatible endpoint. The China platform
+ * (`open.bigmodel.cn`) serves the same model ids at different prices on a
+ * different key, so it is an override rather than a guess: a user on that account
+ * sets `MINDWEAVE_GLM_URL` and everything else works unchanged.
+ */
+const BASE_URL = process.env.MINDWEAVE_GLM_URL ?? "https://api.z.ai/api/paas/v4";
+/** The model a request runs on, matching the shared layer's own fallback so the
+ *  reasoning fields can never be built for a different model than is being called. */
+function modelOf(config) {
+    return config?.model ?? process.env.MINDWEAVE_MODEL ?? DEFAULT_MODEL;
+}
+/**
+ * GLM's reasoning fields.
+ *
+ * `thinking` is ALWAYS sent, both ways. The provider's default is `enabled`, so
+ * omitting the field does not mean "no thinking" — it means the model reasons and
+ * bills for it. Every internal call that never sets `req.model` (the session
+ * summarizer, the page distiller) would quietly pay for reasoning the UI discards.
+ * This is the third provider in this project with that exact default; it is a
+ * family pattern rather than a coincidence.
+ *
+ * `reasoning_effort` rides along only on the model that has the dial. Sending it to
+ * one that does not is a parameter the model has never heard of, not a soft no-op.
+ */
+export function reasoningFields(config) {
+    const thinking = config?.thinking === true;
+    const fields = { thinking: { type: thinking ? "enabled" : "disabled" } };
+    if (thinking && takesEffort(modelOf(config))) {
+        fields.reasoning_effort = config?.effort ?? "high";
+    }
+    return fields;
+}
+/**
+ * Map this provider's own `finish_reason` values onto the shared set.
+ *
+ * The widest extra vocabulary of any driver here, and two of the three matter a
+ * great deal: without a case, `sensitive` and `model_context_window_exceeded` both
+ * fall through to `"end"`, so a refused reply and an overflowed conversation are
+ * each reported to the engine as a clean, complete finish.
+ */
+export function extraStop(reason) {
+    switch (reason) {
+        case "sensitive":
+            return "refused";
+        case "model_context_window_exceeded":
+            return "overflow";
+        case "network_error":
+            return "overloaded";
+        default:
+            return undefined;
+    }
+}
+/**
+ * GLM reports its cache hit as `prompt_tokens_details.cached_tokens`, where
+ * `prompt_tokens` is the FULL prompt including the cached part — so the miss side
+ * is the remainder rather than a separate figure to add.
+ */
+export function cacheSplit(usage) {
+    const details = usage.prompt_tokens_details;
+    const hit = typeof details?.cached_tokens === "number" ? details.cached_tokens : 0;
+    if (hit === 0)
+        return undefined;
+    const prompt = typeof usage.prompt_tokens === "number" ? usage.prompt_tokens : 0;
+    return { hit, miss: Math.max(0, prompt - hit) };
+}
+/** Everything the shared wire layer needs to talk to GLM. */
+export const glmProvider = {
+    label: "GLM",
+    baseUrl: BASE_URL,
+    apiKeyEnv: "ZAI_API_KEY",
+    defaultModel: process.env.MINDWEAVE_MODEL ?? DEFAULT_MODEL,
+    reasoningFields,
+    extraStop,
+    cacheSplit,
+    // Must match the manifest's `bufferedOutputTokens`, which is what core reserves
+    // room for below the context window.
+    bufferedMaxTokens: BUFFERED_OUTPUT_TOKENS,
+};
+/** Ask the model for one turn. */
+export async function toolTurn(req, options = {}) {
+    return compatToolTurn(glmProvider, req, options.signal);
+}
+/** Ask the model for one turn, streaming deltas to `options.onEvent`. */
+export async function streamTurn(req, options = {}) {
+    return compatStreamTurn(glmProvider, req, options.onEvent, options.signal);
+}
+//# sourceMappingURL=client.js.map

package/dist/drivers/glm/client.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"client.js","sourceRoot":"","sources":["../../../src/drivers/glm/client.ts"],"names":[],"mappings":"AAuBA,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAuB,MAAM,yBAAyB,CAAC;AAChG,OAAO,EAAE,sBAAsB,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAEnF;;;;;GAKG;AACH,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,iBAAiB,IAAI,8BAA8B,CAAC;AAEjF;sFACsF;AACtF,SAAS,OAAO,CAAC,MAA+B;IAC9C,OAAO,MAAM,EAAE,KAAK,IAAI,OAAO,CAAC,GAAG,CAAC,eAAe,IAAI,aAAa,CAAC;AACvE,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,eAAe,CAAC,MAA+B;IAC7D,MAAM,QAAQ,GAAG,MAAM,EAAE,QAAQ,KAAK,IAAI,CAAC;IAC3C,MAAM,MAAM,GAA4B,EAAE,QAAQ,EAAE,EAAE,IAAI,EAAE,QAAQ,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC;IAClG,IAAI,QAAQ,IAAI,WAAW,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;QAC7C,MAAM,CAAC,gBAAgB,GAAG,MAAM,EAAE,MAAM,IAAI,MAAM,CAAC;IACrD,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,SAAS,CAAC,MAAc;IACtC,QAAQ,MAAM,EAAE,CAAC;QACf,KAAK,WAAW;YACd,OAAO,SAAS,CAAC;QACnB,KAAK,+BAA+B;YAClC,OAAO,UAAU,CAAC;QACpB,KAAK,eAAe;YAClB,OAAO,YAAY,CAAC;QACtB;YACE,OAAO,SAAS,CAAC;IACrB,CAAC;AACH,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,UAAU,CAAC,KAA8B;IACvD,MAAM,OAAO,GAAG,KAAK,CAAC,qBAA+D,CAAC;IACtF,MAAM,GAAG,GAAG,OAAO,OAAO,EAAE,aAAa,KAAK,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC;IACnF,IAAI,GAAG,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC;IAChC,MAAM,MAAM,GAAG,OAAO,KAAK,CAAC,aAAa,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC;IACjF,OAAO,EAAE,GAAG,EAAE,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,GAAG,GAAG,CAAC,EAAE,CAAC;AAClD,CAAC;AAED,6DAA6D;AAC7D,MAAM,CAAC,MAAM,WAAW,GAAmB;IACzC,KAAK,EAAE,KAAK;IACZ,OAAO,EAAE,QAAQ;IACjB,SAAS,EAAE,aAAa;IACxB,YAAY,EAAE,OAAO,CAAC,GAAG,CAAC,eAAe,IAAI,aAAa;IAC1D,eAAe;IACf,SAAS;IACT,UAAU;IACV,gFAAgF;IAChF,qCAAqC;IACrC,iBAAiB,EAAE,sBAAsB;CAC1C,CAAC;AAEF,kCAAkC;AAClC,MAAM,CAAC,KAAK,UAAU,QAAQ,CAAC,GAAiB,EAAE,UAAuB,EAAE;IACzE,OAAO,cAAc,CAAC,WAAW,EAAE,GAAG,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;AAC1D,CAAC;AAED,yEAAyE;AACzE,MAAM,CAAC,KAAK,UAAU,UAAU,CAAC,GAAiB,EAAE,UAAyB,EAAE;IAC7E,OAAO,gBAAgB,CAAC,WAAW,EAAE,GAAG,EAAE,OAAO,CAAC,OAAO,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;AAC7E,CAAC"}

package/dist/drivers/glm/index.js

@@ -0,0 +1,9 @@
+import { streamTurn, toolTurn } from "./client.js";
+import { glmManifest } from "./manifest.js";
+export const glmDriver = {
+    ...glmManifest,
+    toolTurn,
+    streamTurn,
+};
+export default glmDriver;
+//# sourceMappingURL=index.js.map

package/dist/drivers/glm/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/drivers/glm/index.ts"],"names":[],"mappings":"AAiBA,OAAO,EAAE,UAAU,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AACnD,OAAO,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAE5C,MAAM,CAAC,MAAM,SAAS,GAAW;IAC/B,GAAG,WAAW;IACd,QAAQ;IACR,UAAU;CACX,CAAC;AAEF,eAAe,SAAS,CAAC"}

package/dist/drivers/glm/manifest.js

@@ -0,0 +1,183 @@
+export const GLM_53 = "glm-5.3";
+export const GLM_53_FLASH = "glm-5.3-flash";
+export const GLM_52 = "glm-5.2";
+export const GLM_5 = "glm-5";
+export const GLM_47 = "glm-4.7";
+export const GLM_47_FLASHX = "glm-4.7-flashx";
+/** The model used when nothing is saved and no env override is set. */
+export const DEFAULT_MODEL = GLM_53;
+/**
+ * The models offered by `/model`. First entry is this provider's default, which is
+ * also where `/provider` lands when someone switches to GLM.
+ *
+ * The flagship leads here, unlike the other drivers, because on this provider it is
+ * not the expensive choice: GLM-5.2 runs well under half of Claude Sonnet's rate,
+ * and it is the only model in the lineup with a reasoning dial. The cheaper rungs
+ * are a real ladder below it rather than the sensible middle.
+ */
+export const MODELS = [
+    { id: GLM_53, label: "GLM-5.3", description: "the flagship, three reasoning depths — the default" },
+    { id: GLM_53_FLASH, label: "GLM-5.3 Flash", description: "most of the flagship, a fraction of the price" },
+    { id: GLM_52, label: "GLM-5.2", description: "the previous flagship, with a reasoning dial" },
+    { id: GLM_5, label: "GLM-5", description: "older, a little cheaper again" },
+    { id: GLM_47, label: "GLM-4.7", description: "strong value for everyday work" },
+    { id: GLM_47_FLASHX, label: "GLM-4.7 FlashX", description: "very cheap and quick, for simple work" },
+];
+/**
+ * Whether a model takes the `reasoning_effort` dial. Only GLM-5.2 does; the rest
+ * have thinking as a plain on/off toggle, and sending them an effort would be a
+ * parameter they do not know.
+ */
+export function takesEffort(model) {
+    return model === GLM_52 || alwaysThinks(model);
+}
+/** * Models whose thinking cannot be switched off. * * Offering them an "answer directly" rung would build a menu entry that produces a * request the provider refuses. The ladder and `normalize` both read this. */ export function alwaysThinks(model) { return model === GLM_53 || model === GLM_53_FLASH; }
+/**
+ * The reasoning levels offered by `/think`.
+ *
+ * GLM-5.2's `reasoning_effort` accepts seven values, but they COLLAPSE: `none` and
+ * `minimal` skip thinking, `low` and `medium` both map up to `high`, and `xhigh`
+ * maps up to `max`. So there are two real thinking levels behind seven names.
+ *
+ * Only the two are offered. Listing `low` and `medium` as separate rungs would put
+ * choices in front of the user that provably do nothing — the same reason Fable 5
+ * has no "answer directly" rung on the Anthropic driver. A menu should not contain
+ * a switch that is wired to the same place as the one above it.
+ */
+export function thinkLevels(model) {
+    // GLM-5.3: three documented depths, and no way to answer without reasoning. The
+    // lightest rung is still thinking, so it is not labelled as though it skips it.
+    if (model === GLM_53) {
+        return [
+            { label: "Light", description: "reasons briefly — fastest", thinking: true, effort: "low" },
+            { label: "Thinking", description: "think first, then answer", thinking: true, effort: "high" },
+            { label: "Maximum", description: "maximum reasoning budget", thinking: true, effort: "max" },
+        ];
+    }
+    // GLM-5.3 Flash documents one effort value and no way to disable thinking, so it
+    // gets one honest rung rather than a menu of switches wired to the same place.
+    if (model === GLM_53_FLASH) {
+        return [{ label: "Thinking", description: "always reasons — this model cannot skip it", thinking: true, effort: "max" }];
+    }
+    if (takesEffort(model)) {
+        return [
+            { label: "Standard", description: "answer directly — fastest", thinking: false, effort: "high" },
+            { label: "Thinking", description: "think first, then answer", thinking: true, effort: "high" },
+            { label: "Maximum", description: "maximum reasoning budget", thinking: true, effort: "max" },
+        ];
+    }
+    return [
+        { label: "Standard", description: "answer directly — fastest", thinking: false, effort: "high" },
+        { label: "Thinking", description: "think first, then answer", thinking: true, effort: "high" },
+    ];
+}
+/**
+ * List prices (USD / 1M tokens), international endpoint. Cached input is billed
+ * separately and far lower, which is what keeps a re-sent conversation cheap.
+ *
+ * The China endpoint charges different rates for the same model ids. This driver
+ * targets international, so these are the figures that apply to it.
+ */
+const PRICES = {
+    // Standard rates, NOT the launch promotion that halves Flash until 2026-09-09: a
+    // meter that quotes a discount is wrong the day it ends, and wrong quietly.
+    [GLM_53]: { cacheHit: 0.26, cacheMiss: 1.4, output: 4.4 },
+    [GLM_53_FLASH]: { cacheHit: 0.03, cacheMiss: 0.15, output: 0.5 },
+    [GLM_52]: { cacheHit: 0.26, cacheMiss: 1.4, output: 4.4 },
+    [GLM_5]: { cacheHit: 0.2, cacheMiss: 1, output: 3.2 },
+    [GLM_47]: { cacheHit: 0.11, cacheMiss: 0.6, output: 2.2 },
+    [GLM_47_FLASHX]: { cacheHit: 0.014, cacheMiss: 0.07, output: 0.4 },
+};
+/** Cache-aware list price for a model, falling back to the default model's. */
+export function price(model) {
+    return PRICES[model] ?? PRICES[DEFAULT_MODEL];
+}
+/**
+ * The model's USABLE context window. GLM-5.2 stores 1M, but this is deliberately
+ * the sharp window rather than the storage cap: on BYOK every token in the window is
+ * the user's money on every turn, so anchoring compaction at 1M would mean carrying
+ * an enormous prompt long after it stopped earning its cost. Same reasoning, and the
+ * same figure, as the other drivers.
+ */
+export function contextWindow(_model) {
+    return 200_000;
+}
+/**
+ * The ceiling this driver puts on a single BUFFERED (non-streaming) call. These
+ * models accept up to 128K output, but a non-streaming request that runs long risks
+ * an HTTP timeout, so the buffered path — core's small internal calls — is capped
+ * far lower. `client.ts` sends the same constant, and core reserves exactly this
+ * much room below the window, so the request and the reservation cannot drift apart.
+ */
+export const BUFFERED_OUTPUT_TOKENS = 8_000;
+export function bufferedOutputTokens(_model) {
+    return BUFFERED_OUTPUT_TOKENS;
+}
+/**
+ * Which models can SEE an image.
+ *
+ * GLM-5.3-Flash is natively multimodal — Z.ai documents image (and video) input on its
+ * own id, sent the ordinary OpenAI-compat way (`image_url` content parts), which the
+ * shared transport already produces. So core may attach the bytes and the model reads
+ * them. Every other GLM model here is text-only: pointing an image at one still degrades
+ * before anything is sent, which is what core does with a false answer here. (The older
+ * `glm-*v*` vision line is separate and not offered by this driver.)
+ */
+export function acceptsImages(model) {
+    return model === GLM_53_FLASH;
+}
+/**
+ * Coerce a stored or unknown config onto a model this provider actually serves.
+ *
+ * No per-model rule to enforce beyond the effort dial: every model here can have
+ * thinking turned off. The result is snapped onto a rung `/think` actually lists,
+ * which is also what keeps the request legal — the ladder holds only values the
+ * model accepts, so snapping to it is what stops the shared type's other rungs from
+ * reaching a model that has no dial at all.
+ */
+export function normalize(config) {
+    const model = PRICES[config.model] ? config.model : DEFAULT_MODEL;
+    // Forced on where the provider allows nothing else, so a config carried over from
+    // another model cannot produce a request this one refuses.
+    const thinking = alwaysThinks(model) ? true : config.thinking === true;
+    const effort = takesEffort(model) ? config.effort : "high";
+    return { model, thinking, effort: snapToOfferedRung(model, thinking, effort) };
+}
+/**
+ * Move an effort onto the nearest rung this model's `/think` ladder offers. Ties
+ * break DOWNWARD: an unlisted setting resolves to the cheaper neighbour, because
+ * silently spending more of the user's money is the worse way to be wrong.
+ */
+function snapToOfferedRung(model, thinking, effort) {
+    const ladder = ["low", "medium", "high", "xhigh", "max"];
+    const offered = thinkLevels(model).filter((l) => l.thinking === thinking);
+    if (offered.length === 0)
+        return effort;
+    if (offered.some((l) => l.effort === effort))
+        return effort;
+    const want = ladder.indexOf(effort);
+    let best = offered[0];
+    for (const level of offered) {
+        const d = Math.abs(ladder.indexOf(level.effort) - want);
+        const bestD = Math.abs(ladder.indexOf(best.effort) - want);
+        if (d < bestD || (d === bestD && ladder.indexOf(level.effort) < ladder.indexOf(best.effort))) {
+            best = level;
+        }
+    }
+    return best.effort;
+}
+/** The cheap metadata half of this driver — see `index.ts` for the wire half. */
+export const glmManifest = {
+    id: "glm",
+    label: "GLM",
+    apiKeyEnv: "ZAI_API_KEY",
+    keysUrl: "https://z.ai/manage-apikey/apikey-list",
+    models: MODELS,
+    thinkLevels,
+    price,
+    contextWindow,
+    bufferedOutputTokens,
+    acceptsImages,
+    normalize,
+};
+//# sourceMappingURL=manifest.js.map

package/dist/drivers/glm/manifest.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"manifest.js","sourceRoot":"","sources":["../../../src/drivers/glm/manifest.ts"],"names":[],"mappings":"AA4BA,MAAM,CAAC,MAAM,MAAM,GAAG,SAAS,CAAC;AAChC,MAAM,CAAC,MAAM,YAAY,GAAG,eAAe,CAAC;AAC5C,MAAM,CAAC,MAAM,MAAM,GAAG,SAAS,CAAC;AAChC,MAAM,CAAC,MAAM,KAAK,GAAG,OAAO,CAAC;AAC7B,MAAM,CAAC,MAAM,MAAM,GAAG,SAAS,CAAC;AAChC,MAAM,CAAC,MAAM,aAAa,GAAG,gBAAgB,CAAC;AAE9C,uEAAuE;AACvE,MAAM,CAAC,MAAM,aAAa,GAAG,MAAM,CAAC;AAEpC;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,MAAM,GAAkB;IACnC,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,oDAAoD,EAAE;IACnG,EAAE,EAAE,EAAE,YAAY,EAAE,KAAK,EAAE,eAAe,EAAE,WAAW,EAAE,+CAA+C,EAAE;IAC1G,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,8CAA8C,EAAE;IAC7F,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,WAAW,EAAE,+BAA+B,EAAE;IAC3E,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,gCAAgC,EAAE;IAC/E,EAAE,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,WAAW,EAAE,uCAAuC,EAAE;CACrG,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,WAAW,CAAC,KAAc;IACxC,OAAO,KAAK,KAAK,MAAM,IAAI,YAAY,CAAC,KAAK,CAAC,CAAC;AACjD,CAAC;AACD,sNAAsN,CAAA,MAAM,UAAU,YAAY,CAAC,KAAc,IAAc,OAAO,KAAK,KAAK,MAAM,IAAI,KAAK,KAAK,YAAY,CAAC,CAAA,CAAC;AAElU;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,WAAW,CAAC,KAAc;IACxC,gFAAgF;IAChF,gFAAgF;IAChF,IAAI,KAAK,KAAK,MAAM,EAAE,CAAC;QACrB,OAAO;YACL,EAAE,KAAK,EAAE,OAAO,EAAE,WAAW,EAAE,2BAA2B,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE;YAC3F,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,0BAA0B,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE;YAC9F,EAAE,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,0BAA0B,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE;SAC7F,CAAC;IACJ,CAAC;IACD,iFAAiF;IACjF,+EAA+E;IAC/E,IAAI,KAAK,KAAK,YAAY,EAAE,CAAC;QAC3B,OAAO,CAAC,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,4CAA4C,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC;IAC3H,CAAC;IACD,IAAI,WAAW,CAAC,KAAK,CAAC,EAAE,CAAC;QACvB,OAAO;YACL,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,2BAA2B,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE;YAChG,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,0BAA0B,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE;YAC9F,EAAE,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,0BAA0B,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE;SAC7F,CAAC;IACJ,CAAC;IACD,OAAO;QACL,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,2BAA2B,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE;QAChG,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,0BAA0B,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE;KAC/F,CAAC;AACJ,CAAC;AAED;;;;;;GAMG;AACH,MAAM,MAAM,GAA+B;IACzC,iFAAiF;IACjF,4EAA4E;IAC5E,CAAC,MAAM,CAAC,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,SAAS,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE;IACzD,CAAC,YAAY,CAAC,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,GAAG,EAAE;IAChE,CAAC,MAAM,CAAC,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,SAAS,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE;IACzD,CAAC,KAAK,CAAC,EAAE,EAAE,QAAQ,EAAE,GAAG,EAAE,SAAS,EAAE,CAAC,EAAE,MAAM,EAAE,GAAG,EAAE;IACrD,CAAC,MAAM,CAAC,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,SAAS,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE;IACzD,CAAC,aAAa,CAAC,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,GAAG,EAAE;CACnE,CAAC;AAEF,+EAA+E;AAC/E,MAAM,UAAU,KAAK,CAAC,KAAc;IAClC,OAAO,MAAM,CAAC,KAAK,CAAC,IAAI,MAAM,CAAC,aAAa,CAAE,CAAC;AACjD,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,aAAa,CAAC,MAAe;IAC3C,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAG,KAAK,CAAC;AAE5C,MAAM,UAAU,oBAAoB,CAAC,MAAe;IAClD,OAAO,sBAAsB,CAAC;AAChC,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,aAAa,CAAC,KAAc;IAC1C,OAAO,KAAK,KAAK,YAAY,CAAC;AAChC,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,SAAS,CAAC,MAAmB;IAC3C,MAAM,KAAK,GAAY,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,aAAa,CAAC;IAC3E,kFAAkF;IAClF,2DAA2D;IAC3D,MAAM,QAAQ,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,QAAQ,KAAK,IAAI,CAAC;IACvE,MAAM,MAAM,GAAW,WAAW,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC;IACnE,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,iBAAiB,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,CAAC,EAAE,CAAC;AACjF,CAAC;AAED;;;;GAIG;AACH,SAAS,iBAAiB,CAAC,KAAc,EAAE,QAAiB,EAAE,MAAc;IAC1E,MAAM,MAAM,GAAa,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,CAAC,CAAC;IACnE,MAAM,OAAO,GAAG,WAAW,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,KAAK,QAAQ,CAAC,CAAC;IAC1E,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC;IACxC,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,KAAK,MAAM,CAAC;QAAE,OAAO,MAAM,CAAC;IAE5D,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;IACpC,IAAI,IAAI,GAAG,OAAO,CAAC,CAAC,CAAE,CAAC;IACvB,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC5B,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC;QACxD,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC;QAC3D,IAAI,CAAC,GAAG,KAAK,IAAI,CAAC,CAAC,KAAK,KAAK,IAAI,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;YAC7F,IAAI,GAAG,KAAK,CAAC;QACf,CAAC;IACH,CAAC;IACD,OAAO,IAAI,CAAC,MAAM,CAAC;AACrB,CAAC;AAED,iFAAiF;AACjF,MAAM,CAAC,MAAM,WAAW,GAAmB;IACzC,EAAE,EAAE,KAAK;IACT,KAAK,EAAE,KAAK;IACZ,SAAS,EAAE,aAAa;IACxB,OAAO,EAAE,wCAAwC;IACjD,MAAM,EAAE,MAAM;IACd,WAAW;IACX,KAAK;IACL,aAAa;IACb,oBAAoB;IACpB,aAAa;IACb,SAAS;CACV,CAAC"}

package/dist/drivers/groq/client.js

@@ -0,0 +1,80 @@
+import { compatStreamTurn, compatToolTurn, listModels, standardCacheSplit } from "../openaiCompat/wire.js";
+import { BUFFERED_OUTPUT_TOKENS, DEFAULT_MODEL, takesEffort } from "./manifest.js";
+const BASE_URL = process.env.MINDWEAVE_GROQ_URL ?? "https://api.groq.com/openai/v1";
+/** The model a request runs on, matching the shared layer's own fallback. */
+function modelOf(config) {
+    return config?.model ?? process.env.MINDWEAVE_MODEL ?? DEFAULT_MODEL;
+}
+/**
+ * Groq's reasoning fields.
+ *
+ * `reasoning_effort` here accepts only `none` or `default` — not the graded ladder
+ * every other provider on this layer exposes. Sending `high` is a 400, not a
+ * downgrade, so the shared effort rung is deliberately NOT forwarded: it is
+ * collapsed to the two states this API actually has.
+ *
+ * Sent only for the models that accept the parameter at all; the rest reject it.
+ */
+export function reasoningFields(config) {
+    if (!takesEffort(modelOf(config)))
+        return {};
+    return { reasoning_effort: config?.thinking ? "default" : "none" };
+}
+/**
+ * Turn Groq's live catalogue into picker entries.
+ *
+ * The listing reports a real `context_window` per model, which is the one piece of
+ * per-model truth a discovered provider would otherwise have to guess. It is folded
+ * into the description so the choice is informed, rather than dropped because the
+ * shared `ModelChoice` has nowhere structured to put it.
+ *
+ * Non-chat models are filtered out. The endpoint lists speech and guard models
+ * alongside chat ones, and offering a transcription model in `/model` would be an
+ * entry that fails on first use.
+ */
+export function toChoices(listed) {
+    return listed
+        .filter((m) => !/whisper|tts|guard|prompt-?guard/i.test(m.id))
+        .map((m) => ({
+        id: m.id,
+        label: m.id,
+        description: m.context_window ? `${Math.round(m.context_window / 1000)}K context` : "served on Groq",
+    }))
+        .sort((a, b) => a.id.localeCompare(b.id));
+}
+/**
+ * Fetch the models Groq is serving right now.
+ *
+ * Throws when the key is missing or the request fails, which is the contract the
+ * registry relies on: a throw keeps the previous list, an empty array replaces it.
+ */
+export async function discoverModels() {
+    return toChoices(await listModels(groqProvider));
+}
+/**
+ * Groq caches automatically on every request, with no code change and no opt-in, and
+ * reports the hit as `prompt_tokens_details.cached_tokens`. An earlier version of
+ * this driver claimed it reported nothing, which made a multi-step turn count every
+ * step's whole prompt as fresh and report several times the tokens it used.
+ *
+ * No `extraStop`: the finish reasons are the plain OpenAI set, so there is nothing
+ * extra to map. That absence is deliberate.
+ */
+export const groqProvider = {
+    label: "Groq",
+    baseUrl: BASE_URL,
+    apiKeyEnv: "GROQ_API_KEY",
+    defaultModel: process.env.MINDWEAVE_MODEL ?? DEFAULT_MODEL,
+    reasoningFields,
+    cacheSplit: standardCacheSplit,
+    bufferedMaxTokens: BUFFERED_OUTPUT_TOKENS,
+};
+/** Ask the model for one turn. */
+export async function toolTurn(req, options = {}) {
+    return compatToolTurn(groqProvider, req, options.signal);
+}
+/** Ask the model for one turn, streaming deltas to `options.onEvent`. */
+export async function streamTurn(req, options = {}) {
+    return compatStreamTurn(groqProvider, req, options.onEvent, options.signal);
+}
+//# sourceMappingURL=client.js.map

package/dist/drivers/groq/client.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"client.js","sourceRoot":"","sources":["../../../src/drivers/groq/client.ts"],"names":[],"mappings":"AAmBA,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAE,UAAU,EAAE,kBAAkB,EAAuB,MAAM,yBAAyB,CAAC;AAChI,OAAO,EAAE,sBAAsB,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAEnF,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,kBAAkB,IAAI,gCAAgC,CAAC;AAEpF,6EAA6E;AAC7E,SAAS,OAAO,CAAC,MAA+B;IAC9C,OAAO,MAAM,EAAE,KAAK,IAAI,OAAO,CAAC,GAAG,CAAC,eAAe,IAAI,aAAa,CAAC;AACvE,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,eAAe,CAAC,MAA+B;IAC7D,IAAI,CAAC,WAAW,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;QAAE,OAAO,EAAE,CAAC;IAC7C,OAAO,EAAE,gBAAgB,EAAE,MAAM,EAAE,QAAQ,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC;AACrE,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,SAAS,CAAC,MAAiD;IACzE,OAAO,MAAM;SACV,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,kCAAkC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;SAC7D,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QACX,EAAE,EAAE,CAAC,CAAC,EAAE;QACR,KAAK,EAAE,CAAC,CAAC,EAAE;QACX,WAAW,EAAE,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,cAAc,GAAG,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,gBAAgB;KACrG,CAAC,CAAC;SACF,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;AAC9C,CAAC;AAED;;;;;GAKG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc;IAClC,OAAO,SAAS,CAAC,MAAM,UAAU,CAAC,YAAY,CAAC,CAAC,CAAC;AACnD,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,YAAY,GAAmB;IAC1C,KAAK,EAAE,MAAM;IACb,OAAO,EAAE,QAAQ;IACjB,SAAS,EAAE,cAAc;IACzB,YAAY,EAAE,OAAO,CAAC,GAAG,CAAC,eAAe,IAAI,aAAa;IAC1D,eAAe;IACf,UAAU,EAAE,kBAAkB;IAC9B,iBAAiB,EAAE,sBAAsB;CAC1C,CAAC;AAEF,kCAAkC;AAClC,MAAM,CAAC,KAAK,UAAU,QAAQ,CAAC,GAAiB,EAAE,UAAuB,EAAE;IACzE,OAAO,cAAc,CAAC,YAAY,EAAE,GAAG,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;AAC3D,CAAC;AAED,yEAAyE;AACzE,MAAM,CAAC,KAAK,UAAU,UAAU,CAAC,GAAiB,EAAE,UAAyB,EAAE;IAC7E,OAAO,gBAAgB,CAAC,YAAY,EAAE,GAAG,EAAE,OAAO,CAAC,OAAO,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;AAC9E,CAAC"}

package/dist/drivers/groq/index.js

@@ -0,0 +1,5 @@
+import { streamTurn, toolTurn } from "./client.js";
+import { groqManifest } from "./manifest.js";
+export const groqDriver = { ...groqManifest, toolTurn, streamTurn };
+export default groqDriver;
+//# sourceMappingURL=index.js.map

package/dist/drivers/groq/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/drivers/groq/index.ts"],"names":[],"mappings":"AAeA,OAAO,EAAE,UAAU,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AACnD,OAAO,EAAE,YAAY,EAAE,MAAM,eAAe,CAAC;AAE7C,MAAM,CAAC,MAAM,UAAU,GAAW,EAAE,GAAG,YAAY,EAAE,QAAQ,EAAE,UAAU,EAAE,CAAC;AAE5E,eAAe,UAAU,CAAC"}

package/dist/drivers/groq/manifest.js

@@ -0,0 +1,160 @@
+/**
+ * A minimal fallback list, shown only before discovery has run or when it fails.
+ *
+ * Deliberately short and boring: these are the ids least likely to have been
+ * retired, because a fallback that names a withdrawn model is worse than a fallback
+ * that names two working ones. The live list replaces this entirely.
+ */
+export const MODELS = [
+    { id: "llama-3.3-70b-versatile", label: "Llama 3.3 70B", description: "general purpose, very fast" },
+    { id: "openai/gpt-oss-120b", label: "GPT-OSS 120B", description: "open weights, with a reasoning dial" },
+];
+/** The model used when nothing is saved and no env override is set. */
+export const DEFAULT_MODEL = MODELS[0].id;
+/**
+ * Models known to accept `reasoning_effort`.
+ *
+ * Matched by SUBSTRING rather than exact id, and that is a concession to the
+ * discovered lineup: ids arrive with vendor prefixes and size suffixes that vary
+ * (`openai/gpt-oss-120b`, `qwen/qwen3-32b`), so an exact table would go stale for
+ * the same reason the model list does. A false negative here costs a reasoning dial;
+ * a false positive costs a rejected request, so the list stays narrow.
+ */
+const EFFORT_MODELS = ["gpt-oss", "qwen3"];
+export function takesEffort(model) {
+    const id = model.toLowerCase();
+    return EFFORT_MODELS.some((fragment) => id.includes(fragment));
+}
+/**
+ * The reasoning levels offered by `/think`.
+ *
+ * Groq accepts only `none` and `default` for `reasoning_effort` — not the graded
+ * ladder other providers expose. So there are exactly two states, and offering a
+ * "Deep" or "Maximum" rung would be a menu entry that sends a value the API rejects.
+ * `default` is expressed here as thinking-on; `none` as thinking-off.
+ */
+export function thinkLevels(model) {
+    if (takesEffort(model)) {
+        return [
+            { label: "Standard", description: "answer directly — fastest", thinking: false, effort: "low" },
+            { label: "Thinking", description: "think first, then answer", thinking: true, effort: "high" },
+        ];
+    }
+    return [{ label: "Standard", description: "this model has no reasoning dial", thinking: false, effort: "low" }];
+}
+/**
+ * List prices (USD / 1M tokens), for the models whose rates are published.
+ *
+ * A discovered lineup cannot have a complete price table by definition — a model
+ * that appears tomorrow has no row here. The fallback is deliberately the CHEAPEST
+ * plausible rate rather than an average: this provider serves small open models at
+ * low rates, so guessing high would make a cost figure look alarming for a session
+ * that was in fact inexpensive. Matched by substring, for the same reason as
+ * `takesEffort`.
+ *
+ * Cached input takes a flat 50% discount here — not the ~90% most providers give —
+ * so `cacheHit` is half of `cacheMiss` throughout rather than a tenth.
+ */
+const PRICES = [
+    ["gpt-oss-120b", { cacheHit: 0.075, cacheMiss: 0.15, output: 0.75 }],
+    ["gpt-oss-20b", { cacheHit: 0.05, cacheMiss: 0.1, output: 0.5 }],
+    ["llama-3.3-70b", { cacheHit: 0.295, cacheMiss: 0.59, output: 0.79 }],
+    ["llama-3.1-8b", { cacheHit: 0.025, cacheMiss: 0.05, output: 0.08 }],
+    ["qwen3-32b", { cacheHit: 0.145, cacheMiss: 0.29, output: 0.59 }],
+];
+const UNKNOWN_PRICE = { cacheHit: 0.05, cacheMiss: 0.1, output: 0.5 };
+export function price(model) {
+    const id = model.toLowerCase();
+    return PRICES.find(([fragment]) => id.includes(fragment))?.[1] ?? UNKNOWN_PRICE;
+}
+/**
+ * The model's USABLE context window.
+ *
+ * Groq's `/models` listing reports a real `context_window` per model, and
+ * `client.ts` folds it into the discovered entry. This is the floor used when that
+ * is unavailable — a conservative 128K, since the open models served here range from
+ * 8K to over 250K and over-stating it would let a transcript grow past what the
+ * model can actually attend to.
+ */
+export function contextWindow(_model) {
+    return 128_000;
+}
+/** The ceiling this driver puts on a single BUFFERED call. */
+export const BUFFERED_OUTPUT_TOKENS = 8_000;
+export function bufferedOutputTokens(_model) {
+    return BUFFERED_OUTPUT_TOKENS;
+}
+/** The effort rungs this driver uses; Groq's real vocabulary is two states. */
+const EFFORTS = ["low", "medium", "high", "xhigh", "max"];
+/**
+ * Coerce a stored or unknown config onto something serveable.
+ *
+ * The one thing that CANNOT be done here, and is the difference from every static
+ * driver: reject an unknown model id. A discovered provider's list may be empty
+ * (discovery has not run, or the key is missing), and coercing the user's saved
+ * model onto a fallback in that window would silently change their choice on every
+ * launch. So an unrecognised id is kept as-is and left for the provider to judge.
+ */
+export function normalize(config) {
+    const model = config.model || DEFAULT_MODEL;
+    const thinking = takesEffort(model) ? config.thinking === true : false;
+    const effort = EFFORTS.includes(config.effort) ? config.effort : "low";
+    return { model, thinking, effort: snapToOfferedRung(model, thinking, effort) };
+}
+/** Move an effort onto the nearest rung this model's ladder offers, ties downward. */
+function snapToOfferedRung(model, thinking, effort) {
+    const offered = thinkLevels(model).filter((l) => l.thinking === thinking);
+    if (offered.length === 0)
+        return thinkLevels(model)[0].effort;
+    if (offered.some((l) => l.effort === effort))
+        return effort;
+    const want = EFFORTS.indexOf(effort);
+    let best = offered[0];
+    for (const level of offered) {
+        const d = Math.abs(EFFORTS.indexOf(level.effort) - want);
+        const bestD = Math.abs(EFFORTS.indexOf(best.effort) - want);
+        if (d < bestD || (d === bestD && EFFORTS.indexOf(level.effort) < EFFORTS.indexOf(best.effort)))
+            best = level;
+    }
+    return best.effort;
+}
+/**
+ * Recognise a Groq model id without a list.
+ *
+ * Needed because the provider is derived from the model id and a discovered list is
+ * empty until asked: without this, a saved Groq model would be attributed to the
+ * fallback provider on launch and the session would open on the wrong one. These
+ * fragments are the vendor prefixes Groq's catalogue uses.
+ *
+ * Consulted only AFTER every provider's real list is searched, so it can never take
+ * a model another provider actually serves.
+ */
+export function ownsModel(model) {
+    const id = model.toLowerCase();
+    return ["llama-", "openai/gpt-oss", "qwen/", "moonshotai/", "groq/", "meta-llama/"].some((p) => id.startsWith(p));
+}
+/** The cheap metadata half of this driver — see `index.ts` for the wire half. */
+export const groqManifest = {
+    id: "groq",
+    label: "Groq",
+    apiKeyEnv: "GROQ_API_KEY",
+    keysUrl: "https://console.groq.com/keys",
+    models: MODELS,
+    thinkLevels,
+    price,
+    contextWindow,
+    bufferedOutputTokens,
+    normalize,
+    ownsModel,
+    /**
+     * Discovery lives here rather than on the driver, because the `/model` picker
+     * needs the list before any wire code has been loaded — that is the whole reason
+     * manifests are separate from drivers.
+     *
+     * The import is INSIDE the function, so this module still touches no network and
+     * loads no wire code at import time. A manifest that imported its client at the
+     * top would undo the lazy split for every user of every other provider.
+     */
+    discoverModels: async () => (await import("./client.js")).discoverModels(),
+};
+//# sourceMappingURL=manifest.js.map

package/dist/drivers/groq/manifest.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"manifest.js","sourceRoot":"","sources":["../../../src/drivers/groq/manifest.ts"],"names":[],"mappings":"AAmBA;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,MAAM,GAAkB;IACnC,EAAE,EAAE,EAAE,yBAAyB,EAAE,KAAK,EAAE,eAAe,EAAE,WAAW,EAAE,4BAA4B,EAAE;IACpG,EAAE,EAAE,EAAE,qBAAqB,EAAE,KAAK,EAAE,cAAc,EAAE,WAAW,EAAE,qCAAqC,EAAE;CACzG,CAAC;AAEF,uEAAuE;AACvE,MAAM,CAAC,MAAM,aAAa,GAAG,MAAM,CAAC,CAAC,CAAE,CAAC,EAAE,CAAC;AAE3C;;;;;;;;GAQG;AACH,MAAM,aAAa,GAAG,CAAC,SAAS,EAAE,OAAO,CAAC,CAAC;AAE3C,MAAM,UAAU,WAAW,CAAC,KAAc;IACxC,MAAM,EAAE,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC;IAC/B,OAAO,aAAa,CAAC,IAAI,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,EAAE,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC;AACjE,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,WAAW,CAAC,KAAc;IACxC,IAAI,WAAW,CAAC,KAAK,CAAC,EAAE,CAAC;QACvB,OAAO;YACL,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,2BAA2B,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE;YAC/F,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,0BAA0B,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE;SAC/F,CAAC;IACJ,CAAC;IACD,OAAO,CAAC,EAAE,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,kCAAkC,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC;AAClH,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,MAAM,GAA4C;IACtD,CAAC,cAAc,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;IACpE,CAAC,aAAa,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,SAAS,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE,CAAC;IAChE,CAAC,eAAe,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;IACrE,CAAC,cAAc,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;IACpE,CAAC,WAAW,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;CAClE,CAAC;AAEF,MAAM,aAAa,GAAe,EAAE,QAAQ,EAAE,IAAI,EAAE,SAAS,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE,CAAC;AAElF,MAAM,UAAU,KAAK,CAAC,KAAc;IAClC,MAAM,EAAE,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC;IAC/B,OAAO,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,QAAQ,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,aAAa,CAAC;AAClF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,aAAa,CAAC,MAAe;IAC3C,OAAO,OAAO,CAAC;AACjB,CAAC;AAED,8DAA8D;AAC9D,MAAM,CAAC,MAAM,sBAAsB,GAAG,KAAK,CAAC;AAE5C,MAAM,UAAU,oBAAoB,CAAC,MAAe;IAClD,OAAO,sBAAsB,CAAC;AAChC,CAAC;AAED,+EAA+E;AAC/E,MAAM,OAAO,GAAa,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,CAAC,CAAC;AAEpE;;;;;;;;GAQG;AACH,MAAM,UAAU,SAAS,CAAC,MAAmB;IAC3C,MAAM,KAAK,GAAY,MAAM,CAAC,KAAK,IAAI,aAAa,CAAC;IACrD,MAAM,QAAQ,GAAG,WAAW,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,QAAQ,KAAK,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC;IACvE,MAAM,MAAM,GAAW,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;IAC/E,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,iBAAiB,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,CAAC,EAAE,CAAC;AACjF,CAAC;AAED,sFAAsF;AACtF,SAAS,iBAAiB,CAAC,KAAc,EAAE,QAAiB,EAAE,MAAc;IAC1E,MAAM,OAAO,GAAG,WAAW,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,KAAK,QAAQ,CAAC,CAAC;IAC1E,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,WAAW,CAAC,KAAK,CAAC,CAAC,CAAC,CAAE,CAAC,MAAM,CAAC;IAC/D,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,KAAK,MAAM,CAAC;QAAE,OAAO,MAAM,CAAC;IAE5D,MAAM,IAAI,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;IACrC,IAAI,IAAI,GAAG,OAAO,CAAC,CAAC,CAAE,CAAC;IACvB,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC5B,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC;QACzD,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC;QAC5D,IAAI,CAAC,GAAG,KAAK,IAAI,CAAC,CAAC,KAAK,KAAK,IAAI,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YAAE,IAAI,GAAG,KAAK,CAAC;IAC/G,CAAC;IACD,OAAO,IAAI,CAAC,MAAM,CAAC;AACrB,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,SAAS,CAAC,KAAc;IACtC,MAAM,EAAE,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC;IAC/B,OAAO,CAAC,QAAQ,EAAE,gBAAgB,EAAE,OAAO,EAAE,aAAa,EAAE,OAAO,EAAE,aAAa,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;AACpH,CAAC;AAED,iFAAiF;AACjF,MAAM,CAAC,MAAM,YAAY,GAAmB;IAC1C,EAAE,EAAE,MAAM;IACV,KAAK,EAAE,MAAM;IACb,SAAS,EAAE,cAAc;IACzB,OAAO,EAAE,+BAA+B;IACxC,MAAM,EAAE,MAAM;IACd,WAAW;IACX,KAAK;IACL,aAAa;IACb,oBAAoB;IACpB,SAAS;IACT,SAAS;IACT;;;;;;;;OAQG;IACH,cAAc,EAAE,KAAK,IAAI,EAAE,CAAC,CAAC,MAAM,MAAM,CAAC,aAAa,CAAC,CAAC,CAAC,cAAc,EAAE;CAC3E,CAAC"}

package/dist/drivers/kimi/client.js

@@ -0,0 +1,82 @@
+import { compatStreamTurn, compatToolTurn } from "../openaiCompat/wire.js";
+import { BUFFERED_OUTPUT_TOKENS, DEFAULT_MODEL, surfaceOf } from "./manifest.js";
+/** Moonshot's international endpoint, which is where these model ids are served. */
+const BASE_URL = process.env.MINDWEAVE_KIMI_URL ?? "https://api.moonshot.ai/v1";
+/** The model a request runs on, matching the shared layer's own fallback so the
+ *  reasoning fields can never be built for a different model than is being called. */
+function modelOf(config) {
+    return config?.model ?? process.env.MINDWEAVE_MODEL ?? DEFAULT_MODEL;
+}
+/**
+ * Kimi's reasoning fields — three shapes across one lineup.
+ *
+ *   - K3 has NO `thinking` parameter. It takes a top-level `reasoning_effort`, and
+ *     it reasons on every response whether asked to or not. Sending `thinking` here
+ *     would be a parameter the model does not know.
+ *   - K2.7 Code takes `thinking: {type}` but accepts only `"enabled"`; `"disabled"`
+ *     is a rejected request rather than a soft no-op.
+ *   - K2.6 and K2.5 take `thinking: {type}` both ways, and default to ENABLED — so
+ *     the off switch is sent explicitly, exactly as on Qwen and DeepSeek. Omitting
+ *     it would mean every internal call (the summarizer, the page distiller) quietly
+ *     paying for reasoning the UI then discards.
+ *
+ * `normalize` has already made the config legal for the model, so this only renders
+ * it. The model is resolved the same way the shared layer resolves it, so the two
+ * can never disagree about which model's rules apply.
+ */
+export function reasoningFields(config) {
+    const surface = surfaceOf(modelOf(config));
+    if (surface.takesEffort) {
+        // K3: always reasons; the rung is the only choice. `normalize` has already
+        // clamped it to a value this model accepts.
+        return { reasoning_effort: config?.effort ?? "high" };
+    }
+    if (!surface.canDisableThinking)
+        return { thinking: { type: "enabled" } };
+    return { thinking: { type: config?.thinking ? "enabled" : "disabled" } };
+}
+/**
+ * Kimi reports its cache hit as `cached_tokens` on the usage object, where
+ * `prompt_tokens` is the FULL prompt including the cached part — so the miss side
+ * is the remainder rather than a separate figure to add. The nested
+ * `prompt_tokens_details` shape is read too, since the platform reports it both
+ * ways depending on the model.
+ */
+export function cacheSplit(usage) {
+    const details = usage.prompt_tokens_details;
+    const nested = typeof details?.cached_tokens === "number" ? details.cached_tokens : 0;
+    const flat = typeof usage.cached_tokens === "number" ? usage.cached_tokens : 0;
+    const hit = nested || flat;
+    if (hit === 0)
+        return undefined;
+    const prompt = typeof usage.prompt_tokens === "number" ? usage.prompt_tokens : 0;
+    return { hit, miss: Math.max(0, prompt - hit) };
+}
+/**
+ * Everything the shared wire layer needs to talk to Kimi.
+ *
+ * No `extraStop`: this platform documents exactly the OpenAI vocabulary
+ * (`stop` / `length` / `tool_calls`), so there is nothing extra to map. Stated
+ * rather than left blank, because an absent mapper is otherwise indistinguishable
+ * from a forgotten one.
+ */
+export const kimiProvider = {
+    label: "Kimi",
+    baseUrl: BASE_URL,
+    apiKeyEnv: "MOONSHOT_API_KEY",
+    defaultModel: process.env.MINDWEAVE_MODEL ?? DEFAULT_MODEL,
+    reasoningFields,
+    cacheSplit,
+    // Must match the manifest's `bufferedOutputTokens`, which is what core reserves
+    // room for below the context window.
+    bufferedMaxTokens: BUFFERED_OUTPUT_TOKENS,
+};
+/** Ask the model for one turn. */
+export async function toolTurn(req, options = {}) {
+    return compatToolTurn(kimiProvider, req, options.signal);
+}
+/** Ask the model for one turn, streaming deltas to `options.onEvent`. */
+export async function streamTurn(req, options = {}) {
+    return compatStreamTurn(kimiProvider, req, options.onEvent, options.signal);
+}
+//# sourceMappingURL=client.js.map