@nodetool-ai/agents 0.7.0-rc.23 → 0.8.2

package/dist/capabilities/models.js

@@ -0,0 +1,805 @@
+/**
+ * The `models` capability module — model discovery over the configured
+ * providers.
+ *
+ * Three capabilities that used to be three `Tool` subclasses, one per file.
+ * Wire names, descriptions and schemas are unchanged: a belt builds them
+ * from `models.specs.ts` by name.
+ *
+ * The providers map was a constructor argument and is now `run.providers`. It
+ * is read at call time, so a host that fills the map lazily — the MCP mount
+ * does — still serves the models it resolved after construction.
+ *
+ * `find_model` is where an agent's default model choice is made: with no hint
+ * from the caller it answers with the model that tops the quality leaderboard
+ * for the task (see {@link SCORE_TIERS}), so an agent that just asks for a
+ * capability gets the best model for the job rather than the first one a
+ * provider happens to list.
+ *
+ * Design: docs/tool-class-retirement-design.md § Migration.
+ */
+import { RECOMMENDED_MODELS } from "@nodetool-ai/runtime";
+import { getModelRank, modelRankings, routesFor } from "@nodetool-ai/model-pricing/model-rankings";
+import { findModelSpec, listModelsSpec, listProviderModelsSpec, SUPPORTED_CAPABILITIES, MODEL_TYPES } from "./models.specs.js";
+import { isFunction, isNonBlankString, isNumber, isString } from "../utils/type-guards.js";
+import { promptingSkillFor } from "../model-prompting-skills.js";
+/** Providers this run can reach. Empty when the host wired none. */
+function providersOf(run) {
+    return run.providers ?? {};
+}
+/**
+ * A provider is treated as offering downloaded models when it runs locally.
+ * Neither `find_model` nor `list_models` inspects the on-disk cache, so both
+ * report the same `downloaded` for the same model.
+ *
+ * `huggingface` is not one of them: it is the HF Inference API, a remote call
+ * like any other. Counting it as local gave every HF model the `downloaded`
+ * bonus and put `FLUX.1-schnell` ahead of the fal_ai copy that could actually
+ * run, on a host with no `@huggingface/inference` installed.
+ */
+const LOCAL_PROVIDER_IDS = new Set([
+    "ollama",
+    "lmstudio",
+    "vllm",
+    "llama_cpp",
+    "node_llama_cpp",
+    "whisper_cpp",
+    "whisper_cpp_server"
+]);
+/**
+ * Why a configured provider still cannot serve a call, or `null` when it can.
+ * A provider that answers with a reason is dropped from the ranking and named
+ * in the result's note — ranking a model nothing can run turns a discovery
+ * call into a failed generation call.
+ */
+async function unavailableReasonOf(provider) {
+    // Guarded rather than called outright: a provider from an older build has no
+    // such method, and hiding every provider over that would be a far worse
+    // failure than the one this prevents.
+    if (!isFunction(provider.unavailableReason))
+        return null;
+    try {
+        return await provider.unavailableReason();
+    }
+    catch (e) {
+        return e instanceof Error ? e.message : String(e);
+    }
+}
+/**
+ * The typed-ref property shape a model-typed node property takes. A pick
+ * result's flat fields (`model_id`) do not round-trip into a property — a
+ * live session lost three rounds to exactly that — so every result carries
+ * the assignable form too.
+ */
+const CAPABILITY_REF_TYPE = {
+    generate_message: "language_model",
+    text_to_image: "image_model",
+    image_to_image: "image_model",
+    segment_image: "image_model",
+    text_to_video: "video_model",
+    image_to_video: "video_model",
+    reference_to_video: "video_model",
+    text_to_speech: "tts_model",
+    // A music-typed node property takes a `music_model`, never a `tts_model`.
+    // Handing back the wrong tag made every music ref unassignable: the property
+    // refused it, and the graph validator then checked the id against the TTS
+    // catalog and reported a real music model as one the provider does not offer.
+    text_to_music: "music_model",
+    audio_to_audio: "audio_to_audio_model",
+    automatic_speech_recognition: "asr_model",
+    generate_embedding: "embedding_model"
+};
+function modelRef(capability, provider, id, name) {
+    return { type: CAPABILITY_REF_TYPE[capability], provider, id, name };
+}
+function getRecommendedSet(capability) {
+    const wantedTasks = capabilityToRecommendedTasks(capability);
+    const wantedModalities = capabilityToRecommendedModalities(capability);
+    const ids = new Set();
+    for (const m of RECOMMENDED_MODELS) {
+        const taskMatch = !wantedTasks || (m.task && wantedTasks.has(m.task));
+        const modalityMatch = wantedModalities.has(m.modality);
+        if (modalityMatch && taskMatch && m.provider) {
+            ids.add(`${m.provider}::${m.id}`);
+        }
+    }
+    return ids;
+}
+function capabilityToRecommendedTasks(capability) {
+    switch (capability) {
+        case "text_to_image":
+            return new Set(["text_to_image"]);
+        case "image_to_image":
+            return new Set(["image_to_image"]);
+        case "segment_image":
+            // No recommended entry is a segmentation model, so nothing is boosted.
+            return new Set(["segment"]);
+        case "text_to_video":
+            return new Set(["text_to_video"]);
+        case "image_to_video":
+            return new Set(["image_to_video"]);
+        case "reference_to_video":
+            return new Set(["reference_to_video"]);
+        case "generate_embedding":
+            return new Set(["embedding"]);
+        case "generate_message":
+            return new Set(["text_generation"]);
+        default:
+            // tts/asr have no `task` set on recommended entries — fall back to
+            // modality-only filtering.
+            return null;
+    }
+}
+function capabilityToRecommendedModalities(capability) {
+    switch (capability) {
+        case "text_to_image":
+        case "image_to_image":
+        case "segment_image":
+            return new Set(["image"]);
+        case "text_to_video":
+        case "image_to_video":
+        case "reference_to_video":
+            return new Set(["video"]);
+        case "text_to_speech":
+            return new Set(["tts"]);
+        case "text_to_music":
+            return new Set(["music"]);
+        case "audio_to_audio":
+            return new Set(["audio_to_audio"]);
+        case "automatic_speech_recognition":
+            return new Set(["asr"]);
+        case "generate_embedding":
+        case "generate_message":
+            return new Set(["language"]);
+    }
+}
+async function fetchModelsForCapability(provider, capability) {
+    switch (capability) {
+        case "text_to_image":
+        case "image_to_image":
+        case "segment_image":
+            return await provider.getAvailableImageModels();
+        case "text_to_video":
+        case "image_to_video":
+        case "reference_to_video":
+            return await provider.getAvailableVideoModels();
+        case "text_to_speech":
+            return await provider.getAvailableTTSModels();
+        case "text_to_music":
+            return await provider.getAvailableMusicModels();
+        case "audio_to_audio":
+            return await provider.getAvailableAudioToAudioModels();
+        case "automatic_speech_recognition":
+            return await provider.getAvailableASRModels();
+        case "generate_embedding":
+            return await provider.getAvailableEmbeddingModels();
+        case "generate_message":
+            return await provider.getAvailableLanguageModels();
+    }
+}
+/**
+ * The task a capability asks a model to perform, where one model list serves
+ * two directions. `getAvailableVideoModels` answers with every video endpoint
+ * the provider has — text-to-video, image-to-video, lip-sync, upscalers — so
+ * without this a `text_to_video` search ranked `.../image-to-video` first and
+ * the generation call that followed failed with a 422 from the provider. The
+ * same holds for image models (`text_to_image` vs `image_to_image`).
+ *
+ * It is also the task the leaderboard is read for when the caller names none.
+ * Without that, `pickTaskRank` was free to score a candidate on its *best*
+ * task: a `text_to_music` search ranked ElevenLabs' dialogue model first on
+ * its `text_to_speech` standing, in a leaderboard that has no music task at
+ * all.
+ *
+ * Capabilities whose ranking is meaningless either way (asr, embeddings,
+ * language — none of them ranked) map to `null` and filter nothing.
+ */
+function capabilityTask(capability) {
+    switch (capability) {
+        // The image catalog tags a segmentation endpoint `segment`, which is what
+        // keeps a FLUX from being offered as the answer to "segment this".
+        case "segment_image":
+            return "segment";
+        case "text_to_image":
+        case "image_to_image":
+        case "text_to_video":
+        case "image_to_video":
+        case "reference_to_video":
+        case "text_to_speech":
+        case "text_to_music":
+            return capability;
+        default:
+            return null;
+    }
+}
+/**
+ * Models that declare they can do this capability's task. A model that
+ * declares no tasks at all is kept — the alternative is hiding every model
+ * from a provider whose manifest carries no task information.
+ */
+function forCapabilityTask(candidates, capability) {
+    const task = capabilityTask(capability);
+    if (!task)
+        return { kept: candidates, dropped: 0 };
+    const kept = candidates.filter(({ model }) => taskMatch(model, task));
+    return { kept, dropped: candidates.length - kept.length };
+}
+function taskMatch(model, task) {
+    if (!task)
+        return true;
+    if (!model.supportedTasks || model.supportedTasks.length === 0)
+        return true;
+    return model.supportedTasks.includes(task);
+}
+/**
+ * Lowercase, with every separator collapsed to a single space, so a query
+ * word matches across the punctuation a model id happens to use:
+ * `black-forest-labs/FLUX.1-schnell` → `black forest labs flux 1 schnell`.
+ */
+function searchable(text) {
+    return text
+        .toLowerCase()
+        .replace(/[^a-z0-9]+/g, " ")
+        .trim();
+}
+function queryWords(query) {
+    const words = searchable(query).split(" ").filter(Boolean);
+    return words;
+}
+/** Every word of the query appears somewhere in the model's id or name. */
+function queryMatch(model, words) {
+    if (words.length === 0)
+        return true;
+    const haystack = ` ${searchable(`${model.id} ${model.name}`)} `;
+    return words.every((w) => haystack.includes(w));
+}
+function hintMatch(model, hints) {
+    if (hints.size === 0)
+        return false;
+    const id = searchable(model.id);
+    for (const hint of hints) {
+        const h = searchable(hint);
+        if (h && (id === h || id.includes(h)))
+            return true;
+    }
+    return false;
+}
+/**
+ * The score ladder. Each tier outranks the sum of every tier below it, so the
+ * order between two candidates is a decision rather than an arithmetic
+ * accident:
+ *
+ * | tier | addend | when |
+ * |---|---:|---|
+ * | model hint | 2000 | the caller named this model |
+ * | provider hint | 1000 | the caller named this provider |
+ * | prefer_local | 500 | the caller asked for local and this runs locally |
+ * | leaderboard | 200 + 0…80 | the artifact ranks this model for the task |
+ * | recommended | 100 | `RECOMMENDED_MODELS` pins it and nothing ranks it |
+ * | downloaded | 30 | a local provider serves it |
+ *
+ * What the ladder decides: with no hint from the caller, the default pick is
+ * the model that tops the leaderboard for the task — `find_model`, and so
+ * `nodetool.models.pick`, answer with the best model for the job rather than
+ * with whichever one the static `RECOMMENDED_MODELS` list happens to pin.
+ * The pinned list still orders what the artifact says nothing about: language,
+ * embedding and ASR models, local models, and any model the rankings have
+ * never heard of.
+ *
+ * A caller's own preference still wins: naming a provider or a model, or
+ * asking for local, outranks any leaderboard position. Rankings order the
+ * field; they never overrule the caller.
+ */
+export const SCORE_TIERS = {
+    modelHint: 2000,
+    providerHint: 1000,
+    preferLocal: 500,
+    ranked: 200,
+    recommended: 100,
+    downloaded: 30,
+    /** Nudge against a hosted model when the caller asked for local. */
+    remote: -5
+};
+/** The most a leaderboard *position* adds, on top of {@link SCORE_TIERS}.ranked. */
+export const RANK_BONUS_MAX = 80;
+/**
+ * Capabilities the rankings artifact covers. Language, embedding and ASR
+ * models are unranked by design (§1 of the design doc), so a candidate of one
+ * of those kinds gets no best-task fallback — only an exact task match, which
+ * cannot happen for them either.
+ */
+const RANKED_CAPABILITIES = new Set([
+    "text_to_image",
+    "image_to_image",
+    "text_to_video",
+    "image_to_video",
+    "text_to_speech",
+    "text_to_music"
+]);
+/**
+ * The leaderboard row to report for a candidate: the requested task's, or —
+ * when the caller named no task and the capability is one rankings cover —
+ * the task where the model's `score` is highest.
+ */
+export function pickTaskRank(entry, task, allowBestTask) {
+    if (task) {
+        const exact = entry.tasks?.[task];
+        return exact ? { task, rank: exact } : null;
+    }
+    if (!allowBestTask)
+        return null;
+    let best = null;
+    for (const [name, rank] of Object.entries(entry.tasks ?? {})) {
+        if (!best || rank.score > best.rank.score) {
+            best = { task: name, rank };
+        }
+    }
+    return best;
+}
+/**
+ * Score addend and answer fields for one candidate route. Pure in the
+ * artifact so tests drive it with a fixture instead of mocking the module.
+ */
+export function rankCandidate(provider, modelId, task, allowBestTask, artifact = modelRankings) {
+    const entry = getModelRank(provider, modelId, artifact);
+    if (!entry)
+        return { bonus: 0, fields: {} };
+    const fields = {};
+    if (entry.canonical) {
+        fields.canonical = entry.canonical;
+        const alternates = routesFor(entry.canonical, artifact)
+            .filter((r) => r.provider !== provider || r.modelId !== modelId)
+            .map((r) => ({ provider: r.provider, model_id: r.modelId }));
+        if (alternates.length > 0)
+            fields.alternate_routes = alternates;
+    }
+    const picked = pickTaskRank(entry, task, allowBestTask);
+    if (!picked)
+        return { bonus: 0, fields };
+    fields.ranked_task = picked.task;
+    fields.rank = picked.rank.rank;
+    fields.of = picked.rank.of;
+    return {
+        bonus: Math.round(picked.rank.normalized * RANK_BONUS_MAX),
+        fields
+    };
+}
+/**
+ * One candidate's score and the ranking fields to attach to it. Additive and
+ * pure in the artifact, so the ordering between an explicit hint and a
+ * leaderboard position is pinned by a unit test with a fixture rather than by
+ * reading the constants.
+ */
+export function scoreCandidate(input, artifact = modelRankings) {
+    const { providerId, model } = input;
+    const downloaded = LOCAL_PROVIDER_IDS.has(providerId);
+    const ranking = rankCandidate(providerId, model.id, input.task, input.rankedCapability, artifact);
+    const ranked = ranking.fields.rank !== undefined;
+    let score = 0;
+    if (downloaded)
+        score += SCORE_TIERS.downloaded;
+    if (input.providerHint && providerId === input.providerHint) {
+        score += SCORE_TIERS.providerHint;
+    }
+    if (hintMatch(model, input.modelHints))
+        score += SCORE_TIERS.modelHint;
+    if (input.preferLocal) {
+        score += downloaded ? SCORE_TIERS.preferLocal : SCORE_TIERS.remote;
+    }
+    if (ranked) {
+        // The leaderboard decides among the models it covers, so a candidate it
+        // ranks does not also take the recommended bonus: +100 on top of a 0…80
+        // span would let the static list reorder the leaderboard, which is the
+        // blindness this term exists to fix.
+        score += SCORE_TIERS.ranked + ranking.bonus;
+    }
+    else if (input.recommended) {
+        score += SCORE_TIERS.recommended;
+    }
+    return { score, downloaded, fields: ranking.fields };
+}
+const findModel = {
+    spec: findModelSpec,
+    impl: async (run, params) => {
+        const capability = params["capability"];
+        if (!capability || !SUPPORTED_CAPABILITIES.includes(capability)) {
+            return {
+                capability,
+                total: 0,
+                results: [],
+                error: `capability must be one of: ${SUPPORTED_CAPABILITIES.join(", ")}`
+            };
+        }
+        const task = isString(params["task"])
+            ? params["task"]
+            : undefined;
+        const query = isString(params["query"])
+            ? params["query"]
+            : undefined;
+        const providerHint = isString(params["provider_hint"])
+            ? params["provider_hint"]
+            : undefined;
+        const modelHintRaw = params["model_hint"];
+        const modelHints = new Set(isString(modelHintRaw)
+            ? [modelHintRaw]
+            : Array.isArray(modelHintRaw)
+                ? modelHintRaw.filter((x) => isString(x))
+                : []);
+        const preferLocal = params["prefer_local"] === true;
+        const limit = isNumber(params["limit"]) && params["limit"] > 0
+            ? Math.floor(params["limit"])
+            : 5;
+        const providerEntries = Object.entries(providersOf(run));
+        if (providerEntries.length === 0) {
+            return {
+                capability,
+                total: 0,
+                results: [],
+                note: "No providers configured. Use nodetool.agents.Agent without a model property for AI work."
+            };
+        }
+        const recommendedSet = getRecommendedSet(capability);
+        const candidates = [];
+        const unavailable = [];
+        for (const [providerId, instance] of providerEntries) {
+            let supports;
+            try {
+                supports = instance
+                    .getCapabilities()
+                    .includes(capability);
+            }
+            catch {
+                continue;
+            }
+            if (!supports)
+                continue;
+            const blocked = await unavailableReasonOf(instance);
+            if (blocked) {
+                unavailable.push(`${providerId} (${blocked})`);
+                continue;
+            }
+            let models;
+            try {
+                models = await fetchModelsForCapability(instance, capability);
+            }
+            catch {
+                continue;
+            }
+            for (const m of models) {
+                candidates.push({ providerId, model: m });
+            }
+        }
+        // Drop the models that cannot do what the capability asks before anything
+        // ranks or filters them: an `image_to_video` endpoint is not an answer to
+        // a `text_to_video` search, however well its name matches the query.
+        const forTask = forCapabilityTask(candidates, capability);
+        const capabilityCandidates = forTask.kept;
+        const notes = [];
+        if (unavailable.length > 0) {
+            notes.push(`Skipped providers that cannot run here: ${unavailable.join(", ")}.`);
+        }
+        if (forTask.dropped > 0 && capabilityCandidates.length === 0) {
+            notes.push(`${forTask.dropped} model(s) were found but none declares ${capability}.`);
+        }
+        let words = queryWords(query ?? "");
+        // The `task` filter reads `supportedTasks`. A caller who typed a model name
+        // into it used to get an empty list — or, when no model declares tasks at
+        // all, the unfiltered default ranking — and no way to tell why. A task no
+        // model declares is read as a name search instead.
+        const declaredTasks = new Set();
+        for (const { model } of capabilityCandidates) {
+            for (const t of model.supportedTasks ?? [])
+                declaredTasks.add(t);
+        }
+        let pool;
+        // The task the rank term reads: the caller's, else the one the capability
+        // itself asks for. A `task` no model declares was a name search all along,
+        // so it falls back to the capability's task rather than to no task — which
+        // would let the leaderboard score a candidate on an unrelated task.
+        const defaultRankTask = capabilityTask(capability) ?? undefined;
+        let rankTask = task ?? defaultRankTask;
+        if (task && !declaredTasks.has(task)) {
+            pool = capabilityCandidates;
+            rankTask = defaultRankTask;
+            words = [...words, ...queryWords(task)];
+            notes.push(`No model declares task '${task}'; searched model names for it instead. Use \`query\` to search by name.`);
+        }
+        else {
+            pool = capabilityCandidates.filter(({ model }) => taskMatch(model, task));
+        }
+        let queryMatched;
+        if (words.length > 0) {
+            const matched = pool.filter(({ model }) => queryMatch(model, words));
+            queryMatched = matched.length > 0;
+            if (matched.length > 0) {
+                pool = matched;
+            }
+            else {
+                notes.push(`No model name matched '${words.join(" ")}'. Showing the ranked models for ${capability} instead.`);
+            }
+        }
+        const collected = pool.map(({ providerId, model }) => {
+            const recommended = recommendedSet.has(`${providerId}::${model.id}`);
+            const { score, downloaded, fields } = scoreCandidate({
+                providerId,
+                model,
+                recommended,
+                task: rankTask,
+                providerHint,
+                modelHints,
+                preferLocal,
+                rankedCapability: RANKED_CAPABILITIES.has(capability)
+            });
+            const result = {
+                provider: providerId,
+                model_id: model.id,
+                id: model.id,
+                name: model.name,
+                downloaded,
+                recommended,
+                score,
+                ref: modelRef(capability, providerId, model.id, model.name),
+                ...fields
+            };
+            // Absent, not null, when no shipped guide covers the line: a row that
+            // carries the key at all reads as "there is a guide" to a model
+            // skimming the answer.
+            const promptingSkill = promptingSkillFor(model.id);
+            if (promptingSkill)
+                result.prompting_skill = promptingSkill;
+            return result;
+        });
+        collected.sort((a, b) => {
+            if (b.score !== a.score)
+                return b.score - a.score;
+            if (a.provider !== b.provider)
+                return a.provider.localeCompare(b.provider);
+            return a.model_id.localeCompare(b.model_id);
+        });
+        const results = bestRoutePerModel(collected).slice(0, limit);
+        const answer = {
+            capability,
+            total: collected.length,
+            results
+        };
+        const best = results[0];
+        if (best !== undefined) {
+            answer.ref = best.ref;
+            // Lifted for the same reason `ref` is: callers read the top level and
+            // never dig into `results[0]`, and a prompting guide nobody notices is
+            // a guide nobody reads.
+            if (best.prompting_skill)
+                answer.prompting_skill = best.prompting_skill;
+        }
+        if (queryMatched !== undefined)
+            answer.query_matched = queryMatched;
+        if (notes.length > 0)
+            answer.note = notes.join(" ");
+        return answer;
+    }
+};
+/**
+ * One row per canonical model: its best-scoring route, in the order the sort
+ * left them. Without this a top-5 for `text_to_image` is five routes to the
+ * same leaderboard leader (atlascloud's two endpoints, kie's, openai's) and
+ * the caller sees one model. The routes that drop out are not lost — each
+ * survivor names them in `alternate_routes`.
+ *
+ * A candidate the artifact does not group carries no `canonical` and is never
+ * collapsed: two unranked models are two models.
+ */
+function bestRoutePerModel(results) {
+    const seen = new Set();
+    return results.filter((result) => {
+        if (!result.canonical)
+            return true;
+        if (seen.has(result.canonical))
+            return false;
+        seen.add(result.canonical);
+        return true;
+    });
+}
+/** Names an agent is likely to guess, mapped onto the canonical type. */
+const MODEL_TYPE_ALIASES = {
+    llm: "language",
+    text: "language",
+    language_model: "language",
+    text_generation: "language",
+    image_model: "image",
+    video_model: "video",
+    // The capability names `find_model` takes — a caller who knows one surface
+    // should not have to learn the other's vocabulary to browse.
+    generate_message: "language",
+    text_to_image: "image",
+    image_to_image: "image",
+    text_to_video: "video",
+    image_to_video: "video",
+    text_to_music: "music",
+    generate_embedding: "embedding",
+    speech: "tts",
+    text_to_speech: "tts",
+    audio: "tts",
+    transcription: "asr",
+    automatic_speech_recognition: "asr",
+    embeddings: "embedding"
+};
+/** The capability a provider must report before we ask it for this type. */
+const TYPE_CAPABILITY = {
+    language: "generate_message",
+    image: "text_to_image",
+    video: "text_to_video",
+    tts: "text_to_speech",
+    music: "text_to_music",
+    audio_to_audio: "audio_to_audio",
+    asr: "automatic_speech_recognition",
+    embedding: "generate_embedding"
+};
+async function fetchModelsOfType(provider, type) {
+    switch (type) {
+        case "language":
+            return await provider.getAvailableLanguageModels();
+        case "image":
+            return await provider.getAvailableImageModels();
+        case "video":
+            return await provider.getAvailableVideoModels();
+        case "tts":
+            return await provider.getAvailableTTSModels();
+        case "music":
+            return await provider.getAvailableMusicModels();
+        case "audio_to_audio":
+            return await provider.getAvailableAudioToAudioModels();
+        case "asr":
+            return await provider.getAvailableASRModels();
+        case "embedding":
+            return await provider.getAvailableEmbeddingModels();
+    }
+}
+function normalizeModelType(raw) {
+    if (raw === undefined || raw === null || raw === "")
+        return null;
+    if (!isString(raw))
+        return "invalid";
+    const key = raw.trim().toLowerCase();
+    if (MODEL_TYPES.includes(key))
+        return key;
+    return MODEL_TYPE_ALIASES[key] ?? "invalid";
+}
+const listModels = {
+    spec: listModelsSpec,
+    impl: async (run, params) => {
+        const providers = providersOf(run);
+        const modelType = normalizeModelType(params["model_type"]);
+        if (modelType === "invalid") {
+            return {
+                total: 0,
+                results: [],
+                error: `model_type must be one of: ${MODEL_TYPES.join(", ")}`
+            };
+        }
+        const providerFilterRaw = params["provider"];
+        const providerFilter = isNonBlankString(providerFilterRaw) &&
+            providerFilterRaw.trim().toLowerCase() !== "all"
+            ? providerFilterRaw.trim()
+            : undefined;
+        const downloadedOnly = params["downloaded_only"] === true;
+        const limit = isNumber(params["limit"]) && params["limit"] > 0
+            ? Math.floor(params["limit"])
+            : 50;
+        const entries = Object.entries(providers).filter(([id]) => !providerFilter || id === providerFilter);
+        if (entries.length === 0) {
+            const configured = Object.keys(providers).sort();
+            return {
+                total: 0,
+                results: [],
+                note: providerFilter
+                    ? `Provider '${providerFilter}' is not configured. Configured providers: ${configured.length > 0 ? configured.join(", ") : "none"}.`
+                    : "No providers are configured. Add an API key in Settings → Models & Providers."
+            };
+        }
+        const wantedTypes = modelType ? [modelType] : [...MODEL_TYPES];
+        const collected = [];
+        const unavailable = [];
+        for (const [providerId, instance] of entries) {
+            const downloaded = LOCAL_PROVIDER_IDS.has(providerId);
+            if (downloadedOnly && !downloaded)
+                continue;
+            let capabilities;
+            try {
+                capabilities = instance.getCapabilities();
+            }
+            catch {
+                continue;
+            }
+            const blocked = await unavailableReasonOf(instance);
+            if (blocked) {
+                unavailable.push(`${providerId} (${blocked})`);
+                continue;
+            }
+            for (const type of wantedTypes) {
+                if (!capabilities.includes(TYPE_CAPABILITY[type]))
+                    continue;
+                let models;
+                try {
+                    models = await fetchModelsOfType(instance, type);
+                }
+                catch {
+                    // A provider that can't be reached (no key, server down) drops out
+                    // of the listing rather than failing the whole call.
+                    continue;
+                }
+                for (const model of models) {
+                    if (!model?.id)
+                        continue;
+                    collected.push({
+                        provider: providerId,
+                        model_id: model.id,
+                        id: model.id,
+                        name: model.name ?? model.id,
+                        type,
+                        downloaded
+                    });
+                }
+            }
+        }
+        collected.sort((a, b) => {
+            if (a.provider !== b.provider)
+                return a.provider.localeCompare(b.provider);
+            if (a.type !== b.type)
+                return a.type.localeCompare(b.type);
+            return a.model_id.localeCompare(b.model_id);
+        });
+        const answer = {
+            total: collected.length,
+            truncated: collected.length > limit,
+            results: collected.slice(0, limit)
+        };
+        if (unavailable.length > 0) {
+            answer.note = `Skipped providers that cannot run here: ${unavailable.join(", ")}.`;
+        }
+        return answer;
+    }
+};
+// ---------------------------------------------------------------------------
+// list_provider_models
+// ---------------------------------------------------------------------------
+const listProviderModels = {
+    spec: listProviderModelsSpec,
+    impl: async (run, params) => {
+        const providerId = params["provider"];
+        if (!isString(providerId)) {
+            return { success: false, error: "provider must be a string" };
+        }
+        const provider = providersOf(run)[providerId];
+        if (!provider) {
+            return { success: false, error: `Unknown provider: ${providerId}` };
+        }
+        if (!isFunction(provider.getAvailableLanguageModels)) {
+            return {
+                success: false,
+                error: `Provider ${providerId} does not support model listing`
+            };
+        }
+        try {
+            const models = await provider.getAvailableLanguageModels();
+            return { success: true, provider: providerId, models };
+        }
+        catch (e) {
+            return {
+                success: false,
+                error: `Failed to list models: ${e instanceof Error ? e.message : String(e)}`
+            };
+        }
+    }
+};
+/** Every model capability, in the order `getAllMcpTools` offered them. */
+export const MODEL_CAPABILITIES = [
+    findModel,
+    listModels,
+    listProviderModels
+];
+export const module = {
+    module: "models",
+    exports: MODEL_CAPABILITIES
+};
+export { findModel, listModels, listProviderModels };
+//# sourceMappingURL=models.js.map