@burtson-labs/stealth-core-runtime 1.4.7

package/dist/runtime/modelCapabilities.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"modelCapabilities.d.ts","sourceRoot":"","sources":["../../src/runtime/modelCapabilities.ts"],"names":[],"mappings":"AAEA;;;;;GAKG;AAEH,MAAM,MAAM,SAAS,GAAG,OAAO,GAAG,QAAQ,GAAG,OAAO,CAAC;AAErD,MAAM,WAAW,iBAAiB;IAChC,mDAAmD;IACnD,aAAa,EAAE,MAAM,CAAC;IACtB,kFAAkF;IAClF,gBAAgB,EAAE,OAAO,CAAC;IAC1B,+DAA+D;IAC/D,mBAAmB,EAAE,OAAO,CAAC;IAC7B;;;;OAIG;IACH,cAAc,EAAE,OAAO,CAAC;IACxB,+DAA+D;IAC/D,IAAI,EAAE,SAAS,CAAC;IAChB,2CAA2C;IAC3C,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAkOD;;;GAGG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,iBAAiB,GAAG,IAAI,CAExF;AAED;;;GAGG;AACH,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAWvE;AAeD;;;;;;GAMG;AACH,wBAAsB,4BAA4B,CAChD,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,MAAM,GACd,OAAO,CAAC,iBAAiB,GAAG,IAAI,CAAC,CA2CnC;AAED;;;;GAIG;AACH,MAAM,WAAW,kBAAkB;IACjC;;;oCAGgC;IAChC,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,sEAAsE;IACtE,gBAAgB,EAAE,MAAM,CAAC;IACzB;;;;;wEAKoE;IACpE,WAAW,EAAE,OAAO,CAAC;IACrB;uEACmE;IACnE,iBAAiB,EAAE,MAAM,CAAC;CAC3B;AAED;;;;;;;;;;;GAWG;AACH,wBAAsB,wBAAwB,CAC5C,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,MAAM,EACf,gBAAgB,EAAE,MAAM,GACvB,OAAO,CAAC,kBAAkB,CAAC,CAuC7B;AAED;;GAEG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,SAAS,GAAG,MAAM,CAM3D;AAED;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,IAAI,EAAE,iBAAiB,GAAG,MAAM,CAGrE;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,iBAAiB,GAAG,MAAM,CAOpE;AAED;;;;;;;;;;;;;GAaG;AACH,MAAM,WAAW,oBAAoB;IACnC,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB;;;;;;OAMG;IACH,KAAK,CAAC,EAAE,OAAO,CAAC;CACjB;AAED,wBAAgB,2BAA2B,CAAC,OAAO,EAAE,MAAM,GAAG,oBAAoB,CA+CjF"}

package/dist/runtime/modelCapabilities.js

@@ -0,0 +1,458 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.registerModelCapabilities = registerModelCapabilities;
+exports.getModelCapabilities = getModelCapabilities;
+exports.queryOllamaModelCapabilities = queryOllamaModelCapabilities;
+exports.checkOllamaLoadedContext = checkOllamaLoadedContext;
+exports.getContextFileLimit = getContextFileLimit;
+exports.getContextTokenBudget = getContextTokenBudget;
+exports.getOutputTokenBudget = getOutputTokenBudget;
+exports.resolveOllamaRuntimeOptions = resolveOllamaRuntimeOptions;
+const modelBehavior_1 = require("./modelBehavior");
+/**
+ * Confirmed model inventory as of April 2026.
+ *
+ * Mac local (localhost:11434):
+ * bandit-core:4b/12b-it-qat, gemma3:4b/12b-it-qat, llama3.1, nomic-embed-text
+ *
+ * RTX 5090 node (ollamaNodeUrl):
+ * gemma4:31b, bandit-core:27b-it-qat, gemma3:27b-it-qat, qwen2.5:7b,
+ * deepseek-coder:6.7b, llama3 (NOT 3.1 — 8k ctx), nomic-embed-text
+ */
+const BUILT_IN_PROFILES = [
+    // ── bandit-core fine-tuned variants (text + code only, no vision) ─────────
+    {
+        prefix: 'bandit-core:4b',
+        caps: { contextWindow: 8192, supportsJsonMode: true, supportsToolCalling: false, supportsVision: false, tier: 'small', label: 'Bandit Core 4B' }
+    },
+    // flipped supportsToolCalling on the Gemma-3-derived
+    // bandit-core variants ≥12B. Without it, the loop's `nativeTools`
+    // gate at extension.ts:3445 short-circuits and the model gets only
+    // the system prompt's XML tool-call instructions, which Gemma 3 does
+    // not reliably emit (the `<tool_call>{...}</tool_call>` envelope is
+    // a Qwen-flavored convention not in Gemma's training distribution).
+    // Result bandit-core-1 produced prose only on a
+    // self-eval, zero tool calls. Modern Ollama serves Gemma 3 ≥12B with
+    // a tool-calling chat template; the cloud path's serializeBanditPayload
+    // already forwards tools through to upstream Ollama. 4B stays off
+    // because tool calling is unreliable on small Gemmas.
+    {
+        prefix: 'bandit-core:12b',
+        caps: { contextWindow: 32768, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'medium', label: 'Bandit Core 12B' }
+    },
+    {
+        prefix: 'bandit-core:27b',
+        caps: { contextWindow: 32768, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'medium', label: 'Bandit Core 27B' }
+    },
+    {
+        prefix: 'bandit-core:31b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'large', label: 'Bandit Core 31B' }
+    },
+    // Hosted bandit models (via GatewayApi / api.burtson.ai)
+    {
+        prefix: 'bandit-core-1',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'large', label: 'Bandit Core (hosted, 31B)' }
+    },
+    {
+        prefix: 'bandit-core-2',
+        caps: { contextWindow: 131072, supportsJsonMode: false, supportsToolCalling: false, supportsVision: false, tier: 'large', label: 'Bandit Core 2 (RunPod 70B)' }
+    },
+    {
+        // Gateway alias for Qwen 3.6 27B (repoint from Qwen 2.5 Coder 32B
+        // 2026-04-23 — 2.5 Coder is a code-completion tune, 3.6 is
+        // explicitly agent-trained with 256K context and native multimodal).
+        // Native tool calling is the whole reason we expose it — the
+        // capability profile here must match the underlying model, not
+        // bandit-core-1's text-path profile.
+        prefix: 'bandit-logic',
+        caps: { contextWindow: 262144, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'large', label: 'Bandit Logic (Qwen 3.6 27B)' }
+    },
+    // ── gemma3 — vision capable (all sizes) ──────────────────────────────────
+    // gemma3 supports image input via Ollama's `images: string[]` field.
+    // Mac local: gemma3:4b, gemma3:12b-it-qat. RTX 5090 node: gemma3:27b-it-qat.
+    // These are the primary vision models — no additional pull needed.
+    {
+        prefix: 'gemma3:4b',
+        caps: { contextWindow: 8192, supportsJsonMode: true, supportsToolCalling: false, supportsVision: true, tier: 'small', label: 'Gemma 3 4B' }
+    },
+    {
+        prefix: 'gemma3:12b',
+        caps: { contextWindow: 32768, supportsJsonMode: true, supportsToolCalling: false, supportsVision: true, tier: 'medium', label: 'Gemma 3 12B' }
+    },
+    {
+        prefix: 'gemma3:27b',
+        caps: { contextWindow: 32768, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'medium', label: 'Gemma 3 27B' }
+    },
+    // ── gemma4 — enhanced vision + 128k context (all sizes) ──────────────────
+    // Most-specific prefixes win (longer prefix matched first). The
+    // `e2b` / `e4b` variants are "effective" parameter-count tunes that
+    // sit at the small-tier boundary — that bare
+    // `gemma4:e4b` was being mis-routed to `large` by the catchall
+    // below, getting num_ctx=32768 and outputBudgetTokens=8192 on a 4B
+    // model. The agent then wandered on open-ended creative asks
+    // ("add visual flare") and never committed to a tool call. These
+    // explicit small-tier profiles fix that without affecting the 26B+
+    // sizes.
+    {
+        prefix: 'gemma4:e2b',
+        caps: { contextWindow: 16384, supportsJsonMode: true, supportsToolCalling: false, supportsVision: true, tier: 'small', label: 'Gemma 4 e2B' }
+    },
+    {
+        prefix: 'gemma4:e4b',
+        caps: { contextWindow: 16384, supportsJsonMode: true, supportsToolCalling: false, supportsVision: true, tier: 'small', label: 'Gemma 4 e4B' }
+    },
+    {
+        prefix: 'gemma4:31b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'large', label: 'Gemma 4 31B' }
+    },
+    {
+        prefix: 'gemma4:26b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'medium', label: 'Gemma 4 26B' }
+    },
+    {
+        // Catchall for unknown gemma4 sizes. Defaults to `medium` rather
+        // than `large` — the previous `large` default flagged any
+        // unrecognised size (including the 4B effective tune above
+        // before its explicit profile was added) into the heaviest
+        // budget. Medium is a safer guess; explicit large variants
+        // (31b+) match their own prefix above first.
+        prefix: 'gemma4',
+        caps: { contextWindow: 32768, supportsJsonMode: true, supportsToolCalling: false, supportsVision: true, tier: 'medium', label: 'Gemma 4' }
+    },
+    // ── Llama 3.x (text only) ─────────────────────────────────────────────────
+    {
+        prefix: 'llama3.2-vision',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: false, supportsVision: true, tier: 'medium', label: 'Llama 3.2 Vision' }
+    },
+    {
+        prefix: 'llama3.1',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'medium', label: 'Llama 3.1' }
+    },
+    {
+        prefix: 'llama3',
+        caps: { contextWindow: 8192, supportsJsonMode: true, supportsToolCalling: false, supportsVision: false, tier: 'small', label: 'Llama 3' }
+    },
+    // ── Qwen 3.6 — agent-trained, vision-capable, 256K context ───────────────
+    // Released April 2026 with explicit focus on "agentic coding" and
+    // "repository-level reasoning" — a genuine upgrade over Qwen 2.5 Coder
+    // for tool-use workflows (2.5 Coder is a code-completion tune, not an
+    // agent tune). 27B fits RTX 5090 at ~17GB + ample KV room; 35B is
+    // tighter (~24GB + trim KV). Multimodal: image paste should go through
+    // native vision rather than falling back to OCR.
+    {
+        prefix: 'qwen3.6:35b',
+        caps: { contextWindow: 262144, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'large', label: 'Qwen 3.6 35B' }
+    },
+    {
+        prefix: 'qwen3.6:27b',
+        caps: { contextWindow: 262144, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'large', label: 'Qwen 3.6 27B' }
+    },
+    {
+        prefix: 'qwen3.6',
+        caps: { contextWindow: 262144, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'large', label: 'Qwen 3.6' }
+    },
+    // ── Qwen 2.5 Coder (text + code, no vision) ───────────────────────────────
+    // Most specific prefixes first (longer match wins before shorter prefix).
+    {
+        prefix: 'qwen2.5-coder:72b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'large', label: 'Qwen 2.5 Coder 72B' }
+    },
+    {
+        prefix: 'qwen2.5-coder:32b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'large', label: 'Qwen 2.5 Coder 32B' }
+    },
+    {
+        prefix: 'qwen2.5-coder:14b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'medium', label: 'Qwen 2.5 Coder 14B' }
+    },
+    {
+        prefix: 'qwen2.5-coder',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'medium', label: 'Qwen 2.5 Coder' }
+    },
+    // ── Qwen 2.5 Vision (VL variants) — supportsVision ───────────────────────
+    {
+        prefix: 'qwen2.5vl',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'medium', label: 'Qwen 2.5 VL' }
+    },
+    {
+        prefix: 'qwen2-vl',
+        caps: { contextWindow: 32768, supportsJsonMode: true, supportsToolCalling: true, supportsVision: true, tier: 'medium', label: 'Qwen 2 VL' }
+    },
+    // ── Qwen 2.5 base (non-coder, non-VL) — text only ────────────────────────
+    // qwen2.5:7b confirmed on RTX 5090 node. Not a vision model.
+    {
+        prefix: 'qwen2.5:72b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'large', label: 'Qwen 2.5 72B' }
+    },
+    {
+        prefix: 'qwen2.5:32b',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'large', label: 'Qwen 2.5 32B' }
+    },
+    {
+        prefix: 'qwen2.5',
+        caps: { contextWindow: 131072, supportsJsonMode: true, supportsToolCalling: true, supportsVision: false, tier: 'medium', label: 'Qwen 2.5' }
+    },
+    // ── LLaVA — dedicated vision models (optional pull) ──────────────────────
+    {
+        prefix: 'llava',
+        caps: { contextWindow: 4096, supportsJsonMode: false, supportsToolCalling: false, supportsVision: true, tier: 'small', label: 'LLaVA' }
+    },
+    // ── Small / fast ─────────────────────────────────────────────────────────
+    {
+        prefix: 'deepseek-coder:6.7b',
+        caps: { contextWindow: 16384, supportsJsonMode: true, supportsToolCalling: false, supportsVision: false, tier: 'small', label: 'DeepSeek Coder 6.7B' }
+    },
+];
+const DEFAULT_CAPABILITIES = {
+    contextWindow: 8192,
+    supportsJsonMode: false,
+    supportsToolCalling: false,
+    supportsVision: false,
+    tier: 'small'
+};
+/**
+ * Runtime cache for dynamically discovered model capabilities (e.g. Ollama /api/show).
+ * Used ONLY when no built-in profile matches the model ID. Built-in profiles are
+ * hand-tuned (correct `supportsToolCalling`, `tier`, etc.); the auto-detector
+ * can only infer tier and context-window safely. Letting the cache override
+ * built-ins silently downgraded every known model the moment auto-detection
+ * ran at boot — bandit-logic lost its `supportsToolCalling: true`, nativeTools
+ * gated off, tools were never sent to Ollama, the qwen3.5 parser EOF'd on
+ * text-envelope output, and the fallback couldn't recover.
+ */
+const runtimeCapabilitiesCache = new Map();
+/**
+ * Register runtime-discovered capabilities for a model ID. Only takes effect
+ * when getModelCapabilities() finds no built-in match — explicit profiles win.
+ */
+function registerModelCapabilities(modelId, caps) {
+    if (modelId)
+        runtimeCapabilitiesCache.set(modelId.toLowerCase(), caps);
+}
+/**
+ * Returns capability profile for a given model ID.
+ * Check order: (1) built-in prefix profiles, (2) runtime cache, (3) default.
+ */
+function getModelCapabilities(modelId) {
+    if (!modelId)
+        return DEFAULT_CAPABILITIES;
+    const lower = modelId.toLowerCase();
+    for (const { prefix, caps } of BUILT_IN_PROFILES) {
+        if (lower.startsWith(prefix.toLowerCase())) {
+            return caps;
+        }
+    }
+    const cached = runtimeCapabilitiesCache.get(lower);
+    if (cached)
+        return cached;
+    return DEFAULT_CAPABILITIES;
+}
+/**
+ * Derives a model tier from a parameter-size string returned by Ollama /api/show.
+ * Examples: "4B" → small, "8B" / "12B" / "27B" → medium, "70B" / "72B" → large.
+ */
+function deriveTierFromParamSize(paramSize) {
+    const match = paramSize.match(/(\d+(?:\.\d+)?)\s*[Bb]/);
+    if (!match)
+        return 'small';
+    const billions = parseFloat(match[1]);
+    if (billions <= 5)
+        return 'small';
+    if (billions <= 35)
+        return 'medium';
+    return 'large';
+}
+/**
+ * Queries Ollama /api/show to auto-detect capabilities for a model not in BUILT_IN_PROFILES.
+ * Returns a partial ModelCapabilities object (tier + contextWindow) on success, null on failure.
+ * Silently returns null when Ollama is unreachable or the model is not installed.
+ *
+ * Call this once per model switch and persist via registerModelCapabilities().
+ */
+async function queryOllamaModelCapabilities(modelId, baseUrl) {
+    try {
+        const url = `${baseUrl.replace(/\/$/, '')}/api/show`;
+        const response = await fetch(url, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({ name: modelId }),
+            signal: AbortSignal.timeout(5000)
+        });
+        if (!response.ok)
+            return null;
+        const data = await response.json();
+        const paramSize = data.details?.parameter_size ?? '';
+        const tier = deriveTierFromParamSize(paramSize);
+        // Prefer the model's own declared context length when available.
+        const contextWindow = data.model_info?.['llm.context_length']
+            ?? (tier === 'large' ? 131072 : tier === 'medium' ? 32768 : 8192);
+        const family = (data.details?.family ?? '').toLowerCase();
+        // Trust Ollama's advertised capabilities when present. Hard-coding false
+        // here silently downgraded every tool-calling model that hit auto-detection
+        // before the built-in profile precedence fix landed. Even with that fix,
+        // models without a built-in profile (a new tag, a user-pulled variant)
+        // should report tools accurately when the runtime says so.
+        const caps = new Set((data.capabilities ?? []).map((c) => c.toLowerCase()));
+        const supportsJsonMode = true; // All Ollama-served models support format:"json"
+        const supportsToolCalling = caps.has('tools');
+        const supportsVision = caps.has('vision') ||
+            family.includes('llava') || family.includes('vision') || family.includes('vl');
+        return {
+            tier,
+            contextWindow,
+            supportsJsonMode,
+            supportsToolCalling,
+            supportsVision,
+            label: modelId
+        };
+    }
+    catch {
+        return null;
+    }
+}
+/**
+ * Query Ollama's `/api/ps` for the currently-loaded context length of
+ * `modelId` and compare against our requested num_ctx. Returns a check
+ * the host can render as a one-time tip. Silently returns underweight=
+ * false on any network/parse failure (this is a UX hint, never an error).
+ *
+ * Why /api/ps and not /api/show: /api/show returns the model's NATIVE
+ * context window from the GGUF metadata — that's the maximum the model
+ * could be loaded with, not what's actually loaded right now. /api/ps
+ * lists currently-resident models and reports the loaded context_length
+ * directly. That's the value that determines whether prompts overflow.
+ */
+async function checkOllamaLoadedContext(baseUrl, modelId, requestedContext) {
+    // Suggest a target ~25% above the requested floor so the env var
+    // covers our normal num_ctx + a small headroom for KV growth.
+    const suggested = Math.max(16384, Math.ceil(requestedContext * 1.1 / 1024) * 1024);
+    const fallback = {
+        loadedContext: null,
+        requestedContext,
+        underweight: false,
+        suggestionCommand: `OLLAMA_CONTEXT_LENGTH=${suggested} ollama serve`
+    };
+    try {
+        const url = `${baseUrl.replace(/\/$/, '')}/api/ps`;
+        const response = await fetch(url, { signal: AbortSignal.timeout(3000) });
+        if (!response.ok)
+            return fallback;
+        const data = await response.json();
+        const lowerId = modelId.toLowerCase();
+        const match = (data.models ?? []).find(m => (m.model ?? '').toLowerCase() === lowerId ||
+            (m.name ?? '').toLowerCase() === lowerId);
+        if (!match || typeof match.context_length !== 'number') {
+            return fallback;
+        }
+        const loaded = match.context_length;
+        // Underweight: loaded < 8K (absolute floor below which agent prompts
+        // get truncated) AND loaded < 75% of requested (a 4K env-var clamp
+        // when we asked for 12K — the canonical first-install gotcha).
+        const underweight = loaded < 8192 && loaded < requestedContext * 0.75;
+        return {
+            loadedContext: loaded,
+            requestedContext,
+            underweight,
+            suggestionCommand: `OLLAMA_CONTEXT_LENGTH=${suggested} ollama serve`
+        };
+    }
+    catch {
+        return fallback;
+    }
+}
+/**
+ * Returns the maximum number of context file chunks to include based on tier.
+ */
+function getContextFileLimit(tier) {
+    switch (tier) {
+        case 'large': return 20;
+        case 'medium': return 8;
+        case 'small': return 3;
+    }
+}
+/**
+ * Returns the approximate token budget reserved for context (not prompt/response).
+ */
+function getContextTokenBudget(caps) {
+    // Reserve ~40% of context window for injected file content.
+    return Math.floor(caps.contextWindow * 0.4);
+}
+/**
+ * Approximate per-turn output token budget.
+ *
+ * The tool-use loop uses this to decide whether a model's planned batch of
+ * write/edit calls will exceed the assistant turn's safe output capacity.
+ * When the planned content would push past this number, the loop serialises
+ * execution so each call gets its own iteration — protecting smaller models
+ * from generating malformed JSON in the tail of a multi-file emission, and
+ * giving the user a chance to react between approvals.
+ *
+ * Tier defaults are conservative — they are NOT the model's hard ceiling,
+ * they are the point past which output coherence starts to slip on local
+ * models. A capable hosted model effectively has no binding budget here.
+ *
+ * Tier mapping:
+ * small → 1024 tokens (4B params, ~4 KB of generated content)
+ * medium → 2048 tokens (12B–27B params, ~8 KB)
+ * large → 8192 tokens (31B+ params, hosted strong models)
+ *
+ * Override at runtime by passing `outputBudgetTokens` to the loop directly,
+ * or by registering a tuned profile via `registerModelCapabilities`.
+ */
+function getOutputTokenBudget(caps) {
+    switch (caps.tier) {
+        case 'small': return 1024;
+        case 'medium': return 2048;
+        case 'large': return 8192;
+        default: return 1024;
+    }
+}
+function resolveOllamaRuntimeOptions(modelId) {
+    const caps = getModelCapabilities(modelId);
+    // num_ctx is what we REQUEST from Ollama (not the model's native
+    // maximum — those are routinely 128K+ and would blow RAM). Pick a
+    // value that fits our framework prompt + several tool rounds + a
+    // reasonable user turn on commodity hardware. Tiers measured on:
+    // - small (4B): fits laptop iGPU / Apple M-series w/o swap.
+    // - medium (12B/27B): 16 GB VRAM laptops (MBP Q4).
+    // - large (31B+): RTX 5090 class or cluster; Modelfile can still
+    // override via its own num_ctx parameter.
+    // Anchoring here: the 8-iteration agent runs we saw in April 2026
+    // hitting `hitLimit: true` from accumulated tool results on num_ctx
+    // 16k. Tiers below give the loop ~24k-32k tokens of room before
+    // compaction needs to kick in.
+    let num_ctx;
+    switch (caps.tier) {
+        case 'small':
+            num_ctx = 12288; // 4B — bump from 8k; still fits ~8 GB VRAM
+            break;
+        case 'medium':
+            num_ctx = 24576; // 12B/27B — 24k room for normal agent turns
+            break;
+        case 'large':
+            num_ctx = 32768; // 31B+ — matches cluster bandit-core:31b Modelfile
+            break;
+        default:
+            num_ctx = 12288;
+    }
+    // Clamp to the model's declared native window so we never ask for
+    // more than the model can actually honor.
+    if (caps.contextWindow > 0 && num_ctx > caps.contextWindow) {
+        num_ctx = caps.contextWindow;
+    }
+    // Thinking-mode default now comes from behavior profiles rather than
+    // scattered model-name checks. Users can still override via `/think`
+    // (CLI) or banditStealth.thinkingMode (extension).
+    const behavior = (0, modelBehavior_1.getModelBehaviorProfile)(modelId);
+    const thinkDefault = behavior.prompting.thinking === 'on'
+        ? true
+        : behavior.prompting.thinking === 'off'
+            ? false
+            : undefined;
+    return {
+        num_ctx,
+        keep_alive: -1,
+        ...(thinkDefault !== undefined ? { think: thinkDefault } : {})
+    };
+}
+//# sourceMappingURL=modelCapabilities.js.map

package/dist/runtime/modelCapabilities.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"modelCapabilities.js","sourceRoot":"","sources":["../../src/runtime/modelCapabilities.ts"],"names":[],"mappings":";;AAkQA,8DAEC;AAMD,oDAWC;AAsBD,oEA8CC;AAuCD,4DA2CC;AAKD,kDAMC;AAKD,sDAGC;AAwBD,oDAOC;AA6BD,kEA+CC;AAziBD,mDAA0D;AA8B1D;;;;;;;;;GASG;AACH,MAAM,iBAAiB,GAAuD;IAC5E,6EAA6E;IAC7E;QACE,MAAM,EAAE,gBAAgB;QACxB,IAAI,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,gBAAgB,EAAE;KACjJ;IACD,qDAAqD;IACrD,kEAAkE;IAClE,mEAAmE;IACnE,qEAAqE;IACrE,oEAAoE;IACpE,oEAAoE;IACpE,gDAAgD;IAChD,qEAAqE;IACrE,wEAAwE;IACxE,kEAAkE;IAClE,sDAAsD;IACtD;QACE,MAAM,EAAE,iBAAiB;QACzB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,iBAAiB,EAAE;KACnJ;IACD;QACE,MAAM,EAAE,iBAAiB;QACzB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,iBAAiB,EAAE;KACnJ;IACD;QACE,MAAM,EAAE,iBAAiB;QACzB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,iBAAiB,EAAE;KAClJ;IACD,yDAAyD;IACzD;QACE,MAAM,EAAE,eAAe;QACvB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,2BAA2B,EAAE;KAC5J;IACD;QACE,MAAM,EAAE,eAAe;QACvB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,KAAK,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,4BAA4B,EAAE;KAChK;IACD;QACE,kEAAkE;QAClE,2DAA2D;QAC3D,qEAAqE;QACrE,6DAA6D;QAC7D,+DAA+D;QAC/D,qCAAqC;QACrC,MAAM,EAAE,cAAc;QACtB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,6BAA6B,EAAE;KAC9J;IAED,4EAA4E;IAC5E,qEAAqE;IACrE,6EAA6E;IAC7E,mEAAmE;IACnE;QACE,MAAM,EAAE,WAAW;QACnB,IAAI,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,YAAY,EAAE;KAC5I;IACD;QACE,MAAM,EAAE,YAAY;QACpB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE;KAC/I;IACD;QACE,MAAM,EAAE,YAAY;QACpB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE;KAC9I;IAED,4EAA4E;IAC5E,gEAAgE;IAChE,oEAAoE;IACpE,6CAA6C;IAC7C,+DAA+D;IAC/D,mEAAmE;IACnE,6DAA6D;IAC7D,iEAAiE;IACjE,mEAAmE;IACnE,SAAS;IACT;QACE,MAAM,EAAE,YAAY;QACpB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE;KAC9I;IACD;QACE,MAAM,EAAE,YAAY;QACpB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE;KAC9I;IACD;QACE,MAAM,EAAE,YAAY;QACpB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE;KAC9I;IACD;QACE,MAAM,EAAE,YAAY;QACpB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE;KAC/I;IACD;QACE,iEAAiE;QACjE,0DAA0D;QAC1D,2DAA2D;QAC3D,2DAA2D;QAC3D,2DAA2D;QAC3D,6CAA6C;QAC7C,MAAM,EAAE,QAAQ;QAChB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,SAAS,EAAE;KAC3I;IAED,6EAA6E;IAC7E;QACE,MAAM,EAAE,iBAAiB;QACzB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,kBAAkB,EAAE;KACrJ;IACD;QACE,MAAM,EAAE,UAAU;QAClB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,WAAW,EAAE;KAC9I;IACD;QACE,MAAM,EAAE,QAAQ;QAChB,IAAI,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE;KAC1I;IAED,4EAA4E;IAC5E,kEAAkE;IAClE,uEAAuE;IACvE,sEAAsE;IACtE,kEAAkE;IAClE,uEAAuE;IACvE,iDAAiD;IACjD;QACE,MAAM,EAAE,aAAa;QACrB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,cAAc,EAAE;KAC/I;IACD;QACE,MAAM,EAAE,aAAa;QACrB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,cAAc,EAAE;KAC/I;IACD;QACE,MAAM,EAAE,SAAS;QACjB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE;KAC3I;IAED,6EAA6E;IAC7E,0EAA0E;IAC1E;QACE,MAAM,EAAE,mBAAmB;QAC3B,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,oBAAoB,EAAE;KACtJ;IACD;QACE,MAAM,EAAE,mBAAmB;QAC3B,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,oBAAoB,EAAE;KACtJ;IACD;QACE,MAAM,EAAE,mBAAmB;QAC3B,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,oBAAoB,EAAE;KACvJ;IACD;QACE,MAAM,EAAE,eAAe;QACvB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,gBAAgB,EAAE;KACnJ;IAED,4EAA4E;IAC5E;QACE,MAAM,EAAE,WAAW;QACnB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE;KAC/I;IACD;QACE,MAAM,EAAE,UAAU;QAClB,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,WAAW,EAAE;KAC5I;IAED,4EAA4E;IAC5E,6DAA6D;IAC7D;QACE,MAAM,EAAE,aAAa;QACrB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,cAAc,EAAE;KAChJ;IACD;QACE,MAAM,EAAE,aAAa;QACrB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,cAAc,EAAE;KAChJ;IACD;QACE,MAAM,EAAE,SAAS;QACjB,IAAI,EAAE,EAAE,aAAa,EAAE,MAAM,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,UAAU,EAAE;KAC7I;IAED,4EAA4E;IAC5E;QACE,MAAM,EAAE,OAAO;QACf,IAAI,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,gBAAgB,EAAE,KAAK,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE;KACxI;IAED,4EAA4E;IAC5E;QACE,MAAM,EAAE,qBAAqB;QAC7B,IAAI,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,qBAAqB,EAAE;KACvJ;CACF,CAAC;AAEF,MAAM,oBAAoB,GAAsB;IAC9C,aAAa,EAAE,IAAI;IACnB,gBAAgB,EAAE,KAAK;IACvB,mBAAmB,EAAE,KAAK;IAC1B,cAAc,EAAE,KAAK;IACrB,IAAI,EAAE,OAAO;CACd,CAAC;AAEF;;;;;;;;;GASG;AACH,MAAM,wBAAwB,GAAG,IAAI,GAAG,EAA6B,CAAC;AAEtE;;;GAGG;AACH,SAAgB,yBAAyB,CAAC,OAAe,EAAE,IAAuB;IAChF,IAAI,OAAO;QAAE,wBAAwB,CAAC,GAAG,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,IAAI,CAAC,CAAC;AACzE,CAAC;AAED;;;GAGG;AACH,SAAgB,oBAAoB,CAAC,OAAe;IAClD,IAAI,CAAC,OAAO;QAAE,OAAO,oBAAoB,CAAC;IAC1C,MAAM,KAAK,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC;IACpC,KAAK,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,iBAAiB,EAAE,CAAC;QACjD,IAAI,KAAK,CAAC,UAAU,CAAC,MAAM,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC;YAC3C,OAAO,IAAI,CAAC;QACd,CAAC;IACH,CAAC;IACD,MAAM,MAAM,GAAG,wBAAwB,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;IACnD,IAAI,MAAM;QAAE,OAAO,MAAM,CAAC;IAC1B,OAAO,oBAAoB,CAAC;AAC9B,CAAC;AAED;;;GAGG;AACH,SAAS,uBAAuB,CAAC,SAAiB;IAChD,MAAM,KAAK,GAAG,SAAS,CAAC,KAAK,CAAC,wBAAwB,CAAC,CAAC;IACxD,IAAI,CAAC,KAAK;QAAE,OAAO,OAAO,CAAC;IAC3B,MAAM,QAAQ,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IACtC,IAAI,QAAQ,IAAI,CAAC;QAAE,OAAO,OAAO,CAAC;IAClC,IAAI,QAAQ,IAAI,EAAE;QAAE,OAAO,QAAQ,CAAC;IACpC,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;;;;;GAMG;AACI,KAAK,UAAU,4BAA4B,CAChD,OAAe,EACf,OAAe;IAEf,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,GAAG,OAAO,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,WAAW,CAAC;QACrD,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,GAAG,EAAE;YAChC,MAAM,EAAE,MAAM;YACd,OAAO,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE;YAC/C,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC;YACvC,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,IAAI,CAAC;SAClC,CAAC,CAAC;QACH,IAAI,CAAC,QAAQ,CAAC,EAAE;YAAE,OAAO,IAAI,CAAC;QAC9B,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,EAI/B,CAAC;QACF,MAAM,SAAS,GAAG,IAAI,CAAC,OAAO,EAAE,cAAc,IAAI,EAAE,CAAC;QACrD,MAAM,IAAI,GAAG,uBAAuB,CAAC,SAAS,CAAC,CAAC;QAChD,iEAAiE;QACjE,MAAM,aAAa,GAAG,IAAI,CAAC,UAAU,EAAE,CAAC,oBAAoB,CAAC;eACxD,CAAC,IAAI,KAAK,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QACpE,MAAM,MAAM,GAAG,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;QAC1D,yEAAyE;QACzE,4EAA4E;QAC5E,yEAAyE;QACzE,uEAAuE;QACvE,2DAA2D;QAC3D,MAAM,IAAI,GAAG,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,YAAY,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC;QAC5E,MAAM,gBAAgB,GAAG,IAAI,CAAC,CAAC,iDAAiD;QAChF,MAAM,mBAAmB,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;QAC9C,MAAM,cAAc,GAClB,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC;YAClB,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QACjF,OAAO;YACL,IAAI;YACJ,aAAa;YACb,gBAAgB;YAChB,mBAAmB;YACnB,cAAc;YACd,KAAK,EAAE,OAAO;SACf,CAAC;IACJ,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,IAAI,CAAC;IACd,CAAC;AACH,CAAC;AA2BD;;;;;;;;;;;GAWG;AACI,KAAK,UAAU,wBAAwB,CAC5C,OAAe,EACf,OAAe,EACf,gBAAwB;IAExB,iEAAiE;IACjE,8DAA8D;IAC9D,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,gBAAgB,GAAG,GAAG,GAAG,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC;IACnF,MAAM,QAAQ,GAAuB;QACnC,aAAa,EAAE,IAAI;QACnB,gBAAgB;QAChB,WAAW,EAAE,KAAK;QAClB,iBAAiB,EAAE,yBAAyB,SAAS,eAAe;KACrE,CAAC;IACF,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,GAAG,OAAO,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,SAAS,CAAC;QACnD,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,GAAG,EAAE,EAAE,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACzE,IAAI,CAAC,QAAQ,CAAC,EAAE;YAAE,OAAO,QAAQ,CAAC;QAClC,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,EAE/B,CAAC;QACF,MAAM,OAAO,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC;QACtC,MAAM,KAAK,GAAG,CAAC,IAAI,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CACzC,CAAC,CAAC,CAAC,KAAK,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,KAAK,OAAO;YACzC,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,KAAK,OAAO,CACzC,CAAC;QACF,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,CAAC,cAAc,KAAK,QAAQ,EAAE,CAAC;YACvD,OAAO,QAAQ,CAAC;QAClB,CAAC;QACD,MAAM,MAAM,GAAG,KAAK,CAAC,cAAc,CAAC;QACpC,qEAAqE;QACrE,mEAAmE;QACnE,+DAA+D;QAC/D,MAAM,WAAW,GAAG,MAAM,GAAG,IAAI,IAAI,MAAM,GAAG,gBAAgB,GAAG,IAAI,CAAC;QACtE,OAAO;YACL,aAAa,EAAE,MAAM;YACrB,gBAAgB;YAChB,WAAW;YACX,iBAAiB,EAAE,yBAAyB,SAAS,eAAe;SACrE,CAAC;IACJ,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,QAAQ,CAAC;IAClB,CAAC;AACH,CAAC;AAED;;GAEG;AACH,SAAgB,mBAAmB,CAAC,IAAe;IACjD,QAAQ,IAAI,EAAE,CAAC;QACb,KAAK,OAAO,CAAC,CAAE,OAAO,EAAE,CAAC;QACzB,KAAK,QAAQ,CAAC,CAAC,OAAO,CAAC,CAAC;QACxB,KAAK,OAAO,CAAC,CAAE,OAAO,CAAC,CAAC;IAC1B,CAAC;AACH,CAAC;AAED;;GAEG;AACH,SAAgB,qBAAqB,CAAC,IAAuB;IAC3D,4DAA4D;IAC5D,OAAO,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,aAAa,GAAG,GAAG,CAAC,CAAC;AAC9C,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,SAAgB,oBAAoB,CAAC,IAAuB;IAC1D,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QAClB,KAAK,OAAO,CAAC,CAAE,OAAO,IAAI,CAAC;QAC3B,KAAK,QAAQ,CAAC,CAAC,OAAO,IAAI,CAAC;QAC3B,KAAK,OAAO,CAAC,CAAE,OAAO,IAAI,CAAC;QAC3B,OAAO,CAAC,CAAO,OAAO,IAAI,CAAC;IAC7B,CAAC;AACH,CAAC;AA6BD,SAAgB,2BAA2B,CAAC,OAAe;IACzD,MAAM,IAAI,GAAG,oBAAoB,CAAC,OAAO,CAAC,CAAC;IAC3C,iEAAiE;IACjE,kEAAkE;IAClE,iEAAiE;IACjE,iEAAiE;IACjE,4DAA4D;IAC5D,mDAAmD;IACnD,iEAAiE;IACjE,0CAA0C;IAC1C,kEAAkE;IAClE,oEAAoE;IACpE,gEAAgE;IAChE,+BAA+B;IAC/B,IAAI,OAAe,CAAC;IACpB,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QAClB,KAAK,OAAO;YACV,OAAO,GAAG,KAAK,CAAC,CAAG,2CAA2C;YAC9D,MAAM;QACR,KAAK,QAAQ;YACX,OAAO,GAAG,KAAK,CAAC,CAAG,4CAA4C;YAC/D,MAAM;QACR,KAAK,OAAO;YACV,OAAO,GAAG,KAAK,CAAC,CAAG,mDAAmD;YACtE,MAAM;QACR;YACE,OAAO,GAAG,KAAK,CAAC;IACpB,CAAC;IACD,kEAAkE;IAClE,0CAA0C;IAC1C,IAAI,IAAI,CAAC,aAAa,GAAG,CAAC,IAAI,OAAO,GAAG,IAAI,CAAC,aAAa,EAAE,CAAC;QAC3D,OAAO,GAAG,IAAI,CAAC,aAAa,CAAC;IAC/B,CAAC;IACD,qEAAqE;IACrE,qEAAqE;IACrE,mDAAmD;IACnD,MAAM,QAAQ,GAAG,IAAA,uCAAuB,EAAC,OAAO,CAAC,CAAC;IAClD,MAAM,YAAY,GAAG,QAAQ,CAAC,SAAS,CAAC,QAAQ,KAAK,IAAI;QACvD,CAAC,CAAC,IAAI;QACN,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC,QAAQ,KAAK,KAAK;YACrC,CAAC,CAAC,KAAK;YACP,CAAC,CAAC,SAAS,CAAC;IAChB,OAAO;QACL,OAAO;QACP,UAAU,EAAE,CAAC,CAAC;QACd,GAAG,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC/D,CAAC;AACJ,CAAC"}

package/dist/runtime/modelsDevCatalog.d.ts

@@ -0,0 +1,34 @@
+/**
+ * models.dev catalog adapter.
+ *
+ * https://models.dev/api.json is a community-maintained model metadata
+ * catalog (MIT-licensed, used by sst/opencode). Each provider entry
+ * lists every published model with context window, output cap,
+ * tool-calling support, vision support, and pricing. We use it to
+ * populate the runtime capability registry for openai-compatible
+ * upstreams (LM Studio, llama.cpp, vLLM, OpenRouter, Together, Groq,
+ * DeepSeek, OpenAI proper) that aren't in our hand-curated built-in
+ * profile list — without it those models inherit the conservative
+ * default and miss out on tier-tuned context budgets, output gates,
+ * vision routing, etc.
+ *
+ * Caching:
+ *   - In-memory map for the lifetime of the process.
+ *   - Disk cache at `~/.bandit/cache/models-dev.json` + an `.etag`
+ *     sidecar; revalidated against the server with `If-None-Match`
+ *     once per 24h. 304 Not Modified keeps the on-disk JSON.
+ *   - First-run fetch downloads ~250 KB gzipped (~1.9 MB raw).
+ *     Subsequent launches usually return 304 in <100 ms.
+ *
+ * Failures are silent: when the catalog is unreachable or the model
+ * isn't listed, lookupModelsDevCapabilities returns null and the
+ * caller falls back to whatever default it had.
+ */
+import type { ModelCapabilities } from './modelCapabilities';
+/**
+ * Resolve capabilities for `(baseUrl, modelId)` from the models.dev
+ * catalog. Returns null when the catalog is unavailable, the base URL
+ * doesn't match any known provider, or the model isn't listed.
+ */
+export declare function queryModelsDevCapabilities(modelId: string, baseUrl: string): Promise<ModelCapabilities | null>;
+//# sourceMappingURL=modelsDevCatalog.d.ts.map

package/dist/runtime/modelsDevCatalog.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"modelsDevCatalog.d.ts","sourceRoot":"","sources":["../../src/runtime/modelsDevCatalog.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAKH,OAAO,KAAK,EAAE,iBAAiB,EAAa,MAAM,qBAAqB,CAAC;AA2LxE;;;;GAIG;AACH,wBAAsB,0BAA0B,CAC9C,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,MAAM,GACd,OAAO,CAAC,iBAAiB,GAAG,IAAI,CAAC,CAmBnC"}