@anvaka/vue-llm 0.3.0 → 0.3.2

package/README.md

@@ -7,6 +7,7 @@ Browser-only LLM client + Vue 3 plugin, provider adapters, and lightweight compo
 - LocalStorage-based config store (custom storage adapter supported)
 - Streaming + promise requests via `llmClient.stream()`
 - Normalized usage + USD cost on every response (override built-in rates per app or per model)
+- Automatic prompt caching for Claude (Anthropic + Bedrock) — caches the system+tools prefix, plus the rolling conversation in agent loops; opt out with `promptCache: false`
 - Vue plugin for dependency injection
 - `useLLM()` composable with reactive streaming state
 - Ready-to-use components: `ProviderSelector`, `LLMConfigModal`, `StoredKeysManager`
@@ -157,6 +158,22 @@ const { messages, usage, cost } = await client.runAgentLoop({
 })
 ```
 
+### Prompt caching (Claude)
+
+Anthropic and Bedrock requests are sent with `cache_control` markers by default, so repeated prefixes are read from cache instead of re-billed at full input price. Two prefixes are tagged:
+
+- **System + tools** — on every Claude request. The static prefix recurs identically across calls and runs within the cache TTL (~5 min).
+- **Rolling conversation** — added by `runAgentLoop` only, since it re-sends the whole growing transcript each turn. Iteration *N* reads iterations `1..N-1` from cache and only writes the new turn.
+
+Cache hits and writes surface in the usual `usage` fields (`cachedInputTokens`, `cacheCreationInputTokens`) and are priced via the `cachedInput` / `cacheCreation` rate keys. No setup is required — prefixes under the model's minimum cacheable length simply aren't cached (no error).
+
+Disable it per call or per provider config:
+
+```js
+client.stream({ messages: [...], promptCache: false })   // single request
+configStore.saveConfig('claude', { ...cfg, promptCache: false }) // all requests for this provider
+```
+
 ### Overriding rates
 
 Built-in rates are sourced from public pricing pages and *will* drift. Three ways to override, in priority order:

package/dist/core/LLMClient.js

@@ -120,16 +120,16 @@ class P {
     temperature: n,
     model: o,
     maxTokens: l,
-    enableThinking: p
+    enableThinking: d
   } = {}) {
     if (await this.ensureInitialized(), !this.provider.hasCapability("tools"))
       throw new Error("Configured provider does not support tools");
     const c = t.slice();
-    let d = 0, m = null;
+    let p = 0, m = null;
     const f = { inputTokens: 0, outputTokens: 0, totalTokens: 0, cachedInputTokens: 0, cacheCreationInputTokens: 0, reasoningTokens: 0 };
     let b = !1;
-    for (; d < e; ) {
-      d++, r && r({ type: "iter-start", iter: d });
+    for (; p < e; ) {
+      p++, r && r({ type: "iter-start", iter: p });
       const x = this.validateCapabilities({
         messages: c,
         tools: s,
@@ -137,7 +137,10 @@ class P {
         model: o || this.config.model,
         temperature: n ?? this.config.temperature,
         maxTokens: l ?? this.config.maxTokens ?? 4096,
-        enableThinking: p ?? !1,
+        enableThinking: d ?? !1,
+        // Re-sends the whole growing conversation each turn, so cache the
+        // rolling transcript prefix (Anthropic-family providers honor this).
+        cacheTranscript: !0,
         requestId: this.generateRequestId()
       }), h = await this.provider.streamRequest(c, x, (a) => {
         a.content && r && r({ type: "text-delta", text: a.content }), a.toolCallDelta && r && r({ type: "tool-call-delta", delta: a.toolCallDelta });
@@ -171,7 +174,7 @@ class P {
     }
     m || (m = "max-iters");
     const g = b ? f : null, v = this.costFor(g, { provider: this.config.provider, model: o || this.config.model });
-    return r && r({ type: "stop", reason: m, iterations: d, usage: g, cost: v }), { messages: c, iterations: d, stopReason: m, usage: g, cost: v };
+    return r && r({ type: "stop", reason: m, iterations: p, usage: g, cost: v }), { messages: c, iterations: p, stopReason: m, usage: g, cost: v };
   }
   generateRequestId() {
     return `${Date.now()}-${Math.random().toString(36).substr(2, 9)}`;
@@ -256,8 +259,8 @@ class $ {
     };
     try {
       if (this.targetNode) {
-        const p = r({ ...n, stream: !0 });
-        return await this._streamIntoTarget(e, p, s, i);
+        const d = r({ ...n, stream: !0 });
+        return await this._streamIntoTarget(e, d, s, i);
       }
       const l = r({ ...n, stream: !1 });
       return await this._promiseResponse(e, l, s, i);
@@ -275,8 +278,8 @@ class $ {
       throw new Error("Provider presets require a configured ConfigStore instance");
     const s = (e = (r = this.client).getConfigByName) == null ? void 0 : e.call(r, t);
     if (!s) {
-      const l = (((o = (n = this.client.configStore).getEnabledConfigs) == null ? void 0 : o.call(n)) || []).map((c) => c.name || `${c.provider} (${c.baseUrl})`).filter(Boolean), p = l.length ? ` Available presets: ${l.join(", ")}` : " No configured providers found.";
-      throw new Error(`Provider preset '${t}' not found.${p}`);
+      const l = (((o = (n = this.client.configStore).getEnabledConfigs) == null ? void 0 : o.call(n)) || []).map((c) => c.name || `${c.provider} (${c.baseUrl})`).filter(Boolean), d = l.length ? ` Available presets: ${l.join(", ")}` : " No configured providers found.";
+      throw new Error(`Provider preset '${t}' not found.${d}`);
     }
     const i = U(s.provider, s);
     return await i.initialize(), { provider: i, config: s, cleanup: () => {
@@ -298,9 +301,9 @@ class $ {
       }
   }
   async _streamIntoTarget(t, s, i, r) {
-    var d, m, f, b, g, v, x, h, q, y, C, w, k, a, R, T;
+    var p, m, f, b, g, v, x, h, q, y, C, w, k, a, R, T;
     let e, n;
-    typeof this.targetNode == "string" ? (n = this.targetNode, e = this.contextNode.createChild(n), (d = e.setSelected) == null || d.call(e)) : (e = this.targetNode, n = ((m = e.getText) == null ? void 0 : m.call(e)) || ""), this._applyAttributes(e);
+    typeof this.targetNode == "string" ? (n = this.targetNode, e = this.contextNode.createChild(n), (p = e.setSelected) == null || p.call(e)) : (e = this.targetNode, n = ((m = e.getText) == null ? void 0 : m.call(e)) || ""), this._applyAttributes(e);
     const o = r == null ? void 0 : r.name;
     o && e.addTag && e.addTag("provider", o), (f = e.addLLMLog) == null || f.call(e, this.operationName || "streaming-request", {
       provider: r == null ? void 0 : r.provider,
@@ -308,26 +311,26 @@ class $ {
       messages: t,
       options: s
     }), (b = e.setStreamingState) == null || b.call(e, !0), e._activeRequestId = s.requestId;
-    let l = "", p = null, c = null;
+    let l = "", d = null, c = null;
     try {
       await i.streamRequest(t, s, (_) => {
         var I, S;
-        _.content && (l = _.fullContent, (I = e.setText) == null || I.call(e, l)), _.fullUsage && (c = _.fullUsage), _.finishReason && (p = _.finishReason), _.done && ((S = e.setStreamingState) == null || S.call(e, !1), e._activeRequestId = null);
+        _.content && (l = _.fullContent, (I = e.setText) == null || I.call(e, l)), _.fullUsage && (c = _.fullUsage), _.finishReason && (d = _.finishReason), _.done && ((S = e.setStreamingState) == null || S.call(e, !1), e._activeRequestId = null);
       });
       const u = (v = (g = this.client).costFor) == null ? void 0 : v.call(g, c, { provider: r == null ? void 0 : r.provider, model: s.model });
-      return (x = e.addLLMLog) == null || x.call(e, (this.operationName || "streaming-request") + "-response", { requestId: s.requestId }, { content: l, usage: c, cost: u }), p === "length" ? ((h = e.setAttribute) == null || h.call(e, "_truncated", !0), (q = e.addLog) == null || q.call(e, "Response truncated. Increase Max Tokens.", "warn", { finish_reason: "length" })) : (y = e.setAttribute) == null || y.call(e, "_truncated", !1), await ((C = e.persistNow) == null ? void 0 : C.call(e)), e;
+      return (x = e.addLLMLog) == null || x.call(e, (this.operationName || "streaming-request") + "-response", { requestId: s.requestId }, { content: l, usage: c, cost: u }), d === "length" ? ((h = e.setAttribute) == null || h.call(e, "_truncated", !0), (q = e.addLog) == null || q.call(e, "Response truncated. Increase Max Tokens.", "warn", { finish_reason: "length" })) : (y = e.setAttribute) == null || y.call(e, "_truncated", !1), await ((C = e.persistNow) == null ? void 0 : C.call(e)), e;
     } catch (u) {
       throw (w = e.addLLMLog) == null || w.call(e, (this.operationName || "streaming-request") + "-error", { requestId: s.requestId }, null, u), (k = e.setStreamingState) == null || k.call(e, !1), e._activeRequestId = null, (a = e.setText) == null || a.call(e, `Error: ${u.message}`), (R = e.setAttribute) == null || R.call(e, "_truncated", !1), await ((T = e.persistNow) == null ? void 0 : T.call(e)), u;
     }
   }
   async _promiseResponse(t, s, i, r) {
-    var p, c, d, m, f, b, g;
+    var d, c, p, m, f, b, g;
     const e = i.prepareRequest(t, s), n = await i.makeRequest(e), o = i.processResponse(n);
-    if ((o.finishReason || ((c = (p = n == null ? void 0 : n.choices) == null ? void 0 : p[0]) == null ? void 0 : c.finish_reason) || null) === "length")
+    if ((o.finishReason || ((c = (d = n == null ? void 0 : n.choices) == null ? void 0 : d[0]) == null ? void 0 : c.finish_reason) || null) === "length")
       try {
         this.contextNode.setAttribute("_truncated", !0);
       } catch (v) {
-        (m = (d = this.client.logger) == null ? void 0 : d.warn) == null || m.call(d, "Unable to mark node as truncated", { error: v });
+        (m = (p = this.client.logger) == null ? void 0 : p.warn) == null || m.call(p, "Unable to mark node as truncated", { error: v });
       }
     else
       try {

package/dist/pricing/rates.js

@@ -1,69 +1,79 @@
 const t = {
   openai: {
-    "gpt-5": { input: 0.625, output: 5, cachedInput: 0.125 },
-    "gpt-5-mini": { input: 0.25, output: 2, cachedInput: 0.025 },
-    "gpt-5-nano": { input: 0.05, output: 0.4, cachedInput: 5e-3 },
+    // Flagship + workhorse (April–May 2026 lineup)
+    "gpt-5.5-pro": { input: 5, output: 30, cachedInput: 0.5 },
+    "gpt-5.5": { input: 5, output: 30, cachedInput: 0.5 },
+    "gpt-5.4-pro": { input: 30, output: 180 },
     "gpt-5.4": { input: 2.5, output: 15, cachedInput: 0.25 },
     "gpt-5.4-mini": { input: 0.75, output: 4.5, cachedInput: 0.075 },
     "gpt-5.4-nano": { input: 0.2, output: 1.25, cachedInput: 0.02 },
-    "gpt-4o": { input: 2.5, output: 10, cachedInput: 1.25 },
-    "gpt-4o-mini": { input: 0.15, output: 0.6, cachedInput: 0.075 },
-    "gpt-4.1": { input: 2, output: 8, cachedInput: 0.5 },
-    "gpt-4.1-mini": { input: 0.2, output: 0.8, cachedInput: 0.1 },
-    "gpt-4.1-nano": { input: 0.05, output: 0.2, cachedInput: 0.025 },
-    "gpt-4-turbo": { input: 10, output: 30 },
-    "gpt-4": { input: 30, output: 60 },
-    "gpt-3.5-turbo": { input: 0.5, output: 1 },
-    o1: { input: 15, output: 60, cachedInput: 7.5 },
-    "o1-mini": { input: 0.55, output: 2.2, cachedInput: 0.55 },
+    // GPT-5.x family — still widely deployed
+    "gpt-5.3": { input: 1.75, output: 14, cachedInput: 0.175 },
+    "gpt-5.2-pro": { input: 10.5, output: 84 },
+    "gpt-5.2": { input: 1.75, output: 14, cachedInput: 0.175 },
+    "gpt-5.1": { input: 0.625, output: 5, cachedInput: 0.125 },
+    "gpt-5-chat": { input: 1.25, output: 10, cachedInput: 0.125 },
+    "gpt-5": { input: 0.625, output: 5, cachedInput: 0.125 },
+    "gpt-5-mini": { input: 0.25, output: 2, cachedInput: 0.025 },
+    "gpt-5-nano": { input: 0.05, output: 0.4, cachedInput: 5e-3 },
+    // Reasoning models
+    "o4-mini": { input: 1.1, output: 4.4, cachedInput: 0.275 },
     o3: { input: 2, output: 8, cachedInput: 0.5 },
     "o3-mini": { input: 1.1, output: 4.4, cachedInput: 0.55 },
-    "o4-mini": { input: 1.1, output: 4.4, cachedInput: 0.275 }
+    // Codex variants
+    "gpt-5.1-codex": { input: 1.25, output: 10, cachedInput: 0.125 },
+    "gpt-5.1-codex-mini": { input: 0.25, output: 2, cachedInput: 0.025 },
+    "gpt-5-codex": { input: 1.25, output: 10, cachedInput: 0.125 }
   },
   anthropic: {
+    "claude-opus-4-8": { input: 5, output: 25, cachedInput: 0.5, cacheCreation: 6.25 },
     "claude-opus-4-7": { input: 5, output: 25, cachedInput: 0.5, cacheCreation: 6.25 },
     "claude-opus-4-6": { input: 5, output: 25, cachedInput: 0.5, cacheCreation: 6.25 },
     "claude-sonnet-4-6": { input: 3, output: 15, cachedInput: 0.3, cacheCreation: 3.75 },
     "claude-sonnet-4-5": { input: 3, output: 15, cachedInput: 0.3, cacheCreation: 3.75 },
-    "claude-haiku-4-5": { input: 1, output: 5, cachedInput: 0.1, cacheCreation: 1.25 },
-    "claude-3-5-sonnet": { input: 3, output: 15, cachedInput: 0.3, cacheCreation: 3.75 },
-    "claude-3-5-haiku": { input: 0.8, output: 4, cachedInput: 0.08, cacheCreation: 1 },
-    "claude-3-opus": { input: 15, output: 75 },
-    "claude-3-haiku": { input: 0.25, output: 1.25 }
+    "claude-haiku-4-5": { input: 1, output: 5, cachedInput: 0.1, cacheCreation: 1.25 }
   },
   // Bedrock charges the same per-token rates as Anthropic direct for the
   // Claude inference profiles; the model IDs differ (us.anthropic.* prefix).
   bedrock: {
+    "us.anthropic.claude-opus-4-8": { input: 5, output: 25, cachedInput: 0.5, cacheCreation: 6.25 },
     "us.anthropic.claude-opus-4-7": { input: 5, output: 25, cachedInput: 0.5, cacheCreation: 6.25 },
     "us.anthropic.claude-opus-4-6": { input: 5, output: 25, cachedInput: 0.5, cacheCreation: 6.25 },
     "us.anthropic.claude-sonnet-4-6": { input: 3, output: 15, cachedInput: 0.3, cacheCreation: 3.75 },
     "us.anthropic.claude-sonnet-4-5": { input: 3, output: 15, cachedInput: 0.3, cacheCreation: 3.75 },
-    "us.anthropic.claude-haiku-4-5": { input: 1, output: 5, cachedInput: 0.1, cacheCreation: 1.25 },
-    "us.anthropic.claude-opus-4-1": { input: 15, output: 75 }
+    "us.anthropic.claude-haiku-4-5": { input: 1, output: 5, cachedInput: 0.1, cacheCreation: 1.25 }
   },
   gemini: {
+    // Gemini 3.x — current lineup (May 2026)
+    "gemini-3.5-flash": { input: 1.5, output: 9, cachedInput: 0.15 },
+    "gemini-3.1-pro-preview": { input: 2, output: 12, cachedInput: 0.2 },
+    "gemini-3.1-flash-lite-preview": { input: 0.25, output: 1.5, cachedInput: 0.025 },
+    "gemini-3-pro-preview": { input: 2, output: 12 },
+    "gemini-3-flash-preview": { input: 0.5, output: 3, cachedInput: 0.05 },
+    // Gemini 2.5 — still in active use
     "gemini-2.5-pro": { input: 1, output: 10, cachedInput: 0.125 },
     "gemini-2.5-flash": { input: 0.3, output: 2.5, cachedInput: 0.03 },
-    "gemini-2.5-flash-lite": { input: 0.1, output: 0.4, cachedInput: 0.01 },
-    "gemini-2.0-flash": { input: 0.1, output: 0.4, cachedInput: 0.025 },
-    "gemini-2.0-flash-lite": { input: 0.075, output: 0.3 },
-    "gemini-1.5-pro": { input: 1.25, output: 5 },
-    "gemini-1.5-flash": { input: 0.075, output: 0.3 }
+    "gemini-2.5-flash-lite": { input: 0.1, output: 0.4, cachedInput: 0.01 }
   },
   grok: {
-    "grok-4": { input: 3, output: 15 },
+    // Grok 4.3 (April 30, 2026 release) — current primary, 1M context
+    "grok-4.3": { input: 1.25, output: 2.5, cachedInput: 0.2 },
+    "grok-4.20-0309-reasoning": { input: 1.25, output: 2.5, cachedInput: 0.2 },
+    "grok-4.20-0309-non-reasoning": { input: 1.25, output: 2.5, cachedInput: 0.2 },
+    "grok-4.20-multi-agent-0309": { input: 1.25, output: 2.5, cachedInput: 0.2 },
+    "grok-4.20": { input: 1.25, output: 2.5, cachedInput: 0.2 },
+    "grok-build-0.1": { input: 1, output: 2 },
+    // Cheap fast tier
     "grok-4-fast": { input: 0.2, output: 0.5 },
-    "grok-3": { input: 3, output: 15 },
-    "grok-3-fast": { input: 5, output: 25 },
-    "grok-3-mini": { input: 0.25, output: 0.5 },
-    "grok-code-fast-1": { input: 0.2, output: 1.5 },
-    "grok-2": { input: 2, output: 10 }
+    "grok-code-fast-1": { input: 0.2, output: 1.5 }
   },
   deepseek: {
-    "deepseek-chat": { input: 0.21, output: 0.79, cachedInput: 0.13 },
-    "deepseek-reasoner": { input: 0.55, output: 2 },
+    "deepseek-v4-pro": { input: 0.435, output: 0.87, cachedInput: 36e-4 },
     "deepseek-v4-flash": { input: 0.1, output: 0.2, cachedInput: 28e-4 },
-    "deepseek-v4-pro": { input: 0.435, output: 0.87, cachedInput: 36e-4 }
+    "deepseek-v3.2": { input: 0.252, output: 0.378, cachedInput: 0.0252 },
+    "deepseek-v3.2-thinking": { input: 0.252, output: 0.378, cachedInput: 0.0252 },
+    "deepseek-chat": { input: 0.014, output: 0.028 },
+    "deepseek-reasoner": { input: 0.55, output: 2 }
   },
   // OpenRouter prices vary per upstream model and include their own margin —
   // ship empty; callers can either register their own rates or use

package/dist/providers/AnthropicProvider.js

@@ -1,77 +1,86 @@
-import { BaseProvider as p } from "./BaseProvider.js";
-class y extends p {
+import { BaseProvider as g } from "./BaseProvider.js";
+class b extends g {
   async detectCapabilities() {
-    var e, t, n, o;
-    ((e = this.config.model) != null && e.includes("claude-3") || (t = this.config.model) != null && t.includes("claude-sonnet") || (n = this.config.model) != null && n.includes("claude-opus") || (o = this.config.model) != null && o.includes("claude-haiku")) && (this.capabilities.add("vision"), this.capabilities.add("tools"));
-  }
-  prepareRequest(e, t) {
-    const n = _(e), o = t.model || this.config.model || "claude-3-sonnet-20240229", i = {
-      model: o,
-      max_tokens: t.maxTokens || 1e3,
-      messages: n,
-      stream: t.stream || !1
+    var t, e, o, s;
+    ((t = this.config.model) != null && t.includes("claude-3") || (e = this.config.model) != null && e.includes("claude-sonnet") || (o = this.config.model) != null && o.includes("claude-opus") || (s = this.config.model) != null && s.includes("claude-haiku")) && (this.capabilities.add("vision"), this.capabilities.add("tools"));
+  }
+  prepareRequest(t, e) {
+    const o = m(t), s = e.model || this.config.model || "claude-3-sonnet-20240229", r = {
+      model: s,
+      max_tokens: e.maxTokens || 1e3,
+      messages: o,
+      stream: e.stream || !1
     };
-    o.includes("claude-opus-4-7") || (i.temperature = t.temperature ?? 0.7);
-    const c = e.find((r) => r.role === "system");
-    return c && (i.system = c.content), t.tools && this.capabilities.has("tools") && (i.tools = g(t.tools), t.tool_choice && (i.tool_choice = t.tool_choice)), i;
-  }
-  processMessages(e, t) {
-    return t.images && this.capabilities.has("vision") ? this.addImagesToMessages(e, t.images) : e;
-  }
-  addImagesToMessages(e, t) {
-    const n = e[e.length - 1];
-    if (n && n.role === "user") {
-      const o = [{ type: "text", text: n.content }];
-      t.forEach((i) => {
-        o.push({
+    _(s) || (r.temperature = e.temperature ?? 0.7);
+    const c = t.find((i) => i.role === "system");
+    return c && (r.system = c.content), e.tools && this.capabilities.has("tools") && (r.tools = y(e.tools), e.tool_choice && (r.tool_choice = e.tool_choice)), this.promptCachingEnabled(e) && k(r, e), r;
+  }
+  // Prompt caching is on by default for the Claude/Anthropic-family path
+  // (Bedrock inherits this method) because it's strictly cheaper and degrades
+  // gracefully — a prefix under the model's minimum cacheable length is simply
+  // not cached, with no error. Disable per-call with options.promptCache:false
+  // or per-config with config.promptCache:false.
+  promptCachingEnabled(t) {
+    var e;
+    return (t == null ? void 0 : t.promptCache) !== !1 && ((e = this.config) == null ? void 0 : e.promptCache) !== !1;
+  }
+  processMessages(t, e) {
+    return e.images && this.capabilities.has("vision") ? this.addImagesToMessages(t, e.images) : t;
+  }
+  addImagesToMessages(t, e) {
+    const o = t[t.length - 1];
+    if (o && o.role === "user") {
+      const s = [{ type: "text", text: o.content }];
+      e.forEach((r) => {
+        s.push({
           type: "image",
-          source: { type: "base64", media_type: "image/jpeg", data: typeof i == "string" ? i : i.data }
+          source: { type: "base64", media_type: "image/jpeg", data: typeof r == "string" ? r : r.data }
         });
-      }), n.content = o;
+      }), o.content = s;
     }
-    return e;
+    return t;
   }
-  processResponse(e) {
-    const t = l(e.stop_reason), n = e.content || [], o = n.find((r) => r.type === "text"), c = n.filter((r) => r.type === "tool_use").map((r) => ({ id: r.id, name: r.name, args: r.input || {} }));
+  processResponse(t) {
+    const e = h(t.stop_reason), o = t.content || [], s = o.find((i) => i.type === "text"), c = o.filter((i) => i.type === "tool_use").map((i) => ({ id: i.id, name: i.name, args: i.input || {} }));
     return {
-      content: (o == null ? void 0 : o.text) || "",
-      usage: d(e.usage),
-      finishReason: t,
+      content: (s == null ? void 0 : s.text) || "",
+      usage: f(t.usage),
+      finishReason: e,
       toolCalls: c
     };
   }
-  parseStreamingLine(e) {
-    if (!e.startsWith("data: ")) return null;
-    const t = e.slice(6).trim();
-    if (t === "[DONE]") return { done: !0 };
+  parseStreamingLine(t) {
+    if (!t.startsWith("data: ")) return null;
+    const e = t.slice(6).trim();
+    if (e === "[DONE]") return { done: !0 };
     try {
-      return JSON.parse(t);
+      return JSON.parse(e);
     } catch {
       return null;
     }
   }
-  extractStreamingContent(e) {
-    var t, n, o, i, c, r, h, f;
-    if (e.done) return { done: !0 };
-    if (e.type === "error") {
-      const a = ((t = e.error) == null ? void 0 : t.type) || "anthropic_error", u = new Error(((n = e.error) == null ? void 0 : n.message) || "Anthropic streaming error");
-      throw u.code = a, e.request_id && (u.requestId = e.request_id), u;
+  extractStreamingContent(t) {
+    var e, o, s, r, c, i, d, p;
+    if (t.done) return { done: !0 };
+    if (t.type === "error") {
+      const a = ((e = t.error) == null ? void 0 : e.type) || "anthropic_error", u = new Error(((o = t.error) == null ? void 0 : o.message) || "Anthropic streaming error");
+      throw u.code = a, t.request_id && (u.requestId = t.request_id), u;
     }
-    if (e.type === "message_start")
-      return { content: "", thinking: "", done: !1, usage: d((o = e.message) == null ? void 0 : o.usage), finishReason: null };
-    if (e.type === "content_block_start") {
-      const a = e.content_block;
+    if (t.type === "message_start")
+      return { content: "", thinking: "", done: !1, usage: f((s = t.message) == null ? void 0 : s.usage), finishReason: null };
+    if (t.type === "content_block_start") {
+      const a = t.content_block;
       return (a == null ? void 0 : a.type) === "tool_use" ? {
         content: "",
         done: !1,
-        toolCallDelta: { index: e.index, id: a.id, name: a.name, argsTextDelta: "" }
+        toolCallDelta: { index: t.index, id: a.id, name: a.name, argsTextDelta: "" }
       } : null;
     }
-    return e.type === "content_block_delta" ? ((i = e.delta) == null ? void 0 : i.type) === "text_delta" ? { content: ((c = e.delta) == null ? void 0 : c.text) || "", thinking: "", done: !1, usage: null, finishReason: null } : ((r = e.delta) == null ? void 0 : r.type) === "input_json_delta" ? {
+    return t.type === "content_block_delta" ? ((r = t.delta) == null ? void 0 : r.type) === "text_delta" ? { content: ((c = t.delta) == null ? void 0 : c.text) || "", thinking: "", done: !1, usage: null, finishReason: null } : ((i = t.delta) == null ? void 0 : i.type) === "input_json_delta" ? {
       content: "",
       done: !1,
-      toolCallDelta: { index: e.index, argsTextDelta: ((h = e.delta) == null ? void 0 : h.partial_json) || "" }
-    } : null : e.type === "message_delta" ? { content: "", thinking: "", done: !1, usage: d(e.usage), finishReason: l((f = e.delta) == null ? void 0 : f.stop_reason) } : e.type === "message_stop" ? { content: "", thinking: "", done: !0, usage: null, finishReason: l(e.stop_reason) } : null;
+      toolCallDelta: { index: t.index, argsTextDelta: ((d = t.delta) == null ? void 0 : d.partial_json) || "" }
+    } : null : t.type === "message_delta" ? { content: "", thinking: "", done: !1, usage: f(t.usage), finishReason: h((p = t.delta) == null ? void 0 : p.stop_reason) } : t.type === "message_stop" ? { content: "", thinking: "", done: !0, usage: null, finishReason: h(t.stop_reason) } : null;
   }
   getApiPath() {
     return "/v1/messages";
@@ -86,67 +95,107 @@ class y extends p {
     return this.config.apiKey;
   }
   buildHeaders() {
-    const e = super.buildHeaders();
-    return this.requiresAuth() && (e["anthropic-version"] = "2023-06-01", e["anthropic-dangerous-direct-browser-access"] = "true"), e;
+    const t = super.buildHeaders();
+    return this.requiresAuth() && (t["anthropic-version"] = "2023-06-01", t["anthropic-dangerous-direct-browser-access"] = "true"), t;
   }
   getModelsEndpoint() {
     return `${this.config.baseUrl}/v1/models`;
   }
-  parseModelsResponse(e) {
-    var t;
-    return ((t = e.data) == null ? void 0 : t.map((n) => n.id)) || [];
+  parseModelsResponse(t) {
+    var e;
+    return ((e = t.data) == null ? void 0 : e.map((o) => o.id)) || [];
   }
 }
-function l(s) {
-  if (!s) return null;
-  const e = String(s).toLowerCase();
-  return e === "max_tokens" ? "length" : e;
+function _(n) {
+  return /claude-opus-4-[789]/.test(n);
+}
+function h(n) {
+  if (!n) return null;
+  const t = String(n).toLowerCase();
+  return t === "max_tokens" ? "length" : t;
 }
-function d(s) {
-  if (!s) return null;
-  const e = s.input_tokens ?? 0, t = s.cache_read_input_tokens ?? 0, n = s.cache_creation_input_tokens ?? 0, i = s.input_tokens != null || s.cache_read_input_tokens != null || s.cache_creation_input_tokens != null ? e + t + n : 0, c = s.output_tokens ?? 0, r = { inputTokens: i, outputTokens: c, totalTokens: i + c, raw: s };
-  return s.cache_read_input_tokens != null && (r.cachedInputTokens = t), s.cache_creation_input_tokens != null && (r.cacheCreationInputTokens = n), r;
+function f(n) {
+  if (!n) return null;
+  const t = n.input_tokens ?? 0, e = n.cache_read_input_tokens ?? 0, o = n.cache_creation_input_tokens ?? 0, r = n.input_tokens != null || n.cache_read_input_tokens != null || n.cache_creation_input_tokens != null ? t + e + o : 0, c = n.output_tokens ?? 0, i = { inputTokens: r, outputTokens: c, totalTokens: r + c, raw: n };
+  return n.cache_read_input_tokens != null && (i.cachedInputTokens = e), n.cache_creation_input_tokens != null && (i.cacheCreationInputTokens = o), i;
 }
-function _(s) {
-  const e = [];
-  for (const t of s)
-    if (t.role !== "system") {
-      if (t.role === "assistant" && Array.isArray(t.tool_calls) && t.tool_calls.length) {
-        const n = [];
-        t.content && n.push({ type: "text", text: t.content });
-        for (const o of t.tool_calls)
-          n.push({
+function m(n) {
+  const t = [];
+  for (const e of n)
+    if (e.role !== "system") {
+      if (e.role === "assistant" && Array.isArray(e.tool_calls) && e.tool_calls.length) {
+        const o = [];
+        e.content && o.push({ type: "text", text: e.content });
+        for (const s of e.tool_calls)
+          o.push({
             type: "tool_use",
-            id: o.id,
-            name: o.name,
-            input: o.args || {}
+            id: s.id,
+            name: s.name,
+            input: s.args || {}
           });
-        e.push({ role: "assistant", content: n });
+        t.push({ role: "assistant", content: o });
         continue;
       }
-      if (t.role === "tool") {
-        e.push({
+      if (e.role === "tool") {
+        t.push({
           role: "user",
           content: [{
             type: "tool_result",
-            tool_use_id: t.tool_call_id,
-            content: String(t.content ?? "")
+            tool_use_id: e.tool_call_id,
+            content: String(e.content ?? "")
           }]
         });
         continue;
       }
-      e.push({ role: t.role, content: t.content });
+      t.push({ role: e.role, content: e.content });
+    }
+  return t;
+}
+function y(n) {
+  return n.map((t) => t.type === "function" && t.function ? {
+    name: t.function.name,
+    description: t.function.description,
+    input_schema: t.function.parameters || { type: "object", properties: {} }
+  } : t);
+}
+const l = { type: "ephemeral" };
+function k(n, t) {
+  if (n.system)
+    n.system = x(n.system);
+  else if (Array.isArray(n.tools) && n.tools.length) {
+    const e = n.tools.length - 1;
+    n.tools[e] = { ...n.tools[e], cache_control: l };
+  }
+  if (t != null && t.cacheTranscript) {
+    const e = n.messages;
+    if (Array.isArray(e) && e.length) {
+      const o = e[e.length - 1];
+      o.content = A(o.content);
     }
-  return e;
+  }
+  return n;
+}
+function x(n) {
+  let t;
+  if (typeof n == "string")
+    t = [{ type: "text", text: n }];
+  else if (Array.isArray(n) && n.length)
+    t = n.map((e) => typeof e == "string" ? { type: "text", text: e } : { ...e });
+  else
+    return n;
+  return t[t.length - 1] = { ...t[t.length - 1], cache_control: l }, t;
 }
-function g(s) {
-  return s.map((e) => e.type === "function" && e.function ? {
-    name: e.function.name,
-    description: e.function.description,
-    input_schema: e.function.parameters || { type: "object", properties: {} }
-  } : e);
+function A(n) {
+  if (typeof n == "string")
+    return [{ type: "text", text: n, cache_control: l }];
+  if (Array.isArray(n) && n.length) {
+    const t = n.slice(), e = t.length - 1, o = t[e];
+    return t[e] = typeof o == "string" ? { type: "text", text: o, cache_control: l } : { ...o, cache_control: l }, t;
+  }
+  return n;
 }
 export {
-  y as AnthropicProvider,
-  d as normalizeAnthropicUsage
+  b as AnthropicProvider,
+  f as normalizeAnthropicUsage,
+  _ as samplingParamsRemoved
 };

package/dist/providers/BedrockProvider.js

@@ -1,6 +1,7 @@
 import { BaseProvider as y } from "./BaseProvider.js";
 import { AnthropicProvider as w } from "./AnthropicProvider.js";
-const g = [
+const p = [
+  "us.anthropic.claude-opus-4-8",
   "us.anthropic.claude-opus-4-7",
   "us.anthropic.claude-sonnet-4-6",
   "us.anthropic.claude-haiku-4-5-20251001-v1:0",
@@ -9,10 +10,10 @@ const g = [
   "us.anthropic.claude-sonnet-4-5-20250929-v1:0",
   "us.anthropic.claude-opus-4-1-20250805-v1:0"
 ];
-class R extends w {
-  prepareRequest(r, n) {
-    const t = super.prepareRequest(r, n);
-    return delete t.stream, t.anthropic_version = "bedrock-2023-05-31", t;
+class A extends w {
+  prepareRequest(n, t) {
+    const r = super.prepareRequest(n, t);
+    return delete r.stream, r.anthropic_version = "bedrock-2023-05-31", r;
   }
   buildHeaders() {
     return y.prototype.buildHeaders.call(this);
@@ -26,147 +27,170 @@ class R extends w {
   requiresAuth() {
     return !!this.config.apiKey;
   }
-  async makeRequest(r, n, t) {
-    const e = r.model, d = { ...r };
+  async makeRequest(n, t, r) {
+    const e = n.model, d = { ...n };
     delete d.model;
-    const c = `${this.config.baseUrl}/model/${encodeURIComponent(e)}/invoke`, u = n ? { signal: n } : new AbortController();
-    n || (t = t || this.generateRequestId(), this.activeRequests.set(t, u));
+    const i = `${this.config.baseUrl}/model/${encodeURIComponent(e)}/invoke`, l = t ? { signal: t } : new AbortController();
+    t || (r = r || this.generateRequestId(), this.activeRequests.set(r, l));
     try {
-      const s = await fetch(c, {
+      const o = await fetch(i, {
         method: "POST",
         headers: this.buildHeaders(),
         body: JSON.stringify(d),
-        signal: u.signal || n
+        signal: l.signal || t
       });
-      if (!s.ok) {
-        const l = await s.text();
-        throw new Error(`Bedrock API Error (${s.status}): ${l}`);
+      if (!o.ok) {
+        const h = await o.text();
+        throw new Error(`Bedrock API Error (${o.status}): ${h}`);
       }
-      return s.json();
+      return o.json();
     } finally {
-      !n && t && this.activeRequests.delete(t);
+      !t && r && this.activeRequests.delete(r);
     }
   }
-  async makeStreamingRequest(r, n) {
-    const t = r.model, e = { ...r };
+  async makeStreamingRequest(n, t) {
+    const r = n.model, e = { ...n };
     delete e.model;
-    const d = `${this.config.baseUrl}/model/${encodeURIComponent(t)}/invoke-with-response-stream`, c = await fetch(d, {
+    const d = `${this.config.baseUrl}/model/${encodeURIComponent(r)}/invoke-with-response-stream`, i = await fetch(d, {
       method: "POST",
       headers: this.buildHeaders(),
       body: JSON.stringify(e),
-      signal: n
+      signal: t
     });
-    if (!c.ok) {
-      const i = await c.text();
-      throw new Error(`Bedrock API Error (${c.status}): ${i}`);
+    if (!i.ok) {
+      const c = await i.text();
+      throw new Error(`Bedrock API Error (${i.status}): ${c}`);
     }
-    const u = c.body.getReader(), s = new v(), l = new TextEncoder(), f = new ReadableStream({
-      async pull(i) {
+    const l = i.body.getReader(), o = new v(), h = new TextEncoder(), f = new ReadableStream({
+      async pull(c) {
         try {
-          const { done: a, value: m } = await u.read();
+          const { done: a, value: g } = await l.read();
           if (a) {
-            i.close();
+            c.close();
             return;
           }
-          const p = s.feed(m);
-          for (const h of p) {
-            if (h.kind === "error") {
-              i.error(new Error(h.message));
+          const m = o.feed(g);
+          for (const u of m) {
+            if (u.kind === "error") {
+              c.error(new Error(u.message));
               return;
             }
-            i.enqueue(l.encode(`data: ${JSON.stringify(h.payload)}
+            c.enqueue(h.encode(`data: ${JSON.stringify(u.payload)}
 `));
           }
         } catch (a) {
-          i.error(a);
+          c.error(a);
         }
       },
-      cancel(i) {
+      cancel(c) {
         try {
-          u.cancel(i);
+          l.cancel(c);
         } catch {
         }
       }
     });
     return new Response(f, { status: 200, headers: { "Content-Type": "text/event-stream" } });
   }
-  // Bedrock control plane needs SigV4; Bearer auth can't list models.
-  // Return the hardcoded list so the dropdown still populates.
-  async discoverModels() {
-    return g.slice();
+  // Live discovery is opt-in via config.liveModelDiscovery. A Bedrock API key
+  // authenticates against the control plane with Bearer auth (no SigV4), and
+  // the control-plane host returns permissive CORS headers, so a browser fetch
+  // works. When disabled (default) or on any failure, fall back to the static
+  // list — existing apps keep working with zero surprise network calls.
+  async discoverModels(n = 1e4) {
+    if (!this.config.liveModelDiscovery || !this.getModelsEndpoint())
+      return p.slice();
+    try {
+      const t = await super.discoverModels(n);
+      return t.length ? t : p.slice();
+    } catch {
+      return p.slice();
+    }
   }
+  // The control plane lives at the same host with the `-runtime` segment
+  // stripped: bedrock-runtime.{region}.amazonaws.com (invoke) ->
+  // bedrock.{region}.amazonaws.com (ListInferenceProfiles). Returns null when
+  // no distinct control-plane host can be derived (custom/relative baseUrl),
+  // which disables live discovery and leaves the static fallback in place.
   getModelsEndpoint() {
-    return null;
+    const n = this.config.baseUrl || "", t = n.replace("bedrock-runtime.", "bedrock.");
+    return !t || t === n ? null : `${t}/inference-profiles`;
   }
-  parseModelsResponse() {
-    return g.slice();
+  // ListInferenceProfiles returns
+  //   { inferenceProfileSummaries: [{ inferenceProfileId, ... }] }
+  // Keep only Anthropic Claude profiles (us.* and global.*) — those ids are
+  // exactly what InvokeModel expects in the URL path.
+  parseModelsResponse(n) {
+    const t = n == null ? void 0 : n.inferenceProfileSummaries;
+    if (!Array.isArray(t)) return p.slice();
+    const r = t.map((e) => e == null ? void 0 : e.inferenceProfileId).filter((e) => typeof e == "string" && e.includes("anthropic.claude"));
+    return r.length ? r : p.slice();
   }
 }
 class v {
   constructor() {
     this.buf = new Uint8Array(0), this.textDecoder = new TextDecoder();
   }
-  feed(r) {
-    if (r && r.length) {
-      const t = new Uint8Array(this.buf.length + r.length);
-      t.set(this.buf, 0), t.set(r, this.buf.length), this.buf = t;
+  feed(n) {
+    if (n && n.length) {
+      const r = new Uint8Array(this.buf.length + n.length);
+      r.set(this.buf, 0), r.set(n, this.buf.length), this.buf = r;
     }
-    const n = [];
+    const t = [];
     for (; this.buf.length >= 12; ) {
-      const t = new DataView(this.buf.buffer, this.buf.byteOffset, this.buf.byteLength), e = t.getUint32(0, !1), d = t.getUint32(4, !1);
+      const r = new DataView(this.buf.buffer, this.buf.byteOffset, this.buf.byteLength), e = r.getUint32(0, !1), d = r.getUint32(4, !1);
       if (e < 16 || e > 16 * 1024 * 1024)
         throw new Error(`Bedrock event stream: invalid frame length ${e}`);
       if (this.buf.length < e) break;
-      const c = 12, u = c + d, s = u, l = e - 4, f = k(this.buf.subarray(c, u)), i = this.textDecoder.decode(this.buf.subarray(s, l));
+      const i = 12, l = i + d, o = l, h = e - 4, f = k(this.buf.subarray(i, l)), c = this.textDecoder.decode(this.buf.subarray(o, h));
       this.buf = this.buf.subarray(e);
       let a;
       try {
-        a = JSON.parse(i);
+        a = JSON.parse(c);
       } catch {
         a = null;
       }
-      const m = f[":message-type"], p = f[":exception-type"];
-      if (m === "exception" || p) {
-        const h = p || f[":event-type"] || "BedrockStreamException", b = (a == null ? void 0 : a.message) || i || "unknown error";
-        n.push({ kind: "error", message: `${h}: ${b}` });
+      const g = f[":message-type"], m = f[":exception-type"];
+      if (g === "exception" || m) {
+        const u = m || f[":event-type"] || "BedrockStreamException", b = (a == null ? void 0 : a.message) || c || "unknown error";
+        t.push({ kind: "error", message: `${u}: ${b}` });
         continue;
       }
       if (a && typeof a.bytes == "string")
         try {
-          const h = x(a.bytes), b = JSON.parse(h);
-          n.push({ kind: "event", payload: b });
-        } catch (h) {
-          n.push({ kind: "error", message: `Bedrock event-stream decode failed: ${h.message}` });
+          const u = x(a.bytes), b = JSON.parse(u);
+          t.push({ kind: "event", payload: b });
+        } catch (u) {
+          t.push({ kind: "error", message: `Bedrock event-stream decode failed: ${u.message}` });
         }
     }
-    return n;
+    return t;
   }
 }
-function k(o) {
-  const r = {}, n = new DataView(o.buffer, o.byteOffset, o.byteLength), t = new TextDecoder();
+function k(s) {
+  const n = {}, t = new DataView(s.buffer, s.byteOffset, s.byteLength), r = new TextDecoder();
   let e = 0;
-  for (; e < o.length; ) {
-    const d = o[e];
-    if (e += 1, e + d > o.length) break;
-    const c = t.decode(o.subarray(e, e + d));
-    if (e += d, e >= o.length) break;
-    const u = o[e];
-    if (e += 1, u === 7) {
-      if (e + 2 > o.length) break;
-      const s = n.getUint16(e, !1);
-      if (e += 2, e + s > o.length) break;
-      r[c] = t.decode(o.subarray(e, e + s)), e += s;
+  for (; e < s.length; ) {
+    const d = s[e];
+    if (e += 1, e + d > s.length) break;
+    const i = r.decode(s.subarray(e, e + d));
+    if (e += d, e >= s.length) break;
+    const l = s[e];
+    if (e += 1, l === 7) {
+      if (e + 2 > s.length) break;
+      const o = t.getUint16(e, !1);
+      if (e += 2, e + o > s.length) break;
+      n[i] = r.decode(s.subarray(e, e + o)), e += o;
     } else
       break;
   }
-  return r;
+  return n;
 }
-function x(o) {
-  const r = atob(o), n = new Uint8Array(r.length);
-  for (let t = 0; t < r.length; t++) n[t] = r.charCodeAt(t);
-  return new TextDecoder().decode(n);
+function x(s) {
+  const n = atob(s), t = new Uint8Array(n.length);
+  for (let r = 0; r < n.length; r++) t[r] = n.charCodeAt(r);
+  return new TextDecoder().decode(t);
 }
 export {
-  g as BEDROCK_CLAUDE_MODELS,
-  R as BedrockProvider
+  p as BEDROCK_CLAUDE_MODELS,
+  A as BedrockProvider
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anvaka/vue-llm",
-  "version": "0.3.0",
+  "version": "0.3.2",
   "description": "Browser-only LLM client with provider adapters and Vue 3 components",
   "type": "module",
   "license": "MIT",
@@ -9,7 +9,8 @@
   "scripts": {
     "build": "vite build",
     "dev": "vite build --watch",
-    "test:providers": "node test/providers.mjs"
+    "test:providers": "node test/providers.mjs",
+    "test:caching": "node test/caching.mjs"
   },
   "exports": {
     ".": {