motebit 1.11.0 → 1.11.1

package/dist/index.js

@@ -5126,11 +5126,11 @@ function resolveModelTier(tier, currentModel) {
   if (currentModel.includes("claude") || currentModel.includes("anthropic")) {
     switch (tier) {
       case "strongest":
-        return "claude-opus";
+        return "claude-opus-5";
       case "default":
-        return "claude-sonnet";
+        return "claude-sonnet-5";
       case "fast":
-        return "claude-haiku";
+        return "claude-haiku-4-5";
     }
   }
   if (currentModel.includes("gpt") || currentModel.includes("o1") || currentModel.includes("o3") || currentModel.includes("o4")) {
@@ -5628,6 +5628,9 @@ ${tagged}` : n2;
 function modelSupportsExtendedThinking(model) {
   return /claude-(?:3-7-sonnet|(?:opus|sonnet)-(?:[4-9]|\d\d))/i.test(model);
 }
+function modelRejectsSamplingParams(model) {
+  return /claude-(?:opus-(?:4-[7-9]|[5-9]\b|\d\d)|sonnet-(?:[5-9]\b|\d\d)|fable|mythos)/i.test(model);
+}
 function stripTags(text) {
   return text.replace(/<memory\s+[^>]*>[\s\S]*?<\/memory>/g, "").replace(/<thinking>[\s\S]*?<\/thinking>/g, "").replace(/<state\s+[^>]*\/>/g, "").replace(/<narration\s*>[\s\S]*?<\/narration\s*>/g, "").replace(/\[EXTERNAL_DATA[^\]]*\][\s\S]*?\[\/EXTERNAL_DATA\]/g, "").replace(/\[MEMORY_DATA\][\s\S]*?\[\/MEMORY_DATA\]/g, "").replace(/\[EXTERNAL_DATA[^\]]*\]/g, "").replace(/\[\/EXTERNAL_DATA\]/g, "").replace(/\[MEMORY_DATA\]/g, "").replace(/\[\/MEMORY_DATA\]/g, "").replace(/\*[^*]+\*/g, "").replace(/\n{3,}/g, "\n\n").replace(/\s{2,}/g, " ").trim();
 }
@@ -5761,11 +5764,15 @@ var init_core = __esm({
         const body = {
           model: this.config.model,
           max_tokens: this.config.max_tokens ?? 4096,
-          // Only send temperature when the user explicitly configured it. Claude
-          // Opus 4.7+ deprecates the parameter entirely and returns HTTP 400 when
-          // it's present. Omitting it lets each model use its own default; users
-          // who want to tune sampling still can via `config.temperature`.
-          ...this.config.temperature !== void 0 && { temperature: this.config.temperature },
+          // Only send temperature when the user explicitly configured it AND the
+          // model still accepts it — the Opus 4.7+/Claude-5 family returns HTTP
+          // 400 when the parameter is present at all (witnessed live 2026-07-29:
+          // the CLI's personality default 0.7 400'd every claude-opus-5 turn).
+          // Omitting it lets each model use its own default; users on models
+          // that accept sampling still tune via `config.temperature`.
+          ...this.config.temperature !== void 0 && !modelRejectsSamplingParams(this.config.model) && {
+            temperature: this.config.temperature
+          },
           // Cacheable system blocks (see buildSystemBlocks) — the static prefix
           // caches at 1/10th input cost, the lever for cost on the agentic loop.
           system: this.buildSystemBlocks(contextPack),
@@ -5828,11 +5835,15 @@ var init_core = __esm({
         const body = {
           model: this.config.model,
           max_tokens: this.config.max_tokens ?? 4096,
-          // Only send temperature when the user explicitly configured it. Claude
-          // Opus 4.7+ deprecates the parameter entirely and returns HTTP 400 when
-          // it's present. Omitting it lets each model use its own default; users
-          // who want to tune sampling still can via `config.temperature`.
-          ...this.config.temperature !== void 0 && { temperature: this.config.temperature },
+          // Only send temperature when the user explicitly configured it AND the
+          // model still accepts it — the Opus 4.7+/Claude-5 family returns HTTP
+          // 400 when the parameter is present at all (witnessed live 2026-07-29:
+          // the CLI's personality default 0.7 400'd every claude-opus-5 turn).
+          // Omitting it lets each model use its own default; users on models
+          // that accept sampling still tune via `config.temperature`.
+          ...this.config.temperature !== void 0 && !modelRejectsSamplingParams(this.config.model) && {
+            temperature: this.config.temperature
+          },
           // Cacheable system blocks (see buildSystemBlocks) — same caching lever as
           // generate(); the streaming path is the one the chat surface actually uses.
           system: this.buildSystemBlocks(contextPack),
@@ -5990,7 +6001,7 @@ var init_core = __esm({
         if (!et || !modelSupportsExtendedThinking(this.config.model))
           return;
         const budget = Math.max(1024, Math.floor(et.budgetTokens));
-        body.thinking = { type: "enabled", budget_tokens: budget };
+        body.thinking = modelRejectsSamplingParams(this.config.model) ? { type: "adaptive" } : { type: "enabled", budget_tokens: budget };
         const configured = typeof body.max_tokens === "number" ? body.max_tokens : 4096;
         body.max_tokens = Math.max(configured, budget + 4096);
         delete body.temperature;
@@ -25493,7 +25504,7 @@ var init_config2 = __esm({
   "src/config.ts"() {
     "use strict";
     init_esm_shims();
-    VERSION = true ? "1.11.0" : "0.0.0-dev";
+    VERSION = true ? "1.11.1" : "0.0.0-dev";
     CONFIG_DIR = process.env["MOTEBIT_CONFIG_DIR"] ?? path2.join(os.homedir(), ".motebit");
     CONFIG_PATH = path2.join(CONFIG_DIR, "config.json");
     RELAY_DIR = path2.join(CONFIG_DIR, "relay");
@@ -57120,8 +57131,8 @@ function deniedIntentKey(toolName, args) {
   return `${toolName}:${canonical}`;
 }
 function stripDisplayTags(text) {
-  const clean2 = text.replace(/<memory\s+[^>]*>[\s\S]*?<\/memory>/g, "").replace(/<thinking>[\s\S]*?<\/thinking>/g, "").replace(/<state\s+[^>]*\/>/g, "").replace(/<parameter\s+[^>]*>[\s\S]*?<\/parameter>/g, "").replace(/<\/?(?:artifact|function_calls|invoke|antml)[^>]*>/g, "").replace(/\[EXTERNAL_DATA[^\]]*\][\s\S]*?\[\/EXTERNAL_DATA\]/g, "").replace(/\[MEMORY_DATA\][\s\S]*?\[\/MEMORY_DATA\]/g, "").replace(/\[EXTERNAL_DATA[^\]]*\]/g, "").replace(/\[\/EXTERNAL_DATA\]/g, "").replace(/\[MEMORY_DATA\]/g, "").replace(/\[\/MEMORY_DATA\]/g, "").replace(/\*{1,3}/g, "").replace(/ {2,}/g, " ");
-  for (const tag2 of ["<memory", "<thinking", "<parameter"]) {
+  const clean2 = text.replace(/<memory\s+[^>]*>[\s\S]*?<\/memory>/g, "").replace(/<thinking>[\s\S]*?<\/thinking>/g, "").replace(/<narration>[\s\S]*?<\/narration>/g, "").replace(/<state\s+[^>]*\/>/g, "").replace(/<parameter\s+[^>]*>[\s\S]*?<\/parameter>/g, "").replace(/<\/?(?:artifact|function_calls|invoke|antml)[^>]*>/g, "").replace(/\[EXTERNAL_DATA[^\]]*\][\s\S]*?\[\/EXTERNAL_DATA\]/g, "").replace(/\[MEMORY_DATA\][\s\S]*?\[\/MEMORY_DATA\]/g, "").replace(/\[EXTERNAL_DATA[^\]]*\]/g, "").replace(/\[\/EXTERNAL_DATA\]/g, "").replace(/\[MEMORY_DATA\]/g, "").replace(/\[\/MEMORY_DATA\]/g, "").replace(/\*{1,3}/g, "").replace(/ {2,}/g, " ");
+  for (const tag2 of ["<memory", "<thinking", "<parameter", "<narration"]) {
     const lastOpen2 = clean2.lastIndexOf(tag2);
     if (lastOpen2 !== -1) {
       const closeTag = `</${tag2.slice(1)}>`;
@@ -57168,6 +57179,28 @@ var init_streaming = __esm({
        * exit.
        */
       deniedIntents = /* @__PURE__ */ new Map();
+      /**
+       * Tools the human refused THIS exchange — since their last message (#470).
+       *
+       * The exact-args ledger above deliberately lets a reworked proposal ask
+       * once more ("a model that meaningfully reworks its arguments gets to ask
+       * once more; the prompt text it just received tells it not to"). That bet
+       * lost, live, 2026-07-29: a weak model reworded the denied paid hire
+       * trivially — folding one more sentence of the user's own words into the
+       * prompt — and re-prompted seconds after the "no", with the terminal
+       * REFUSED_BY_HUMAN instruction in its context. Model compliance cannot
+       * carry a money boundary.
+       *
+       * So the exact ledger gets a complement, not a replacement: within one
+       * exchange, a human denial is terminal for the TOOL, however the arguments
+       * are reworded. The set clears on the next genuine user message
+       * (`beginExchange`) — a human-initiated follow-up can always re-open the
+       * line it closed, so no future request is silently swallowed. Fuzzy arg
+       * matching was rejected for the same reason exact was chosen: a false
+       * match breaks the approval gate; tool-until-next-message cannot
+       * mis-classify.
+       */
+      deniedToolsThisExchange = /* @__PURE__ */ new Set();
       approvalExpiredCallback = null;
       /** Fired when a NEW turn voids a pending approval (#462) — the owning surface renders the void. */
       approvalVoidedCallback = null;
@@ -57230,6 +57263,16 @@ var init_streaming = __esm({
       onApprovalVoided(cb) {
         this.approvalVoidedCallback = cb;
       }
+      /**
+       * Mark the start of a genuine user turn (#470). The exchange-scoped denial
+       * brake stops re-asks BETWEEN user messages; the user's own next message
+       * always releases it — the human can re-open any line they closed, and
+       * only the human can. Callers: user-initiated turns only, never proactive
+       * or system-triggered ones (a consolidation tick must not launder a "no").
+       */
+      beginExchange() {
+        this.deniedToolsThisExchange.clear();
+      }
       /** Shared stream processing — extracts state tags, handles tool/approval/injection chunks. */
       async *processStream(stream, userMessage, runId, options) {
         let result = null;
@@ -57342,6 +57385,27 @@ var init_streaming = __esm({
               type: "text",
               text: `
 [refused already \u2014 not asking again: ${chunk.name}]
+`
+            };
+            continue;
+          }
+          if (chunk.type === "approval_request" && this.deniedToolsThisExchange.has(chunk.name)) {
+            this.deps.injectIntermediateMessages({
+              role: "assistant",
+              content: `[tool_use: ${chunk.name}(${JSON.stringify(chunk.args)})]`
+            }, {
+              role: "user",
+              content: `[tool_result: ${JSON.stringify({
+                ok: false,
+                error: "REFUSED_BY_HUMAN_THIS_EXCHANGE",
+                terminal: true,
+                message: `The human refused a ${chunk.name} call moments ago in this same exchange. Rewording the arguments does not make it a new request \u2014 this proposal was NOT shown to them. Do not propose ${chunk.name} again until the human sends a new message. Tell the user plainly what you wanted to do and that they declined, then ask what they would like instead.`
+              })}]`
+            });
+            yield {
+              type: "text",
+              text: `
+[you already declined ${chunk.name} this turn \u2014 not asking again]
 `
             };
             continue;
@@ -57439,6 +57503,7 @@ var init_streaming = __esm({
           } else {
             const key = deniedIntentKey(pending.toolName, pending.args);
             this.deniedIntents.set(key, (this.deniedIntents.get(key) ?? 0) + 1);
+            this.deniedToolsThisExchange.add(pending.toolName);
             this.deps.injectIntermediateMessages({
               role: "assistant",
               content: `[tool_use: ${pending.toolName}(${JSON.stringify(pending.args)})]`
@@ -60750,6 +60815,7 @@ var init_motebit_runtime = __esm({
         this.preemptCycleForUserMessage();
         this._lastUserMessageAt = Date.now();
         this._currentTypedIntent = options?.userActionAttestation ?? null;
+        this.streaming.beginExchange();
         const voidedApproval = this.streaming.voidPendingApproval();
         this.state.pushUpdate({ processing: 0.9, attention: 0.8 });
         this.behavior.setSpeaking(true);
@@ -106408,9 +106474,9 @@ ${summary}`);
     }
     case "model": {
       const MODEL_ALIASES = {
-        opus: "claude-opus-4-6-20250414",
-        sonnet: "claude-sonnet-4-5-latest",
-        haiku: "claude-haiku-4-5-20251001",
+        opus: "claude-opus-5",
+        sonnet: "claude-sonnet-5",
+        haiku: "claude-haiku-4-5",
         "gpt-5.4": "gpt-5.4",
         "gpt-5.4-mini": "gpt-5.4-mini",
         "gpt-5.4-nano": "gpt-5.4-nano",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "motebit",
-  "version": "1.11.0",
+  "version": "1.11.1",
   "type": "module",
   "description": "Reference runtime and operator console for sovereign AI agents — REPL, daemon, delegation, MCP server. Persistent Ed25519 identity, signed execution receipts, governance at the boundary.",
   "main": "./dist/index.js",
@@ -76,30 +76,30 @@
     "typescript": "^5.6.0",
     "vitest": "^4.1.10",
     "@motebit/ai-core": "0.0.0-private",
-    "@motebit/behavior-engine": "0.0.0-private",
     "@motebit/core-identity": "0.0.0-private",
     "@motebit/crypto": "3.19.1",
+    "@motebit/behavior-engine": "0.0.0-private",
     "@motebit/encryption": "0.0.0-private",
     "@motebit/event-log": "0.0.0-private",
-    "@motebit/gradient": "0.0.0-private",
     "@motebit/identity-file": "0.0.0-private",
     "@motebit/mcp-client": "0.0.0-private",
+    "@motebit/gradient": "0.0.0-private",
     "@motebit/mcp-server": "0.0.0-private",
-    "@motebit/memory-graph": "0.0.0-private",
     "@motebit/persistence": "0.0.0-private",
+    "@motebit/memory-graph": "0.0.0-private",
     "@motebit/protocol": "3.15.1",
     "@motebit/planner": "0.0.0-private",
-    "@motebit/policy": "0.0.0-private",
     "@motebit/privacy-layer": "0.0.0-private",
+    "@motebit/policy": "0.0.0-private",
     "@motebit/relay": "0.0.0-private",
     "@motebit/relay-client": "0.0.0-private",
-    "@motebit/runtime": "0.0.0-private",
     "@motebit/runtime-host": "0.0.0-private",
     "@motebit/sdk": "2.5.4",
     "@motebit/self-knowledge": "0.0.0-private",
-    "@motebit/skills": "0.0.0-private",
+    "@motebit/runtime": "0.0.0-private",
     "@motebit/state-vector": "0.0.0-private",
     "@motebit/sync-engine": "0.0.0-private",
+    "@motebit/skills": "0.0.0-private",
     "@motebit/tools": "0.0.0-private",
     "@motebit/verify": "1.8.12",
     "@motebit/voice": "0.0.0-private",