omnius 1.0.646 → 1.0.647

package/dist/index.js

@@ -628934,6 +628934,67 @@ var init_personality = __esm({
   }
 });
 
+// packages/orchestrator/dist/ste-communication-policy.js
+function ensureSteCommunicationPolicy(messages2) {
+  const hasPolicy = messages2.some((message2) => message2.role === "system" && typeof message2.content === "string" && message2.content.includes(STE_POLICY_MARKER));
+  if (hasPolicy)
+    return [...messages2];
+  return [
+    {
+      role: "system",
+      content: STE_COMMUNICATION_POLICY
+    },
+    ...messages2
+  ];
+}
+var STE_POLICY_MARKER, STE_COMMUNICATION_POLICY;
+var init_ste_communication_policy = __esm({
+  "packages/orchestrator/dist/ste-communication-policy.js"() {
+    "use strict";
+    STE_POLICY_MARKER = "[ASD-STE100 ISSUE 9 COMMUNICATION POLICY]";
+    STE_COMMUNICATION_POLICY = `${STE_POLICY_MARKER}
+
+Use ASD-STE100 Issue 9 for all natural-language communication.
+Apply this policy to external replies and internal agent communication.
+
+- Use approved words when possible.
+- Treat necessary project and domain terms as technical terms.
+- Use one term for each concept.
+- Use American English spelling.
+- Do not use contractions, slang, jargon, or idioms.
+- Use active voice.
+- Use passive voice only when the agent is unknown.
+- Use simple verb tenses.
+- Use direct verbs.
+- Avoid phrasal verbs.
+- Keep each instruction at 20 words or fewer.
+- Give one instruction in each sentence.
+- Put a condition before its instruction.
+- Keep each descriptive sentence at 25 words or fewer.
+- Keep one topic in each paragraph.
+- Keep each paragraph at six sentences or fewer.
+- Use a vertical list for complex information.
+- Use clear connecting words.
+- Do not use semicolons.
+- Identify the level of each safety risk.
+- Start each safety instruction with a clear command or condition.
+- Explain the possible result of each safety risk.
+
+Preserve exact content in these items:
+
+- Code and code blocks
+- JSON, schemas, keys, and API data
+- Commands, paths, identifiers, and formal names
+- User input and source quotations
+- Logs, tool output, and test output
+- Required machine formats
+
+Use STE for all text that explains an exact item.
+If an exact format conflicts with STE, use the exact format.
+Do not claim certified ASD-STE100 conformance without a qualified human review.`;
+  }
+});
+
 // packages/orchestrator/dist/reference-contracts.js
 import { createHash as createHash37 } from "node:crypto";
 function compactSource(source, maxChars) {
@@ -648985,7 +649046,7 @@ function classifyThinkOutcome(raw) {
   }
   return null;
 }
-var PRESERVE_MODEL_VISIBLE_HISTORY, TOOL_SUBSETS, TOOL_AUTO_DEMOTE_TURNS, LEGACY_ACTION_REASON_KEYS, CLAIM_TASK_STOPWORDS, SYSTEM_PROMPT, SYSTEM_PROMPT_MEDIUM, SYSTEM_PROMPT_SMALL, VISUAL_TOOLS, AUDIO_TOOLS, SOCIAL_TOOLS, SPATIAL_TOOLS, CODE_TOOLS, AgenticRunner, OllamaAgenticBackend;
+var PRESERVE_MODEL_VISIBLE_HISTORY, TOOL_SUBSETS, TOOL_AUTO_DEMOTE_TURNS, LEGACY_ACTION_REASON_KEYS, CLAIM_TASK_STOPWORDS, SYSTEM_PROMPT_COMMON, SYSTEM_PROMPT, SYSTEM_PROMPT_MEDIUM, SYSTEM_PROMPT_SMALL, VISUAL_TOOLS, AUDIO_TOOLS, SOCIAL_TOOLS, SPATIAL_TOOLS, CODE_TOOLS, AgenticRunner, OllamaAgenticBackend;
 var init_agenticRunner = __esm({
   "packages/orchestrator/dist/agenticRunner.js"() {
     "use strict";
@@ -649008,6 +649069,7 @@ var init_agenticRunner = __esm({
     init_ollama_pool();
     init_personality();
     init_promptLoader();
+    init_ste_communication_policy();
     init_reference_contracts();
     init_phase_skill_guidance();
     init_deliveryCoverage();
@@ -649183,9 +649245,18 @@ var init_agenticRunner = __esm({
       "new",
       "old"
     ]);
-    SYSTEM_PROMPT = loadPrompt("agentic/system-large.md");
-    SYSTEM_PROMPT_MEDIUM = loadPrompt("agentic/system-medium.md");
-    SYSTEM_PROMPT_SMALL = loadPrompt("agentic/system-small.md");
+    SYSTEM_PROMPT_COMMON = `${STE_COMMUNICATION_POLICY}
+
+${loadPrompt("agentic/system-common.md")}`;
+    SYSTEM_PROMPT = `${SYSTEM_PROMPT_COMMON}
+
+${loadPrompt("agentic/system-large.md")}`;
+    SYSTEM_PROMPT_MEDIUM = `${SYSTEM_PROMPT_COMMON}
+
+${loadPrompt("agentic/system-medium.md")}`;
+    SYSTEM_PROMPT_SMALL = `${SYSTEM_PROMPT_COMMON}
+
+${loadPrompt("agentic/system-small.md")}`;
     VISUAL_TOOLS = /* @__PURE__ */ new Set([
       "vision",
       "camera_capture",
@@ -676242,6 +676313,10 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
         return `${serviceBaseUrl.replace(/\/+$/, "")}${path16}`;
       }
       async chatCompletion(request) {
+        request = {
+          ...request,
+          messages: ensureSteCommunicationPolicy(request.messages)
+        };
         if (this._provider.protocol === "anthropic-messages") {
           return this._anthropicChatCompletion(request);
         }
@@ -676268,7 +676343,8 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
         }
         const requestMessages2 = cleanedMessages;
         const responseFormat = request.responseFormat ?? request.response_format;
-        const isOllama = this._provider.protocol === "ollama" && shouldUseOllamaPoolForBaseUrl(this.baseUrl);
+        const isOllamaTransport = this._provider.protocol === "ollama";
+        const useOllamaPool = isOllamaTransport && shouldUseOllamaPoolForBaseUrl(this.baseUrl);
         const body = {
           model: this.model,
           messages: requestMessages2,
@@ -676276,20 +676352,20 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
           temperature: request.temperature,
           max_tokens: effectiveMaxTokens
         };
-        if (isOllama) {
+        if (isOllamaTransport) {
           body["think"] = effectiveThink;
         }
         if (responseFormat !== void 0) {
           body["response_format"] = responseFormat;
         }
         const reqNumCtx = request.numCtx ?? request.num_ctx;
-        if (isOllama && Number.isFinite(reqNumCtx) && (reqNumCtx ?? 0) > 0) {
+        if (isOllamaTransport && Number.isFinite(reqNumCtx) && (reqNumCtx ?? 0) > 0) {
           const opts = body["options"] ?? {};
           opts["num_ctx"] = reqNumCtx;
           body["options"] = opts;
           body["num_ctx"] = reqNumCtx;
         }
-        let poolSlot = isOllama ? await getOllamaPool({ baseInstanceUrl: this.baseUrl }).acquire({
+        let poolSlot = useOllamaPool ? await getOllamaPool({ baseInstanceUrl: this.baseUrl }).acquire({
           model: this.model,
           queuePolicy: request.poolQueuePolicy,
           queueTimeoutMs: request.poolQueueTimeoutMs
@@ -676333,7 +676409,7 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
         ]);
         let nativePreferredAttempted = false;
         try {
-          if (isOllama && request.preferNativeOllamaChat === true) {
+          if (isOllamaTransport && request.preferNativeOllamaChat === true) {
             nativePreferredAttempted = true;
             try {
               const native = await this.nativeOllamaChatCompletion(request, {
@@ -676406,7 +676482,7 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
           const independentOutcome = effectiveThink !== true ? classifyThinkOutcome(responseText2) : null;
           const emptyIndependentOutcome = independentOutcome === "empty_after_strip" || independentOutcome === "unclosed_think";
           const expectsVisibleContent = responseFormat !== void 0 || (request.tools ?? []).length === 0;
-          const shouldProbeNativeOnEmpty = isOllama && !nativePreferredAttempted && expectsVisibleContent && emptyIndependentOutcome;
+          const shouldProbeNativeOnEmpty = isOllamaTransport && !nativePreferredAttempted && expectsVisibleContent && emptyIndependentOutcome;
           const shouldRecoverFromEmpty = request.disableEmptyContentRecovery !== true && expectsVisibleContent && emptyIndependentOutcome;
           const justSuppressed = this._thinkSuppressed && this._thinkFailStreak === _OllamaAgenticBackend._thinkFailThreshold;
           const shouldRetryThinkGuard = outcome !== null && effectiveThink === true && (justSuppressed || outcome === "empty_after_strip" || outcome === "unclosed_think");
@@ -676443,12 +676519,12 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
               temperature: request.temperature,
               max_tokens: request.maxTokens
             };
-            if (isOllama)
+            if (isOllamaTransport)
               retryBody["think"] = false;
             if (responseFormat !== void 0 && shouldRetryThinkGuard && !shouldRecoverFromEmpty) {
               retryBody["response_format"] = responseFormat;
             }
-            if (isOllama && Number.isFinite(reqNumCtx) && (reqNumCtx ?? 0) > 0) {
+            if (isOllamaTransport && Number.isFinite(reqNumCtx) && (reqNumCtx ?? 0) > 0) {
               const retryOptsBody = retryBody["options"] ?? {};
               retryOptsBody["num_ctx"] = reqNumCtx;
               retryBody["options"] = retryOptsBody;
@@ -676516,6 +676592,10 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
         }
       }
       async nativeOllamaChatCompletion(request, transport = {}) {
+        request = {
+          ...request,
+          messages: ensureSteCommunicationPolicy(request.messages)
+        };
         const cleanedMessages = applyMemoryPrefixToMessages(normalizeMessagesForStrictOpenAI(sanitizeHistoryThink(request.messages, {
           preserveModelVisibleHistory: request[PRESERVE_MODEL_VISIBLE_HISTORY] === true
         })), request.memoryPrefix);
@@ -676628,6 +676708,10 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
        * Ollama pool routing as non-stream completions.
        */
       async *chatCompletionStream(request) {
+        request = {
+          ...request,
+          messages: ensureSteCommunicationPolicy(request.messages)
+        };
         if (this._provider.protocol === "anthropic-messages") {
           const result = await this._anthropicChatCompletion(request);
           const message2 = result.choices[0]?.message;
@@ -676691,20 +676775,21 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
           stream: true,
           stream_options: { include_usage: true }
         };
-        const isOllama = this._provider.protocol === "ollama" && shouldUseOllamaPoolForBaseUrl(this.baseUrl);
-        if (isOllama)
+        const isOllamaTransport = this._provider.protocol === "ollama";
+        const useOllamaPool = isOllamaTransport && shouldUseOllamaPoolForBaseUrl(this.baseUrl);
+        if (isOllamaTransport)
           body["think"] = effectiveThink;
         if (responseFormat !== void 0) {
           body["response_format"] = responseFormat;
         }
         const reqNumCtx = request.numCtx ?? request.num_ctx;
-        if (isOllama && Number.isFinite(reqNumCtx) && (reqNumCtx ?? 0) > 0) {
+        if (isOllamaTransport && Number.isFinite(reqNumCtx) && (reqNumCtx ?? 0) > 0) {
           const opts = body["options"] ?? {};
           opts["num_ctx"] = reqNumCtx;
           body["options"] = opts;
           body["num_ctx"] = reqNumCtx;
         }
-        let poolSlot = isOllama ? await getOllamaPool({ baseInstanceUrl: this.baseUrl }).acquire({
+        let poolSlot = useOllamaPool ? await getOllamaPool({ baseInstanceUrl: this.baseUrl }).acquire({
           model: this.model,
           queuePolicy: request.poolQueuePolicy,
           queueTimeoutMs: request.poolQueueTimeoutMs
@@ -676736,7 +676821,7 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
           }
         }
         try {
-          if (isOllama && request.preferNativeOllamaChat === true) {
+          if (isOllamaTransport && request.preferNativeOllamaChat === true) {
             const nativeOptions = {
               temperature: request.temperature,
               num_predict: effectiveMaxTokens
@@ -677255,6 +677340,7 @@ var init_nexusBackend = __esm({
   "packages/orchestrator/dist/nexusBackend.js"() {
     "use strict";
     init_textSanitize();
+    init_ste_communication_policy();
     NexusAgenticBackend = class _NexusAgenticBackend {
       sendFn;
       model;
@@ -677288,7 +677374,7 @@ var init_nexusBackend = __esm({
         return messages2.map((message2) => typeof message2.content === "string" ? { ...message2, content: stripNoThinkPromptDirectives(message2.content) } : message2);
       }
       requestMessages(request, effectiveThink) {
-        return this.noThinkMessages(request.messages);
+        return this.noThinkMessages(ensureSteCommunicationPolicy(request.messages));
       }
       applyOptionalRequestFields(daemonArgs, request) {
         const responseFormat = request.responseFormat ?? request.response_format;
@@ -684526,6 +684612,8 @@ __export(dist_exports3, {
   RequestFingerprintCache: () => RequestFingerprintCache,
   RetryController: () => RetryController,
   STALENESS_THRESHOLD_DEFAULTS: () => STALENESS_THRESHOLD_DEFAULTS,
+  STE_COMMUNICATION_POLICY: () => STE_COMMUNICATION_POLICY,
+  STE_POLICY_MARKER: () => STE_POLICY_MARKER,
   ScoutRunner: () => ScoutRunner,
   SessionMetrics: () => SessionMetrics,
   StreamingToolExecutor: () => StreamingToolExecutor,
@@ -684645,6 +684733,7 @@ __export(dist_exports3, {
   discoverPlugins: () => discoverPlugins3,
   discoverSystemOllamaModelStore: () => discoverSystemOllamaModelStore,
   effectiveContextWindow: () => effectiveContextWindow,
+  ensureSteCommunicationPolicy: () => ensureSteCommunicationPolicy,
   estimateConservativeRequestInputTokens: () => estimateConservativeRequestInputTokens,
   evaluateDeliveryCoverage: () => evaluateDeliveryCoverage,
   evaluateExecutorStep: () => evaluateExecutorStep,
@@ -684884,6 +684973,7 @@ var init_dist8 = __esm({
     init_mergeRunner();
     init_retryController();
     init_agenticRunner();
+    init_ste_communication_policy();
     init_request_budget();
     init_memory_compiler();
     init_memory_compiler_replay();
@@ -714982,7 +715072,7 @@ var init_task_templates = __esm({
           "git_info",
           "codebase_map"
         ],
-        outputGuidance: "Deliver working code with tests passing. Summarize what was changed and why."
+        outputGuidance: "Deliver working code. Confirm that the applicable tests pass. Identify each change and its reason."
       },
       document: {
         type: "document",
@@ -714997,7 +715087,7 @@ var init_task_templates = __esm({
           "create_structured_file",
           "memory_read"
         ],
-        outputGuidance: "Deliver a complete, well-structured document. Specify the output format (Markdown, PDF, etc.)."
+        outputGuidance: "Deliver a complete and well-structured document. Identify the output format."
       },
       analysis: {
         type: "analysis",
@@ -715016,7 +715106,7 @@ var init_task_templates = __esm({
           "create_structured_file",
           "shell"
         ],
-        outputGuidance: "Deliver a structured analysis with findings, data tables, and actionable recommendations."
+        outputGuidance: "Deliver a structured analysis. Include findings, necessary data tables, and specific recommendations."
       },
       plan: {
         type: "plan",
@@ -715033,12 +715123,12 @@ var init_task_templates = __esm({
           "file_write",
           "git_info"
         ],
-        outputGuidance: "Deliver a structured plan with phases, milestones, risks, and action items."
+        outputGuidance: "Deliver a structured plan. Include phases, milestones, risks, and specific actions."
       },
       general: {
         type: "general",
         label: "General Task",
-        description: "Tasks that don't clearly fit another category.",
+        description: "Tasks that do not clearly fit another category.",
         systemPromptAddition: loadPrompt2("general.md"),
         recommendedTools: [
           "file_read",
@@ -715048,7 +715138,7 @@ var init_task_templates = __esm({
           "web_search",
           "memory_read"
         ],
-        outputGuidance: "Deliver complete results with a clear summary of what was done."
+        outputGuidance: "Deliver complete results. Give a clear summary of the completed work."
       }
     };
   }
@@ -720885,6 +720975,7 @@ var init_status_bar = __esm({
           row: screenRow,
           lineText: hit.lineText
         })) {
+          this.invalidateRowCountCache();
           if (this.active && !isOverlayActive() && !this._suspendContentLayer) {
             this.repaintContent();
           }
@@ -759451,46 +759542,55 @@ function buildRealtimeSystemPrompt(opts) {
   const voiceLimit = opts.maxVoiceChars ?? DEFAULT_VOICE_CHARS;
   const maxReplyWords = clampInt2(opts.maxReplyWords, DEFAULT_REALTIME_MAX_REPLY_WORDS, 8, 80);
   const sections = [
+    STE_COMMUNICATION_POLICY,
     "[Omnius realtime conversation mode]",
     [
-      "Purpose: short, natural, back-and-forth spoken dialogue for ASR/TTS pipelines.",
+      "Purpose: short and direct spoken dialog for ASR and TTS systems.",
       `Surface: ${surface}`,
       opts.sessionId ? `Session: ${opts.sessionId}` : "",
       `Workspace: ${repoRoot}`
     ].filter(Boolean).join("\n"),
     [
       "Interaction basis:",
-      "- Use dialogue alignment (Pickering/Garrod), grounding and repair, adjacency-pair response, and turn-taking timing principles (Wilson oscillator timing).",
-      "- Treat the latest user utterance as the live turn; if it interrupts or corrects earlier context, follow the latest turn.",
-      "- Listen for human cues in the provided words and conversation state; do not run local keyword classifiers."
+      "- Use dialog alignment, grounding, repair, adjacent responses, and turn timing.",
+      "- Treat the latest user statement as the current turn.",
+      "- If that statement corrects earlier context, use the correction.",
+      "- Use the supplied words and conversation state.",
+      "- Do not use a local keyword classifier."
     ].join("\n"),
     [
       "Phone reply contract:",
       `- Produce one natural spoken turn, normally ${maxReplyWords} words or fewer.`,
-      "- Use one sentence when possible; two short sentences only when repair or confirmation needs it.",
-      "- Lead with the answer. Do not preface with status, analysis, summaries, or implementation narration.",
-      "- When the caller requests an exact literal response, emit only that literal text: no preface, suffix, punctuation, or commentary.",
-      "- No markdown, bullets, tables, headings, citations, inline code, code blocks, JSON, or labels like 'Assistant:'.",
-      "- Sound like a person on a live call: brief acknowledgment, direct answer, one focused follow-up only if needed.",
-      "- If the ASR text is garbled or underspecified, ask a single compact repair question.",
-      "- Do not invent app modes, method names, settings, or implementation details when the caller has not supplied them.",
-      "- Do not mention ASR, TTS, prompts, realtime mode, hidden reasoning, tools, or policy unless the caller explicitly asks.",
-      "- If a request needs work outside this text-only exchange, say the next handoff in one short sentence."
+      "- Use one sentence when possible.",
+      "- Use two short sentences only for repair or confirmation.",
+      "- Start with the answer.",
+      "- Do not add status, analysis, summaries, or implementation details before the answer.",
+      "- If the caller requests exact text, return only that text.",
+      "- Do not add punctuation or commentary to requested exact text.",
+      "- Do not use document formatting.",
+      "- Use a brief acknowledgment and a direct answer.",
+      "- Ask one focused question only when necessary.",
+      "- If the ASR text is unclear, ask one short repair question.",
+      "- Do not invent modes, names, settings, or implementation details.",
+      "- Do not mention system operation unless the caller requests that information.",
+      "- If the request needs other work, identify the next transfer in one short sentence."
     ].join("\n"),
     soul ? `Project SOUL.md (${basename35(soul.path)}), compacted for realtime:
 ${blockText2(soul.content, soulLimit)}` : [
       "No project SOUL.md found. Default soul:",
-      "- pragmatic, direct, concrete",
-      "- low-fluff, human-paced, evidence-aware",
-      "- truthful about uncertainty without overexplaining"
+      "- Use practical, direct, and concrete language.",
+      "- Use a natural speaking rate.",
+      "- Base each claim on evidence.",
+      "- State uncertainty without unnecessary explanation."
     ].join("\n"),
     voice ? `Project voice profile (${basename35(voice.path)}), compacted for realtime:
 ${blockText2(voice.content, voiceLimit)}` : [
       "Default realtime voice:",
-      "- conversational, brief, and proportional",
-      "- phone-call natural: contractions, plain words, no written-document structure",
-      "- contractions are fine when natural",
-      "- no list formatting unless the user asks for a list"
+      "- Use brief and proportional conversation.",
+      "- Use plain words.",
+      "- Do not use contractions.",
+      "- Do not use document structure.",
+      "- Do not use a list unless the user requests a list."
     ].join("\n")
   ];
   return sections.join("\n\n");
@@ -759565,7 +759665,7 @@ function wordParts(text3) {
 function finalizeRealtimeReply(text3, opts = {}) {
   const maxWords = clampInt2(opts.maxReplyWords, DEFAULT_REALTIME_MAX_REPLY_WORDS, 8, 80);
   let clean7 = stripHiddenThinking(String(text3 ?? "")).replace(/```[\s\S]*?```/g, "").split("\n").map((line) => line.replace(/^\s*(?:[-*]+|\d+[.)])\s+/, "").trim()).filter(Boolean).join(" ").replace(/^(?:assistant|omnius|agent)\s*:\s*/i, "").replace(/`([^`]+)`/g, "$1").replace(/\s+/g, " ").trim();
-  if (!clean7) return "I didn't catch that. Can you say it again?";
+  if (!clean7) return "I did not understand that. Please say it again.";
   const sentences = clean7.match(/[^.!?]+[.!?]+(?=\s|$)|[^.!?]+$/g) ?? [clean7];
   const selected = [];
   let words = 0;
@@ -759590,6 +759690,7 @@ function finalizeRealtimeReply(text3, opts = {}) {
 var DEFAULT_REALTIME_HISTORY_MESSAGES, DEFAULT_REALTIME_MAX_TOKENS, DEFAULT_REALTIME_MAX_REPLY_WORDS, DEFAULT_SOUL_CHARS, DEFAULT_VOICE_CHARS;
 var init_realtime = __esm({
   "packages/cli/src/realtime.ts"() {
+    init_dist8();
     DEFAULT_REALTIME_HISTORY_MESSAGES = 8;
     DEFAULT_REALTIME_MAX_TOKENS = 120;
     DEFAULT_REALTIME_MAX_REPLY_WORDS = 36;
@@ -759791,9 +759892,9 @@ function loadPersistedSessions() {
   return report2;
 }
 function buildSystemPrompt(cwd4) {
-  const parts = [];
+  const parts = [STE_COMMUNICATION_POLICY];
   parts.push(
-    "You are Open Agent (Omnius), an AI coding assistant running locally via Ollama. You have access to the user's workspace and can discuss code, files, and projects. Be helpful, concise, and technically precise. When asked about files or code, describe what you know from the conversation context."
+    "You are Open Agent, also named Omnius. You are an AI coding assistant. You can use the user's workspace. Give concise and precise technical information. For a file or code question, use facts from the conversation context."
   );
   parts.push(
     `\\nEnvironment: ${process.platform}, Node ${process.version}, CWD: ${cwd4}`
@@ -760185,6 +760286,7 @@ var sessions2, inFlight, SESSION_TTL_MS, INFERENCE_ROLES, PARTIAL_TAIL_BUDGET;
 var init_chat_session = __esm({
   "packages/cli/src/api/chat-session.ts"() {
     init_secret_redactor();
+    init_dist8();
     init_session_quality();
     sessions2 = /* @__PURE__ */ new Map();
     inFlight = /* @__PURE__ */ new Map();
@@ -769026,6 +769128,7 @@ function createInferenceLiveBlockState(label = "Model inference", detail = "", o
     // This block exists specifically to expose the generated reasoning stream.
     // Callers may still hide only that section with Ctrl+O.
     showThinking: options2.showThinking ?? true,
+    readMode: false,
     version: 0
   };
 }
@@ -769050,6 +769153,11 @@ function setInferenceThinkingVisible(state5, visible) {
   state5.showThinking = visible;
   state5.version++;
 }
+function setInferenceReadMode(state5, enabled2) {
+  if (state5.readMode === enabled2) return;
+  state5.readMode = enabled2;
+  state5.version++;
+}
 function finishInferenceLiveBlock(state5, success, detail, now2 = Date.now()) {
   if (state5.status !== "running") return;
   if (detail) appendInferenceHandling(state5, detail);
@@ -769061,7 +769169,7 @@ function finishInferenceLiveBlock(state5, success, detail, now2 = Date.now()) {
 function inferenceLiveBlockFingerprint(state5, now2 = Date.now()) {
   const end = state5.finishedAt ?? now2;
   const sec = Math.floor(Math.max(0, end - state5.startedAt) / 1e3);
-  return `${state5.status}|${state5.phase}|${state5.version}|${state5.showThinking ? 1 : 0}|${sec}`;
+  return `${state5.status}|${state5.phase}|${state5.version}|${state5.showThinking ? 1 : 0}|${state5.readMode ? 1 : 0}|${sec}`;
 }
 function buildInferenceLiveBlockLines(state5, width, now2 = Date.now()) {
   const w = Math.max(36, width);
@@ -769099,8 +769207,8 @@ function buildInferenceLiveBlockLines(state5, width, now2 = Date.now()) {
   if (visualRows.length === 0) {
     visualRows.push("handling › waiting for first generated token");
   }
-  const hidden = Math.max(0, visualRows.length - state5.maxRows);
-  const visibleRows = visualRows.slice(-state5.maxRows);
+  const hidden = state5.readMode ? 0 : Math.max(0, visualRows.length - state5.maxRows);
+  const visibleRows = state5.readMode ? visualRows : visualRows.slice(-state5.maxRows);
   const rows = [top];
   if (hidden > 0) {
     rows.push(
@@ -769111,6 +769219,8 @@ function buildInferenceLiveBlockLines(state5, width, now2 = Date.now()) {
     );
   }
   for (const row2 of visibleRows) rows.push(contentRow2(row2, inner));
+  const readModeLabel = state5.readMode ? INFERENCE_EXIT_READ_MODE_LABEL : hidden > 0 ? `${INFERENCE_READ_MODE_LABEL} · show ${hidden} earlier row${hidden === 1 ? "" : "s"}` : INFERENCE_READ_MODE_LABEL;
+  rows.push(contentRow2(`↳ ${readModeLabel}`, inner));
   rows.push(bottom);
   return rows;
 }
@@ -769170,9 +769280,11 @@ function formatElapsed2(ms) {
   const minutes = Math.floor(seconds / 60);
   return `${minutes}m${String(seconds % 60).padStart(2, "0")}s`;
 }
-var ANSI_OR_OSC_RE, InferenceLiveBlockController;
+var INFERENCE_READ_MODE_LABEL, INFERENCE_EXIT_READ_MODE_LABEL, ANSI_OR_OSC_RE, InferenceLiveBlockController;
 var init_inference_live_block = __esm({
   "packages/cli/src/tui/inference-live-block.ts"() {
+    INFERENCE_READ_MODE_LABEL = "read mode";
+    INFERENCE_EXIT_READ_MODE_LABEL = "exit read mode";
     ANSI_OR_OSC_RE = /\x1B\[[0-?]*[ -/]*[@-~]|\x1B\].*?(?:\x07|\x1B\\)/g;
     InferenceLiveBlockController = class {
       constructor(host2) {
@@ -769195,6 +769307,15 @@ var init_inference_live_block = __esm({
           id2,
           (width) => buildInferenceLiveBlockLines(state5, width)
         );
+        this.host.registerDynamicBlockClickHandler?.(id2, ({ lineText }) => {
+          const plain = lineText.replace(ANSI_OR_OSC_RE, "").toLowerCase();
+          if (!plain.includes(`↳ ${INFERENCE_READ_MODE_LABEL}`) && !plain.includes(`↳ ${INFERENCE_EXIT_READ_MODE_LABEL}`)) {
+            return false;
+          }
+          setInferenceReadMode(state5, !state5.readMode);
+          this.host.refreshDynamicBlocks?.();
+          return true;
+        });
         this.host.appendDynamicBlock(id2);
         this.schedule(key, block);
         return state5;
@@ -783959,10 +784080,11 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
             } else if (chunk.type === "finish") {
               finishReason = chunk.finishReason;
             } else if (chunk.type === "usage") {
+              const chunkUsage = chunk.usage;
               usage = {
-                prompt_tokens: chunk.promptTokens,
-                completion_tokens: chunk.completionTokens,
-                total_tokens: chunk.totalTokens
+                prompt_tokens: chunkUsage?.promptTokens,
+                completion_tokens: chunkUsage?.completionTokens,
+                total_tokens: chunkUsage?.totalTokens
               };
             }
           }
@@ -787845,6 +787967,7 @@ ${conversationStream}`
             });
             let accumulated = "";
             let streamError;
+            let inferenceSucceeded = false;
             const sessionKey = this.sessionKeyForMessage(msg);
             const inferenceId = this.registerTelegramInference(
               "chat-fast-path",
@@ -787858,11 +787981,28 @@ ${conversationStream}`
                 try {
                   let lastTokenEmitMs = 0;
                   for await (const chunk of stream) {
+                    if (chunk.type === "tool_call_delta") {
+                      const contractDelta = `${chunk.toolCallName ? `name=${chunk.toolCallName}
+arguments=` : ""}` + (chunk.toolCallArgs ?? "");
+                      if (contractDelta) {
+                        this.telegramInferenceBlocks?.append(
+                          inferenceId,
+                          "tool_args",
+                          getSecretRedactor().redactText(contractDelta)
+                        );
+                      }
+                      continue;
+                    }
                     if (chunk.type !== "content") continue;
                     const piece = chunk.content;
                     if (!piece) continue;
                     if (chunk.thinking) {
                       this.bumpTelegramInferenceTokens(inferenceId, 0, 1);
+                      this.telegramInferenceBlocks?.append(
+                        inferenceId,
+                        "thinking",
+                        getSecretRedactor().redactText(piece)
+                      );
                       if (this.telegramThinkingVisible) {
                         const preview = piece.slice(0, 120);
                         this.tuiWrite(
@@ -787874,6 +788014,11 @@ ${conversationStream}`
                       }
                     } else {
                       this.bumpTelegramInferenceTokens(inferenceId, 1, 0);
+                      this.telegramInferenceBlocks?.append(
+                        inferenceId,
+                        "content",
+                        getSecretRedactor().redactText(piece)
+                      );
                       accumulated += piece;
                       const now2 = Date.now();
                       if (now2 - lastTokenEmitMs > 120) {
@@ -787893,6 +788038,10 @@ ${conversationStream}`
                 } catch (err) {
                   streamError = err;
                   accumulated = "";
+                  this.telegramInferenceBlocks?.handling(
+                    inferenceId,
+                    `stream failed; falling back to unary response: ${err instanceof Error ? err.message : String(err)}`
+                  );
                 }
               }
               if (!accumulated.trim()) {
@@ -787914,7 +788063,17 @@ ${conversationStream}`
                 const fullExtracted = extractPartialTelegramReplyJson(accumulated);
                 if (fullExtracted) await onToken(fullExtracted);
               }
+              this.telegramInferenceBlocks?.handling(
+                inferenceId,
+                "response received; handing returned contract to parser"
+              );
+              inferenceSucceeded = true;
             } finally {
+              this.telegramInferenceBlocks?.finish(
+                inferenceId,
+                inferenceSucceeded,
+                inferenceSucceeded ? void 0 : "inference ended before a response contract was available"
+              );
               this.deregisterTelegramInference(inferenceId);
             }
             const extracted = extractFinalTelegramReplyJson(accumulated);
@@ -799633,7 +799792,8 @@ ${entry.fullContent}`
   const inferenceBlocks = statusBar?.isActive && !isNeovimActive() ? new InferenceLiveBlockController({
     registerDynamicBlock: (id2, render2) => statusBar.registerDynamicBlock(id2, render2),
     appendDynamicBlock: (id2) => statusBar.appendDynamicBlock(id2),
-    refreshDynamicBlocks: () => statusBar.refreshDynamicBlocks()
+    refreshDynamicBlocks: () => statusBar.refreshDynamicBlocks(),
+    registerDynamicBlockClickHandler: (id2, handler) => statusBar.registerDynamicBlockClickHandler(id2, handler)
   }) : null;
   _mainInferenceBlocksRef = inferenceBlocks;
   const mainInferenceBlockKey = "main";
@@ -804194,7 +804354,8 @@ Respond concisely and safely. Remember: you are talking to the general public.`;
       telegramBridge.setLiveInferenceHost({
         registerDynamicBlock: (id2, render2) => statusBar.registerDynamicBlock(id2, render2),
         appendDynamicBlock: (id2) => statusBar.appendDynamicBlock(id2),
-        refreshDynamicBlocks: () => statusBar.refreshDynamicBlocks()
+        refreshDynamicBlocks: () => statusBar.refreshDynamicBlocks(),
+        registerDynamicBlockClickHandler: (id2, handler) => statusBar.registerDynamicBlockClickHandler(id2, handler)
       });
       telegramBridge.setSubAgentViewCallbacks({
         onRegister: (id2, label, taskSummary) => {