lynkr 9.14.0 → 9.14.2

package/src/clients/gpt-utils.js

@@ -25,7 +25,7 @@ function stringSimilarity(s1, s2) {
   // Tokenize by whitespace and common delimiters
   const tokenize = (s) => new Set(
     s.toLowerCase()
-      .split(/[\s\-_\/\.\,\:\;]+/)
+      .split(/[\s\-_/.,:;]+/)
       .filter(t => t.length > 0)
   );
 

package/src/clients/openai-format.js

@@ -113,21 +113,48 @@ function convertOpenAIToAnthropic(openaiRequest) {
         });
       }
     } else if (msg.role === "tool") {
-      // OpenAI tool response → Anthropic tool_result
-      const previousMsg = anthropicMessages[anthropicMessages.length - 1];
-
-      // Tool results must follow assistant message with tool_use
-      // Add as separate user message with tool_result
-      anthropicMessages.push({
-        role: "user",
-        content: [
-          {
-            type: "tool_result",
-            tool_use_id: msg.tool_call_id,
-            content: msg.content
-          }
-        ]
-      });
+      // OpenAI tool response → Anthropic tool_result. Two Anthropic rules
+      // apply (violating either 400s the request):
+      //   1. Every tool_use_id must match a tool_use block in the preceding
+      //      assistant message — orphans are rejected as "unexpected
+      //      tool_use_id".
+      //   2. All tool_results for one assistant turn must arrive in a SINGLE
+      //      user message — OpenAI sends parallel tool results as consecutive
+      //      role:"tool" messages, which must merge, not split.
+      const lastAssistant = anthropicMessages.findLast?.(m => m.role === "assistant")
+        ?? [...anthropicMessages].reverse().find(m => m.role === "assistant");
+      const hasMatchingToolUse = Array.isArray(lastAssistant?.content) &&
+        lastAssistant.content.some(b => b?.type === "tool_use" && b.id === msg.tool_call_id);
+
+      if (!hasMatchingToolUse) {
+        // Orphaned result (history truncated or ids rewritten by the client):
+        // degrade to plain text so the request survives instead of 400ing.
+        logger.warn({ toolCallId: msg.tool_call_id }, "Orphaned tool result — no matching tool_use in prior assistant message; converting to text");
+        const text = typeof msg.content === "string" ? msg.content : JSON.stringify(msg.content);
+        anthropicMessages.push({
+          role: "user",
+          content: [{ type: "text", text: `[Tool result ${msg.tool_call_id}]: ${text}` }]
+        });
+        continue;
+      }
+
+      const resultBlock = {
+        type: "tool_result",
+        tool_use_id: msg.tool_call_id,
+        content: msg.content
+      };
+
+      const prev = anthropicMessages[anthropicMessages.length - 1];
+      const prevIsToolResultCarrier = prev?.role === "user" &&
+        Array.isArray(prev.content) &&
+        prev.content.length > 0 &&
+        prev.content.every(b => b?.type === "tool_result");
+
+      if (prevIsToolResultCarrier) {
+        prev.content.push(resultBlock);
+      } else {
+        anthropicMessages.push({ role: "user", content: [resultBlock] });
+      }
     }
   }
 
@@ -261,7 +288,10 @@ function convertAnthropicToOpenAI(anthropicResponse, model = "claude-3-5-sonnet-
 
   // Build OpenAI response
   // Ensure ID has the chatcmpl- prefix that OpenAI clients expect
-  const responseId = id && id.startsWith("chatcmpl-") ? id : `chatcmpl-${Date.now()}`;
+  // Preserve the upstream id whatever its prefix (msg_… from Anthropic-shaped
+  // providers) — clients don't validate the prefix and a stable id keeps
+  // request tracing intact across the format conversion.
+  const responseId = id || `chatcmpl-${Date.now()}`;
   const openaiResponse = {
     id: responseId,
     object: "chat.completion",
@@ -277,11 +307,23 @@ function convertAnthropicToOpenAI(anthropicResponse, model = "claude-3-5-sonnet-
         finish_reason: mapStopReason(stop_reason)
       }
     ],
-    usage: {
-      prompt_tokens: usage?.input_tokens || 0,
-      completion_tokens: usage?.output_tokens || 0,
-      total_tokens: (usage?.input_tokens || 0) + (usage?.output_tokens || 0)
-    }
+    // OpenAI semantics: prompt_tokens INCLUDES cached tokens (Anthropic's
+    // input_tokens excludes cache reads/writes). Dropping cache tokens made
+    // clients like opencode undercount context — with prompt caching on,
+    // most of a turn's input is cache-read (issue: sidebar token drift).
+    usage: (() => {
+      const cacheRead = usage?.cache_read_input_tokens || 0;
+      const cacheWrite = usage?.cache_creation_input_tokens || 0;
+      const promptTokens = (usage?.input_tokens || 0) + cacheRead + cacheWrite;
+      const completionTokens = usage?.output_tokens || 0;
+      const u = {
+        prompt_tokens: promptTokens,
+        completion_tokens: completionTokens,
+        total_tokens: promptTokens + completionTokens
+      };
+      if (cacheRead > 0) u.prompt_tokens_details = { cached_tokens: cacheRead };
+      return u;
+    })()
   };
 
   // Add citations if present
@@ -427,10 +469,18 @@ function convertAnthropicStreamChunkToOpenAI(chunk, model = "claude-3-5-sonnet-2
           finish_reason: mapStopReason(stopReason)
         }
       ],
+      // Anthropic's message_delta usage carries output_tokens and (on newer
+      // API versions) cumulative input_tokens — forward whatever is present
+      // instead of hardcoding prompt_tokens: 0.
       usage: usage ? {
-        prompt_tokens: 0, // Not available in streaming
+        prompt_tokens: (usage.input_tokens || 0) +
+          (usage.cache_read_input_tokens || 0) +
+          (usage.cache_creation_input_tokens || 0),
         completion_tokens: usage.output_tokens || 0,
-        total_tokens: usage.output_tokens || 0
+        total_tokens: (usage.input_tokens || 0) +
+          (usage.cache_read_input_tokens || 0) +
+          (usage.cache_creation_input_tokens || 0) +
+          (usage.output_tokens || 0)
       } : undefined
     };
   } else if (eventType === "message_stop") {

package/src/clients/provider-capabilities.js

@@ -1,5 +1,3 @@
-const config = require("../config");
-
 const NATIVE_THINKING_PROVIDERS = new Set(["azure-anthropic", "databricks"]);
 
 const NATIVE_THINKING_BEDROCK_MODELS = [

package/src/clients/resilience.js

@@ -195,7 +195,6 @@ class CockatielCircuitBreaker {
     this._lastStateChange = Date.now();
 
     // Recreate the policy to reset state
-    const oldPolicy = this._policy;
     this._policy = circuitBreaker(handleAll, {
       breaker: new ConsecutiveBreaker(this.failureThreshold),
       halfOpenAfter: this.halfOpenAfter,

package/src/clients/responses-format.js

@@ -10,61 +10,6 @@
 const logger = require("../logger");
 const { repairToolCallIds } = require("./tool-call-repair");
 
-/**
- * Map client tool names back to Lynkr tool names
- * Used when processing incoming function_call messages from various AI coding clients
- * Supports: Codex CLI, Cline (VS Code), Continue.dev
- * @param {string} clientToolName - Client tool name (e.g., shell_command, execute_command, read_file)
- * @returns {string} Lynkr tool name (e.g., Bash, Read)
- */
-function mapClientToolToLynkr(clientToolName) {
-  const reverseMapping = {
-    // ============== CODEX CLI ==============
-    "shell": "Bash",
-    "shell_command": "Bash",
-    "read_file": "Read",
-    "write_file": "Write",
-    "apply_patch": "Edit",
-    "glob_file_search": "Glob",
-    "rg": "Grep",
-    "list_dir": "ListDir",
-
-    // ============== CLINE (VS Code) ==============
-    "execute_command": "Bash",
-    // "read_file" already mapped above
-    "write_to_file": "Write",
-    "replace_in_file": "Edit",
-    "search_files": "Grep",
-    "list_files": "ListDir",
-
-    // ============== KILO CODE (Fork of Cline) ==============
-    // Most tools same as Cline, but apply_diff is different
-    "apply_diff": "Edit",
-    "codebase_search": "Grep",
-    "delete_file": "Bash",  // No direct equivalent, use Bash rm
-    "browser_action": "WebFetch",  // Approximate mapping
-
-    // ============== CONTINUE.DEV ==============
-    "run_terminal_command": "Bash",
-    // "read_file" already mapped above
-    "create_new_file": "Write",
-    "edit_existing_file": "Edit",
-    "exact_search": "Grep",
-    "read_currently_open_file": "Read",
-
-    // ============== Lowercase Lynkr tools (pass-through) ==============
-    "bash": "Bash",
-    "read": "Read",
-    "write": "Write",
-    "edit": "Edit",
-    "glob": "Glob",
-    "grep": "Grep",
-    "listdir": "ListDir"
-  };
-
-  return reverseMapping[clientToolName] || clientToolName;
-}
-
 /**
  * Convert Responses API request to Chat Completions format
  * @param {Object} responsesRequest - Responses API format request
@@ -189,7 +134,6 @@ function convertResponsesToChat(responsesRequest) {
             }, "Converted multimodal content array to string");
           } else {
             // No text found, keep as array (might be image-only)
-            content = content;
           }
         }
 

package/src/clients/routing.js

@@ -15,7 +15,7 @@ const config = require('../config');
 
 // Synchronous version for benchmarking/tests
 // (when tiers are disabled, routing is purely static)
-function determineProviderSync(payload) {
+function determineProviderSync(_payload) {
   const primaryProvider = config.modelProvider?.type || 'databricks';
   const defaultModel = config.modelProvider?.defaultModel || 'databricks-claude-sonnet-4-5';
 

package/src/config/index.js

@@ -545,7 +545,6 @@ const agentsEnabled = process.env.AGENTS_ENABLED === "true";
 // LLM Audit logging configuration
 const auditEnabled = process.env.LLM_AUDIT_ENABLED === "true"; // default false
 const auditLogFile = process.env.LLM_AUDIT_LOG_FILE ?? path.join(process.cwd(), "logs", "llm-audit.log");
-const auditMaxContentLength = Number.parseInt(process.env.LLM_AUDIT_MAX_CONTENT_LENGTH ?? "5000", 10); // Legacy fallback
 const auditMaxSystemLength = Number.parseInt(process.env.LLM_AUDIT_MAX_SYSTEM_LENGTH ?? "2000", 10);
 const auditMaxUserLength = Number.parseInt(process.env.LLM_AUDIT_MAX_USER_LENGTH ?? "3000", 10);
 const auditMaxResponseLength = Number.parseInt(process.env.LLM_AUDIT_MAX_RESPONSE_LENGTH ?? "3000", 10);

package/src/context/budget.js

@@ -223,7 +223,7 @@ function finalizeBudgetEnforcement(original, optimized, strategy, initialCheck,
 /**
  * Compress system prompt aggressively for budget enforcement
  */
-function compressSystemPromptAggressively(systemPromptContent, payload) {
+function compressSystemPromptAggressively(systemPromptContent, _payload) {
   let text = typeof systemPromptContent === 'string'
     ? systemPromptContent
     : systemPrompt.flattenBlocks(systemPromptContent);

package/src/context/tool-result-compressor.js

@@ -696,7 +696,7 @@ function compressDirectoryListing(text, opts = {}) {
   // skips this check — real listings can contain names (spaces, unicode)
   // the shape regex would reject.
   if (!opts.trusted) {
-    const pathLines = lines.filter(l => /^[.\w\/-]+\.\w+$/.test(l.trim()) || /^[.\w\/-]+\/$/.test(l.trim()) || /^[-drwx]{10}/.test(l.trim()));
+    const pathLines = lines.filter(l => /^[.\w/-]+\.\w+$/.test(l.trim()) || /^[.\w/-]+\/$/.test(l.trim()) || /^[-drwx]{10}/.test(l.trim()));
     if (pathLines.length < lines.length * 0.6) return null;
   }
 
@@ -823,7 +823,7 @@ function compressLintOutput(text, opts = {}) {
   for (const line of clean.split("\n")) {
     // ESLint stylish prints each file as a standalone path line above its
     // issue block — track it so issues group per file.
-    if (/^\S+[\/\\]\S+\.\w+$/.test(line.trim()) && !line.includes(":")) {
+    if (/^\S+[/\\]\S+\.\w+$/.test(line.trim()) && !line.includes(":")) {
       currentFile = line.trim();
       continue;
     }
@@ -935,8 +935,6 @@ function compressLargeFile(text) {
 
   // Extract structural skeleton
   const skeleton = [];
-  let inBlock = false;
-  let braceDepth = 0;
 
   for (let i = 0; i < lines.length; i++) {
     const line = lines[i];
@@ -960,7 +958,7 @@ function compressLargeFile(text) {
     }
 
     // Keep decorators/attributes
-    if (/^[@#\[]/.test(trimmed)) {
+    if (/^[@#[]/.test(trimmed)) {
       skeleton.push(line);
       continue;
     }

package/src/dashboard/api.js

@@ -2,7 +2,6 @@ const config = require('../config');
 const telemetry = require('../routing/telemetry');
 const { getUsage } = require('../usage/aggregator');
 const metrics = require('../metrics');
-const { getMetricsCollector } = require('../observability/metrics');
 const { TIER_DEFINITIONS } = require('../routing/model-tiers');
 
 // Per-provider type + whether its credentials/endpoint are actually present.

package/src/headroom/client.js

@@ -115,27 +115,48 @@ async function compressMessages(messages, tools = [], options = {}) {
   }
 
   try {
-    const controller = new AbortController();
-    const timeout = setTimeout(() => controller.abort(), headroomConfig.timeoutMs);
-
-    const response = await fetch(`${headroomConfig.endpoint}/compress`, {
-      method: "POST",
-      headers: { "Content-Type": "application/json" },
-      body: JSON.stringify({
-        messages,
-        tools,
-        model: options.model || "claude-3-5-sonnet-20241022",
-        model_limit: options.modelLimit || 200000,
-        mode: options.mode || headroomConfig.mode,
-        token_budget: options.tokenBudget,
-        query_context: options.queryContext,
-        preserve_recent_turns: options.preserveRecentTurns,
-        target_ratio: options.targetRatio,
-      }),
-      signal: controller.signal,
+    const requestBody = JSON.stringify({
+      messages,
+      tools,
+      model: options.model || "claude-3-5-sonnet-20241022",
+      model_limit: options.modelLimit || 200000,
+      mode: options.mode || headroomConfig.mode,
+      token_budget: options.tokenBudget,
+      query_context: options.queryContext,
+      preserve_recent_turns: options.preserveRecentTurns,
+      target_ratio: options.targetRatio,
     });
 
-    clearTimeout(timeout);
+    // The sidecar (uvicorn) closes idle keep-alive sockets after ~5s, and
+    // undici can reuse a pooled socket in that exact window → ECONNRESET
+    // surfaces as "fetch failed". One retry gets a fresh connection, so the
+    // race never costs us the compression. Aborts (real timeouts) still
+    // propagate immediately.
+    let response;
+    for (let attempt = 0; ; attempt++) {
+      const controller = new AbortController();
+      const timeout = setTimeout(() => controller.abort(), headroomConfig.timeoutMs);
+      try {
+        response = await fetch(`${headroomConfig.endpoint}/compress`, {
+          method: "POST",
+          headers: { "Content-Type": "application/json" },
+          body: requestBody,
+          signal: controller.signal,
+        });
+        break;
+      } catch (fetchErr) {
+        if (attempt === 0 && fetchErr.name !== "AbortError") {
+          logger.debug(
+            { error: fetchErr.message, cause: fetchErr.cause?.code || fetchErr.cause?.message },
+            "Headroom fetch failed, retrying on a fresh connection"
+          );
+          continue;
+        }
+        throw fetchErr;
+      } finally {
+        clearTimeout(timeout);
+      }
+    }
 
     if (!response.ok) {
       const errorText = await response.text();
@@ -179,7 +200,10 @@ async function compressMessages(messages, tools = [], options = {}) {
     if (err.name === "AbortError") {
       logger.warn({ timeoutMs: headroomConfig.timeoutMs }, "Headroom compression timed out");
     } else {
-      logger.warn({ error: err.message }, "Headroom compression failed, using original");
+      logger.warn(
+        { error: err.message, cause: err.cause?.code || err.cause?.message },
+        "Headroom compression failed, using original"
+      );
     }
 
     return {

package/src/logger/oversized-error-stream.js

@@ -66,7 +66,7 @@ function createOversizedErrorStream(config) {
 		},
 		final(callback) {
 			// Close all open file handles when stream ends
-			for (const [sessionId, { writeStream }] of sessionFiles.entries()) {
+			for (const [_sessionId, { writeStream }] of sessionFiles.entries()) {
 				writeStream.end();
 			}
 			sessionFiles.clear();

package/src/memory/extractor.js

@@ -26,14 +26,14 @@ const EXTRACTION_PATTERNS = {
   ],
 
   entity: [
-    /(?:file|function|class|module|component|package|library)\s+['"`]?([A-Za-z0-9_./\-]+)['"`]?/gi,
-    /in\s+['"`]([A-Za-z0-9_./\-]+\.(?:js|ts|py|java|go|rs|cpp|c|h))['"`]/gi,
+    /(?:file|function|class|module|component|package|library)\s+['"`]?([A-Za-z0-9_./-]+)['"`]?/gi,
+    /in\s+['"`]([A-Za-z0-9_./-]+\.(?:js|ts|py|java|go|rs|cpp|c|h))['"`]/gi,
     /`([A-Za-z0-9_]+(?:\.[A-Za-z0-9_]+)*)`/g, // Code references
   ],
 
   relationship: [
     /([A-Za-z0-9_]+)\s+(?:depends on|imports|uses|extends|implements|inherits from)\s+([A-Za-z0-9_]+)/gi,
-    /([A-Za-z0-9_./\-]+)\s+(?:calls|invokes|references)\s+([A-Za-z0-9_./\-]+)/gi,
+    /([A-Za-z0-9_./-]+)\s+(?:calls|invokes|references)\s+([A-Za-z0-9_./-]+)/gi,
   ],
 };
 

package/src/memory/search.js

@@ -30,7 +30,7 @@ function sanitizeKeyword(keyword) {
   
   // Remove ALL FTS5 special characters
   const sanitized = keyword
-    .replace(/[*()<>\-:\[\]{}|^~,.;!?'"@#$%&+=/\\\\]/g, ' ')
+    .replace(/[*()<>\-:[\]{}|^~,.;!?'"@#$%&+=/\\\\]/g, ' ')
     .replace(/\s+/g, ' ')
     .trim();
   
@@ -97,7 +97,7 @@ function prepareFTS5Query(query) {
   // - @ symbol → "syntax error near @"
   // - Quotes (") → can break string quoting
   // - Parentheses, brackets → break grouping syntax
-  cleaned = cleaned.replace(/[*()<>\-:\[\]{}|^~,.;!?'"@#$%&+=/\\\\]/g, ' ');
+  cleaned = cleaned.replace(/[*()<>\-:[\]{}|^~,.;!?'"@#$%&+=/\\\\]/g, ' ');
   cleaned = cleaned.replace(/\s+/g, ' ').trim();
 
   if (!cleaned) {

package/src/memory/store.js

@@ -122,12 +122,6 @@ const upsertEntityStmt = db.prepare(`
     properties = @properties
 `);
 
-const getEntityStmt = db.prepare(`
-  SELECT id, entity_type, entity_name, first_seen_at, last_seen_at, occurrence_count, properties
-  FROM memory_entities
-  WHERE entity_type = ? AND entity_name = ?
-`);
-
 const getAllEntitiesStmt = db.prepare(`
   SELECT id, entity_type, entity_name, first_seen_at, last_seen_at, occurrence_count, properties
   FROM memory_entities
@@ -176,19 +170,6 @@ function toMemory(row) {
   };
 }
 
-function toEntity(row) {
-  if (!row) return null;
-  return {
-    id: row.id,
-    entityType: row.entity_type,
-    entityName: row.entity_name,
-    firstSeenAt: row.first_seen_at,
-    lastSeenAt: row.last_seen_at,
-    occurrenceCount: row.occurrence_count ?? 1,
-    properties: parseJSON(row.properties, {}),
-  };
-}
-
 /**
  * Create a new memory
  */
@@ -400,16 +381,20 @@ function trackEntity(options) {
 }
 
 /**
- * Get an entity
+ * Get an entity by name, optionally scoped to a type. Names can collide
+ * across types (e.g. a "file" and a "topic" both named "router"), so
+ * without a type filter the most recently seen entity wins,
+ * deterministically.
  */
-function getEntity(name) {
-  // Since we only have the name, we need to search across all entity types
+function getEntity(name, type = null) {
   const stmt = db.prepare(`
     SELECT id, entity_type, entity_name, first_seen_at, last_seen_at, occurrence_count, properties
     FROM memory_entities
-    WHERE entity_name = ?
+    WHERE entity_name = ? ${type ? "AND entity_type = ?" : ""}
+    ORDER BY last_seen_at DESC
+    LIMIT 1
   `);
-  const row = stmt.get(name);
+  const row = type ? stmt.get(name, type) : stmt.get(name);
   if (!row) return null;
 
   return {

package/src/memory/tools.js

@@ -340,6 +340,19 @@ async function memory_stats(args, context = {}) {
       };
     }
 
+    // Surface tracked entities — the extractor writes them on every
+    // conversation, but until this they were collected and never readable.
+    let entities = [];
+    try {
+      entities = store.getAllEntities(10).map(e => ({
+        name: e.name,
+        type: e.type,
+        count: e.count,
+      }));
+    } catch (err) {
+      logger.debug({ err: err.message }, 'Entity stats unavailable');
+    }
+
     return {
       ok: true,
       content: JSON.stringify({
@@ -350,6 +363,7 @@ async function memory_stats(args, context = {}) {
         recentCount: stats.recentCount,
         importantCount: stats.importantCount,
         sessionId: stats.sessionId || 'global',
+        topEntities: entities,
       }, null, 2),
     };
   } catch (err) {

package/src/observability/metrics.js

@@ -1,6 +1,3 @@
-const config = require("../config");
-const logger = require("../logger");
-
 /**
  * High-performance metrics collection
  *

package/src/orchestrator/index.js

@@ -13,12 +13,11 @@ const { applyToonCompression } = require("../context/toon");
 const { applyGcfCompression } = require("../context/gcf");
 const { compressMessages: headroomCompress, isEnabled: isHeadroomEnabled } = require("../headroom");
 const { createAuditLogger } = require("../logger/audit-logger");
-const { getResolvedIp, runWithDnsContext } = require("../clients/dns-logger");
 const { getShuttingDown } = require("../api/health");
 const { tryPreflight, buildSatisfiedResponse: buildPreflightResponse } = require("./preflight");
 const { detectBypass, buildBypassResponse } = require("./bypass");
 const crypto = require("crypto");
-const { getSemanticCache, isSemanticCacheEnabled } = require("../cache/semantic");
+const { getSemanticCache } = require("../cache/semantic");
 const { areSimilarToolCalls } = require("../clients/gpt-utils");
 const { getModelRegistrySync } = require("../routing/model-registry");
 const sessionAffinity = require("../routing/session-affinity");
@@ -416,37 +415,6 @@ function countToolCallsInHistory(messages) {
   return { toolUseCount, toolResultCount, lastUserTextIndex };
 }
 
-/**
- * Inject a "stop looping" instruction if there are too many tool calls in history.
- * This helps prevent infinite loops when the model keeps calling tools instead of responding.
- *
- * @param {Array} messages - The conversation messages
- * @param {number} threshold - Max tool results before injection (default: 5)
- * @returns {Array} - Messages with stop instruction injected if needed
- */
-function injectToolLoopStopInstruction(messages, threshold = 5) {
-  if (!Array.isArray(messages)) return messages;
-
-  const { toolResultCount } = countToolCallsInHistory(messages);
-
-  if (toolResultCount >= threshold) {
-    logger.warn({
-      toolResultCount,
-      threshold,
-    }, "[ToolLoopGuard] Too many tool results in conversation - injecting stop instruction");
-
-    // Inject instruction to stop tool calls and provide a final answer
-    const stopInstruction = {
-      role: "user",
-      content: `⚠️ IMPORTANT: You have already executed ${toolResultCount} tool calls in this conversation. This is likely an infinite loop. STOP calling tools immediately and provide a direct text response to the user based on the information you have gathered. If you cannot complete the task, explain why. DO NOT call any more tools. Your response must contain your actual findings — NOT an acknowledgment or restatement of this warning.`,
-    };
-
-    return [...messages, stopInstruction];
-  }
-
-  return messages;
-}
-
 // === CROSS-REQUEST TOOL CALL DEDUP TRACKING ===
 // These helpers track tool call signatures across multiple HTTP requests within
 // the same session (client/passthrough mode). The inner-loop detection in
@@ -718,52 +686,6 @@ function normaliseToolChoice(choice) {
   return undefined;
 }
 
-/**
- * Strip thinking-style reasoning from Ollama model outputs
- * Patterns to remove:
- * - Lines starting with bullet points (●, •, -, *)
- * - Explanatory reasoning before the actual response
- * - Multiple newlines used to separate thinking from response
- */
-function stripThinkingBlocks(text) {
-  if (typeof text !== "string") return text;
-
-  const lines = text.split("\n");
-  const cleanedLines = [];
-  let inThinkingBlock = false;
-  let consecutiveEmptyLines = 0;
-
-  for (const line of lines) {
-    const trimmed = line.trim();
-
-    // Detect thinking block markers (bullet points followed by reasoning)
-    if (/^[●•\-\*]\s/.test(trimmed)) {
-      inThinkingBlock = true;
-      continue;
-    }
-
-    // Empty lines might separate thinking from response
-    if (trimmed === "") {
-      consecutiveEmptyLines++;
-      // If we've seen 2+ empty lines, likely end of thinking block
-      if (consecutiveEmptyLines >= 2) {
-        inThinkingBlock = false;
-      }
-      continue;
-    }
-
-    consecutiveEmptyLines = 0;
-
-    if (inThinkingBlock) {
-      continue;
-    }
-
-    cleanedLines.push(line);
-  }
-
-  return cleanedLines.join("\n").trim();
-}
-
 /**
  * Convert legacy Ollama /api/chat response to Anthropic Messages format.
  * Used when Ollama < v0.14.0 (no native Anthropic endpoint).
@@ -1392,7 +1314,6 @@ async function runAgentLoop({
   requestedModel,
   wantsThinking,
   session,
-  cwd,
   options,
   cacheKey,
   providerType,