lynkr 9.14.0 → 9.14.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "lynkr",
-  "version": "9.14.0",
+  "version": "9.14.2",
   "files": [
     "index.js",
     "install.sh",
@@ -38,12 +38,12 @@
     "start:supervised": "while true; do node index.js 2>&1 | npx pino-pretty --sync; code=$?; echo \"[supervisor] lynkr exited ($code) — restarting in 3s\"; sleep 3; done",
     "lint": "eslint src index.js",
     "test": "npm run test:unit && npm run test:performance",
-    "test:unit": "DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com LOG_FILE_ENABLED=false node --test test/routing.test.js test/hybrid-routing-integration.test.js test/retry-logic.test.js test/sse-transformer.test.js test/passthrough-stream.test.js test/passthrough-mode.test.js test/openrouter-error-resilience.test.js test/format-conversion.test.js test/azure-openai-config.test.js test/azure-openai-format-conversion.test.js test/azure-openai-routing.test.js test/azure-openai-streaming.test.js test/azure-openai-error-resilience.test.js test/azure-openai-integration.test.js test/openai-integration.test.js test/atlas-integration.test.js test/toon-compression.test.js test/gcf-compression.test.js test/llamacpp-integration.test.js test/resilience.test.js test/telemetry-routing.test.js test/memory/store.test.js test/memory/surprise.test.js test/memory/extractor.test.js test/memory/search.test.js test/memory/retriever.test.js test/memory/distiller.test.js test/memory/distiller-freeze.test.js test/memory/wiki.test.js test/memory/skills-cache.test.js test/memory/tencentdb-launcher.test.js test/distill.test.js test/large-payload.test.js test/prompt-cache-injection.test.js test/risk-analyzer.test.js test/interaction-block.test.js test/preflight.test.js test/token-reduction.test.js test/session-affinity.test.js test/cache-state.test.js test/cache-switch-cost.test.js test/lens-recommendations.test.js test/model-registry-cost.test.js test/output-format-guard.test.js test/tier-fallback.test.js test/wrap.test.js test/init.test.js test/tool-call-response-metadata.test.js test/degradation.test.js test/routing-telemetry-columns.test.js test/sticky-routing.test.js test/knn-ambiguous-escalate.test.js test/deescalator.test.js test/client-profiles.test.js test/strip-internal-fields.test.js test/complexity-tool-subtraction.test.js test/bandit.test.js test/routing-propensity.test.js test/reward-pipeline.test.js test/knn-cold-start.test.js test/calibration.test.js test/feedback-loop.test.js test/session-fingerprint.test.js test/side-request-guards.test.js test/verifier.test.js test/intent-score.test.js test/difficulty-classifier.test.js test/classifier-setup.test.js test/usage-stats.test.js test/loop-guard.test.js test/moonshot-model-mapping.test.js test/baidu-model-mapping.test.js",
-    "test:memory": "DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/memory/store.test.js test/memory/surprise.test.js test/memory/extractor.test.js test/memory/search.test.js test/memory/retriever.test.js test/memory/distiller.test.js test/memory/distiller-freeze.test.js test/memory/wiki.test.js test/memory/skills-cache.test.js test/memory/tencentdb-launcher.test.js",
-    "test:new-features": "DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/passthrough-mode.test.js test/openrouter-error-resilience.test.js test/format-conversion.test.js",
-    "test:performance": "DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/hybrid-routing-performance.test.js && DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/performance-tests.js",
-    "test:benchmark": "DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/performance-benchmark.js",
-    "test:quick": "DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/routing.test.js",
+    "test:unit": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com LOG_FILE_ENABLED=false node --test test/routing.test.js test/hybrid-routing-integration.test.js test/retry-logic.test.js test/sse-transformer.test.js test/passthrough-stream.test.js test/passthrough-mode.test.js test/openrouter-error-resilience.test.js test/format-conversion.test.js test/azure-openai-config.test.js test/azure-openai-format-conversion.test.js test/azure-openai-routing.test.js test/azure-openai-streaming.test.js test/azure-openai-error-resilience.test.js test/azure-openai-integration.test.js test/openai-integration.test.js test/atlas-integration.test.js test/toon-compression.test.js test/gcf-compression.test.js test/llamacpp-integration.test.js test/resilience.test.js test/telemetry-routing.test.js test/memory/store.test.js test/memory/surprise.test.js test/memory/extractor.test.js test/memory/search.test.js test/memory/retriever.test.js test/memory/distiller.test.js test/memory/distiller-freeze.test.js test/memory/wiki.test.js test/memory/skills-cache.test.js test/memory/tencentdb-launcher.test.js test/distill.test.js test/large-payload.test.js test/prompt-cache-injection.test.js test/risk-analyzer.test.js test/interaction-block.test.js test/preflight.test.js test/token-reduction.test.js test/session-affinity.test.js test/cache-state.test.js test/cache-switch-cost.test.js test/lens-recommendations.test.js test/model-registry-cost.test.js test/output-format-guard.test.js test/tier-fallback.test.js test/wrap.test.js test/init.test.js test/tool-call-response-metadata.test.js test/degradation.test.js test/routing-telemetry-columns.test.js test/sticky-routing.test.js test/knn-ambiguous-escalate.test.js test/deescalator.test.js test/client-profiles.test.js test/strip-internal-fields.test.js test/complexity-tool-subtraction.test.js test/bandit.test.js test/routing-propensity.test.js test/reward-pipeline.test.js test/knn-cold-start.test.js test/calibration.test.js test/feedback-loop.test.js test/session-fingerprint.test.js test/side-request-guards.test.js test/verifier.test.js test/intent-score.test.js test/difficulty-classifier.test.js test/classifier-setup.test.js test/usage-stats.test.js test/loop-guard.test.js test/moonshot-model-mapping.test.js test/baidu-model-mapping.test.js",
+    "test:memory": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/memory/store.test.js test/memory/surprise.test.js test/memory/extractor.test.js test/memory/search.test.js test/memory/retriever.test.js test/memory/distiller.test.js test/memory/distiller-freeze.test.js test/memory/wiki.test.js test/memory/skills-cache.test.js test/memory/tencentdb-launcher.test.js",
+    "test:new-features": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/passthrough-mode.test.js test/openrouter-error-resilience.test.js test/format-conversion.test.js",
+    "test:performance": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/hybrid-routing-performance.test.js && DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/performance-tests.js",
+    "test:benchmark": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/performance-benchmark.js",
+    "test:quick": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/routing.test.js",
     "test:all": "npm run test:unit && npm run test:performance && npm run test:benchmark"
   },
   "keywords": [
@@ -90,32 +90,22 @@
     "node": ">=20.0.0"
   },
   "dependencies": {
-    "@azure/openai": "^2.0.0",
-    "@babel/parser": "^7.29.0",
-    "@babel/traverse": "^7.29.0",
     "@blackwell-systems/gcf": "2.4.0",
-    "@tencentdb-agent-memory/memory-tencentdb": "^1.0.1",
     "@toon-format/toon": "^2.1.0",
     "cockatiel": "^3.2.1",
-    "compression": "^1.7.4",
     "diff": "^5.2.0",
     "dotenv": "^16.4.5",
     "express": "^5.1.0",
     "express-rate-limit": "^8.2.1",
-    "fast-glob": "^3.3.2",
-    "graphify": "^1.0.0",
     "hnswlib-node": "^3.0.0",
     "js-tiktoken": "^1.0.20",
-    "js-yaml": "^4.1.1",
-    "openai": "^6.14.0",
     "pino": "^8.17.2",
     "pino-http": "^8.6.0",
-    "pino-roll": "^4.0.0",
-    "undici": "^6.22.0"
+    "pino-roll": "^4.0.0"
   },
   "optionalDependencies": {
     "better-sqlite3": "^12.11.1",
-    "dockerode": "^4.0.2",
+    "dockerode": "^5.0.1",
     "tree-sitter": "^0.21.1",
     "tree-sitter-javascript": "^0.21.0",
     "tree-sitter-python": "^0.21.0",
@@ -127,7 +117,5 @@
     "nodemon": "^3.1.0",
     "pino-pretty": "^10.2.0"
   },
-  "allowScripts": {
-    "@tencentdb-agent-memory/memory-tencentdb@1.0.1": true
-  }
+  "allowScripts": {}
 }

package/src/api/health.js

@@ -4,7 +4,6 @@ try {
 } catch {
   Database = null;
 }
-const { invokeModel } = require("../clients/databricks");
 const logger = require("../logger");
 const config = require("../config");
 const { getHeadroomManager } = require("../headroom");

package/src/api/openai-router.js

@@ -355,12 +355,36 @@ async function forwardAnthropicStreamAsOpenAIChunks(res, stream, requestedModel)
   const created = Math.floor(Date.now() / 1000);
   const model = requestedModel || "lynkr";
 
-  const chunk = (delta, finish = null) => `data: ${JSON.stringify({
+  const chunk = (delta, finish = null, usage = null) => `data: ${JSON.stringify({
     id, object: "chat.completion.chunk", created, model,
     system_fingerprint: "fp_lynkr",
     choices: [{ index: 0, delta, logprobs: null, finish_reason: finish }],
+    ...(usage ? { usage } : {}),
   })}\n\n`;
 
+  // Usage accounting (issue: opencode sidebar showed wrong token counts —
+  // this path never emitted usage at all). Anthropic sends input/cache
+  // tokens in message_start and output tokens in message_delta; accumulate
+  // and attach to the final chunk in OpenAI shape.
+  const usageAcc = { input: 0, cacheRead: 0, cacheWrite: 0, output: 0 };
+  const captureUsage = (u) => {
+    if (!u) return;
+    if (u.input_tokens != null) usageAcc.input = u.input_tokens;
+    if (u.cache_read_input_tokens != null) usageAcc.cacheRead = u.cache_read_input_tokens;
+    if (u.cache_creation_input_tokens != null) usageAcc.cacheWrite = u.cache_creation_input_tokens;
+    if (u.output_tokens != null) usageAcc.output = u.output_tokens;
+  };
+  const finalUsage = () => {
+    const promptTokens = usageAcc.input + usageAcc.cacheRead + usageAcc.cacheWrite;
+    const u = {
+      prompt_tokens: promptTokens,
+      completion_tokens: usageAcc.output,
+      total_tokens: promptTokens + usageAcc.output,
+    };
+    if (usageAcc.cacheRead > 0) u.prompt_tokens_details = { cached_tokens: usageAcc.cacheRead };
+    return u;
+  };
+
   const mapStop = (r) => r === "tool_use" ? "tool_calls" : r === "max_tokens" ? "length" : "stop";
 
   // Normalize Node Readable / web ReadableStream / async iterable to chunks.
@@ -394,6 +418,7 @@ async function forwardAnthropicStreamAsOpenAIChunks(res, stream, requestedModel)
 
         switch (ev.type) {
           case "message_start":
+            captureUsage(ev.message?.usage);
             if (!started) { started = true; res.write(chunk({ role: "assistant", content: "" })); }
             break;
           case "content_block_start":
@@ -419,10 +444,11 @@ async function forwardAnthropicStreamAsOpenAIChunks(res, stream, requestedModel)
             break;
           }
           case "message_delta":
+            captureUsage(ev.usage);
             if (ev.delta?.stop_reason) finishReason = mapStop(ev.delta.stop_reason);
             break;
           case "message_stop":
-            res.write(chunk({}, finishReason || "stop"));
+            res.write(chunk({}, finishReason || "stop", finalUsage()));
             res.write("data: [DONE]\n\n");
             res.end();
             return;
@@ -447,7 +473,7 @@ async function forwardAnthropicStreamAsOpenAIChunks(res, stream, requestedModel)
     // truncated stream: send an error chunk, never a fabricated "stop".
     if (!res.writableEnded) {
       if (finishReason) {
-        res.write(chunk({}, finishReason));
+        res.write(chunk({}, finishReason, finalUsage()));
         res.write("data: [DONE]\n\n");
       } else {
         logger.warn("Upstream stream ended without message_stop — sending error chunk to OpenAI-surface client");
@@ -468,7 +494,11 @@ async function forwardAnthropicStreamAsOpenAIChunks(res, stream, requestedModel)
 
 router.post("/chat/completions", async (req, res) => {
   const startTime = Date.now();
-  const sessionId = req.headers["x-session-id"] || req.headers["authorization"]?.split(" ")[1] || "openai-session";
+  // Issue #95: use sessionMiddleware's id (header → body id → content
+  // fingerprint → UUID). Falling back to the raw Authorization token here
+  // collapsed every conversation from a static-key client into ONE session
+  // (defeating WS1 pinning) and persisted the bearer token as a store key.
+  const sessionId = req.sessionId;
 
   try {
     if (!req.body || typeof req.body !== 'object') {
@@ -730,7 +760,10 @@ router.post("/chat/completions", async (req, res) => {
             delta: {},
             logprobs: null,
             finish_reason: openaiResponse.choices[0].finish_reason
-          }]
+          }],
+          // Clients (opencode et al.) read token accounting from the final
+          // chunk — omitting it here left streamed turns uncounted.
+          usage: openaiResponse.usage
         };
 
         logger.debug({ chunk: "finish", finishReason: openaiResponse.choices[0].finish_reason }, "Sending finish chunk");
@@ -835,6 +868,24 @@ router.post("/chat/completions", async (req, res) => {
         finishReason: openaiResponse.choices[0].finish_reason
       }, "=== OPENAI CHAT COMPLETION RESPONSE ===");
 
+      // Surface the routed reality: which model actually answered and what
+      // it cost (LiteLLM-style headers). Clients that can't price a dynamic
+      // router (opencode #24113) get the truth out-of-band.
+      const routingMeta = result.body?._routingMeta;
+      if (routingMeta?.model) {
+        res.set("x-lynkr-routed-model", String(routingMeta.model));
+        if (routingMeta.provider) res.set("x-lynkr-routed-provider", String(routingMeta.provider));
+        try {
+          const { computeCostUsd } = require("../clients/databricks");
+          const cost = computeCostUsd(
+            routingMeta.model,
+            result.body?.usage?.input_tokens || 0,
+            result.body?.usage?.output_tokens || 0
+          );
+          if (cost != null) res.set("x-lynkr-response-cost", cost.toFixed(6));
+        } catch { /* cost pricing unavailable — headers are best-effort */ }
+      }
+
       res.json(openaiResponse);
     }
 
@@ -1039,7 +1090,6 @@ function getConfiguredProviders() {
 router.get("/models", (req, res) => {
   try {
     const providers = getConfiguredProviders();
-    const primaryProvider = config.modelProvider?.type || "databricks";
     const timestamp = Math.floor(Date.now() / 1000);
     const models = [];
     const seenModelIds = new Set();
@@ -1476,7 +1526,7 @@ router.post("/embeddings", async (req, res) => {
   const startTime = Date.now();
 
   try {
-    const { input, model, encoding_format } = req.body;
+    const { input, model, encoding_format: encodingFormat } = req.body;
 
     if (!input) {
       return res.status(400).json({
@@ -1488,6 +1538,16 @@ router.post("/embeddings", async (req, res) => {
       });
     }
 
+    if (encodingFormat && encodingFormat !== "float" && encodingFormat !== "base64") {
+      return res.status(400).json({
+        error: {
+          message: `Invalid encoding_format: ${encodingFormat}. Must be "float" or "base64".`,
+          type: "invalid_request_error",
+          code: "invalid_parameter"
+        }
+      });
+    }
+
     const inputs = Array.isArray(input) ? input : [input];
 
     logger.debug({
@@ -1561,6 +1621,20 @@ router.post("/embeddings", async (req, res) => {
       totalTokens: embeddingResponse.usage?.total_tokens || 0
     }, "=== EMBEDDINGS RESPONSE ===");
 
+    // Providers return float arrays; honor encoding_format=base64 (OpenAI
+    // spec: base64 of the little-endian float32 buffer) instead of silently
+    // returning floats for every request.
+    if (encodingFormat === "base64" && Array.isArray(embeddingResponse.data)) {
+      for (const item of embeddingResponse.data) {
+        if (Array.isArray(item.embedding)) {
+          item.embedding = Buffer.from(new Float32Array(item.embedding).buffer).toString("base64");
+        }
+      }
+      if ("encoding_format" in embeddingResponse) {
+        embeddingResponse.encoding_format = "base64";
+      }
+    }
+
     res.json(embeddingResponse);
 
   } catch (error) {
@@ -1588,7 +1662,10 @@ router.post("/embeddings", async (req, res) => {
  */
 router.post("/responses", async (req, res) => {
   const startTime = Date.now();
-  const sessionId = req.headers["x-session-id"] || req.headers["authorization"]?.split(" ")[1] || "responses-session";
+  // Issue #95: same as /chat/completions — trust sessionMiddleware, never
+  // the Authorization token. fingerprintSessionId understands the Responses
+  // API's `input` field, so fingerprinting works on this route too.
+  const sessionId = req.sessionId;
 
   try {
     const { convertResponsesToChat, convertChatToResponses } = require("../clients/responses-format");
@@ -1768,7 +1845,7 @@ router.post("/responses", async (req, res) => {
 
         // SSE comment keepalive (spec-compliant, ignored by all clients)
         const keepalive = setInterval(() => {
-          try { res.write(`: keepalive\n\n`); } catch {}
+          try { res.write(`: keepalive\n\n`); } catch { /* client gone — best-effort keepalive */ }
         }, 2000);
 
         let result;
@@ -1875,21 +1952,25 @@ router.post("/responses", async (req, res) => {
                   if (taskArgs.prompt && taskArgs.prompt.toLowerCase().includes("read")) {
                     exploreCmd = "cat README.md 2>/dev/null && cat package.json 2>/dev/null";
                   }
-                } catch {}
+                } catch (err) {
+                  logger.debug({ err: err?.message }, "Failed to parse task tool arguments — using default explore command");
+                }
               }
               converted.push({ ...tc, function: { name: "shell", arguments: JSON.stringify({ command: ["bash", "-c", exploreCmd] }) } });
               continue;
             }
 
             let args = {};
-            try { args = JSON.parse(tc.function?.arguments || "{}"); } catch {}
+            try { args = JSON.parse(tc.function?.arguments || "{}"); } catch (err) {
+              logger.debug({ err: err?.message }, "Failed to parse tool call arguments — treating as empty object");
+            }
 
             // Already a shell command — normalise to array format
             if (ln === "shell" || ln === "bash") {
               let cmd = args.command || "";
               // Handle string-encoded arrays: '["bash","-c","ls"]'
               if (typeof cmd === "string" && cmd.trim().startsWith("[")) {
-                try { cmd = JSON.parse(cmd); } catch {}
+                try { cmd = JSON.parse(cmd); } catch { /* not a JSON array — keep as plain string */ }
               }
               // Extract the actual command from ["bash", "-c", "actual command"]
               if (Array.isArray(cmd) && cmd.length >= 3 && cmd[0] === "bash" && cmd[1] === "-c") {
@@ -2100,7 +2181,7 @@ router.post("/responses", async (req, res) => {
             sequence_number: 1
           })}\n\n`
         ].join("");
-        try { res.end(errorPayload); } catch { try { res.end(); } catch {} }
+        try { res.end(errorPayload); } catch { try { res.end(); } catch { /* res already ended — best-effort */ } }
       }
 
     } else {

package/src/api/providers-handler.js

@@ -21,7 +21,6 @@ const router = express.Router();
  */
 function getConfiguredProviders() {
   const providers = [];
-  const timestamp = Math.floor(Date.now() / 1000);
 
   // Check Databricks
   if (config.databricks?.url && config.databricks?.apiKey) {

package/src/api/router.js

@@ -450,7 +450,7 @@ async function handleOauthPassthrough(req, res, opts = {}) {
       try { reader.cancel(); } catch { /* already dead */ }
       try { res.write(SSE_STALL_EVENT); } catch { /* client gone */ }
     } finally {
-      try { reader.releaseLock(); } catch {}
+      try { reader.releaseLock(); } catch { /* lock already released */ }
     }
     res.end();
   } else {
@@ -495,7 +495,7 @@ async function handleOauthPassthrough(req, res, opts = {}) {
       const tier = opts.tier || {};
       let parsedResponse = null;
       if (contentType.includes("application/json")) {
-        try { parsedResponse = JSON.parse(responseTextForObservability); } catch {}
+        try { parsedResponse = JSON.parse(responseTextForObservability); } catch { /* not JSON — telemetry proceeds without a parsed body */ }
       } else if (contentType.includes("text/event-stream")) {
         // Extract a usable response object from the SSE stream by finding the
         // final message_delta / message_stop events.
@@ -1027,7 +1027,10 @@ router.post("/v1/messages", rateLimiter, async (req, res, next) => {
     try {
       const { getMetricsCollector } = require("../observability/metrics");
       getMetricsCollector().recordRequest("POST", "/v1/messages", null, null);
-    } catch (_) {}
+    } catch (err) {
+      // If this fails silently, the wrap UI's "No requests tracked" bug returns.
+      logger.debug({ err: err?.message }, "observability collector recordRequest failed");
+    }
 
     messagesRequestCount += 1;
 

package/src/cache/prompt.js

@@ -19,13 +19,6 @@ const fs = require("fs");
 const config = require("../config");
 const logger = require("../logger");
 
-function cloneValue(value) {
-  if (typeof structuredClone === "function") {
-    return structuredClone(value);
-  }
-  return JSON.parse(JSON.stringify(value));
-}
-
 function normaliseObject(value) {
   if (value === null || typeof value !== "object") return value;
   if (Array.isArray(value)) {

package/src/clients/bedrock-utils.js

@@ -116,7 +116,7 @@ function convertAnthropicToBedrockFormat(body, modelFamily) {
         },
       };
 
-    case "llama":
+    case "llama": {
       // Meta Llama format
       const llamaPrompt = messagesToPrompt(body.messages);
       return {
@@ -125,6 +125,7 @@ function convertAnthropicToBedrockFormat(body, modelFamily) {
         temperature: body.temperature || 0.7,
         top_p: body.top_p || 0.9,
       };
+    }
 
     case "jurassic":
       // AI21 Jurassic format

package/src/clients/codex-process.js

@@ -279,7 +279,7 @@ class CodexProcess {
    * Reject all pending requests
    */
   _rejectAllPending(error) {
-    for (const [id, pending] of this.pendingRequests) {
+    for (const [_id, pending] of this.pendingRequests) {
       clearTimeout(pending.timeout);
       pending.reject(error);
     }

package/src/clients/codex-utils.js

@@ -7,8 +7,6 @@
  * @module clients/codex-utils
  */
 
-const logger = require("../logger");
-
 /**
  * Extract text content from Anthropic message format
  * Handles both string content and content block arrays

package/src/clients/databricks.js

@@ -165,7 +165,7 @@ async function performJsonRequest(url, { headers = {}, body, retryableStatusesOv
   });
 }
 
-async function invokeDatabricks(body, incomingHeaders = {}) {
+async function invokeDatabricks(body, _incomingHeaders = {}) {
   if (!config.databricks?.url) {
     throw new Error("Databricks configuration is missing required URL.");
   }
@@ -423,12 +423,20 @@ function _liftLeakedThinkingBlocks(response) {
   const newContent = [];
   let lifted = 0;
   for (const block of payload.content) {
-    if (block?.type === "text" && typeof block.text === "string" && block.text.includes("<think>")) {
+    if (block?.type === "text" && typeof block.text === "string" &&
+        (block.text.includes("<think>") || block.text.includes("</think>"))) {
       const thoughts = [];
       let m;
       while ((m = thinkRegex.exec(block.text)) !== null) thoughts.push(m[1].trim());
       thinkRegex.lastIndex = 0;
-      const cleaned = block.text.replace(thinkRegex, "").trim();
+      let cleaned = block.text.replace(thinkRegex, "").trim();
+      // Orphan closing tag: GLM-family templates open <think> in the prompt,
+      // so the completion may carry reasoning terminated by a bare </think>.
+      const orphanIdx = cleaned.indexOf("</think>");
+      if (orphanIdx !== -1) {
+        thoughts.push(cleaned.slice(0, orphanIdx).trim());
+        cleaned = cleaned.slice(orphanIdx + "</think>".length).trim();
+      }
       const merged = thoughts.filter(Boolean).join("\n\n");
       if (merged) {
         newContent.push({ type: "thinking", thinking: merged });
@@ -489,7 +497,7 @@ async function _throwIfOllamaError(response, modelName) {
   return response;
 }
 
-async function invokeOllama(body, incomingHeaders = {}) {
+async function invokeOllama(body, _incomingHeaders = {}) {
   if (!config.ollama?.endpoint) {
     throw new Error("Ollama endpoint is not configured.");
   }
@@ -754,7 +762,7 @@ async function invokeOllama(body, incomingHeaders = {}) {
   );
 }
 
-async function invokeOpenRouter(body, incomingHeaders = {}) {
+async function invokeOpenRouter(body, _incomingHeaders = {}) {
   if (!config.openrouter?.endpoint || !config.openrouter?.apiKey) {
     throw new Error("OpenRouter endpoint or API key is not configured.");
   }
@@ -791,6 +799,13 @@ async function invokeOpenRouter(body, incomingHeaders = {}) {
     stream: body.stream ?? false
   };
 
+  // Issue #95: forward the Lynkr session id so OpenRouter's meta-routers
+  // (auto/pareto) can apply session stickiness — without it every call in a
+  // pinned session re-ranks fresh and one agentic task can span 6+ models.
+  if (body._sessionId) {
+    openRouterBody.session_id = body._sessionId;
+  }
+
   // Add tools - inject standard tools if client didn't send any (passthrough mode)
   let toolsToSend = body.tools;
   let toolsInjected = false;
@@ -820,7 +835,7 @@ async function invokeOpenRouter(body, incomingHeaders = {}) {
 // Eden AI is an OpenAI-compatible gateway (provider/model naming, EU/GDPR).
 // It speaks the same wire format as OpenRouter, so this mirrors invokeOpenRouter
 // and reuses the shared Anthropic<->OpenAI converters.
-async function invokeEdenAI(body, incomingHeaders = {}) {
+async function invokeEdenAI(body, _incomingHeaders = {}) {
   if (!config.edenai?.endpoint || !config.edenai?.apiKey) {
     throw new Error("Eden AI endpoint or API key is not configured.");
   }
@@ -905,7 +920,7 @@ function derivePromptCacheKey(body) {
   return "conv-" + crypto.createHash("sha1").update(text.slice(0, 4000)).digest("hex").slice(0, 32);
 }
 
-async function invokeAzureOpenAI(body, incomingHeaders = {}) {
+async function invokeAzureOpenAI(body, _incomingHeaders = {}) {
   if (!config.azureOpenAI?.endpoint || !config.azureOpenAI?.apiKey) {
     throw new Error("Azure OpenAI endpoint or API key is not configured.");
   }
@@ -1350,7 +1365,7 @@ async function invokeAzureOpenAI(body, incomingHeaders = {}) {
 }
 
 
-async function invokeOpenAI(body, incomingHeaders = {}) {
+async function invokeOpenAI(body, _incomingHeaders = {}) {
   if (!config.openai?.apiKey) {
     throw new Error("OpenAI API key is not configured.");
   }
@@ -1488,7 +1503,7 @@ async function invokeAtlas(body) {
   }, "Atlas Cloud");
 }
 
-async function invokeLlamaCpp(body, incomingHeaders = {}) {
+async function invokeLlamaCpp(body, _incomingHeaders = {}) {
   if (!config.llamacpp?.endpoint) {
     throw new Error("llama.cpp endpoint is not configured.");
   }
@@ -1634,7 +1649,7 @@ async function invokeLlamaCpp(body, incomingHeaders = {}) {
   return result;
 }
 
-async function invokeLMStudio(body, incomingHeaders = {}) {
+async function invokeLMStudio(body, _incomingHeaders = {}) {
   if (!config.lmstudio?.endpoint) {
     throw new Error("LM Studio endpoint is not configured.");
   }
@@ -1761,7 +1776,7 @@ function normalizeBodyForConverse(body) {
   return normalized;
 }
 
-async function invokeBedrock(body, incomingHeaders = {}) {
+async function invokeBedrock(body, _incomingHeaders = {}) {
   if (!config.bedrock?.apiKey) {
     throw new Error(
       "AWS Bedrock requires AWS_BEDROCK_API_KEY (Bearer token). " +
@@ -1774,11 +1789,9 @@ async function invokeBedrock(body, incomingHeaders = {}) {
 
   // Inject standard tools if needed
   let toolsToSend = body.tools;
-  let toolsInjected = false;
 
   if (!Array.isArray(toolsToSend) || toolsToSend.length === 0) {
     toolsToSend = STANDARD_TOOLS;
-    toolsInjected = true;
     logger.debug({
       injectedToolCount: STANDARD_TOOLS.length,
       injectedToolNames: STANDARD_TOOL_NAMES,
@@ -1947,7 +1960,7 @@ async function invokeBedrock(body, incomingHeaders = {}) {
  * Z.AI offers GLM models through an Anthropic-compatible API at ~1/7 the cost.
  * Minimal transformation needed - mostly passthrough with model mapping.
  */
-async function invokeZai(body, incomingHeaders = {}) {
+async function invokeZai(body, _incomingHeaders = {}) {
   if (!config.zai?.apiKey) {
     throw new Error("Z.AI API key is not configured. Set ZAI_API_KEY in your .env file.");
   }
@@ -2152,7 +2165,7 @@ async function invokeZai(body, incomingHeaders = {}) {
  * Moonshot offers Kimi models through an OpenAI-compatible chat completions API.
  * Uses native system role support (unlike Z.AI which merges into user message).
  */
-async function invokeMoonshot(body, incomingHeaders = {}) {
+async function invokeMoonshot(body, _incomingHeaders = {}) {
   if (!config.moonshot?.apiKey) {
     throw new Error("Moonshot API key is not configured. Set MOONSHOT_API_KEY in your .env file.");
   }
@@ -2314,7 +2327,7 @@ async function invokeMoonshot(body, incomingHeaders = {}) {
  * modelMap and pinned-params logic (see Moonshot's kimi-k* precedent) once
  * real traffic surfaces 400s.
  */
-async function invokeBaidu(body, incomingHeaders = {}) {
+async function invokeBaidu(body, _incomingHeaders = {}) {
   if (!config.baidu?.apiKey) {
     throw new Error("Baidu API key is not configured. Set BAIDU_API_KEY in your .env file.");
   }
@@ -2450,12 +2463,34 @@ function convertOpenAIToAnthropic(response) {
   const hasToolCalls = Array.isArray(message.tool_calls) && message.tool_calls.length > 0;
 
   // Extract text content and reasoning from thinking models
-  const textContent = typeof message.content === 'string' ? message.content : '';
+  let textContent = typeof message.content === 'string' ? message.content : '';
   const reasoningContent = typeof message.reasoning_content === 'string' ? message.reasoning_content : '';
 
-  // Emit reasoning_content as a proper thinking block (not discarded)
-  if (reasoningContent) {
-    content.push({ type: "thinking", thinking: reasoningContent });
+  // GLM-family models leak reasoning into `content` in two shapes:
+  //   1. matched <think>...</think> pairs;
+  //   2. an ORPHAN closing </think> — GLM's chat template opens <think>
+  //      inside the prompt, so the completion starts mid-reasoning and only
+  //      emits the closer (observed live 2026-08-23, baidu/glm-5.2:
+  //      "Hi. What need?</think>Hi. What need?" reached the client verbatim).
+  let leakedThinking = '';
+  if (textContent.includes('</think>')) {
+    if (textContent.includes('<think>')) {
+      const thoughts = [];
+      textContent = textContent
+        .replace(/<think>([\s\S]*?)<\/think>/g, (_, t) => { thoughts.push(t.trim()); return ''; })
+        .trim();
+      leakedThinking = thoughts.filter(Boolean).join('\n\n');
+    } else {
+      const idx = textContent.indexOf('</think>');
+      leakedThinking = textContent.slice(0, idx).trim();
+      textContent = textContent.slice(idx + '</think>'.length).trim();
+    }
+  }
+
+  // Emit reasoning (side-channel and/or leaked) as a proper thinking block
+  const combinedThinking = [reasoningContent, leakedThinking].filter(Boolean).join('\n\n');
+  if (combinedThinking) {
+    content.push({ type: "thinking", thinking: combinedThinking });
   }
 
   if (textContent) {
@@ -2551,7 +2586,7 @@ function sanitizeSchemaForGemini(schema) {
  * (generativelanguage.googleapis.com), not Google Cloud Vertex AI.
  * Converts Anthropic format to Gemini format and back.
  */
-async function invokeVertex(body, incomingHeaders = {}) {
+async function invokeVertex(body, _incomingHeaders = {}) {
   const apiKey = config.vertex?.apiKey;
 
   if (!apiKey) {
@@ -2796,7 +2831,7 @@ function convertGeminiToAnthropic(response, requestedModel) {
   };
 }
 
-async function invokeCodex(body, incomingHeaders = {}) {
+async function invokeCodex(body, _incomingHeaders = {}) {
   const { getCodexProcess } = require("./codex-process");
   const { convertAnthropicToCodexPrompt, convertCodexResponseToAnthropic } = require("./codex-utils");
 
@@ -2920,7 +2955,6 @@ const LYNKR_BADGE_PREFIX_RE = /^\*\[Lynkr\][^*\n]*\*\s*/;
 function stripLynkrBadges(messages) {
   if (!Array.isArray(messages)) return messages;
   let mutated = false;
-  let badgeCount = 0;
   const out = messages.map((msg) => {
     if (msg?.role !== 'assistant') return msg;
 
@@ -2931,7 +2965,6 @@ function stripLynkrBadges(messages) {
       if (!LYNKR_BADGE_PREFIX_RE.test(msg.content)) return msg;
       const stripped = msg.content.replace(LYNKR_BADGE_PREFIX_RE, '');
       mutated = true;
-      badgeCount++;
       // Badge-only content must not become an empty string — Anthropic
       // rejects empty assistant content (this is the interrupted-response
       // continuation case, where the partial text WAS just the badge).
@@ -2950,7 +2983,6 @@ function stripLynkrBadges(messages) {
       for (const b of msg.content) {
         if (b?.type === 'text' && typeof b.text === 'string' && LYNKR_BADGE_PREFIX_RE.test(b.text)) {
           changed = true;
-          badgeCount++;
           const strippedText = b.text.replace(LYNKR_BADGE_PREFIX_RE, '');
           if (strippedText.trim()) rebuilt.push({ ...b, text: strippedText });
           // badge-only block → drop
@@ -3853,6 +3885,8 @@ function destroyHttpAgents() {
 
 module.exports = {
   invokeModel,
+  computeCostUsd,
+  convertOpenAIToAnthropic,
   invokeAzureAnthropic,
   invokeZai,
   invokeOllama,