lynkr 9.10.0 → 9.11.0

package/bin/lynkr-init.js

@@ -236,6 +236,13 @@ const BASELINE_ENV = {
   HEADROOM_CCR: 'true',
   HEADROOM_CCR_TTL: '300',
 
+  // ── TencentDB-Agent-Memory sidecar (team memory hub) ──────────────────
+  // Launches agentmemory/memory-core + memory-hub containers on
+  // `lynkr start` (Panel UI :8125). First start pulls two Docker Hub
+  // images; skips gracefully when Docker isn't running. See .env.example
+  // for all knobs.
+  TENCENTDB_MEMORY_ENABLED: 'true',
+
   // ── Memory + token tracking ───────────────────────────────────────────
   MEMORY_ENABLED: 'true',
   MEMORY_RETRIEVAL_LIMIT: '5',

package/config/model-tiers.json

@@ -53,7 +53,7 @@
           "GLM-4-Flash"
         ],
         "moonshot": [
-          "kimi-k2-turbo-preview"
+          "kimi-k2.6"
         ]
       }
     },
@@ -108,7 +108,7 @@
           "GLM-4.7"
         ],
         "moonshot": [
-          "kimi-k2-turbo-preview"
+          "kimi-k2.6"
         ]
       }
     },
@@ -157,7 +157,7 @@
           "GLM-4.7"
         ],
         "moonshot": [
-          "kimi-k2-turbo-preview"
+          "kimi-k2.6"
         ]
       }
     },
@@ -203,8 +203,8 @@
           "deepseek-r1"
         ],
         "moonshot": [
-          "kimi-k2-thinking",
-          "kimi-k2-turbo-preview"
+          "kimi-k3",
+          "kimi-k2.6"
         ]
       }
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "lynkr",
-  "version": "9.10.0",
+  "version": "9.11.0",
   "files": [
     "index.js",
     "install.sh",
@@ -36,8 +36,8 @@
     "dev": "nodemon index.js",
     "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/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/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/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: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: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/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/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: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",

package/public/dashboard.html

@@ -493,6 +493,35 @@ const App = {
         ` : emptyState('No routing data for last 24h')}
       `, 'mb-6')}
 
+      <!-- Cache Economics (cache-aware routing receipt) -->
+      ${card(`
+        <h3 class="text-sm font-semibold text-slate-300 mb-4">Cache Economics (last ${d.window || '24h'})</h3>
+        ${d.cacheEconomics && d.cacheEconomics.decisions > 0 ? `
+          <div class="grid grid-cols-2 lg:grid-cols-4 gap-4">
+            <div class="text-center">
+              <p class="text-2xl font-bold text-green-400">${fmt.usd(d.cacheEconomics.totalDollarsSaved)}</p>
+              <p class="text-xs text-slate-500 mt-1">Cache $ saved by routing</p>
+              <p class="text-xs text-slate-600">(holds + cleared switches)</p>
+            </div>
+            <div class="text-center">
+              <p class="text-2xl font-bold">${fmt.num(d.cacheEconomics.holds)}</p>
+              <p class="text-xs text-slate-500 mt-1">Switches held</p>
+              <p class="text-xs text-slate-600">(warm prefix protected)</p>
+            </div>
+            <div class="text-center">
+              <p class="text-2xl font-bold">${fmt.num(d.cacheEconomics.switches)}</p>
+              <p class="text-xs text-slate-500 mt-1">Switches cleared</p>
+              <p class="text-xs text-slate-600">(break-even ≤ remaining turns)</p>
+            </div>
+            <div class="text-center">
+              <p class="text-2xl font-bold text-amber-400">${fmt.usd(d.cacheEconomics.dollarsSavedByHolds)}</p>
+              <p class="text-xs text-slate-500 mt-1">Saved by holding pins</p>
+              <p class="text-xs text-slate-600">(cache re-writes avoided)</p>
+            </div>
+          </div>
+        ` : emptyState('No gated switch decisions yet — appears once a pinned session hits a downgrade decision')}
+      `, 'mb-6')}
+
       <!-- Provider Stats -->
       ${card(`
         <h3 class="text-sm font-semibold text-slate-300 mb-3">Provider Stats (last ${d.window || '24h'})</h3>

package/src/api/router.js

@@ -579,6 +579,30 @@ function _buildStreamBadge(tier) {
   return `*[Lynkr] ${tier.tier || "—"} → ${tier.model || "—"} (${tier.provider || "—"}) · score ${tier.score ?? "—"}${pin}*\n\n`;
 }
 
+/**
+ * Terminal upstream errors carry no content array — without this, the synth
+ * SSE writers render an empty message and the turn goes silent in the TUI
+ * (only the routing badge shows). Returns visible-text content blocks for an
+ * error result, using the fallback chain's failure trail when the provider
+ * layer recorded one (so "moonshot: quota exhausted" isn't hidden behind the
+ * last rung's unrelated error), or null for normal responses.
+ */
+function _buildFailureBlocks(result, msg) {
+  if (result.terminationReason !== "api_error" && Array.isArray(msg?.content)) return null;
+  const trail = Array.isArray(msg?._lynkr_failure_trail) ? msg._lynkr_failure_trail : null;
+  const lastErr = msg?.error?.message || msg?.message
+    || (typeof msg === "string" ? msg : JSON.stringify(msg ?? {}).slice(0, 200));
+  const trailText = (trail || [String(lastErr)]).map((t) => `  · ${t}`).join("\n");
+  const quotaHit = result.status === 429 || /rate.?limit|quota|TPD|TPM/i.test(trailText);
+  return [{
+    type: "text",
+    text:
+      `⚠ [Lynkr] ${quotaHit ? "Provider quota exhausted — no fallback could serve this request" : "No provider could serve this request"} (HTTP ${result.status ?? "?"}).\n\n` +
+      `${trailText}\n` +
+      (quotaHit ? `\nAll fallback candidates failed after the quota hit. Wait for the quota window to reset or repoint the affected tier in .env.\n` : ""),
+  }];
+}
+
 /**
  * Scan the entire payload for tool_use or tool_result content blocks.
  * Unlike session-affinity.payloadHasToolHistory (which checks only the last
@@ -1611,6 +1635,9 @@ router.post("/v1/messages", rateLimiter, async (req, res, next) => {
       // tool_use with no tool_use block).
       let contentBlocks = (msg.content || []).slice();
 
+      const failureBlocks = _buildFailureBlocks(result, msg);
+      if (failureBlocks) contentBlocks = failureBlocks;
+
       // When LYNKR_VISIBLE_ROUTING=true, prepend a one-line routing badge so
       // users can see which tier/provider/model handled the request inside
       // Claude Code's TUI (TUI only renders content blocks; unknown top-level
@@ -1807,6 +1834,9 @@ router.post("/v1/messages", rateLimiter, async (req, res, next) => {
       // tool_use with no tool_use block).
       let contentBlocks = (msg.content || []).slice();
 
+      const failureBlocks = _buildFailureBlocks(result, msg);
+      if (failureBlocks) contentBlocks = failureBlocks;
+
       // When LYNKR_VISIBLE_ROUTING=true, prepend a one-line routing badge so
       // users can see which tier/provider/model handled the request inside
       // Claude Code's TUI (TUI only renders content blocks; unknown top-level
@@ -1969,6 +1999,10 @@ router.post("/v1/messages", rateLimiter, async (req, res, next) => {
     }
 
     metrics.recordResponse(result.status);
+    // finalBody is usually a JSON string; Express defaults string bodies to
+    // text/html, which Claude Code rejects as a malformed 200. Only default —
+    // an upstream content-type copied from result.headers above still wins.
+    if (!res.getHeader("Content-Type")) res.type("application/json");
     res.status(result.status).send(finalBody);
   } catch (error) {
     next(error);

package/src/clients/databricks.js

@@ -560,9 +560,20 @@ async function invokeOllama(body, incomingHeaders = {}) {
     if (body.tool_choice) ollamaBody.tool_choice = body.tool_choice;
     if (body.metadata) ollamaBody.metadata = body.metadata;
 
-    // Tools (already Anthropic format — no conversion needed)
+    // Tools: normalize to Anthropic shape. Callers that stayed OpenAI-format
+    // upstream (OpenWorker, aisuite, any client hitting /v1/chat/completions)
+    // otherwise ship {type:"function", function:{...}} which Ollama's
+    // /v1/messages endpoint rejects with 400 "can't find closing '}' symbol".
     if (supportsTools && Array.isArray(toolsToSend) && toolsToSend.length > 0) {
-      ollamaBody.tools = toolsToSend;
+      ollamaBody.tools = toolsToSend.map(tool =>
+        tool?.type === "function" && tool.function
+          ? {
+              name: tool.function.name,
+              description: tool.function.description || "",
+              input_schema: tool.function.parameters || { type: "object", properties: {} },
+            }
+          : tool
+      );
     }
 
     if (config.ollama.keepAlive !== undefined) {
@@ -1934,6 +1945,21 @@ async function invokeZai(body, incomingHeaders = {}) {
       }, "=== INJECTING STANDARD TOOLS (Z.AI Anthropic) ===");
     }
 
+    // Normalize to Anthropic tool shape — the /anthropic/v1/messages endpoint
+    // rejects OpenAI-nested tools ({type:"function", function:{name,...}}) with
+    // 422 missing body.tools[0].name. Callers that stayed OpenAI-format upstream
+    // (OpenWorker, aisuite, any client hitting /v1/chat/completions with an
+    // unusual tool shape) get unwrapped here.
+    zaiBody.tools = zaiBody.tools.map(tool =>
+      tool?.type === "function" && tool.function
+        ? {
+            name: tool.function.name,
+            description: tool.function.description || "",
+            input_schema: tool.function.parameters || { type: "object", properties: {} },
+          }
+        : tool
+    );
+
     headers = {
       "Content-Type": "application/json",
       "x-api-key": config.zai.apiKey,
@@ -2019,13 +2045,17 @@ async function invokeMoonshot(body, incomingHeaders = {}) {
 
   // Model mapping: Anthropic names → Moonshot/Kimi names
   const modelMap = {
-    "claude-sonnet-4-5-20250929": "kimi-k2-turbo-preview",
-    "claude-sonnet-4-5": "kimi-k2-turbo-preview",
-    "claude-sonnet-4.5": "kimi-k2-turbo-preview",
-    "claude-3-5-sonnet": "kimi-k2-turbo-preview",
-    "claude-haiku-4-5-20251001": "kimi-k2-turbo-preview",
-    "claude-haiku-4-5": "kimi-k2-turbo-preview",
-    "claude-3-haiku": "kimi-k2-turbo-preview",
+    "claude-sonnet-4-5-20250929": "kimi-k2.6",
+    "claude-sonnet-4-5": "kimi-k2.6",
+    "claude-sonnet-4.5": "kimi-k2.6",
+    "claude-3-5-sonnet": "kimi-k2.6",
+    "claude-haiku-4-5-20251001": "kimi-k2.6",
+    "claude-haiku-4-5": "kimi-k2.6",
+    "claude-3-haiku": "kimi-k2.6",
+    // Retired ids (gone from /v1/models as of 2026-07: "Not found the model")
+    // that older tier configs may still reference.
+    "kimi-k2-turbo-preview": "kimi-k2.6",
+    "kimi-k2-thinking": "kimi-k3",
     // moonshot-v1-auto 400s with "tokenization failed" (its server-side auto
     // context-size pass fails on large tool-bearing payloads). Remap to a
     // fixed model that's broadly available on api.moonshot.ai.
@@ -2033,7 +2063,12 @@ async function invokeMoonshot(body, incomingHeaders = {}) {
   };
 
   const requestedModel = body._tierModel || body.model || config.moonshot.model;
-  let mappedModel = modelMap[requestedModel] || config.moonshot.model || "kimi-k2-turbo-preview";
+  // Honor tier-selected Moonshot ids (e.g. TIER_REASONING=moonshot:kimi-k2.6)
+  // instead of silently swapping in the .env default model.
+  let mappedModel = modelMap[requestedModel]
+    || (/^(kimi-|moonshot-v1)/i.test(requestedModel || "") ? requestedModel : null)
+    || config.moonshot.model
+    || "kimi-k3";
   // Guard against the deprecated auto model arriving via config too.
   if (mappedModel === "moonshot-v1-auto") mappedModel = "moonshot-v1-128k";
 
@@ -2047,18 +2082,18 @@ async function invokeMoonshot(body, incomingHeaders = {}) {
     messages.unshift({ role: "system", content: systemContent });
   }
 
-  // kimi-k2.x (k2.5 / k2.6 …) are thinking models that only accept
-  // temperature: 1 — any other value 400s with "invalid temperature".
-  const isKimiThinking = /^kimi-k2/i.test(mappedModel);
+  // Every current kimi-k* model (k2.5, k2.6, k2.7-code, k3, …) pins sampling
+  // params: temperature must be 1 — any other value 400s with "invalid
+  // temperature: only 1 is allowed for this model" (probed live 2026-07-23).
+  // Only the moonshot-v1-* models accept caller-supplied values.
+  const isKimiPinned = /^kimi-k/i.test(mappedModel);
 
   const moonshotBody = {
     model: mappedModel,
     messages,
     max_tokens: body.max_tokens || 16384,
-    // kimi-k2.x thinking models pin sampling params: temperature must be 1
-    // and top_p must be 0.95 — any other value 400s.
-    temperature: isKimiThinking ? 1 : (body.temperature ?? 0.7),
-    top_p: isKimiThinking ? 0.95 : (body.top_p ?? 1.0),
+    temperature: isKimiPinned ? 1 : (body.temperature ?? 0.7),
+    top_p: isKimiPinned ? 0.95 : (body.top_p ?? 1.0),
     stream: false,  // Force non-streaming - OpenAI SSE to Anthropic SSE conversion not implemented
   };
 
@@ -2678,6 +2713,43 @@ function stripLynkrBadges(messages) {
   return mutated ? out : messages;
 }
 
+// Single source of truth for provider dispatch. Both the initial call and the
+// fallback ladder MUST route through invokeProvider — the previous duplicated
+// if/else chains drifted (the fallback chain was missing ollama, bedrock,
+// lmstudio and codex), and unknown providers silently dialed databricks,
+// which under tier routing is a deliberate dead number (see .env note about
+// the 2026-07-07 self-loop incident). Add new providers HERE, nowhere else;
+// the registry-coverage unit test fails if this map and
+// SUPPORTED_MODEL_PROVIDERS fall out of sync.
+const PROVIDER_INVOKERS = {
+  "azure-openai": invokeAzureOpenAI,
+  "azure-anthropic": invokeAzureAnthropic,
+  ollama: invokeOllama,
+  openrouter: invokeOpenRouter,
+  edenai: invokeEdenAI,
+  openai: invokeOpenAI,
+  llamacpp: invokeLlamaCpp,
+  lmstudio: invokeLMStudio,
+  bedrock: invokeBedrock,
+  zai: invokeZai,
+  vertex: invokeVertex,
+  moonshot: invokeMoonshot,
+  codex: invokeCodex,
+};
+
+function invokeProvider(provider, body, incomingHeaders) {
+  const invoke = PROVIDER_INVOKERS[provider];
+  if (!invoke) {
+    // databricks itself, or an unknown name. Loud, not silent: under tier
+    // routing this path used to swallow misconfigured providers.
+    if (provider !== "databricks") {
+      logger.warn({ provider }, "No invoker registered for provider — defaulting to databricks");
+    }
+    return invokeDatabricks(body, incomingHeaders);
+  }
+  return invoke(body, incomingHeaders);
+}
+
 async function invokeModel(body, options = {}) {
   const { determineProviderSmart, isFallbackEnabled, getFallbackProvider } = require("./routing");
   const metricsCollector = getMetricsCollector();
@@ -2834,34 +2906,7 @@ async function invokeModel(body, options = {}) {
 
   try {
     const result = await breaker.execute(async () => {
-      if (initialProvider === "azure-openai") {
-        return await invokeAzureOpenAI(body, incomingHeaders);
-      } else if (initialProvider === "azure-anthropic") {
-        return await invokeAzureAnthropic(body, incomingHeaders);
-      } else if (initialProvider === "ollama") {
-        return await invokeOllama(body, incomingHeaders);
-      } else if (initialProvider === "openrouter") {
-        return await invokeOpenRouter(body, incomingHeaders);
-      } else if (initialProvider === "edenai") {
-        return await invokeEdenAI(body, incomingHeaders);
-      } else if (initialProvider === "openai") {
-        return await invokeOpenAI(body, incomingHeaders);
-      } else if (initialProvider === "llamacpp") {
-        return await invokeLlamaCpp(body, incomingHeaders);
-      } else if (initialProvider === "lmstudio") {
-        return await invokeLMStudio(body, incomingHeaders);
-      } else if (initialProvider === "bedrock") {
-        return await invokeBedrock(body, incomingHeaders);
-      } else if (initialProvider === "zai") {
-        return await invokeZai(body, incomingHeaders);
-      } else if (initialProvider === "vertex") {
-        return await invokeVertex(body, incomingHeaders);
-      } else if (initialProvider === "moonshot") {
-        return await invokeMoonshot(body, incomingHeaders);
-      } else if (initialProvider === "codex") {
-        return await invokeCodex(body, incomingHeaders);
-      }
-      return await invokeDatabricks(body, incomingHeaders);
+      return await invokeProvider(initialProvider, body, incomingHeaders);
     });
 
     const latency = Date.now() - startTime;
@@ -2954,6 +2999,7 @@ async function invokeModel(body, options = {}) {
       candidates: routingResult.candidates ?? null,
       pinned: routingResult.pinned ? 1 : 0,
       switch_reason: routingResult.switch_reason ?? null,
+      cache_decision: routingResult._cacheDecision ?? null,
     });
 
     // WS5.4 — feedback loop (success path).
@@ -3035,6 +3081,14 @@ async function invokeModel(body, options = {}) {
     healthTracker.recordFailure(initialProvider, err, err.status);
     getLatencyTracker().record(initialProvider, routingDecision?.model, failLatency);
 
+    // Accumulate every provider failure for this request so the terminal
+    // error can tell the client WHY nothing answered (quota exhausted vs
+    // context overflow vs dead config) instead of only the last rung's
+    // error. Rendered by the SSE writer when the whole chain fails.
+    const failureTrail = [
+      `${initialProvider}${routingDecision?.model ? ` (${routingDecision.model})` : ""}: ${err.status ? `HTTP ${err.status} — ` : ""}${err.message}`,
+    ];
+
     // undici wraps every network failure as "TypeError: fetch failed" — the
     // actionable detail (ENOTFOUND vs ECONNRESET vs ETIMEDOUT vs TLS) lives
     // in err.cause. Without this line the z-ai breaker trips on repeated
@@ -3097,6 +3151,7 @@ async function invokeModel(body, options = {}) {
               status: attempt.status,
               error: attempt.json?.error?.message || attempt.text?.slice(0, 120) || null,
             }, "[TierFallback] Candidate answered with an error response, trying next");
+            failureTrail.push(`${cand.provider} (${cand.model}): HTTP ${attempt.status} — ${attempt.json?.error?.message || attempt.text?.slice(0, 120) || "error response"}`);
             continue;
           }
 
@@ -3127,6 +3182,7 @@ async function invokeModel(body, options = {}) {
             { toProvider: cand.provider, error: innerErr.message },
             "[TierFallback] Candidate failed, trying next"
           );
+          failureTrail.push(`${cand.provider} (${cand.model}): ${innerErr.message}`);
         }
       }
       logger.warn(
@@ -3175,6 +3231,7 @@ async function invokeModel(body, options = {}) {
         candidates: routingResult.candidates ?? null,
         pinned: routingResult.pinned ? 1 : 0,
         switch_reason: routingResult.switch_reason ?? null,
+      cache_decision: routingResult._cacheDecision ?? null,
       });
 
       // WS5.4 — feedback loop (primary-failed, no fallback). Low quality
@@ -3194,6 +3251,7 @@ async function invokeModel(body, options = {}) {
         },
       });
 
+      err.lynkrFailureTrail = failureTrail;
       throw err;
     }
 
@@ -3221,26 +3279,7 @@ async function invokeModel(body, options = {}) {
       const fallbackStart = Date.now();
 
       const fallbackResult = await fallbackBreaker.execute(async () => {
-        if (fallbackProvider === "azure-openai") {
-          return await invokeAzureOpenAI(body, incomingHeaders);
-        } else if (fallbackProvider === "azure-anthropic") {
-          return await invokeAzureAnthropic(body, incomingHeaders);
-        } else if (fallbackProvider === "openrouter") {
-          return await invokeOpenRouter(body, incomingHeaders);
-        } else if (fallbackProvider === "edenai") {
-          return await invokeEdenAI(body, incomingHeaders);
-        } else if (fallbackProvider === "openai") {
-          return await invokeOpenAI(body, incomingHeaders);
-        } else if (fallbackProvider === "llamacpp") {
-          return await invokeLlamaCpp(body, incomingHeaders);
-        } else if (fallbackProvider === "zai") {
-          return await invokeZai(body, incomingHeaders);
-        } else if (fallbackProvider === "vertex") {
-          return await invokeVertex(body, incomingHeaders);
-        } else if (fallbackProvider === "moonshot") {
-          return await invokeMoonshot(body, incomingHeaders);
-        }
-        return await invokeDatabricks(body, incomingHeaders);
+        return await invokeProvider(fallbackProvider, body, incomingHeaders);
       });
 
       const fallbackLatency = Date.now() - fallbackStart;
@@ -3311,6 +3350,7 @@ async function invokeModel(body, options = {}) {
         candidates: routingResult.candidates ?? null,
         pinned: routingResult.pinned ? 1 : 0,
         switch_reason: routingResult.switch_reason ?? null,
+      cache_decision: routingResult._cacheDecision ?? null,
       });
 
       // WS5.4 — feedback loop (fallback success). The served provider
@@ -3330,6 +3370,16 @@ async function invokeModel(body, options = {}) {
         },
       });
 
+      // The last rung also answered with an error — thread the whole trail
+      // into the error payload so the client sees every failure, not just
+      // this one (e.g. "moonshot: quota" matters more than "azure: 404").
+      if (fallbackResult?.ok === false) {
+        failureTrail.push(`${fallbackProvider}: HTTP ${fallbackResult.status} — ${fallbackResult.json?.error?.message || String(fallbackResult.text || "").slice(0, 120) || "error response"}`);
+        if (fallbackResult.json && typeof fallbackResult.json === "object") {
+          fallbackResult.json._lynkr_failure_trail = failureTrail;
+        }
+      }
+
       return {
         ...fallbackResult,
         actualProvider: fallbackProvider,
@@ -3368,6 +3418,7 @@ async function invokeModel(body, options = {}) {
         candidates: routingResult.candidates ?? null,
         pinned: routingResult.pinned ? 1 : 0,
         switch_reason: routingResult.switch_reason ?? null,
+      cache_decision: routingResult._cacheDecision ?? null,
       });
 
       // WS5.4 — feedback loop (double failure). quality=0 is a hard
@@ -3394,6 +3445,8 @@ async function invokeModel(body, options = {}) {
       }, "Both primary and fallback provider failed");
 
       // Return fallback error (more actionable than Ollama error)
+      failureTrail.push(`${fallbackProvider}: ${fallbackErr.message}`);
+      fallbackErr.lynkrFailureTrail = failureTrail;
       throw fallbackErr;
     }
   }
@@ -3459,6 +3512,8 @@ module.exports = {
   invokeAzureAnthropic,
   invokeZai,
   invokeOllama,
+  invokeMoonshot,
+  PROVIDER_INVOKERS,
   stripLynkrBadges,
   destroyHttpAgents,
   normalizeBodyForConverse,

package/src/clients/gpt-utils.js

@@ -61,7 +61,14 @@ function areSimilarToolCalls(call1, call2) {
 
   // Only search-style tools get fuzzy matching; mutating tools with
   // near-identical args may be intentional repeats.
-  const searchTools = ['grep', 'glob', 'search', 'find', 'read', 'bash', 'shell'];
+  // 'read' is deliberately NOT fuzzy-matched: its argument is a file path, and
+  // absolute paths in one repo share nearly every Jaccard token
+  // (/Users/x/project/src/…), so reads of DIFFERENT files scored ≥0.8 and
+  // merged into one "repeated call" signature (live incident: an opencode
+  // code-trace reading server.js, openai-router.js and orchestrator/index.js
+  // was flagged as a loop). A genuine re-read of the same file is caught by
+  // the exact-match branch above.
+  const searchTools = ['grep', 'glob', 'search', 'find', 'bash', 'shell'];
   const toolName = (name1 || '').toLowerCase();
   const isSearchTool = searchTools.some(t => toolName.includes(t));
 

package/src/clients/ollama-utils.js

@@ -128,17 +128,26 @@ function convertAnthropicToolsToOllama(anthropicTools) {
     return [];
   }
 
-  return anthropicTools.map(tool => ({
-    type: "function",
-    function: {
-      name: tool.name,
-      description: tool.description || "",
-      parameters: tool.input_schema || {
-        type: "object",
-        properties: {},
+  return anthropicTools.map(tool => {
+    // Shape-detect: already OpenAI/Ollama-format tools pass through untouched.
+    // Without this, an OpenAI-shape tool ({type:"function", function:{...}})
+    // reaching this helper produces {function:{name:undefined,...}}, which
+    // Ollama's legacy /api/chat endpoint rejects.
+    if (tool?.type === "function" && tool.function) {
+      return tool;
+    }
+    return {
+      type: "function",
+      function: {
+        name: tool.name,
+        description: tool.description || "",
+        parameters: tool.input_schema || {
+          type: "object",
+          properties: {},
+        },
       },
-    },
-  }));
+    };
+  });
 }
 
 module.exports = {