nodal-agents 0.4.2 → 0.4.4

package/runner.js

@@ -1106,7 +1106,10 @@ var init_connector_catalog = __esm({
 function findModelCatalogEntry(provider, modelId) {
   return MODEL_CATALOG[provider]?.find((e) => e.modelId === modelId);
 }
-var MODEL_CATALOG;
+function modelContextWindow(provider, modelId) {
+  return findModelCatalogEntry(provider, modelId)?.contextWindow ?? DEFAULT_CONTEXT_WINDOW;
+}
+var MODEL_CATALOG, DEFAULT_CONTEXT_WINDOW;
 var init_model_catalog = __esm({
   "../../packages/shared/src/model-catalog.ts"() {
     "use strict";
@@ -1115,127 +1118,167 @@ var init_model_catalog = __esm({
         {
           modelId: "claude-opus-4-8",
           label: "Claude Opus 4.8",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 2e5
         },
         {
           modelId: "claude-sonnet-4-6",
           label: "Claude Sonnet 4.6",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 2e5
         },
         {
           modelId: "claude-haiku-4-5-20251001",
           label: "Claude Haiku 4.5",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 2e5
         }
       ],
       openai: [
-        { modelId: "gpt-5", label: "GPT-5", capabilities: { tools: true, forcedToolChoice: true } },
+        {
+          modelId: "gpt-5",
+          label: "GPT-5",
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 4e5
+        },
         {
           modelId: "gpt-5-mini",
           label: "GPT-5 mini",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 4e5
         }
       ],
       google: [
         {
           modelId: "gemini-2.0-flash",
           label: "Gemini 2.0 Flash",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1048576
         },
         {
           modelId: "gemini-2.5-pro",
           label: "Gemini 2.5 Pro",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1048576
         }
       ],
       groq: [
         {
           modelId: "llama-3.3-70b-versatile",
           label: "Llama 3.3 70B",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 131072
         }
       ],
       mistral: [
         {
           modelId: "mistral-large-latest",
           label: "Mistral Large",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 131072
         }
       ],
       // OpenRouter models are namespaced by sub-vendor (anthropic/, deepseek/, …).
       // The UI groups them by that vendor (see modelGroupLabel). Tested + working
-      // routes; all accept a forced tool_choice.
+      // routes. `forcedToolChoice` is per-model: most accept tool_choice:'required';
+      // MiniMax M3 does not (some of its OpenRouter endpoints reject the forced
+      // value), so it runs on 'auto' + the runtime floor.
       openrouter: [
         // Anthropic
         {
           modelId: "anthropic/claude-haiku-4.5",
           label: "Claude Haiku 4.5",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 2e5
         },
         {
           modelId: "anthropic/claude-opus-4.7",
           label: "Claude Opus 4.7",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1e6
         },
         {
           modelId: "anthropic/claude-opus-4.7-fast",
           label: "Claude Opus 4.7 (fast)",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1e6
         },
         {
           modelId: "anthropic/claude-opus-4.8",
           label: "Claude Opus 4.8",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1e6
         },
         {
           modelId: "anthropic/claude-opus-4.8-fast",
           label: "Claude Opus 4.8 (fast)",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1e6
         },
         {
           modelId: "anthropic/claude-sonnet-4.6",
           label: "Claude Sonnet 4.6",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1e6
         },
         // DeepSeek
         {
           modelId: "deepseek/deepseek-v3.2",
           label: "DeepSeek V3.2",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 131072
         },
         {
           modelId: "deepseek/deepseek-v4-flash",
           label: "DeepSeek V4 Flash",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1048576
         },
         {
           modelId: "deepseek/deepseek-v4-pro",
           label: "DeepSeek V4 Pro",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1048576
         },
         // Google
         {
           modelId: "google/gemini-3.1-flash-lite-preview",
           label: "Gemini 3.1 Flash Lite (preview)",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1048576
         },
         {
           modelId: "google/gemini-3.1-pro-preview",
           label: "Gemini 3.1 Pro (preview)",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1048576
         },
         {
           modelId: "google/gemini-3.5-flash",
           label: "Gemini 3.5 Flash",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 1048576
         },
         {
           modelId: "google/gemma-4-31b-it",
           label: "Gemma 4 31B-IT",
-          capabilities: { tools: true, forcedToolChoice: true }
+          capabilities: { tools: true, forcedToolChoice: true },
+          contextWindow: 262144
+        },
+        // MiniMax
+        {
+          modelId: "minimax/minimax-m3",
+          label: "MiniMax M3",
+          // A reasoning model. Some of its OpenRouter endpoints reject a FORCED
+          // tool_choice ('required') → we send 'auto' (forcedToolChoice:false).
+          // reasoning:true makes the provider return reasoning_details so the runner
+          // can round-trip them across tool-call turns.
+          capabilities: { tools: true, forcedToolChoice: false, reasoning: true },
+          contextWindow: 1048576
         }
       ]
     };
+    DEFAULT_CONTEXT_WINDOW = 128e3;
   }
 });
 
@@ -10946,7 +10989,8 @@ function buildGroqModel(config) {
 }
 
 // ../../packages/llm/src/providers/openrouter.ts
-import { createOpenAICompatible as createOpenAICompatible2 } from "@ai-sdk/openai-compatible";
+init_src();
+import { createOpenRouter } from "@openrouter/ai-sdk-provider";
 import { wrapLanguageModel } from "ai";
 
 // ../../packages/llm/src/providers/parsers.ts
@@ -11174,15 +11218,21 @@ function buildOpenRouterModel(config) {
     throw new ProviderConfigError("openrouter provider requires an apiKey");
   }
   const baseURL = config.baseURL ?? PROVIDER_PRESETS.openrouter.defaultBaseURL;
-  const provider = createOpenAICompatible2({
-    name: "openrouter",
-    baseURL,
+  const provider = createOpenRouter({
     apiKey: config.apiKey,
+    baseURL,
+    // 'strict' is the documented mode for the first-party OpenRouter API
+    // ('compatible' is for 3rd-party proxies). We hit openrouter.ai directly.
+    compatibility: "strict",
     // Normalise non-spec responses (e.g. DeepSeek V4 returning function.arguments
-    // as an object instead of a JSON string) before AI SDK's Zod schema sees them.
+    // as an object instead of a JSON string) before the SDK's Zod schema sees them.
     fetch: createTolerantFetch()
   });
-  const base = provider(config.model);
+  const isReasoning = findModelCatalogEntry("openrouter", config.model)?.capabilities.reasoning;
+  const base = provider.chat(
+    config.model,
+    isReasoning ? { extraBody: { reasoning: { enabled: true } } } : void 0
+  );
   const middleware = middlewareForFamily(detectAgenticFamily(config.model));
   if (middleware) {
     return wrapLanguageModel({ model: base, middleware });
@@ -27426,13 +27476,17 @@ async function completeJob(db, jobId, result, toolsUsed = [], stats, messages) {
     }
   }).where(eq4(agentJobs.id, jobId));
 }
-async function failJob(db, jobId, errorCode, stats) {
+async function failJob(db, jobId, errorCode, stats, messages) {
   const now = /* @__PURE__ */ new Date();
   await db.update(agentJobs).set({
     status: "failed",
     error: errorCode,
     completedAt: now,
     updatedAt: now,
+    // Persist the transcript on failure (mirrors completeJob) so a failed job
+    // is DIAGNOSABLE, not opaque. The resume/guard failure paths used to lose
+    // it entirely. Omitted ⇒ the stored messages are left untouched.
+    ...messages ? { messages } : {},
     ...stats && {
       inputTokens: stats.inputTokens,
       outputTokens: stats.outputTokens,
@@ -27654,6 +27708,40 @@ function truncateForContext(value) {
 
 [... truncated: ${dropped} chars dropped (total ${value.length}) ...]`;
 }
+var EVICTED_TOOL_RESULT_MARKER = "[earlier tool result elided to fit the context window \u2014 re-run the tool if you need this data again]";
+function compactOldToolResults(messages, keepRecentToolMessages) {
+  const toolMsgIdx = [];
+  messages.forEach((m, i) => {
+    if (m.role === "tool") toolMsgIdx.push(i);
+  });
+  const evictCount = toolMsgIdx.length - keepRecentToolMessages;
+  if (evictCount <= 0) return { messages: [...messages], evicted: 0 };
+  const evictIdx = new Set(toolMsgIdx.slice(0, evictCount));
+  let evicted = 0;
+  const out = messages.map((m, i) => {
+    if (!evictIdx.has(i) || m.role !== "tool" || !Array.isArray(m.content)) return m;
+    const content = m.content.map((p) => {
+      if (p.type !== "tool-result") return p;
+      if (p.output.type === "text" && p.output.value === EVICTED_TOOL_RESULT_MARKER) return p;
+      evicted += 1;
+      return { ...p, output: { type: "text", value: EVICTED_TOOL_RESULT_MARKER } };
+    });
+    return { ...m, content };
+  });
+  return { messages: out, evicted };
+}
+function describeLlmError(err) {
+  if (!(err instanceof Error)) return "unknown_error";
+  const e = err;
+  const body = typeof e.responseBody === "string" && e.responseBody.trim().length > 0 ? e.responseBody : "";
+  const dataStr = !body && e.data != null ? JSON.stringify(e.data) : "";
+  const detail = body || dataStr;
+  if (detail) {
+    const status = typeof e.statusCode === "number" ? `${e.statusCode} ` : "";
+    return `${status}${err.message}: ${detail}`.slice(0, 400);
+  }
+  return err.message.slice(0, 200);
+}
 function stableStringify(value) {
   return JSON.stringify(value, (_key, val) => {
     if (val && typeof val === "object" && !Array.isArray(val)) {
@@ -27692,6 +27780,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
     });
     return { status: "failed", error: "chain_limit_exceeded" };
   }
+  let messages = Array.isArray(job.messages) ? job.messages : [];
   let inputTokens = job.inputTokens ?? 0;
   let outputTokens = job.outputTokens ?? 0;
   let turn = job.turn ?? 0;
@@ -27704,13 +27793,13 @@ async function executeJob(jobId, deps, _runnerEnv) {
   });
   await setJobStatus(db, jobId, "processing");
   if (!job.agentId) {
-    await failJob(db, jobId, "agent_not_found", runStats());
+    await failJob(db, jobId, "agent_not_found", runStats(), messages);
     return { status: "failed", error: "agent_not_found" };
   }
   const agentRows = await db.select().from(agents).where(eq4(agents.id, job.agentId)).limit(1);
   const agentRow = agentRows[0];
   if (!agentRow || !agentRow.active) {
-    await failJob(db, jobId, "agent_not_found", runStats());
+    await failJob(db, jobId, "agent_not_found", runStats(), messages);
     return { status: "failed", error: "agent_not_found" };
   }
   const agent = {
@@ -27728,7 +27817,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
   const wsRows = await db.select({ label: agentWorkspaces.label, path: agentWorkspaces.path }).from(agentWorkspaces).where(eq4(agentWorkspaces.agentId, agentRow.id)).orderBy(agentWorkspaces.position, agentWorkspaces.label);
   const agentWorkspacesList = wsRows;
   if (!agentRow.llmKeyId) {
-    await failJob(db, jobId, "agent_no_llm_configured", runStats());
+    await failJob(db, jobId, "agent_no_llm_configured", runStats(), messages);
     return { status: "failed", error: "agent_no_llm_configured" };
   }
   {
@@ -27743,7 +27832,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
     );
     if (!resolved.ok) {
       const code = resolved.reason === "agent_no_llm_configured" ? "agent_no_llm_configured" : `llm_key_invalid:${resolved.detail}`;
-      await failJob(db, jobId, code, runStats());
+      await failJob(db, jobId, code, runStats(), messages);
       return { status: "failed", error: code };
     }
     llmClient = resolved.client;
@@ -27778,7 +27867,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
   let toolDefs;
   const memoryBuiltins = ALWAYS_ON_TOOLS.filter((n) => n !== "return_result").map((n) => registry.get(n)).filter((t) => t !== void 0);
   const capabilityTools = [];
-  if (agentRow.telegramBotToken) {
+  if (agentRow.telegramBotToken && job.chatId) {
     capabilityTools.push(createTelegramSendMessageTool());
   }
   const mcpClosers = [];
@@ -27931,7 +28020,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
     }
   } catch (err) {
     const errorCode = err instanceof Error ? err.message : "whitelist_computation_failed";
-    await failJob(db, jobId, errorCode, runStats());
+    await failJob(db, jobId, errorCode, runStats(), messages);
     return { status: "failed", error: errorCode };
   }
   const ruleRows = await db.select().from(approvalRules).where(eq4(approvalRules.entityId, job.entityId ?? ""));
@@ -27950,7 +28039,6 @@ async function executeJob(jobId, deps, _runnerEnv) {
     const mcpToolNames = await listWorkspaceMcpToolNames(db, job.entityId ?? "");
     authoringToolsSuffix = mcpToolNames.length > 0 ? ` When a skill needs an MCP tool, reference one of these EXACT names available in this workspace (built-in and connector tools are bare snake_case and need no namespace): ${mcpToolNames.join(", ")}.` : " This workspace has no MCP servers connected yet \u2014 a skill should only reference built-in tools (bare snake_case) or none.";
   }
-  let messages = Array.isArray(job.messages) ? job.messages : [];
   if (messages.length === 0) {
     messages = [{ role: "user", content: job.task }];
   }
@@ -28092,6 +28180,18 @@ async function executeJob(jobId, deps, _runnerEnv) {
     const n = raw ? Number(raw) : NaN;
     return Number.isFinite(n) && n > 0 ? n : DEFAULT_LIMITS.maxTotalTokensPerJob;
   })();
+  const compactionThreshold = (() => {
+    const ctxWindow = modelContextWindow(llmClient.config.provider, llmClient.config.model);
+    const fracRaw = Number(process.env["LLM_COMPACTION_FRACTION"]);
+    const frac = Number.isFinite(fracRaw) && fracRaw > 0 && fracRaw < 1 ? fracRaw : 0.7;
+    const absRaw = Number(process.env["LLM_COMPACTION_TOKENS"]);
+    const abs = Number.isFinite(absRaw) && absRaw > 0 ? absRaw : 12e4;
+    return Math.min(Math.floor(frac * ctxWindow), abs);
+  })();
+  const compactionKeepRecentToolMsgs = (() => {
+    const raw = Number(process.env["LLM_COMPACTION_KEEP_TURNS"]);
+    return Number.isFinite(raw) && raw >= 1 ? Math.floor(raw) : 3;
+  })();
   const recentTurnSignatures = [];
   const maxNoProgressRepeats = (() => {
     const raw = process.env["MAX_NO_PROGRESS_REPEATS"];
@@ -28112,7 +28212,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
         return { status: "cancelled" };
       }
       if (turn > DEFAULT_LIMITS.maxTurns) {
-        await failJob(db, jobId, "turn_limit_exceeded", runStats());
+        await failJob(db, jobId, "turn_limit_exceeded", runStats(), messages);
         return { status: "failed", error: "turn_limit_exceeded" };
       }
       validateMessageStructure(messages);
@@ -28141,9 +28241,21 @@ async function executeJob(jobId, deps, _runnerEnv) {
       outputTokens += Number.isFinite(completionT) ? completionT : 0;
       if (inputTokens + outputTokens > maxTotalTokensPerJob) {
         trace("token_budget_exceeded", { turn, inputTokens, outputTokens, maxTotalTokensPerJob });
-        await failJob(db, jobId, "token_budget_exceeded", runStats());
+        await failJob(db, jobId, "token_budget_exceeded", runStats(), messages);
         return { status: "failed", error: "token_budget_exceeded" };
       }
+      if (promptT > compactionThreshold) {
+        const ev = compactOldToolResults(messages, compactionKeepRecentToolMsgs);
+        if (ev.evicted > 0) {
+          messages = ev.messages;
+          trace("context_compacted", {
+            turn,
+            promptTokens: promptT,
+            threshold: compactionThreshold,
+            evictedToolResults: ev.evicted
+          });
+        }
+      }
       const rawToolCalls = response.toolCalls ?? [];
       trace("llm_call_done", {
         turn,
@@ -28155,17 +28267,21 @@ async function executeJob(jobId, deps, _runnerEnv) {
       consecutiveDeliveryOnlyTurns = isDeliveryOnlyTurn ? consecutiveDeliveryOnlyTurns + 1 : 0;
       if (consecutiveDeliveryOnlyTurns > DEFAULT_LIMITS.maxConsecutiveDeliveryTurns) {
         trace("delivery_spam_guard", { turn, consecutiveDeliveryOnlyTurns });
-        await failJob(db, jobId, "delivery_spam_guard", runStats());
+        await failJob(db, jobId, "delivery_spam_guard", runStats(), messages);
         return { status: "failed", error: "delivery_spam_guard" };
       }
+      const reasoningParts = response.reasoning ?? [];
       const assistantMsg = {
         role: "assistant",
-        content: rawToolCalls.length > 0 ? rawToolCalls.map((tc) => ({
-          type: "tool-call",
-          toolCallId: tc.toolCallId,
-          toolName: tc.toolName,
-          input: tc.input
-        })) : response.text || ""
+        content: rawToolCalls.length > 0 ? [
+          ...reasoningParts,
+          ...rawToolCalls.map((tc) => ({
+            type: "tool-call",
+            toolCallId: tc.toolCallId,
+            toolName: tc.toolName,
+            input: tc.input
+          }))
+        ] : reasoningParts.length > 0 ? [...reasoningParts, { type: "text", text: response.text || "" }] : response.text || ""
       };
       messages = [...messages, assistantMsg];
       const returnResultCall = rawToolCalls.find((tc) => tc.toolName === "return_result");
@@ -28188,7 +28304,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
               continue;
             }
             trace("telegram_not_delivered", { turn, via: "text_branch" });
-            await failJob(db, jobId, "telegram_not_delivered", runStats());
+            await failJob(db, jobId, "telegram_not_delivered", runStats(), messages);
             return { status: "failed", error: "telegram_not_delivered" };
           }
           await completeJob(db, jobId, textContent, toolsUsed, runStats(), messages);
@@ -28200,7 +28316,8 @@ async function executeJob(jobId, deps, _runnerEnv) {
           messages = messages.slice(0, -1);
           continue;
         }
-        await failJob(db, jobId, "no_tool_calls_no_text", runStats());
+        messages = messages.slice(0, -1);
+        await failJob(db, jobId, "no_tool_calls_no_text", runStats(), messages);
         return { status: "failed", error: "no_tool_calls_no_text" };
       }
       const rawCallBlocks = rawToolCalls.map((tc) => ({
@@ -28235,7 +28352,13 @@ async function executeJob(jobId, deps, _runnerEnv) {
         toolsUsed = [.../* @__PURE__ */ new Set([...toolsUsed, call.name])];
         const toolDef = toolMap.get(call.name);
         if (!toolDef) {
-          await failJob(db, jobId, `whitelist_violation:${call.name}`, runStats());
+          await failJob(
+            db,
+            jobId,
+            `whitelist_violation:${call.name}`,
+            runStats(),
+            messages
+          );
           return { status: "failed", error: `whitelist_violation:${call.name}` };
         }
         if (call.name.startsWith("assign_")) {
@@ -28470,7 +28593,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
             continue;
           }
           trace("unresolved_tool_failure", { turn, stuck });
-          await failJob(db, jobId, "unresolved_tool_failure", runStats());
+          await failJob(db, jobId, "unresolved_tool_failure", runStats(), messages);
           return { status: "failed", error: "unresolved_tool_failure" };
         }
         const taskRows = await db.select({ id: agentTasks.id }).from(agentTasks).where(eq4(agentTasks.rootJobId, jobId));
@@ -28495,7 +28618,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
             continue;
           }
           trace("telegram_not_delivered", { turn, via: "return_result_branch" });
-          await failJob(db, jobId, "telegram_not_delivered", runStats());
+          await failJob(db, jobId, "telegram_not_delivered", runStats(), messages);
           return { status: "failed", error: "telegram_not_delivered" };
         }
         const finalResult = "";
@@ -28545,7 +28668,7 @@ async function executeJob(jobId, deps, _runnerEnv) {
         }
         if (recentTurnSignatures.length === maxNoProgressRepeats && recentTurnSignatures.every((s) => s === turnSignature)) {
           trace("no_progress_detected", { turn, repeats: recentTurnSignatures.length });
-          await failJob(db, jobId, "no_progress_detected", runStats());
+          await failJob(db, jobId, "no_progress_detected", runStats(), messages);
           return { status: "failed", error: "no_progress_detected" };
         }
       }
@@ -28561,30 +28684,36 @@ async function executeJob(jobId, deps, _runnerEnv) {
   } catch (err) {
     trace("catch", {
       errName: err instanceof Error ? err.name : "unknown",
-      errMsg: err instanceof Error ? err.message.slice(0, 200) : String(err)
+      errMsg: describeLlmError(err)
     });
     if (err instanceof ToolCallLimitExceededError) {
-      await failJob(db, jobId, err.code, runStats());
+      await failJob(db, jobId, err.code, runStats(), messages);
       return { status: "failed", error: err.code };
     }
     if (err instanceof ChainLimitExceededError) {
-      await failJob(db, jobId, err.code, runStats());
+      await failJob(db, jobId, err.code, runStats(), messages);
       return { status: "failed", error: err.code };
     }
     if (err instanceof DelegationDepthExceededError) {
-      await failJob(db, jobId, err.code, runStats());
+      await failJob(db, jobId, err.code, runStats(), messages);
       return { status: "failed", error: err.code };
     }
     if (err instanceof QuotaExhaustedError) {
-      await failJob(db, jobId, "quota_exhausted", runStats());
+      await failJob(db, jobId, "quota_exhausted", runStats(), messages);
       return { status: "failed", error: "quota_exhausted" };
     }
     if (err instanceof AllProvidersFailedError) {
-      await failJob(db, jobId, err.code, runStats());
+      await failJob(db, jobId, err.code, runStats(), messages);
       return { status: "failed", error: err.code };
     }
     if (err instanceof MessageStructureError) {
-      await failJob(db, jobId, `message_structure_invalid:${err.code}`, runStats());
+      await failJob(
+        db,
+        jobId,
+        `message_structure_invalid:${err.code}`,
+        runStats(),
+        messages
+      );
       return { status: "failed", error: `message_structure_invalid:${err.code}` };
     }
     if (err instanceof Error) {
@@ -28594,12 +28723,12 @@ async function executeJob(jobId, deps, _runnerEnv) {
       if (unavailableMatch) {
         const toolName = unavailableMatch[1] ?? "unknown_tool";
         const code = `whitelist_violation:${toolName}`;
-        await failJob(db, jobId, code, runStats());
+        await failJob(db, jobId, code, runStats(), messages);
         return { status: "failed", error: code };
       }
     }
-    const errorCode = err instanceof Error ? err.message.slice(0, 200) : "unknown_error";
-    await failJob(db, jobId, errorCode, runStats());
+    const errorCode = describeLlmError(err);
+    await failJob(db, jobId, errorCode, runStats(), messages);
     return { status: "failed", error: errorCode };
   } finally {
     for (const close of mcpClosers) {
@@ -29208,6 +29337,7 @@ var CHAT_TOOLS = {
     inputSchema: z89.object({ instruction: z89.string().min(1).max(4e3) })
   }
 };
+var ESCALATION_RECHECK = "Re-read your previous reply. If it committed to performing an action \u2014 running, launching, sending, fetching, creating, configuring, delegating, or any task or tool use \u2014 then your text ALONE did nothing: call the run_task tool NOW with a clear, self-contained instruction. If your reply was pure conversation, a question, or simply recalling a fact, do not call any tool \u2014 the conversation is complete.";
 async function runChatTurn(opts) {
   const { deps, entityId, agentId, conversationId, message } = opts;
   const db = deps.db;
@@ -29249,8 +29379,43 @@ async function runChatTurn(opts) {
     origin: "dashboard",
     surface: "chat"
   });
-  const rows = await db.select({ role: chatMessages.role, content: chatMessages.content }).from(chatMessages).where(eq4(chatMessages.conversationId, conversationId)).orderBy(desc(chatMessages.createdAt)).limit(HISTORY_LIMIT);
-  const messages = rows.reverse().map((r) => ({ role: r.role, content: r.content }));
+  const rows = await db.select({
+    role: chatMessages.role,
+    content: chatMessages.content,
+    jobId: chatMessages.jobId,
+    jobTask: agentJobs.task
+  }).from(chatMessages).leftJoin(agentJobs, eq4(chatMessages.jobId, agentJobs.id)).where(eq4(chatMessages.conversationId, conversationId)).orderBy(desc(chatMessages.createdAt)).limit(HISTORY_LIMIT);
+  const messages = [];
+  for (const r of rows.reverse()) {
+    if (r.role === "assistant" && r.jobId) {
+      const toolCallId = `hist-${r.jobId}`;
+      messages.push({
+        role: "assistant",
+        content: [
+          ...r.content ? [{ type: "text", text: r.content }] : [],
+          {
+            type: "tool-call",
+            toolCallId,
+            toolName: "run_task",
+            input: { instruction: r.jobTask ?? "" }
+          }
+        ]
+      });
+      messages.push({
+        role: "tool",
+        content: [
+          {
+            type: "tool-result",
+            toolCallId,
+            toolName: "run_task",
+            output: { type: "text", value: "Task dispatched." }
+          }
+        ]
+      });
+    } else {
+      messages.push({ role: r.role, content: r.content });
+    }
+  }
   let text22 = "";
   let runTask;
   try {
@@ -29261,7 +29426,23 @@ async function runChatTurn(opts) {
     });
     text22 = (response.text ?? "").trim();
     runTask = (response.toolCalls ?? []).find((tc) => tc.toolName === "run_task");
-  } catch {
+  } catch (err) {
+    console.warn(`[run-chat-turn] tools call failed (${agentRow.slug}):`, err.message);
+  }
+  if (!runTask && text22) {
+    try {
+      const recheck = await llmClient.generateText({
+        system: systemPrompt,
+        messages: [
+          ...messages,
+          { role: "assistant", content: text22 },
+          { role: "user", content: ESCALATION_RECHECK }
+        ],
+        tools: CHAT_TOOLS
+      });
+      runTask = (recheck.toolCalls ?? []).find((tc) => tc.toolName === "run_task");
+    } catch {
+    }
   }
   if (runTask) {
     const instruction = String(runTask.input?.instruction ?? "").trim() || message;

package/web/.next/BUILD_ID

@@ -1 +1 @@
-ZuUX-HBTQOhLf0tFI6JQI
+PV5qpGylrVW3jZ7zKHuhE

package/web/.next/build-manifest.json

@@ -4,8 +4,8 @@
   ],
   "devFiles": [],
   "lowPriorityFiles": [
-    "static/ZuUX-HBTQOhLf0tFI6JQI/_buildManifest.js",
-    "static/ZuUX-HBTQOhLf0tFI6JQI/_ssgManifest.js"
+    "static/PV5qpGylrVW3jZ7zKHuhE/_buildManifest.js",
+    "static/PV5qpGylrVW3jZ7zKHuhE/_ssgManifest.js"
   ],
   "rootMainFiles": [
     "static/chunks/webpack-3741bf0a7636d65e.js",

package/web/.next/prerender-manifest.json

@@ -78,8 +78,8 @@
   "dynamicRoutes": {},
   "notFoundRoutes": [],
   "preview": {
-    "previewModeId": "3ee52df782a2df2a4ee50789902ca7be",
-    "previewModeSigningKey": "89ca14b4b6d0a92f00c85b444ae7555c09e000f7447bbab9332647fa03c9e2ae",
-    "previewModeEncryptionKey": "c5e0ed4849dcf791d73cc1bb97004c7e4b13127cb4b820419f39aa7aea22021e"
+    "previewModeId": "ab66490d94469ea63e5777d5e5465b39",
+    "previewModeSigningKey": "3b725fc347ffafb5d6ab5906e8f5f39cb4523c75a6a3769d4e6779e014736265",
+    "previewModeEncryptionKey": "356915c85cb0a6b84c6f46dc1bceb092c3ed009fbca03946585cea052e5c754c"
   }
 }