From a5f420bfb4ff9f8d75b72ea6df865acaa544e6eb Mon Sep 17 00:00:00 2001
From: Alex Schmitt <actus7@gmail.com>
Date: Wed, 19 Aug 2026 15:50:03 -0300
Subject: [PATCH] fix(chat): harden auto agent workflow

---
 components/chat/chat-page.tsx                |  31 +++-
 lib/chat-utils.test.ts                       |  15 ++
 lib/chat-utils.ts                            |  14 ++
 lib/harness/harness.test.ts                  | 106 +++++++++++++-
 server/harness/core-tools.ts                 |  22 ++-
 server/harness/model-stream.ts               |  22 +++
 server/harness/prompt.ts                     |  19 ++-
 server/harness/runtime.ts                    | 143 ++++++++++++++++++-
 server/lib/routing/complexity-scorer.test.ts |  35 +++++
 server/lib/routing/complexity-scorer.ts      |  24 ++++
 server/providers/nvidianim.ts                |  13 +-
 server/tests/openai-compatible.test.ts       |  11 +-
 12 files changed, 437 insertions(+), 18 deletions(-)

diff --git a/components/chat/chat-page.tsx b/components/chat/chat-page.tsx
index b78e121..316c7ff 100644
--- a/components/chat/chat-page.tsx
+++ b/components/chat/chat-page.tsx
@@ -186,6 +186,7 @@ import {
   getUserMessageText,
   hydrateChatMessage,
   isHydratedAttachmentPart,
+  normalizeConversationTitle,
   parseApiErrorResponse,
   persistMessagesForConversation,
   releaseAttachmentPreview,
@@ -194,6 +195,7 @@ import {
   resolveModelSelectPlaceholder,
   resolveStickToBottom,
   resolveStreamErrorContent,
+  shouldUseHarnessRuntime,
   trimConversation,
   validateAttachmentCompatibility,
   type ChatMessage,
@@ -734,7 +736,7 @@ export function ChatPage() {
       modelId: selectedModelId || undefined,
       projectId: activeProjectId ?? undefined,
       providerId: selectedProviderId || undefined,
-      title: titleSeed || "Nova conversa",
+      title: normalizeConversationTitle(titleSeed),
     })
 
     setActiveConversationId(response.conversation.id)
@@ -1491,9 +1493,10 @@ export function ChatPage() {
         })
         setBrowserProviderAuthState("signed-in")
       } else if (
-        !temporaryChat &&
-        selectedProviderId !== AUTO_PROVIDER_ID &&
-        selectedModel?.capabilities.tools === true
+        shouldUseHarnessRuntime({
+          supportsTools: selectedModel?.capabilities.tools === true,
+          temporaryChat,
+        })
       ) {
         harnessConversationId = await ensureConversationId(
           (text || currentAttachments[0]?.fileName || "Nova conversa").slice(
@@ -1588,6 +1591,12 @@ export function ChatPage() {
                   updateAssistantToolCall(assistantMessageId, call)
                 }
               }
+              if (
+                event.type === "assistant/message" &&
+                typeof event.payload.modelLabel === "string"
+              ) {
+                effectiveModelLabel = event.payload.modelLabel
+              }
             },
           )
           fullText += result.text
@@ -2292,7 +2301,10 @@ export function ChatPage() {
               value={selectedProviderId}
               onValueChange={setSelectedProviderId}
             >
-              <SelectTrigger className="h-8 w-auto max-w-[min(200px,55vw)] shrink-0 text-xs sm:min-w-[140px] sm:max-w-[200px]">
+              <SelectTrigger
+                aria-label="Selecionar provider"
+                className="h-8 w-auto max-w-[min(200px,55vw)] shrink-0 text-xs sm:min-w-[140px] sm:max-w-[200px]"
+              >
                 <SelectValue placeholder="Provider" />
               </SelectTrigger>
               <SelectContent>
@@ -2402,7 +2414,10 @@ export function ChatPage() {
                   models.length === 0
                 }
               >
-                <SelectTrigger className="h-8 w-auto max-w-[min(240px,60vw)] shrink-0 text-xs sm:min-w-[140px] sm:max-w-[240px]">
+                <SelectTrigger
+                  aria-label="Selecionar modelo"
+                  className="h-8 w-auto max-w-[min(240px,60vw)] shrink-0 text-xs sm:min-w-[140px] sm:max-w-[240px]"
+                >
                   <SelectValue
                     placeholder={resolveModelSelectPlaceholder({
                       hasModels: !!selectedProvider?.hasModels,
@@ -3277,6 +3292,7 @@ export function ChatPage() {
                     <TooltipTrigger asChild>
                       <span>
                         <Button
+                          aria-label="Anexar arquivo"
                           variant="ghost"
                           size="icon-xs"
                           className="size-8 md:size-6"
@@ -3310,6 +3326,7 @@ export function ChatPage() {
                       value={activeProjectId ?? "__none__"}
                     >
                       <SelectTrigger
+                        aria-label="Selecionar projeto da conversa"
                         className="h-7 w-auto max-w-[150px] shrink-0 border-dashed text-[11px]"
                         title="Projeto do contexto da conversa"
                       >
@@ -3336,6 +3353,7 @@ export function ChatPage() {
                 </div>
                 {pending ? (
                   <Button
+                    aria-label="Parar geração"
                     size="sm"
                     variant="destructive"
                     className="h-7 text-xs"
@@ -3346,6 +3364,7 @@ export function ChatPage() {
                   </Button>
                 ) : (
                   <InputGroupButton
+                    aria-label="Enviar mensagem"
                     size="sm"
                     disabled={
                       (!input.trim() && attachments.length === 0) ||
diff --git a/lib/chat-utils.test.ts b/lib/chat-utils.test.ts
index 5310534..142e3c4 100644
--- a/lib/chat-utils.test.ts
+++ b/lib/chat-utils.test.ts
@@ -10,16 +10,31 @@ import {
   formatMessageTimestamp,
   getUserMessageText,
   hydrateChatMessage,
+  normalizeConversationTitle,
   parseApiErrorResponse,
   resolveAssistantModelLabel,
   resolveModelFallbackFromHeaders,
   resolveModelSelectPlaceholder,
   resolveStickToBottom,
   resolveStreamErrorContent,
+  shouldUseHarnessRuntime,
   STREAM_INTERRUPTED_NOTE,
   validateAttachmentCompatibility,
 } from "./chat-utils";
 
+describe("conversation runtime helpers", () => {
+  it("normalizes long prompts into valid conversation titles", () => {
+    expect(normalizeConversationTitle(`  ${"x".repeat(250)}  `)).toBe("x".repeat(200));
+    expect(normalizeConversationTitle("   ")).toBe("Nova conversa");
+  });
+
+  it("uses the harness for every durable tool-capable provider, including Auto", () => {
+    expect(shouldUseHarnessRuntime({ supportsTools: true, temporaryChat: false })).toBe(true);
+    expect(shouldUseHarnessRuntime({ supportsTools: false, temporaryChat: false })).toBe(false);
+    expect(shouldUseHarnessRuntime({ supportsTools: true, temporaryChat: true })).toBe(false);
+  });
+});
+
 const MODELS: ProviderModel[] = [
   { capabilities: { documents: true, images: false }, id: "gpt-4o", name: "GPT-4o" },
   { capabilities: { documents: true, images: true }, id: "gpt-4o-mini", name: "GPT-4o mini" },
diff --git a/lib/chat-utils.ts b/lib/chat-utils.ts
index 9381095..644ad3c 100644
--- a/lib/chat-utils.ts
+++ b/lib/chat-utils.ts
@@ -106,6 +106,20 @@ export const EMPTY_STATE_PROMPTS = [
   "Me ajude a diagnosticar um erro 500.",
 ] as const
 
+export const MAX_CONVERSATION_TITLE_LENGTH = 200
+
+export function normalizeConversationTitle(titleSeed: string): string {
+  const title = titleSeed.trim() || "Nova conversa"
+  return title.slice(0, MAX_CONVERSATION_TITLE_LENGTH)
+}
+
+export function shouldUseHarnessRuntime(input: {
+  supportsTools: boolean
+  temporaryChat: boolean
+}): boolean {
+  return !input.temporaryChat && input.supportsTools
+}
+
 export type ChatRequestError = Error & {
   status?: number
   suppressToast?: boolean
diff --git a/lib/harness/harness.test.ts b/lib/harness/harness.test.ts
index 04dd0b6..d877ffd 100644
--- a/lib/harness/harness.test.ts
+++ b/lib/harness/harness.test.ts
@@ -5,17 +5,100 @@ import { join } from "node:path"
 import { parseChatStream } from "../chat-stream"
 import { consumeHarnessStream } from "./client"
 import type { HarnessEvent } from "./contracts"
-import { compactModelMessages } from "../../server/harness/prompt"
+import {
+  compactModelMessages,
+  HARNESS_COMPLETION_GUIDANCE,
+  HARNESS_ITERATIVE_DELIVERY_GUIDANCE,
+  HARNESS_NO_PROJECT_GUIDANCE,
+  HARNESS_TOOL_USE_GUIDANCE,
+} from "../../server/harness/prompt"
 import { HarnessRegistry } from "../../server/harness/registry"
-import { consumeHarnessModelResponse } from "../../server/harness/model-stream"
+import {
+  consumeHarnessModelResponse,
+  isModelOutputLimitFinishReason,
+} from "../../server/harness/model-stream"
 import { assertPublicHttpUrl } from "../../server/harness/core-tools"
 import { collectHarnessEventPages } from "../../server/harness/events"
 import { toHarnessJson } from "../../server/harness/json"
 import { assertSecureMcpUrl, mcpToolRisk, readMcpRpcResponse, safeMcpToolName } from "../../server/harness/mcp"
 import { parseSingleMessagePart } from "../../server/lib/conversation-attachments"
-import { limitHarnessToolCalls, MAX_TOOL_CALLS_PER_STEP } from "../../server/harness/runtime"
+import {
+  limitHarnessToolCalls,
+  MAX_TOOL_CALLS_PER_STEP,
+  selectHarnessToolSchemas,
+} from "../../server/harness/runtime"
 
 describe("harness stream protocol", () => {
+  it("keeps ordinary multi-role refinement from pausing for internal tracking approval", () => {
+    expect(HARNESS_TOOL_USE_GUIDANCE).toContain("creator/critic")
+    expect(HARNESS_TOOL_USE_GUIDANCE).toContain("goal_write")
+    expect(HARNESS_TOOL_USE_GUIDANCE).toContain("pause for approval")
+  })
+
+  it("recognizes provider finish reasons that require automatic continuation", () => {
+    expect(isModelOutputLimitFinishReason("length")).toBe(true)
+    expect(isModelOutputLimitFinishReason("max_tokens")).toBe(true)
+    expect(isModelOutputLimitFinishReason("max-output-tokens")).toBe(true)
+    expect(isModelOutputLimitFinishReason("stop")).toBe(false)
+    expect(HARNESS_COMPLETION_GUIDANCE).toContain("numbered round")
+    expect(HARNESS_NO_PROJECT_GUIDANCE).toContain("project_file_write")
+    expect(HARNESS_ITERATIVE_DELIVERY_GUIDANCE).toContain("final round")
+  })
+
+  it("hides unusable stateful tools for an ordinary multi-role artifact request", () => {
+    const tool = (name: string) => ({
+      function: { description: name, name, parameters: { type: "object" as const } },
+      type: "function" as const,
+    })
+    const selected = selectHarnessToolSchemas({
+      messages: [{ content: "Atue como criador e crítico e crie um jogo", role: "user" }],
+      projectId: null,
+      tools: [
+        tool("goal_write"),
+        tool("todo_write"),
+        tool("subagent"),
+        tool("project_file_write"),
+        tool("web_search"),
+        tool("memory_search"),
+        tool("session_event_search"),
+      ],
+    })
+    expect(selected.map((item) => item.function.name)).toEqual([])
+  })
+
+  it("exposes stateful tools only for explicit persistence or delegation requests", () => {
+    const tool = (name: string) => ({
+      function: { description: name, name, parameters: { type: "object" as const } },
+      type: "function" as const,
+    })
+    const selected = selectHarnessToolSchemas({
+      messages: [{ content: "Salve um plano persistente e delegue a um subagente", role: "user" }],
+      projectId: "project-1",
+      tools: [tool("plan_write"), tool("subagent"), tool("project_file_write")],
+    })
+    expect(selected.map((item) => item.function.name)).toEqual([
+      "plan_write",
+      "subagent",
+      "project_file_write",
+    ])
+  })
+
+  it("exposes web tools only when the user explicitly requests research", () => {
+    const tool = (name: string) => ({
+      function: { description: name, name, parameters: { type: "object" as const } },
+      type: "function" as const,
+    })
+    const selected = selectHarnessToolSchemas({
+      messages: [{ content: "Pesquise na web e cite fontes", role: "user" }],
+      projectId: null,
+      tools: [tool("web_search"), tool("web_fetch"), tool("goal_write")],
+    })
+    expect(selected.map((item) => item.function.name)).toEqual([
+      "web_search",
+      "web_fetch",
+    ])
+  })
+
   it("parses streamed harness events and terminal status", async () => {
     const events: HarnessEvent[] = [
       {
@@ -155,6 +238,23 @@ describe("harness stream protocol", () => {
     })
   })
 
+  it("keeps the effective Auto routing metadata", async () => {
+    const result = await consumeHarnessModelResponse(
+      new Response('0:"ok"\nd:{"finishReason":"stop"}\n', {
+        headers: {
+          "x-modelhub-model": "qwen3.5-27b",
+          "x-modelhub-provider": "groq",
+          "x-modelhub-tier": "reasoning",
+        },
+      }),
+    )
+    expect(result.routing).toEqual({
+      modelId: "qwen3.5-27b",
+      providerId: "groq",
+      tier: "reasoning",
+    })
+  })
+
   it("persists streamed deltas in source order even when writes have different latency", async () => {
     const observed: string[] = []
     await consumeHarnessModelResponse(
diff --git a/server/harness/core-tools.ts b/server/harness/core-tools.ts
index 8251379..c4ef11f 100644
--- a/server/harness/core-tools.ts
+++ b/server/harness/core-tools.ts
@@ -97,7 +97,27 @@ export async function safeFetchPublicHttpUrl(
   const headers = Object.fromEntries(new Headers(init.headers).entries())
   headers.host = url.host
   const pinnedLookup: LookupFunction = (_hostname, _options, callback) => {
-    callback(null, address, family)
+    const wantsAll =
+      typeof _options === "object" &&
+      _options !== null &&
+      "all" in _options &&
+      _options.all === true
+    if (wantsAll) {
+      ;(
+        callback as (
+          error: NodeJS.ErrnoException | null,
+          addresses: Array<{ address: string; family: number }>,
+        ) => void
+      )(null, [{ address, family }])
+      return
+    }
+    ;(
+      callback as (
+        error: NodeJS.ErrnoException | null,
+        resolvedAddress: string,
+        resolvedFamily: number,
+      ) => void
+    )(null, address, family)
   }
   return new Promise<Response>((resolve, reject) => {
     const request = (url.protocol === "https:" ? httpsRequest : httpRequest)(
diff --git a/server/harness/model-stream.ts b/server/harness/model-stream.ts
index c7e7a3d..d799336 100644
--- a/server/harness/model-stream.ts
+++ b/server/harness/model-stream.ts
@@ -3,10 +3,21 @@ import type { HarnessToolCall } from "../../lib/harness/contracts"
 
 export type HarnessModelResult = {
   finishReason: string
+  routing?: {
+    modelId: string
+    providerId: string
+    tier: string | null
+  }
   text: string
   toolCalls: HarnessToolCall[]
 }
 
+export function isModelOutputLimitFinishReason(finishReason: string): boolean {
+  return /^(?:length|max[_-]?(?:output[_-]?)?tokens?|token[_-]?limit)$/i.test(
+    finishReason.trim(),
+  )
+}
+
 export async function consumeHarnessModelResponse(
   response: Response,
   onTextDelta?: (delta: string) => void | Promise<void>,
@@ -36,8 +47,19 @@ export async function consumeHarnessModelResponse(
   await pendingDeltaWrites
 
   if (parsed.errorMessage) throw new Error(parsed.errorMessage)
+  const routingProviderId = response.headers.get("x-modelhub-provider")
+  const routingModelId = response.headers.get("x-modelhub-model")
   return {
     finishReason: parsed.finishReason ?? (toolCalls.size > 0 ? "tool-calls" : "stop"),
+    ...(routingProviderId && routingModelId
+      ? {
+          routing: {
+            modelId: routingModelId,
+            providerId: routingProviderId,
+            tier: response.headers.get("x-modelhub-tier"),
+          },
+        }
+      : {}),
     text: parsed.text,
     toolCalls: [...toolCalls.values()],
   }
diff --git a/server/harness/prompt.ts b/server/harness/prompt.ts
index ac4ca31..2778b64 100644
--- a/server/harness/prompt.ts
+++ b/server/harness/prompt.ts
@@ -2,6 +2,18 @@ import { prisma } from "../lib/db"
 import { listAllHarnessEvents } from "./events"
 import type { HarnessRegistry } from "./registry"
 
+export const HARNESS_TOOL_USE_GUIDANCE =
+  "Do not call goal_write, plan_write, todo_write, or subagent merely to restate, track, or role-play an ordinary user task. These tools pause for approval. Complete requested creator/critic or multi-role refinement directly unless the user explicitly asks to persist tracking state or delegate a separate task."
+
+export const HARNESS_COMPLETION_GUIDANCE =
+  "Before ending, verify that every explicitly requested numbered round, section, and deliverable is present. If the latest assistant content is visibly truncated or the requested format is incomplete, continue exactly where it stopped without repeating completed content."
+
+export const HARNESS_NO_PROJECT_GUIDANCE =
+  "No active project is attached. Do not call project_context, project_file_read, or project_file_write. Deliver requested code and artifacts directly in the response."
+
+export const HARNESS_ITERATIVE_DELIVERY_GUIDANCE =
+  "For multi-round refinement, keep non-final artifact deliveries concise: state concrete changes and show only the changed excerpts or patch. Reserve the full self-contained artifact for the final round unless the user explicitly requires the complete artifact in every round."
+
 export type ModelMessage = {
   content: string | unknown[]
   name?: string
@@ -156,10 +168,15 @@ export async function buildHarnessSystemPrompt(input: {
   return [
     "You are ModelHub's agent runtime. Work in explicit, bounded steps and use tools when they materially improve the answer.",
     "Never claim that a tool ran unless a tool result is present. Respect approval denials and capability limits.",
+    HARNESS_TOOL_USE_GUIDANCE,
+    HARNESS_COMPLETION_GUIDANCE,
+    HARNESS_ITERATIVE_DELIVERY_GUIDANCE,
     settings?.customInstructionsAbout ? `User context:\n${settings.customInstructionsAbout}` : "",
     settings?.customInstructionsStyle ? `Response preferences:\n${settings.customInstructionsStyle}` : "",
     memories.length ? `Saved memories:\n${memories.map((memory) => `- ${memory.content}`).join("\n")}` : "",
-    project ? `Active project: ${project.name}\n${project.instructions ?? ""}` : "",
+    project
+      ? `Active project: ${project.name}\n${project.instructions ?? ""}`
+      : HARNESS_NO_PROJECT_GUIDANCE,
     skills.length
       ? `Active skills:\n${skills.map((skill) => `## ${skill.name}\n${skill.content}`).join("\n\n")}`
       : "",
diff --git a/server/harness/runtime.ts b/server/harness/runtime.ts
index 9694ea7..0a6cc0a 100644
--- a/server/harness/runtime.ts
+++ b/server/harness/runtime.ts
@@ -3,7 +3,11 @@ import type { HarnessEvent, HarnessRunStatus, HarnessToolCall } from "../../lib/
 import { prisma } from "../lib/db"
 import { appendHarnessEvent, ensureLegacyMessageEvents, HarnessLeaseLostError, serializeHarnessEvent } from "./events"
 import { toHarnessJson } from "./json"
-import { consumeHarnessModelResponse } from "./model-stream"
+import {
+  consumeHarnessModelResponse,
+  isModelOutputLimitFinishReason,
+  type HarnessModelResult,
+} from "./model-stream"
 import { createMcpPlugin } from "./mcp"
 import { buildHarnessSystemPrompt, compactModelMessages, deriveMessagesFromEvents, type ModelMessage } from "./prompt"
 import { HarnessRegistry, type HarnessToolDefinition } from "./registry"
@@ -13,9 +17,73 @@ const EXECUTION_SLICE_MS = 50_000
 const LEASE_MS = 60_000
 export const MAX_TOOL_CALLS_PER_STEP = 16
 export const MAX_SUBAGENTS_PER_RUN = 4
+export const MAX_OUTPUT_LIMIT_CONTINUATIONS = 8
 const APPROVAL_EXECUTION_MS = 60_000
 const activeToolExecutions = new Map<string, Set<AbortController>>()
 
+type OpenAiHarnessToolSchema = ReturnType<HarnessRegistry["openAiSchemas"]>[number]
+
+const PERSISTENT_COORDINATION_REQUEST =
+  /\b(goal_write|plan_write|todo_write|salv(?:e|ar)|persist(?:a|ir|e)|registre|atualize)\b[\s\S]{0,80}\b(objetivo|goal|plano|plan|tarefas?|todos?)\b/i
+const SUBAGENT_REQUEST =
+  /\b(subagent|subagente|deleg(?:ate|ar|ue)|separate agent|agente separado)\b/i
+const WEB_TOOL_REQUEST =
+  /\b(search|research|look up|browse|web|internet|source|url|pesquis\w*|busc\w*|procure|consulte|fontes?)\b|https?:\/\//i
+const CONTEXT_SEARCH_REQUEST =
+  /\b(memory|memories|mem[oó]ria|history|hist[oó]rico|conversation events?|eventos? da conversa|previous conversation|conversa anterior)\b/i
+
+export function selectHarnessToolSchemas(input: {
+  messages: ModelMessage[]
+  projectId: string | null
+  tools: OpenAiHarnessToolSchema[]
+}): OpenAiHarnessToolSchema[] {
+  const latestUserContent = [...input.messages]
+    .reverse()
+    .find((message) => message.role === "user")?.content
+  const latestUserText =
+    typeof latestUserContent === "string"
+      ? latestUserContent
+      : JSON.stringify(latestUserContent ?? "")
+  const exposePersistentCoordination =
+    PERSISTENT_COORDINATION_REQUEST.test(latestUserText)
+  const exposeSubagent = SUBAGENT_REQUEST.test(latestUserText)
+  const exposeWebTools = WEB_TOOL_REQUEST.test(latestUserText)
+  const exposeContextSearch = CONTEXT_SEARCH_REQUEST.test(latestUserText)
+
+  return input.tools.filter((tool) => {
+    const name = tool.function.name
+    if (!input.projectId && name.startsWith("project_")) return false
+    if (
+      ["goal_write", "plan_write", "todo_write"].includes(name) &&
+      !exposePersistentCoordination
+    ) {
+      return false
+    }
+    if (name === "subagent" && !exposeSubagent) return false
+    if (["web_fetch", "web_search"].includes(name) && !exposeWebTools) {
+      return false
+    }
+    if (
+      ["memory_search", "session_event_search"].includes(name) &&
+      !exposeContextSearch
+    ) {
+      return false
+    }
+    return true
+  })
+}
+
+function resolvedHarnessModelLabel(
+  requestedModel: string,
+  routing: HarnessModelResult["routing"],
+): string {
+  if (!routing || requestedModel !== "auto") return requestedModel
+  const tier = routing.tier
+    ? `${routing.tier.charAt(0).toUpperCase()}${routing.tier.slice(1)}`
+    : "Auto"
+  return `Auto · ${tier} (${routing.providerId}/${routing.modelId})`
+}
+
 export function abortActiveHarnessRun(runId: string): void {
   for (const controller of activeToolExecutions.get(runId) ?? []) controller.abort("Harness run cancelled")
 }
@@ -365,6 +433,7 @@ async function completeAssistantRun(input: {
           content: input.content,
           finishReason: input.finishReason,
           messageId: input.assistantMessageId,
+          modelLabel: input.model,
           toolCalls: [],
         }),
         runId: input.runId,
@@ -533,6 +602,7 @@ export async function runHarness(input: RunHarnessInput): Promise<HarnessRunStat
     let messages = input.initialMessages?.length ? input.initialMessages : await deriveMessagesFromEvents(run.conversationId)
     const systemPrompt = await buildHarnessSystemPrompt({ projectId: run.conversation.projectId, registry, userId: run.userId })
     let currentStepCount = run.stepCount
+    let outputLimitContinuations = 0
     const unresolved = await unresolvedToolCalls(run.id)
     if (unresolved) {
       const stepNumber = currentStepCount + 1
@@ -602,7 +672,11 @@ export async function runHarness(input: RunHarnessInput): Promise<HarnessRunStat
           model: run.modelId ?? "auto",
           stream: true,
           tool_choice: "auto",
-          tools: registry.openAiSchemas(),
+          tools: selectHarnessToolSchemas({
+            messages,
+            projectId: run.conversation.projectId,
+            tools: registry.openAiSchemas(),
+          }),
         },
         signal: input.signal,
       })
@@ -636,14 +710,75 @@ export async function runHarness(input: RunHarnessInput): Promise<HarnessRunStat
       })
       await flushDurableChunks()
       await renewRunLease(run.id, input.leaseToken)
+      const resultModelLabel = resolvedHarnessModelLabel(
+        run.modelId ?? "auto",
+        result.routing,
+      )
       if (result.toolCalls.length === 0) {
+        if (
+          isModelOutputLimitFinishReason(result.finishReason) &&
+          outputLimitContinuations < MAX_OUTPUT_LIMIT_CONTINUATIONS &&
+          stepNumber < run.maxSteps
+        ) {
+          outputLimitContinuations += 1
+          await emit(
+            {
+              conversationId: run.conversationId,
+              payload: {
+                content: result.text,
+                continuationRequired: true,
+                finishReason: result.finishReason,
+                messageId: assistantMessageId,
+                modelLabel: resultModelLabel,
+                toolCalls: [],
+              },
+              runId: run.id,
+              stepId,
+              turnId,
+              type: "assistant/message",
+            },
+            input.onEvent,
+            input.leaseToken,
+          )
+          messages.push({ content: result.text, role: "assistant" })
+          messages.push({
+            content:
+              "Continue exactly from the last emitted character. Do not repeat completed content. Finish every section, numbered round, and deliverable required by the original user request.",
+            role: "system",
+          })
+          await emit(
+            {
+              conversationId: run.conversationId,
+              payload: {
+                finishReason: result.finishReason,
+                outputLimitContinuation: outputLimitContinuations,
+                stepNumber,
+              },
+              runId: run.id,
+              stepId,
+              turnId,
+              type: "step/end",
+            },
+            input.onEvent,
+            input.leaseToken,
+          )
+          const advanced = await prisma.agentRun.updateMany({
+            where: { id: run.id, leaseToken: input.leaseToken, status: "running" },
+            data: {
+              leaseExpiresAt: new Date(Date.now() + LEASE_MS),
+              stepCount: stepNumber,
+            },
+          })
+          if (advanced.count !== 1) throw new HarnessLeaseLostError()
+          continue
+        }
         await completeAssistantRun({
           assistantMessageId,
           content: result.text,
           conversationId: run.conversationId,
           finishReason: result.finishReason,
           leaseToken: input.leaseToken,
-          model: run.modelId ?? "auto",
+          model: resultModelLabel,
           onEvent: input.onEvent,
           runId: run.id,
           stepId,
@@ -723,7 +858,7 @@ export async function runHarness(input: RunHarnessInput): Promise<HarnessRunStat
           content: result.text,
           conversationId: run.conversationId,
           id: assistantMessageId,
-          model: run.modelId ?? "auto",
+          model: resultModelLabel,
           toolCalls: await projectedToolCalls(run.id),
         })
         await setRunStatus(run.id, input.leaseToken, "waiting_approval")
diff --git a/server/lib/routing/complexity-scorer.test.ts b/server/lib/routing/complexity-scorer.test.ts
index 29c66aa..c6aa70c 100644
--- a/server/lib/routing/complexity-scorer.test.ts
+++ b/server/lib/routing/complexity-scorer.test.ts
@@ -78,6 +78,41 @@ describe('scoreComplexity', () => {
     expect(result.rawScore).toBeGreaterThan(40)
   })
 
+  it('routes autonomous iterative multi-agent creation to reasoning', () => {
+    const result = scoreComplexity(msg(`
+      Atue como um sistema multiagente autônomo de refinamento iterativo composto por dois papéis:
+      Agente Criador e Agente Crítico. Crie o jogo fapibird.
+      Realize 3 rodadas consecutivas. Em cada rodada apresente Crítica, Ajustes e Entrega.
+      Encerre somente após a terceira rodada com o artefato totalmente polido.
+    `))
+    expect(result.tier).toBe('reasoning')
+    expect(result.signals).toEqual(expect.arrayContaining([
+      'autonomous_multi_agent',
+      'iterative_refinement',
+      'complex_deliverable',
+      'autonomous_workflow_floor',
+    ]))
+    expect(result.confidence).toBeGreaterThanOrEqual(0.9)
+  })
+
+  it('routes the full Portuguese creator/critic workflow to reasoning', () => {
+    const result = scoreComplexity(msg(`
+      Atue como um sistema multiagente autônomo de refinamento iterativo composto por dois papéis:
+      Agente Criador: Responsável por gerar o rascunho inicial e implementar todas as melhorias estruturais, visuais e conceituais exigidas a cada ciclo.
+      Agente Crítico: Um revisor rigoroso e implacável de padrão "AAA" que não aceita soluções genéricas, superficiais ou incompletas. Seu foco é apontar falhas de profundidade, narrativa, acabamento, consistência e detalhes ausentes.
+      Objetivo da tarefa: criar o jogo fapibird.
+      Realize 3 rodadas consecutivas de refinamento autônomo. Em cada rodada, apresente Crítica, Ajustes e Entrega. O loop se encerra após a terceira rodada com a versão final totalmente polida e comentada.
+    `))
+    expect(result.tier).toBe('reasoning')
+    expect(result.rawScore).toBeGreaterThanOrEqual(50)
+    expect(result.signals).toEqual(expect.arrayContaining([
+      'autonomous_multi_agent',
+      'iterative_refinement',
+      'quality_review',
+      'autonomous_workflow_floor',
+    ]))
+  })
+
   it('detects formal_logic signal', () => {
     const result = scoreComplexity(msg('Therefore, we can conclude that P implies Q.'))
     expect(result.signals).toContain('formal_logic')
diff --git a/server/lib/routing/complexity-scorer.ts b/server/lib/routing/complexity-scorer.ts
index 31bdb57..97c9671 100644
--- a/server/lib/routing/complexity-scorer.ts
+++ b/server/lib/routing/complexity-scorer.ts
@@ -67,6 +67,7 @@ const MULTI_STEP_KEYWORDS = [
   'step by step', 'passo a passo', 'first.*then', 'sequentially', 'sequencialmente',
   'prove that', 'prove ', 'demonstrate', 'demonstre', 'derive ', 'derive that',
   'walk me through', 'explain in detail', 'in depth', 'em detalhes',
+  'iterative refinement', 'refinamento iterativo', 'rodadas consecutivas', 'consecutive rounds',
 ]
 
 const REASONING_KEYWORDS = [
@@ -86,6 +87,10 @@ const MATH_PATTERNS = [
 const CODE_PATTERN = /```[\s\S]{20,}/
 const INLINE_CODE_PATTERN = /`[^`]+`/g
 const CODE_REQUEST_PATTERN = /\b(c[oó]digo|code|python|typescript|javascript|fun[cç][aã]o|function|script|programa|implemente|implement|escreva|write)\b/i
+const AUTONOMOUS_MULTI_AGENT_PATTERN = /\b(multi[-\s]?agents?|multiagentes?|autonomous agents?|sistema (?:de )?agentes|agente criador|creator agent|agente cr[ií]tico|critic agent)\b/i
+const ITERATIVE_REFINEMENT_PATTERN = /\b(iterative refinement|refinamento iterativo|(?:three|tr[eê]s|\d+)\s+(?:consecutive\s+)?(?:rounds?|rodadas?|iterations?|itera[cç][oõ]es|cycles?|ciclos?))\b/i
+const COMPLEX_DELIVERABLE_PATTERN = /\b(create|build|develop|implement|criar|crie|construir|construa|desenvolver|desenvolva|implementar|implemente)\b[\s\S]{0,120}\b(game|jogo|app|aplicativo|artefato|artifact|interface|site|canvas)\b/i
+const QUALITY_REVIEW_PATTERN = /\b(critic|cr[ií]tic[oa]|reviewer|revisor|rigoroso|implac[aá]vel|polid[oa]|production-ready|padr[aã]o\s+aaa)\b/i
 const ARCHITECTURE_REQUEST_PATTERN = /\b(arquitetura|architecture|architect|design a system)\b/i
 const DECISION_REQUEST_PATTERN = /\b(como decidir|decidir|escolher|choose|select|selecionar|balancear|otimizar|optimize)\b/i
 const DECISION_OBJECTIVE_PATTERN = /\b(custo|cost|lat[eê]ncia|latency|taxa de erro|error rate|or[cç]amento|budget|fallback|confiabilidade|reliability|qualidade|quality)\b/gi
@@ -211,6 +216,17 @@ export function scoreComplexity(
     score += 4; signals.push('structured_output')
   }
 
+  // 14. Orquestração autônoma e refinamento iterativo exigem modelos que
+  // consigam manter papéis, estado e critérios de parada por vários ciclos.
+  const autonomousMultiAgent = AUTONOMOUS_MULTI_AGENT_PATTERN.test(lastText)
+  const iterativeRefinement = ITERATIVE_REFINEMENT_PATTERN.test(lastText)
+  const complexDeliverable = COMPLEX_DELIVERABLE_PATTERN.test(lastText)
+  const qualityReview = QUALITY_REVIEW_PATTERN.test(lastText)
+  if (autonomousMultiAgent) { score += 20; signals.push('autonomous_multi_agent') }
+  if (iterativeRefinement) { score += 15; signals.push('iterative_refinement') }
+  if (complexDeliverable) { score += 12; signals.push('complex_deliverable') }
+  if (qualityReview) { score += 8; signals.push('quality_review') }
+
   score = Math.min(score, 100)
   let tier = tierFromScore(score)
   let confidence = confidenceFromScore(score)
@@ -258,6 +274,14 @@ export function scoreComplexity(
     signals.push('multi_objective_reasoning')
   }
 
+  // Um workflow multiagente com ciclos de crítica e refinamento é uma tarefa
+  // de raciocínio mesmo sem vocabulário técnico clássico.
+  if (autonomousMultiAgent && iterativeRefinement) {
+    tier = floorTier(tier, 'reasoning')
+    confidence = Math.max(confidence, 0.9)
+    signals.push('autonomous_workflow_floor')
+  }
+
   // Tools ativas exigem um modelo que saiba usá-las → piso "standard".
   if (options.hasTools && tier === 'simple' && !signals.includes('heartbeat')) {
     tier = 'standard'
diff --git a/server/providers/nvidianim.ts b/server/providers/nvidianim.ts
index b62e215..6e81355 100644
--- a/server/providers/nvidianim.ts
+++ b/server/providers/nvidianim.ts
@@ -9,10 +9,19 @@ export function isNvidiaNimChatModel(model: { id: string }): boolean {
   return !NON_CHAT_MODEL_RE.test(model.id)
 }
 
+const NVIDIA_NIM_MODEL_ALIASES: Record<string, string> = {
+  'meta/codellama-70b': 'nvidia/nemotron-3-ultra-550b-a55b',
+  'nvidia/llama-3.1-nemotron-ultra-253b-v1': 'nvidia/nemotron-3-ultra-550b-a55b',
+}
+
+export function resolveNvidiaNimModelId(modelId: string): string {
+  return NVIDIA_NIM_MODEL_ALIASES[modelId] ?? modelId
+}
+
 export const models = [
   { capabilities: { documents: true, images: false, tools: true }, id: 'nvidia/nemotron-3-super-120b-a12b', name: 'Nemotron 3 Super 120B A12B' },
+  { capabilities: { documents: true, images: false, tools: true }, id: 'nvidia/nemotron-3-ultra-550b-a55b', name: 'Nemotron 3 Ultra 550B A55B' },
   { capabilities: { documents: true, images: false, tools: true }, id: 'nvidia/llama-3.3-nemotron-super-49b-v1.5', name: 'Nemotron Super 49B v1.5' },
-  { capabilities: { documents: true, images: false, tools: true }, id: 'nvidia/llama-3.1-nemotron-ultra-253b-v1', name: 'Nemotron Ultra 253B' },
   { capabilities: { documents: true, images: false, tools: true }, id: 'nvidia/llama-3.1-nemotron-70b-instruct', name: 'Nemotron 70B Instruct' },
   { capabilities: { documents: true, images: false, tools: true }, id: 'nvidia/nemotron-3-nano-30b-a3b', name: 'Nemotron-3 Nano 30B' },
   { capabilities: { documents: true, images: false, tools: true }, id: 'nvidia/nemotron-nano-9b-v2', name: 'Nemotron Nano 9B v2' },
@@ -86,7 +95,7 @@ const app = createProviderApp({
         fallbackModelIds: FALLBACK_MODEL_IDS,
         isModelUnavailableError: isNvidiaNimModelUnavailableError,
       },
-      { messages, modelId, rawBody },
+      { messages, modelId: resolveNvidiaNimModelId(modelId), rawBody },
       credentials,
     ),
   fetchModels: createOpenAiFetchModels({
diff --git a/server/tests/openai-compatible.test.ts b/server/tests/openai-compatible.test.ts
index b1933aa..275a18c 100644
--- a/server/tests/openai-compatible.test.ts
+++ b/server/tests/openai-compatible.test.ts
@@ -11,7 +11,7 @@ vi.mock("../env", () => ({}));
 vi.mock("@/lib/auth/server", () => ({ auth: { getSession: vi.fn().mockResolvedValue({ data: null }) } }));
 
 const { chatViaOpenAiCompatible, createOpenAiFetchModels } = await import("../lib/openai-compatible");
-const { buildNvidiaNimBody, isNvidiaNimChatModel, isNvidiaNimModelUnavailableError, testNvidiaNimCredentials } = await import("../providers/nvidianim");
+const { buildNvidiaNimBody, isNvidiaNimChatModel, isNvidiaNimModelUnavailableError, resolveNvidiaNimModelId, testNvidiaNimCredentials } = await import("../providers/nvidianim");
 
 describe("OpenAI-compatible provider helpers", () => {
   const originalFetch = globalThis.fetch;
@@ -214,4 +214,13 @@ describe("OpenAI-compatible provider helpers", () => {
       rawBody: {},
     })).not.toHaveProperty("stream_options");
   });
+
+  it("maps retired NVIDIA NIM model IDs before the upstream request", () => {
+    expect(resolveNvidiaNimModelId("nvidia/llama-3.1-nemotron-ultra-253b-v1"))
+      .toBe("nvidia/nemotron-3-ultra-550b-a55b");
+    expect(resolveNvidiaNimModelId("meta/codellama-70b"))
+      .toBe("nvidia/nemotron-3-ultra-550b-a55b");
+    expect(resolveNvidiaNimModelId("openai/gpt-oss-20b"))
+      .toBe("openai/gpt-oss-20b");
+  });
 });