@agent-native/core 0.99.3 → 0.100.1

package/corpus/core/src/agent/production-agent.ts

@@ -48,6 +48,10 @@ import {
   type ReasoningEffort,
 } from "../shared/reasoning-effort.js";
 import { actionPreparationContinuationNote } from "./action-continuation-guidance.js";
+import {
+  buildSystemManifestSections,
+  readContextXraySystemSections,
+} from "./context-xray/manifest.js";
 import {
   AGENT_CHAT_BACKGROUND_RUN_FIELD,
   AGENT_CHAT_PROCESS_RUN_PATH,
@@ -1799,6 +1803,12 @@ export interface AgentLoopToolResultSummary {
 
 export interface AgentLoopFinalResponseGuardContext {
   messages: EngineMessage[];
+  /**
+   * Stable text from the real user request that started this turn. Unlike the
+   * trailing entry in `messages`, this never points at an internal continuation
+   * or a final-guard corrective retry.
+   */
+  requestText?: string;
   assistantContent: EngineContentPart[];
   text: string;
   toolCalls: AgentLoopToolCallSummary[];
@@ -2049,6 +2059,17 @@ function findCurrentTurnStartForContinuation(
   return 0;
 }
 
+/** Resolve the real user request behind any internal continuation messages. */
+export function resolveFinalResponseGuardRequestText(
+  messages: EngineMessage[],
+): string | undefined {
+  const message = messages[findCurrentTurnStartForContinuation(messages)];
+  if (!message || message.role !== "user") return undefined;
+  const text = textFromEngineMessage(message).trim();
+  if (text.startsWith(AGENT_INTERNAL_CONTINUE_PROMPT)) return undefined;
+  return text || undefined;
+}
+
 /**
  * First message index that is safe to start a trimmed window on. A window must
  * not begin with a tool-result-only user message — that would orphan it from
@@ -2913,6 +2934,7 @@ export async function runAgentLoop(opts: {
   tools: EngineTool[];
   availableTools?: EngineTool[];
   messages: EngineMessage[];
+  systemSections?: import("../shared/context-xray.js").ContextManifestSystemSection[];
   actions: Record<string, ActionEntry>;
   send: (event: AgentChatEvent) => void;
   signal: AbortSignal;
@@ -2932,6 +2954,12 @@ export async function runAgentLoop(opts: {
   executionMode?: AgentExecutionMode;
   maxIterations?: number;
   finalResponseGuard?: AgentLoopFinalResponseGuard;
+  /**
+   * Stable real-user request text preserved across internal continuation
+   * attempts. Callers normally omit this; continuation wrappers freeze it
+   * before appending their synthetic user messages.
+   */
+  finalResponseGuardRequestText?: string;
   threadId?: string;
   turnId?: string;
   /**
@@ -2967,6 +2995,9 @@ export async function runAgentLoop(opts: {
     send,
     signal,
   } = opts;
+  const finalResponseGuardRequestText =
+    opts.finalResponseGuardRequestText ??
+    resolveFinalResponseGuardRequestText(messages);
   const availableToolMap = new Map(
     (availableTools ?? tools).map((tool) => [tool.name, tool]),
   );
@@ -3106,6 +3137,7 @@ export async function runAgentLoop(opts: {
         turnId: opts.turnId,
         model,
         messages,
+        systemSections: opts.systemSections,
       });
 
       // Observational Memory (consumer): for long threads that have already been
@@ -3680,11 +3712,41 @@ export async function runAgentLoop(opts: {
         appendAgentLoopContinuation(messages, "max_tokens");
         continue;
       }
+
+      // Some providers (notably OpenAI Responses for gpt-5+) can complete a
+      // successful turn with reasoning content but no text or tool call. App
+      // final-answer guards validate an actual draft; letting them see this
+      // contentless turn can misclassify our synthetic recovery message and
+      // replace the shared retry with an unrelated app fallback.
+      const hasEmptyFinalResponse =
+        collectTextParts(assistantContentForHistory).trim().length === 0 &&
+        streamedAssistantText.trim().length === 0;
+      if (hasEmptyFinalResponse) {
+        if (emptyFinalResponseRetries < EMPTY_FINAL_RESPONSE_RETRY_LIMIT) {
+          emptyFinalResponseRetries += 1;
+          effectiveMaxOutputTokens =
+            resolveEmptyResponseRetryMaxOutputTokens(model);
+          effectiveReasoningEffort = stepDownReasoningEffort(
+            effectiveReasoningEffort,
+          );
+          send({ type: "clear" });
+          appendAgentLoopContinuation(messages, "max_tokens");
+          continue;
+        }
+        send({ type: "clear" });
+        send({
+          type: "text",
+          text: "The model returned an empty response. This usually means reasoning used the full output-token budget. Try again, or pick a different model from the model menu.",
+        });
+        break;
+      }
+
       let guard: Awaited<ReturnType<AgentLoopFinalResponseGuard>> | null = null;
       if (opts.finalResponseGuard) {
         try {
           guard = await opts.finalResponseGuard({
             messages,
+            requestText: finalResponseGuardRequestText,
             assistantContent: assistantContentForHistory,
             text: collectTextParts(assistantContentForHistory),
             toolCalls: [...toolCallHistory],
@@ -3697,7 +3759,6 @@ export async function runAgentLoop(opts: {
           throw err;
         }
       }
-      let guardEmittedFallback = false;
       if (guard) {
         const retryMessage =
           typeof guard === "string" ? guard : guard.retryMessage;
@@ -3730,45 +3791,12 @@ export async function runAgentLoop(opts: {
         }
         send({ type: "clear" });
         send({ type: "text", text: fallbackMessage ?? retryMessage });
-        guardEmittedFallback = true;
       } else {
         flushUnstreamedAssistantText();
       }
-      // Some providers (notably OpenAI Responses for gpt-5+) can stream a
-      // successful turn that contains only reasoning content and zero output
-      // text — typically when reasoning consumes the entire output-token
-      // budget. Without a final text part the SSE stream still ends with a
-      // clean `done`, which renders as a totally empty assistant bubble.
-      // Retry so a reasoning-budget miss can still finish; each retry raises
-      // the token ceiling and steps reasoning effort down a tier so it's not
-      // just re-issuing the identical doomed request. If retries also come
-      // back empty, surface a plain-language error.
-      const hasEmptyFinalResponse =
-        !guardEmittedFallback &&
-        collectTextParts(assistantContentForHistory).trim().length === 0 &&
-        streamedAssistantText.trim().length === 0;
-      if (hasEmptyFinalResponse) {
-        if (emptyFinalResponseRetries < EMPTY_FINAL_RESPONSE_RETRY_LIMIT) {
-          emptyFinalResponseRetries += 1;
-          effectiveMaxOutputTokens =
-            resolveEmptyResponseRetryMaxOutputTokens(model);
-          effectiveReasoningEffort = stepDownReasoningEffort(
-            effectiveReasoningEffort,
-          );
-          send({ type: "clear" });
-          appendAgentLoopContinuation(messages, "max_tokens");
-          continue;
-        }
-        send({ type: "clear" });
-        send({
-          type: "text",
-          text: "The model returned an empty response. This usually means reasoning used the full output-token budget. Try again, or pick a different model from the model menu.",
-        });
-      } else {
-        emptyFinalResponseRetries = 0;
-        effectiveMaxOutputTokens = opts.maxOutputTokens;
-        effectiveReasoningEffort = opts.reasoningEffort;
-      }
+      emptyFinalResponseRetries = 0;
+      effectiveMaxOutputTokens = opts.maxOutputTokens;
+      effectiveReasoningEffort = opts.reasoningEffort;
       break;
     }
 
@@ -4965,6 +4993,9 @@ export async function runAgentLoopWithMainChatInternalContinuations(
     maxContinuations?: number;
   },
 ): Promise<Awaited<ReturnType<typeof runAgentLoop>>> {
+  const finalResponseGuardRequestText =
+    opts.finalResponseGuardRequestText ??
+    resolveFinalResponseGuardRequestText(opts.messages);
   const usage: Awaited<ReturnType<typeof runAgentLoop>> = {
     inputTokens: 0,
     outputTokens: 0,
@@ -5010,7 +5041,11 @@ export async function runAgentLoopWithMainChatInternalContinuations(
     };
 
     try {
-      const nextUsage = await runAgentLoop({ ...opts, send });
+      const nextUsage = await runAgentLoop({
+        ...opts,
+        send,
+        finalResponseGuardRequestText,
+      });
       addUsage(nextUsage);
     } catch (err) {
       // Preserve exact prior behavior unless explicitly opted in: an aborted
@@ -6669,6 +6704,31 @@ export function createProductionAgentHandler(
       availableRequestTools,
       options.initialToolNames,
     );
+    // System sections are emitted by the prompt builder once per request. Tool
+    // schemas become known just after prompt setup, so append their measured
+    // contribution here and reuse the immutable result for every loop pass.
+    const contextXraySystemSections = [
+      ...readContextXraySystemSections(event),
+      ...(requestTools.length > 0
+        ? await buildSystemManifestSections([
+            {
+              label: "Action and MCP tool schemas",
+              provenance: "tools",
+              governance: "required",
+              content: requestTools
+                .map((tool) =>
+                  JSON.stringify({
+                    name: tool.name,
+                    description: tool.description,
+                    parameters: tool.inputSchema,
+                  }),
+                )
+                .join("\n"),
+              sourceRef: { scope: "tools" },
+            },
+          ])
+        : []),
+    ];
     setupMark("actions");
     // DIAGNOSTIC-ONLY: action/tool resolution + engine-tool filtering finished.
     workerStep("action_tool_setup");
@@ -7648,6 +7708,7 @@ export function createProductionAgentHandler(
           tools: requestTools,
           availableTools: availableRequestTools,
           messages,
+          systemSections: contextXraySystemSections,
           actions: requestActions,
           send,
           signal,
@@ -7664,6 +7725,7 @@ export function createProductionAgentHandler(
           executionMode: requestMode,
           maxIterations: loopSettings.maxIterations,
           finalResponseGuard: options.finalResponseGuard,
+          finalResponseGuardRequestText: messageToPersist,
           ...(options.toolLimits ? { toolLimits: options.toolLimits } : {}),
           ...(threadId
             ? { threadId: effectiveThreadId, turnId: effectiveTurnId }

package/corpus/core/src/agent/run-loop-with-resume.ts

@@ -28,6 +28,7 @@ import {
   isResumableEngineError,
   continuationReasonForResumableError,
   lastUnfinishedPreparingActionToolFromEvents,
+  resolveFinalResponseGuardRequestText,
   type AgentLoopContinuationReason,
 } from "./production-agent.js";
 import { resolveRunSoftTimeoutMs } from "./run-manager.js";
@@ -203,8 +204,12 @@ export async function runAgentLoopDirectWithSoftTimeout(
   softTimeoutMs?: number,
   timeoutOptions?: ResolveRunSoftTimeoutOptions,
 ): Promise<Awaited<ReturnType<typeof runAgentLoop>>> {
+  const finalResponseGuardRequestText =
+    opts.finalResponseGuardRequestText ??
+    resolveFinalResponseGuardRequestText(opts.messages);
+  const stableOpts = { ...opts, finalResponseGuardRequestText };
   const timeoutMs = resolveRunSoftTimeoutMs(softTimeoutMs, timeoutOptions);
-  if (timeoutMs <= 0) return runAgentLoop(opts);
+  if (timeoutMs <= 0) return runAgentLoop(stableOpts);
 
   const upstreamSignal = opts.signal;
   const usage: Awaited<ReturnType<typeof runAgentLoop>> = {
@@ -260,7 +265,7 @@ export async function runAgentLoopDirectWithSoftTimeout(
 
     try {
       const nextUsage = await runAgentLoop({
-        ...opts,
+        ...stableOpts,
         send,
         signal: controller.signal,
       });

package/corpus/core/src/client/AgentPanel.tsx

@@ -79,7 +79,7 @@ const MultiTabAssistantChatLazy = lazy(() =>
   })),
 );
 import { useQuery, useQueryClient } from "@tanstack/react-query";
-import { useLocation, useNavigate } from "react-router";
+import { Link, useLocation, useNavigate } from "react-router";
 
 import type { AgentChatSurfaceKind } from "./agent-chat-adapter.js";
 import {
@@ -595,6 +595,8 @@ export interface AgentPanelProps extends Omit<
   showPageNewChatButton?: boolean;
   /** Allow the sidebar settings view to render inside this panel. Default: true. */
   allowSettingsMode?: boolean;
+  /** Optional link shown in Resources and Settings modes for the full Agent page. */
+  agentPageHref?: string;
   /** Capability gate for source edits and CLI access. */
   codeAccess?: AgentPanelCodeAccess;
 }
@@ -745,6 +747,7 @@ function AgentPanelInner({
   showTabBar = true,
   showPageNewChatButton = false,
   allowSettingsMode = true,
+  agentPageHref,
   codeAccess,
   ...assistantChatProps
 }: AgentPanelProps) {
@@ -1807,11 +1810,11 @@ function AgentPanelInner({
             ".agent-tabs-scroll::-webkit-scrollbar{display:none;}" +
             `[data-agent-fullscreen='true'] .agent-thread-content,` +
             `[data-agent-fullscreen='true'] .agent-running-activity{` +
-            `max-width:${FULLSCREEN_CONTENT_MAX_PX}px;` +
+            `max-width:${FULLSCREEN_CHAT_COLUMN_MAX_PX}px;` +
             `margin-left:auto;margin-right:auto;width:100%;}` +
             `[data-agent-fullscreen='true'] .agent-composer-area,` +
             `[data-agent-fullscreen='true'] .agent-plan-mode-callout{` +
-            `max-width:${FULLSCREEN_COMPOSER_MAX_PX}px;` +
+            `max-width:${FULLSCREEN_CHAT_COLUMN_MAX_PX}px;` +
             `margin-left:auto;margin-right:auto;width:100%;}`,
         }}
       />
@@ -1933,7 +1936,18 @@ function AgentPanelInner({
 
       {/* Resources view */}
       {mode === "resources" && (
-        <div className="flex-1 min-h-0">
+        <div className="flex flex-1 flex-col min-h-0">
+          {agentPageHref && (
+            <div className="flex shrink-0 justify-end border-b border-border px-3 py-1.5">
+              <Link
+                to={agentPageHref}
+                className="inline-flex cursor-pointer items-center gap-1 text-[11px] font-medium text-muted-foreground hover:text-foreground"
+              >
+                {t("agentPanel.openFullView")}
+                <IconExternalLink className="h-3 w-3" />
+              </Link>
+            </div>
+          )}
           <Suspense
             fallback={
               <div className="flex h-full flex-col min-h-0">
@@ -1954,6 +1968,17 @@ function AgentPanelInner({
       {/* Settings / Setup view */}
       {mode === "settings" && (
         <div className="flex flex-col flex-1 min-h-0">
+          {agentPageHref && (
+            <div className="flex shrink-0 justify-end border-b border-border px-3 py-1.5">
+              <Link
+                to={agentPageHref}
+                className="inline-flex cursor-pointer items-center gap-1 text-[11px] font-medium text-muted-foreground hover:text-foreground"
+              >
+                {t("agentPanel.openFullView")}
+                <IconExternalLink className="h-3 w-3" />
+              </Link>
+            </div>
+          )}
           <Suspense
             fallback={
               <div className="p-3 space-y-2">
@@ -1987,10 +2012,8 @@ const SIDEBAR_MAX = 700;
 const SIDEBAR_ANIMATION_MS = 260;
 const SIDEBAR_OVERLAY_Z_INDEX = 70;
 const SIDEBAR_FULLSCREEN_Z_INDEX = 90;
-/** Max width of the centered chat column in fullscreen mode (Claude-style). */
-const FULLSCREEN_CONTENT_MAX_PX = 570;
-/** Max width of the centered composer in fullscreen mode. */
-const FULLSCREEN_COMPOSER_MAX_PX = 684;
+/** Shared max width of the centered fullscreen chat column and composer. */
+const FULLSCREEN_CHAT_COLUMN_MAX_PX = 684;
 
 function ResizeHandle({
   position,
@@ -2594,6 +2617,8 @@ export interface AgentSidebarProps {
   browserTabId?: string;
   /** Keep chat thread selection in URL state. */
   threadUrlSync?: MultiTabAssistantChatProps["threadUrlSync"];
+  /** Optional link shown in Resources and Settings modes for the full Agent page. */
+  agentPageHref?: string;
 }
 
 /**
@@ -2624,6 +2649,7 @@ export function AgentSidebar({
   showScopeBadge,
   browserTabId,
   threadUrlSync,
+  agentPageHref,
 }: AgentSidebarProps) {
   const initialWidth = defaultSidebarWidth ?? sidebarWidth ?? 380;
   const [open, setOpen] = useState(
@@ -3160,6 +3186,7 @@ export function AgentSidebar({
             showScopeBadge={showScopeBadge}
             browserTabId={browserTabId}
             threadUrlSync={threadUrlSync}
+            agentPageHref={agentPageHref}
             allowSettingsMode={false}
           />
         </div>