Harden Agent mode: plan persistence, transient-error retry, error cards

Persist the agent's set_plan checklist on the session so reopening a
chat mid-task shows where the plan stood; silently retry a completion
once when the failure is transient (network blip, rate limit, 5xx)
while still failing fast on auth/validation errors; and render failed
tool results as visually distinct error cards so a broken step is
scannable instead of buried in result bodies.

Author: voidstackloop

app/src/sessions-store.ts

@@ -16,6 +16,9 @@ export interface ChatSession {
     systemPrompt?: string | null;
     agentMode?: boolean;
     agentWorkspace?: string | null;
+    // The agent's last set_plan checklist for this conversation, persisted so
+    // reopening a chat mid-task shows where the plan stood.
+    planSteps?: { text: string; done: boolean }[];
     tags?: string[];
     createdAt: string;
     updatedAt: string;
@@ -63,7 +66,7 @@ export function updateSession(
     partial: Partial<
         Pick<
             ChatSession,
-            "title" | "model" | "messages" | "params" | "projectId" | "systemPrompt" | "agentMode" | "agentWorkspace" | "tags"
+            "title" | "model" | "messages" | "params" | "projectId" | "systemPrompt" | "agentMode" | "agentWorkspace" | "planSteps" | "tags"
         >
     >
 ): ChatSession | null {

frontend/src/lib/transient-errors.test.ts

@@ -0,0 +1,22 @@
+import { describe, it, expect } from "vitest";
+import { isTransientError } from "./transient-errors";
+
+describe("isTransientError", () => {
+    it("treats network-level failures as transient", () => {
+        expect(isTransientError("Can't reach Ollama at http://127.0.0.1:11434 — is it running? (fetch failed)")).toBe(true);
+        expect(isTransientError("request to https://api.openai.com failed, reason: ECONNRESET")).toBe(true);
+        expect(isTransientError("The request timed out after 30s")).toBe(true);
+    });
+
+    it("treats rate limits and server errors as transient", () => {
+        expect(isTransientError("OpenAI request failed (HTTP 529): overloaded_error")).toBe(true);
+        expect(isTransientError("Rate limit exceeded, HTTP 429")).toBe(true);
+        expect(isTransientError("Gemini request failed (HTTP 503)")).toBe(true);
+    });
+
+    it("does not retry auth or validation errors", () => {
+        expect(isTransientError("Incorrect API key provided (HTTP 401)")).toBe(false);
+        expect(isTransientError("model 'gpt-99' not found (HTTP 404)")).toBe(false);
+        expect(isTransientError("Agent mode isn't supported yet for Gemini")).toBe(false);
+    });
+});

frontend/src/lib/transient-errors.ts

@@ -0,0 +1,9 @@
+// Whether a provider error is worth one silent automatic retry — network
+// blips, timeouts, rate limits, and 5xx-class server hiccups usually succeed
+// on a second attempt, whereas auth/validation errors ("invalid API key",
+// "model not found") never will and should surface immediately.
+export function isTransientError(message: string): boolean {
+    return /timed? ?out|timeout|ECONNRESET|ECONNREFUSED|ENOTFOUND|EAI_AGAIN|fetch failed|network error|HTTP 5\d\d|\(5\d\d\)|\b429\b|rate limit|overloaded|temporarily unavailable|try again/i.test(
+        message
+    );
+}

frontend/src/lib/translations.ts

@@ -24,6 +24,8 @@ export interface Dictionary {
     allow: string;
     deny: string;
     toolResult: string;
+    toolFailed: string;
+    transientErrorRetrying: string;
     alwaysAllowThisSession: string;
     agentStep: string;
     agentStepTooltip: string;
@@ -320,6 +322,8 @@ export const en: Dictionary = {
     allow: "Allow",
     deny: "Deny",
     toolResult: "result",
+    toolFailed: "failed",
+    transientErrorRetrying: "Temporary provider error — retrying once...",
     alwaysAllowThisSession: "Always allow this session",
     agentStep: "Agent step",
     agentStepTooltip: "How many automatic tool-result → model-continuation round trips have happened for this turn.",
@@ -626,6 +630,8 @@ export const tr: Dictionary = {
     allow: "İzin ver",
     deny: "Reddet",
     toolResult: "sonucu",
+    toolFailed: "başarısız oldu",
+    transientErrorRetrying: "Geçici sağlayıcı hatası — bir kez yeniden deneniyor...",
     alwaysAllowThisSession: "Bu oturumda her zaman izin ver",
     agentStep: "Ajan adımı",
     agentStepTooltip: "Bu tur için kaç otomatik araç sonucu → model devamı gidiş-dönüşü gerçekleşti.",

frontend/src/pages/Chat.tsx

@@ -69,6 +69,7 @@ import { ScreenshotPickerDialog } from "@/components/screenshot-picker-dialog";
 import { speakText, stopSpeaking } from "@/lib/tts";
 import { computeLineDiff } from "@/lib/diff";
 import { useToast } from "@/components/toast";
+import { isTransientError } from "@/lib/transient-errors";
 import type {
     ChatMessage,
     OllamaModel,
@@ -223,11 +224,20 @@ const MessageBubble = memo(function MessageBubble({
     const { t } = useI18n();
 
     if (m.role === "tool") {
+        // Tool failures get a visually distinct card — an agent run that hit
+        // an error mid-way should be scannable at a glance, not require
+        // reading every result body to find where things went wrong.
+        const isError = m.content.startsWith("Error:") || m.content === "The user denied this tool call.";
         return (
             <div className="flex flex-col items-start">
-                <div className="max-w-[85%] rounded-lg border border-border bg-muted/50 px-3 py-2 font-mono text-xs text-muted-foreground">
-                    <div className="mb-1 font-sans font-medium text-foreground">
-                        🔧 {m.toolName} {t.toolResult}
+                <div
+                    className={cn(
+                        "max-w-[85%] rounded-lg border px-3 py-2 font-mono text-xs text-muted-foreground",
+                        isError ? "border-destructive/40 bg-destructive/5" : "border-border bg-muted/50"
+                    )}
+                >
+                    <div className={cn("mb-1 font-sans font-medium", isError ? "text-destructive" : "text-foreground")}>
+                        {isError ? "⚠️" : "🔧"} {m.toolName} {isError ? t.toolFailed : t.toolResult}
                     </div>
                     <pre className="max-h-48 overflow-auto whitespace-pre-wrap">{m.content}</pre>
                 </div>
@@ -492,7 +502,7 @@ export default function Chat() {
             setAgentWorkspace(session.agentWorkspace ?? null);
             setPendingToolCalls([]);
             setAgentStepCount(0);
-            setPlanSteps([]);
+            setPlanSteps(session.planSteps ?? []);
             setAutoApprovedTools(new Set());
             setWriteDiffPreviews({});
             setUndoMessage(null);
@@ -832,7 +842,7 @@ export default function Chat() {
     async function runCompletion(
         history: ChatMessage[],
         baseMessages: ChatMessage[],
-        opts: { isFirstMessage: boolean; titleSource: string }
+        opts: { isFirstMessage: boolean; titleSource: string; attempt?: number }
     ) {
         const parsed = parseModelRef(model);
         if (!parsed || !sessionId) return;
@@ -883,6 +893,18 @@ export default function Chat() {
         const result = await promise;
         setActiveRequestId(null);
 
+        // One silent retry for errors that usually clear on their own
+        // (network blips, rate limits, 5xx) — but never for a user-initiated
+        // stop, and never more than once, so a genuinely down provider still
+        // fails fast instead of looping.
+        if (result.error && !result.aborted && (opts.attempt ?? 0) === 0 && isTransientError(result.error)) {
+            toast.info(t.transientErrorRetrying);
+            setIsStreaming(false);
+            window.api.app.setBusy(false);
+            await new Promise((resolve) => setTimeout(resolve, 1500));
+            return runCompletion(history, baseMessages, { ...opts, attempt: 1 });
+        }
+
         if (result.error) {
             setMessages((m) => {
                 const next = [...m];
@@ -999,6 +1021,7 @@ export default function Chat() {
             done: Boolean((s as { done?: unknown })?.done),
         }));
         setPlanSteps(steps);
+        if (sessionId) window.api.sessions.update(sessionId, { planSteps: steps });
         resolveToolCall(call, "Plan updated.");
     }
 
@@ -1098,6 +1121,7 @@ export default function Chat() {
         setImageAttachments([]);
         setAgentStepCount(0);
         setPlanSteps([]);
+        window.api.sessions.update(sessionId, { planSteps: [] });
         await runCompletion(history, baseMessages, { isFirstMessage, titleSource });
     }
 

frontend/src/types/electron.d.ts

@@ -247,6 +247,7 @@ export interface ChatSession {
   systemPrompt?: string | null;
   agentMode?: boolean;
   agentWorkspace?: string | null;
+  planSteps?: { text: string; done: boolean }[];
   tags?: string[];
   createdAt: string;
   updatedAt: string;
@@ -391,6 +392,7 @@ export interface ElectronApi {
           | "systemPrompt"
           | "agentMode"
           | "agentWorkspace"
+          | "planSteps"
           | "tags"
         >
       >