Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/workflows/verification-workflow.ts, line 63:

<comment>Missing error boundary: if `ragSearch()` or `extractIntent()` throws, the generator crashes without yielding an SSE error event, leaving the client mid-stream with no graceful error signal.</comment>

<file context>
@@ -43,10 +45,84 @@ export type WorkflowReturn = {
+  };
+
+  const ragStarted = Date.now();
+  const rag = await ragSearch({ rag: deps.rag }, { request });
+
+  yield {
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/steps/intent-extraction.ts, line 90:

<comment>Objective-code validation should trim candidate strings before regex matching; otherwise whitespace-padded valid codes are downgraded to the default fallback code.</comment>

<file context>
@@ -27,7 +28,73 @@ export interface IntentExtractionOutput {
+
+function firstValidObjectiveCode(candidates: Array<string | null | undefined>): string {
+  for (const candidate of candidates) {
+    if (candidate !== undefined && candidate !== null && isObjectiveCode(candidate)) {
+      return candidate;
+    }
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/server/sse/progress-stream.ts, line 11:

<comment>Unhandled exceptions from the workflow generator terminate the SSE stream without sending a protocol-level `error` event.</comment>

<file context>
@@ -2,11 +2,88 @@
+  events: AsyncGenerator<ProgressEvent, unknown, void>,
 ): Promise<void> {
-  throw new Error("pipeProgressToSse: not implemented yet");
+  for await (const event of events) {
+    await stream.writeSSE(toWireEvent(event));
+  }
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/tools/math-engine-client.ts, line 94:

<comment>When the request times out, `controller.abort()` fires without a reason, causing `fetch` to reject with a generic `AbortError` (`DOMException: The operation was aborted`). The caller gets no indication of which endpoint or timeout duration caused the failure, making runtime debugging harder.</comment>

<file context>
@@ -68,7 +68,45 @@ export interface MathEngineClientOptions {
+  timeoutMs: number,
+): Promise<T> {
+  const controller = new AbortController();
+  const timeout = setTimeout(() => controller.abort(), timeoutMs);
+
+  try {
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/steps/intent-extraction.ts, line 98:

<comment>The objective code regex is declared inline in `isObjectiveCode`, duplicating the same pattern already defined in `intent.schema.ts` and `strategy.schema.ts`. Export a shared constant (e.g., from `intent.schema.ts`) and reuse it here to prevent pattern drift.</comment>

<file context>
@@ -27,7 +28,73 @@ export interface IntentExtractionOutput {
+  return "9수00-00";
+}
+
+function isObjectiveCode(value: string): boolean {
+  return /^(9수|10공수)\d{2}-\d{2}$/.test(value);
 }
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/server/sse/progress-stream.ts, line 27:

<comment>`"info"` step status is incorrectly mapped to `"completed"`. The `StepStatusSchema` includes `"info"` as a valid status (alongside `"start"` and `"done"`), but the ternary `event.status === "start" ? "started" : "completed"` lumps `"info"` into `"completed"`. The SSE consumer will interpret an `"info"` event as step completion, which misrepresents the actual workflow state.</comment>

<file context>
@@ -2,11 +2,88 @@
+        data: JSON.stringify({
+          index: stepIndex(event.step),
+          name: event.step,
+          status: event.status === "start" ? "started" : "completed",
+          summary: stepSummary(event),
+          raw: event,
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/web/app/app/new/verify/view.tsx, line 172:

<comment>useEffect called after early return — violates React Rules of Hooks. If `valid` toggles between renders, React will detect a hook count mismatch, causing an error or state corruption.</comment>

<file context>
@@ -169,6 +169,15 @@ export function VerifyView({ grade, topic, mode, dims }: Props) {
   const isDone = stream.status === "done";
   const showPreview = stream.previewLatex !== null;
 
+  useEffect(() => {
+    if (isDone && stream.candidates.length > 0) {
+      window.localStorage.setItem(
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/web/components/math/latex-renderer.tsx, line 64:

<comment>The `**` exponent regex greedily captures only the first safe token into the exponent group. Inputs like `x ** (y+1)` produce garbled LaTeX (`x^{(y}+1)` instead of `x^{y+1}` or `x^{(y+1)}`).</comment>

<file context>
@@ -56,6 +58,14 @@ function renderLatex(latex: string, block: boolean): string {
+function normalizeMathNotation(input: string): string {
+  return input
+    .replace(/\bsqrt\(([^()]+)\)/g, "\\sqrt{$1}")
+    .replace(/([A-Za-z0-9)\]}]+)\s*\*\*\s*([A-Za-z0-9({[]+)/g, "$1^{$2}")
+    .replace(/([0-9A-Za-z)\]}]+)\s*\*\s*([A-Za-z({[]+)/g, "$1$2")
+    .replace(/\^\{([^{}]+)\}\s*\)/g, "^{$1})");
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/steps/sympy-verification.ts, line 85:

<comment>`extractEquation` regex captures natural-language text as part of the equation because `A-Za-z` in the character class allows English words. E.g., "What is the solution to 2x+3=7?" extracts "What is the solution to 2x+3=7" instead of just "2x+3=7", causing a misleading `math_engine_error` rather than `equation_not_found`.</comment>

<file context>
@@ -16,8 +16,87 @@ export interface SympyVerificationOutput {
+
+export function extractEquation(text: string): string | null {
+  const normalized = normalizeMathText(text);
+  const match = normalized.match(/[A-Za-z0-9+\-*/^().\s]+=[A-Za-z0-9+\-*/^().\s]+/);
+  return match?.[0]?.trim() ?? null;
+}
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/agents/generator-agent.ts, line 58:

<comment>`inferred_intent` is sourced from model output (except objective_code), which can make objective mapping fail due to intent-dimension drift. Use the requested intent as the canonical value here.</comment>

<file context>
@@ -25,9 +30,45 @@ export interface GeneratorAgent {
+        source_refs: result.object.source_refs.length
+          ? result.object.source_refs
+          : input.refs.map((ref) => ref.item_id),
+        inferred_intent: {
+          ...result.object.inferred_intent,
+          objective_code: input.intent.objective_code,
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/tools/schema-loader.ts, line 77:

<comment>Validate or sanitize `code` before resolving the strategy file path to prevent directory traversal outside `strategiesDir`.</comment>

<file context>
@@ -22,7 +31,73 @@ export interface FsStrategyLoaderOptions {
+  strategiesDir: string,
+  code: string,
+): Promise<Strategy | null> {
+  const path = resolve(strategiesDir, `${code}.yaml`);
+  let file: string;
+  try {
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/tools/prompt-loader.ts, line 88:

<comment>Path traversal risk in `readPrompt`: the `id` parameter is interpolated directly into `resolve(promptsDir, \`${id}.md\`)` without verifying the resolved path stays inside `promptsDir`. An `id` like `../../etc/somefile` could read arbitrary `.md` files anywhere on the filesystem. This is a defense-in-depth concern even though current callers pass controlled IDs.</comment>

<file context>
@@ -37,7 +58,51 @@ export interface FsPromptLoaderOptions {
+}
+
+async function readPrompt(promptsDir: string, id: string): Promise<LoadedPrompt> {
+  const path = resolve(promptsDir, `${id}.md`);
+  const file = await readFile(path, "utf8");
+  const parsed = matter(file);
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/web/app/app/new/result/view.tsx, line 89:

<comment>Using an unscoped `openmath:last-results` cache can show stale problems for unrelated grade/topic/mode selections.</comment>

<file context>
@@ -83,10 +83,27 @@ export function ResultView({
+  const [liveProblems, setLiveProblems] = useState<ResultProblem[] | null>(null);
+
+  useEffect(() => {
+    const raw = window.localStorage.getItem("openmath:last-results");
+    if (raw === null) return;
+    try {
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/src/steps/sympy-verification.ts, line 89:

<comment>`normalizeMathText` does not replace `\cdot` with `*`. Since `\` is not in the `extractEquation` character class, equations containing `\cdot` produce garbled extractions like `"cdot b = c"` instead of `"a * b = c"`, leading to spurious math engine errors.</comment>

<file context>
@@ -16,8 +16,87 @@ export interface SympyVerificationOutput {
+  return match?.[0]?.trim() ?? null;
+}
+
+export function normalizeMathText(text: string): string {
+  return text
+    .replace(/\\\((.*?)\\\)/g, "$1")
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/.env.example, line 6:

<comment>The provider example now uses `openai-compatible`, but the documented provider options in the same file still mention `cliproxy` (an invalid value). This can cause `.env` setup failures.</comment>

<file context>
@@ -3,13 +3,12 @@ MATH_ENGINE_URL=http://localhost:8000
 
 # LLM Provider: "openai" or "cliproxy"
-LLM_PROVIDER=openai
+LLM_PROVIDER=openai-compatible
+LLM_BASE_URL=http://localhost:8080/v1
+LLM_API_KEY=dummy-key
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At packages/agent/data/README.md, line 34:

<comment>Avoid committing a developer-specific absolute corpus path in docs; use a portable placeholder path to prevent misleading setup instructions.</comment>

<file context>
@@ -5,23 +5,43 @@
+런타임 corpus는 `openmath-rag-record-v1` JSONL이다. 기본 로컬 경로:
+
+```text
+/Users/tw81512/dev/AI_HUB_data/rag_problem_generation_dataset/openmath_rag_records.jsonl
+```
+
</file context>