pulsemcp-cms-admin-mcp-server 0.7.0 → 0.7.2

package/build/shared/src/exam-result-store.js

@@ -1,18 +1,50 @@
 import { randomUUID } from 'crypto';
+import { readFileSync, writeFileSync, unlinkSync, readdirSync, mkdirSync, existsSync } from 'fs';
+import { join } from 'path';
+import { tmpdir } from 'os';
 /**
- * Maximum number of results to keep in memory. Oldest results are evicted
- * when this limit is reached (FIFO). Each result can be 60KB+ for servers
- * with many tools, so 100 entries ≈ 6MB worst case.
+ * Extract exam_id from a proctor exam stream line, checking both the
+ * data payload and top-level fields. The API may place exam_id in
+ * either location depending on the exam type.
+ */
+export function extractExamId(line) {
+    const data = line.data;
+    return (data?.exam_id ||
+        line.exam_id ||
+        data?.exam_type ||
+        line.exam_type ||
+        'unknown');
+}
+/**
+ * Extract status from a proctor exam stream line, checking both the
+ * data payload and top-level fields.
+ */
+export function extractStatus(line) {
+    const data = line.data;
+    return data?.status || line.status || 'unknown';
+}
+/**
+ * Maximum number of results to keep on disk. Oldest results are evicted
+ * when this limit is reached (FIFO by insertion order).
  */
 const MAX_RESULTS = 100;
+const STORE_DIR = join(tmpdir(), 'pulsemcp-exam-results');
+const FILE_SUFFIX = '.json';
 /**
- * In-memory store for proctor exam results.
+ * File-based store for proctor exam results.
  *
- * When `run_exam_for_mirror` completes, the full result is stored here
- * and a UUID `result_id` is returned. This avoids dumping large payloads
- * (~60KB+ for servers with many tools) into the LLM context.
+ * When `run_exam_for_mirror` completes, the full result is written to a
+ * JSON file in /tmp/ and a UUID `result_id` is returned. This avoids
+ * dumping large payloads (~60KB+ for servers with many tools) into the
+ * LLM context, and survives across tool calls without relying on
+ * in-memory state.
  *
- * Eviction: When the store exceeds MAX_RESULTS entries, the oldest result
+ * Files are named with a zero-padded sequence number prefix so that
+ * lexicographic sorting preserves insertion order for FIFO eviction.
+ * The sequence counter is initialized from existing files on disk so
+ * that new entries sort after old ones even across process restarts.
+ *
+ * Eviction: When the store exceeds MAX_RESULTS files, the oldest result
  * is evicted (FIFO). Results are also deleted after successful save via
  * `save_results_for_mirror`.
  *
@@ -21,38 +53,118 @@ const MAX_RESULTS = 100;
  * - Pass `result_id` to `save_results_for_mirror` instead of the full payload
  */
 class ExamResultStore {
-    results = new Map();
+    seq;
+    constructor() {
+        this.seq = this.initSeqFromDisk();
+    }
+    /**
+     * Scan existing files to find the highest sequence number and start
+     * one past it. This ensures new files always sort after existing ones,
+     * even across process restarts.
+     */
+    initSeqFromDisk() {
+        this.ensureDir();
+        const files = readdirSync(STORE_DIR)
+            .filter((f) => f.endsWith(FILE_SUFFIX) && f.length > FILE_SUFFIX.length)
+            .sort();
+        if (files.length === 0)
+            return 0;
+        const lastFile = files[files.length - 1];
+        const seqStr = lastFile.slice(0, 10);
+        const parsed = parseInt(seqStr, 10);
+        return isNaN(parsed) ? 0 : parsed + 1;
+    }
+    ensureDir() {
+        if (!existsSync(STORE_DIR)) {
+            mkdirSync(STORE_DIR, { recursive: true, mode: 0o700 });
+        }
+    }
+    /** Filename format: {seq}-{uuid}.json — seq is zero-padded for lexicographic ordering */
+    fileName(seq, resultId) {
+        return `${String(seq).padStart(10, '0')}-${resultId}${FILE_SUFFIX}`;
+    }
+    extractResultId(fileName) {
+        // Format: 0000000001-<uuid>.json
+        return fileName.slice(11, -FILE_SUFFIX.length);
+    }
+    listResultFiles() {
+        this.ensureDir();
+        return readdirSync(STORE_DIR)
+            .filter((f) => f.endsWith(FILE_SUFFIX) && f.length > FILE_SUFFIX.length)
+            .sort(); // Lexicographic sort gives insertion order via zero-padded seq
+    }
+    findFileForResult(resultId) {
+        const files = this.listResultFiles();
+        return files.find((f) => this.extractResultId(f) === resultId);
+    }
     store(mirrorIds, runtimeId, examType, lines) {
+        this.ensureDir();
         const resultId = randomUUID();
-        // Evict oldest entries if at capacity (Map preserves insertion order)
-        while (this.results.size >= MAX_RESULTS) {
-            const oldestKey = this.results.keys().next().value;
-            if (oldestKey !== undefined) {
-                this.results.delete(oldestKey);
+        const seqNum = this.seq++;
+        // Evict oldest entries if at capacity (files are sorted by seq prefix)
+        const files = this.listResultFiles();
+        const toEvict = files.length - MAX_RESULTS + 1;
+        for (let i = 0; i < toEvict; i++) {
+            try {
+                unlinkSync(join(STORE_DIR, files[i]));
+            }
+            catch {
+                // ignore
             }
         }
-        this.results.set(resultId, {
+        const stored = {
             result_id: resultId,
             mirror_ids: mirrorIds,
             runtime_id: runtimeId,
             exam_type: examType,
             lines,
             stored_at: new Date().toISOString(),
+        };
+        writeFileSync(join(STORE_DIR, this.fileName(seqNum, resultId)), JSON.stringify(stored), {
+            encoding: 'utf-8',
+            mode: 0o600,
         });
         return resultId;
     }
     get(resultId) {
-        return this.results.get(resultId);
+        const file = this.findFileForResult(resultId);
+        if (!file)
+            return undefined;
+        try {
+            const content = readFileSync(join(STORE_DIR, file), 'utf-8');
+            return JSON.parse(content);
+        }
+        catch {
+            return undefined;
+        }
     }
     delete(resultId) {
-        return this.results.delete(resultId);
+        const file = this.findFileForResult(resultId);
+        if (!file)
+            return false;
+        try {
+            unlinkSync(join(STORE_DIR, file));
+            return true;
+        }
+        catch {
+            return false;
+        }
     }
     get size() {
-        return this.results.size;
+        return this.listResultFiles().length;
     }
-    /** For testing only */
+    /** For testing only — removes all result files and resets the sequence counter */
     clear() {
-        this.results.clear();
+        this.ensureDir();
+        for (const file of this.listResultFiles()) {
+            try {
+                unlinkSync(join(STORE_DIR, file));
+            }
+            catch {
+                // ignore
+            }
+        }
+        this.seq = 0;
     }
 }
 /** Singleton instance shared across all tool factories */

package/build/shared/src/tools/get-exam-result.js

@@ -49,7 +49,7 @@ Typical usage:
                     content: [
                         {
                             type: 'text',
-                            text: `No stored result found for result_id "${validatedArgs.result_id}". Results are stored in-memory and may have been lost if the server restarted.`,
+                            text: `No stored result found for result_id "${validatedArgs.result_id}". The result file may have been cleaned up or the /tmp directory cleared.`,
                         },
                     ],
                     isError: true,

package/build/shared/src/tools/run-exam-for-mirror.js

@@ -1,5 +1,5 @@
 import { z } from 'zod';
-import { examResultStore } from '../exam-result-store.js';
+import { examResultStore, extractExamId, extractStatus } from '../exam-result-store.js';
 const PARAM_DESCRIPTIONS = {
     mirror_ids: 'Array of unofficial mirror IDs to run exams against. Mirrors without saved mcp_json configs will be skipped.',
     runtime_id: 'The Fly Machines runtime ID to use for running the exam containers (e.g., "fly-machines-v1")',
@@ -66,7 +66,7 @@ Available exam types:
 
 Mirrors without saved mcp_json configurations are automatically skipped.
 
-Results are stored server-side and a \`result_id\` UUID is returned. The response includes a truncated summary (status, tool names/counts, errors) that fits within MCP size limits. Use \`get_exam_result\` to drill into full details, or pass the \`result_id\` directly to \`save_results_for_mirror\`.
+Results are stored server-side in a local file and a \`result_id\` UUID is returned. The response includes a truncated summary (status, tool names/counts, errors) that fits within MCP size limits. Use \`get_exam_result\` to drill into full details, or pass the \`result_id\` directly to \`save_results_for_mirror\`.
 
 Use cases:
 - Test if an unofficial mirror's MCP server is working correctly before linking it
@@ -119,15 +119,18 @@ Use cases:
                         case 'log':
                             content += `[LOG] ${line.message || JSON.stringify(line)}\n`;
                             break;
-                        case 'exam_result':
-                            content += `\n**Exam Result** (Mirror: ${line.mirror_id || 'unknown'})\n`;
-                            content += `  Exam: ${line.exam_id || line.exam_type || 'unknown'}\n`;
-                            content += `  Status: ${line.status || 'unknown'}\n`;
-                            if (line.data) {
-                                const truncatedData = truncateExamResultData(line.data);
+                        case 'exam_result': {
+                            const data = line.data;
+                            const mirrorId = line.mirror_id ?? data?.mirror_id ?? 'unknown';
+                            content += `\n**Exam Result** (Mirror: ${mirrorId})\n`;
+                            content += `  Exam: ${extractExamId(line)}\n`;
+                            content += `  Status: ${extractStatus(line)}\n`;
+                            if (data) {
+                                const truncatedData = truncateExamResultData(data);
                                 content += `  Data: ${JSON.stringify(truncatedData, null, 2)}\n`;
                             }
                             break;
+                        }
                         case 'summary':
                             content += `\n**Summary**\n`;
                             content += `  Total: ${line.total || 0}\n`;

package/build/shared/src/tools/save-results-for-mirror.js

@@ -1,9 +1,9 @@
 import { z } from 'zod';
-import { examResultStore } from '../exam-result-store.js';
+import { examResultStore, extractExamId, extractStatus } from '../exam-result-store.js';
 const PARAM_DESCRIPTIONS = {
     mirror_id: 'The ID of the unofficial mirror to save results for',
     runtime_id: 'The runtime ID that was used to run the exams',
-    result_id: 'The UUID returned by run_exam_for_mirror. When provided, the server retrieves the full result from the in-memory store — no need to pass the results array. This is the preferred approach.',
+    result_id: 'The UUID returned by run_exam_for_mirror. When provided, the server retrieves the full result from the local file store — no need to pass the results array. This is the preferred approach.',
     results: 'Array of exam results to save. Each result must include exam_id, status, and optional data. Only needed if result_id is not provided.',
     exam_id: 'The exam identifier (e.g., "auth-check", "init-tools-list")',
     status: 'The result status (e.g., "pass", "fail", "error", "skip")',
@@ -32,7 +32,7 @@ export function saveResultsForMirror(_server, clientFactory) {
         name: 'save_results_for_mirror',
         description: `Save proctor exam results for an unofficial mirror.
 
-**Preferred**: Pass the \`result_id\` returned by \`run_exam_for_mirror\`. The full result is retrieved from the in-memory store server-side — no need to pass the large results payload through the LLM context.
+**Preferred**: Pass the \`result_id\` returned by \`run_exam_for_mirror\`. The full result is retrieved from the local file store server-side — no need to pass the large results payload through the LLM context.
 
 **Fallback**: Pass results directly (as before) if result_id is not available.
 
@@ -88,20 +88,39 @@ Typical workflow:
                             content: [
                                 {
                                     type: 'text',
-                                    text: `No stored result found for result_id "${validatedArgs.result_id}". Results are stored in-memory and may have been lost if the server restarted. Pass the results array directly instead.`,
+                                    text: `No stored result found for result_id "${validatedArgs.result_id}". The result file may have been cleaned up or the /tmp directory cleared. Pass the results array directly instead.`,
                                 },
                             ],
                             isError: true,
                         };
                     }
-                    // Extract exam_result lines from stored data
+                    // Extract exam_result lines from stored data.
+                    // The exam_id may live at the top level of the stream line OR inside
+                    // line.data (the actual result payload). Prefer the data payload to
+                    // avoid reading from potentially incomplete display metadata.
+                    //
+                    // The real proctor API returns line.data as a metadata wrapper:
+                    //   { mirror_id, exam_id, status, result: { status, output: {...} } }
+                    // The actual output lives inside line.data.result. When we pass the
+                    // entire line.data as `data`, the output ends up nested too deeply
+                    // (result.data.result.output) and the backend saves empty output.
+                    // Use line.data.result when present so that `output` is at the
+                    // expected depth (result.data.output).
                     results = stored.lines
                         .filter((line) => line.type === 'exam_result')
-                        .map((line) => ({
-                        exam_id: (line.exam_id || line.exam_type || 'unknown'),
-                        status: (line.status || 'unknown'),
-                        ...(line.data ? { data: line.data } : {}),
-                    }));
+                        .map((line) => {
+                        const data = line.data;
+                        // Prefer the nested result object (contains output, input, etc.)
+                        // over the full data wrapper (contains metadata like mirror_id)
+                        const resultData = data?.result && typeof data.result === 'object' && !Array.isArray(data.result)
+                            ? data.result
+                            : data;
+                        return {
+                            exam_id: extractExamId(line),
+                            status: extractStatus(line),
+                            ...(resultData ? { data: resultData } : {}),
+                        };
+                    });
                     if (!runtimeId) {
                         runtimeId = stored.runtime_id;
                     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pulsemcp-cms-admin-mcp-server",
-  "version": "0.7.0",
+  "version": "0.7.2",
   "description": "Local implementation of PulseMCP CMS Admin MCP server",
   "mcpName": "com.pulsemcp.servers/pulsemcp-cms-admin",
   "main": "build/index.js",

package/shared/exam-result-store.d.ts

@@ -8,13 +8,31 @@ export interface StoredExamResult {
     stored_at: string;
 }
 /**
- * In-memory store for proctor exam results.
+ * Extract exam_id from a proctor exam stream line, checking both the
+ * data payload and top-level fields. The API may place exam_id in
+ * either location depending on the exam type.
+ */
+export declare function extractExamId(line: ProctorExamStreamLine): string;
+/**
+ * Extract status from a proctor exam stream line, checking both the
+ * data payload and top-level fields.
+ */
+export declare function extractStatus(line: ProctorExamStreamLine): string;
+/**
+ * File-based store for proctor exam results.
+ *
+ * When `run_exam_for_mirror` completes, the full result is written to a
+ * JSON file in /tmp/ and a UUID `result_id` is returned. This avoids
+ * dumping large payloads (~60KB+ for servers with many tools) into the
+ * LLM context, and survives across tool calls without relying on
+ * in-memory state.
  *
- * When `run_exam_for_mirror` completes, the full result is stored here
- * and a UUID `result_id` is returned. This avoids dumping large payloads
- * (~60KB+ for servers with many tools) into the LLM context.
+ * Files are named with a zero-padded sequence number prefix so that
+ * lexicographic sorting preserves insertion order for FIFO eviction.
+ * The sequence counter is initialized from existing files on disk so
+ * that new entries sort after old ones even across process restarts.
  *
- * Eviction: When the store exceeds MAX_RESULTS entries, the oldest result
+ * Eviction: When the store exceeds MAX_RESULTS files, the oldest result
  * is evicted (FIFO). Results are also deleted after successful save via
  * `save_results_for_mirror`.
  *
@@ -23,12 +41,25 @@ export interface StoredExamResult {
  * - Pass `result_id` to `save_results_for_mirror` instead of the full payload
  */
 declare class ExamResultStore {
-    private results;
+    private seq;
+    constructor();
+    /**
+     * Scan existing files to find the highest sequence number and start
+     * one past it. This ensures new files always sort after existing ones,
+     * even across process restarts.
+     */
+    private initSeqFromDisk;
+    private ensureDir;
+    /** Filename format: {seq}-{uuid}.json — seq is zero-padded for lexicographic ordering */
+    private fileName;
+    private extractResultId;
+    private listResultFiles;
+    private findFileForResult;
     store(mirrorIds: number[], runtimeId: string, examType: string, lines: ProctorExamStreamLine[]): string;
     get(resultId: string): StoredExamResult | undefined;
     delete(resultId: string): boolean;
     get size(): number;
-    /** For testing only */
+    /** For testing only — removes all result files and resets the sequence counter */
     clear(): void;
 }
 /** Singleton instance shared across all tool factories */

package/shared/exam-result-store.js

@@ -1,18 +1,50 @@
 import { randomUUID } from 'crypto';
+import { readFileSync, writeFileSync, unlinkSync, readdirSync, mkdirSync, existsSync } from 'fs';
+import { join } from 'path';
+import { tmpdir } from 'os';
 /**
- * Maximum number of results to keep in memory. Oldest results are evicted
- * when this limit is reached (FIFO). Each result can be 60KB+ for servers
- * with many tools, so 100 entries ≈ 6MB worst case.
+ * Extract exam_id from a proctor exam stream line, checking both the
+ * data payload and top-level fields. The API may place exam_id in
+ * either location depending on the exam type.
+ */
+export function extractExamId(line) {
+    const data = line.data;
+    return (data?.exam_id ||
+        line.exam_id ||
+        data?.exam_type ||
+        line.exam_type ||
+        'unknown');
+}
+/**
+ * Extract status from a proctor exam stream line, checking both the
+ * data payload and top-level fields.
+ */
+export function extractStatus(line) {
+    const data = line.data;
+    return data?.status || line.status || 'unknown';
+}
+/**
+ * Maximum number of results to keep on disk. Oldest results are evicted
+ * when this limit is reached (FIFO by insertion order).
  */
 const MAX_RESULTS = 100;
+const STORE_DIR = join(tmpdir(), 'pulsemcp-exam-results');
+const FILE_SUFFIX = '.json';
 /**
- * In-memory store for proctor exam results.
+ * File-based store for proctor exam results.
  *
- * When `run_exam_for_mirror` completes, the full result is stored here
- * and a UUID `result_id` is returned. This avoids dumping large payloads
- * (~60KB+ for servers with many tools) into the LLM context.
+ * When `run_exam_for_mirror` completes, the full result is written to a
+ * JSON file in /tmp/ and a UUID `result_id` is returned. This avoids
+ * dumping large payloads (~60KB+ for servers with many tools) into the
+ * LLM context, and survives across tool calls without relying on
+ * in-memory state.
  *
- * Eviction: When the store exceeds MAX_RESULTS entries, the oldest result
+ * Files are named with a zero-padded sequence number prefix so that
+ * lexicographic sorting preserves insertion order for FIFO eviction.
+ * The sequence counter is initialized from existing files on disk so
+ * that new entries sort after old ones even across process restarts.
+ *
+ * Eviction: When the store exceeds MAX_RESULTS files, the oldest result
  * is evicted (FIFO). Results are also deleted after successful save via
  * `save_results_for_mirror`.
  *
@@ -21,38 +53,118 @@ const MAX_RESULTS = 100;
  * - Pass `result_id` to `save_results_for_mirror` instead of the full payload
  */
 class ExamResultStore {
-    results = new Map();
+    seq;
+    constructor() {
+        this.seq = this.initSeqFromDisk();
+    }
+    /**
+     * Scan existing files to find the highest sequence number and start
+     * one past it. This ensures new files always sort after existing ones,
+     * even across process restarts.
+     */
+    initSeqFromDisk() {
+        this.ensureDir();
+        const files = readdirSync(STORE_DIR)
+            .filter((f) => f.endsWith(FILE_SUFFIX) && f.length > FILE_SUFFIX.length)
+            .sort();
+        if (files.length === 0)
+            return 0;
+        const lastFile = files[files.length - 1];
+        const seqStr = lastFile.slice(0, 10);
+        const parsed = parseInt(seqStr, 10);
+        return isNaN(parsed) ? 0 : parsed + 1;
+    }
+    ensureDir() {
+        if (!existsSync(STORE_DIR)) {
+            mkdirSync(STORE_DIR, { recursive: true, mode: 0o700 });
+        }
+    }
+    /** Filename format: {seq}-{uuid}.json — seq is zero-padded for lexicographic ordering */
+    fileName(seq, resultId) {
+        return `${String(seq).padStart(10, '0')}-${resultId}${FILE_SUFFIX}`;
+    }
+    extractResultId(fileName) {
+        // Format: 0000000001-<uuid>.json
+        return fileName.slice(11, -FILE_SUFFIX.length);
+    }
+    listResultFiles() {
+        this.ensureDir();
+        return readdirSync(STORE_DIR)
+            .filter((f) => f.endsWith(FILE_SUFFIX) && f.length > FILE_SUFFIX.length)
+            .sort(); // Lexicographic sort gives insertion order via zero-padded seq
+    }
+    findFileForResult(resultId) {
+        const files = this.listResultFiles();
+        return files.find((f) => this.extractResultId(f) === resultId);
+    }
     store(mirrorIds, runtimeId, examType, lines) {
+        this.ensureDir();
         const resultId = randomUUID();
-        // Evict oldest entries if at capacity (Map preserves insertion order)
-        while (this.results.size >= MAX_RESULTS) {
-            const oldestKey = this.results.keys().next().value;
-            if (oldestKey !== undefined) {
-                this.results.delete(oldestKey);
+        const seqNum = this.seq++;
+        // Evict oldest entries if at capacity (files are sorted by seq prefix)
+        const files = this.listResultFiles();
+        const toEvict = files.length - MAX_RESULTS + 1;
+        for (let i = 0; i < toEvict; i++) {
+            try {
+                unlinkSync(join(STORE_DIR, files[i]));
+            }
+            catch {
+                // ignore
             }
         }
-        this.results.set(resultId, {
+        const stored = {
             result_id: resultId,
             mirror_ids: mirrorIds,
             runtime_id: runtimeId,
             exam_type: examType,
             lines,
             stored_at: new Date().toISOString(),
+        };
+        writeFileSync(join(STORE_DIR, this.fileName(seqNum, resultId)), JSON.stringify(stored), {
+            encoding: 'utf-8',
+            mode: 0o600,
         });
         return resultId;
     }
     get(resultId) {
-        return this.results.get(resultId);
+        const file = this.findFileForResult(resultId);
+        if (!file)
+            return undefined;
+        try {
+            const content = readFileSync(join(STORE_DIR, file), 'utf-8');
+            return JSON.parse(content);
+        }
+        catch {
+            return undefined;
+        }
     }
     delete(resultId) {
-        return this.results.delete(resultId);
+        const file = this.findFileForResult(resultId);
+        if (!file)
+            return false;
+        try {
+            unlinkSync(join(STORE_DIR, file));
+            return true;
+        }
+        catch {
+            return false;
+        }
     }
     get size() {
-        return this.results.size;
+        return this.listResultFiles().length;
     }
-    /** For testing only */
+    /** For testing only — removes all result files and resets the sequence counter */
     clear() {
-        this.results.clear();
+        this.ensureDir();
+        for (const file of this.listResultFiles()) {
+            try {
+                unlinkSync(join(STORE_DIR, file));
+            }
+            catch {
+                // ignore
+            }
+        }
+        this.seq = 0;
     }
 }
 /** Singleton instance shared across all tool factories */

package/shared/tools/get-exam-result.js

@@ -49,7 +49,7 @@ Typical usage:
                     content: [
                         {
                             type: 'text',
-                            text: `No stored result found for result_id "${validatedArgs.result_id}". Results are stored in-memory and may have been lost if the server restarted.`,
+                            text: `No stored result found for result_id "${validatedArgs.result_id}". The result file may have been cleaned up or the /tmp directory cleared.`,
                         },
                     ],
                     isError: true,

package/shared/tools/run-exam-for-mirror.js

@@ -1,5 +1,5 @@
 import { z } from 'zod';
-import { examResultStore } from '../exam-result-store.js';
+import { examResultStore, extractExamId, extractStatus } from '../exam-result-store.js';
 const PARAM_DESCRIPTIONS = {
     mirror_ids: 'Array of unofficial mirror IDs to run exams against. Mirrors without saved mcp_json configs will be skipped.',
     runtime_id: 'The Fly Machines runtime ID to use for running the exam containers (e.g., "fly-machines-v1")',
@@ -66,7 +66,7 @@ Available exam types:
 
 Mirrors without saved mcp_json configurations are automatically skipped.
 
-Results are stored server-side and a \`result_id\` UUID is returned. The response includes a truncated summary (status, tool names/counts, errors) that fits within MCP size limits. Use \`get_exam_result\` to drill into full details, or pass the \`result_id\` directly to \`save_results_for_mirror\`.
+Results are stored server-side in a local file and a \`result_id\` UUID is returned. The response includes a truncated summary (status, tool names/counts, errors) that fits within MCP size limits. Use \`get_exam_result\` to drill into full details, or pass the \`result_id\` directly to \`save_results_for_mirror\`.
 
 Use cases:
 - Test if an unofficial mirror's MCP server is working correctly before linking it
@@ -119,15 +119,18 @@ Use cases:
                         case 'log':
                             content += `[LOG] ${line.message || JSON.stringify(line)}\n`;
                             break;
-                        case 'exam_result':
-                            content += `\n**Exam Result** (Mirror: ${line.mirror_id || 'unknown'})\n`;
-                            content += `  Exam: ${line.exam_id || line.exam_type || 'unknown'}\n`;
-                            content += `  Status: ${line.status || 'unknown'}\n`;
-                            if (line.data) {
-                                const truncatedData = truncateExamResultData(line.data);
+                        case 'exam_result': {
+                            const data = line.data;
+                            const mirrorId = line.mirror_id ?? data?.mirror_id ?? 'unknown';
+                            content += `\n**Exam Result** (Mirror: ${mirrorId})\n`;
+                            content += `  Exam: ${extractExamId(line)}\n`;
+                            content += `  Status: ${extractStatus(line)}\n`;
+                            if (data) {
+                                const truncatedData = truncateExamResultData(data);
                                 content += `  Data: ${JSON.stringify(truncatedData, null, 2)}\n`;
                             }
                             break;
+                        }
                         case 'summary':
                             content += `\n**Summary**\n`;
                             content += `  Total: ${line.total || 0}\n`;

package/shared/tools/save-results-for-mirror.d.ts

@@ -17,7 +17,7 @@ export declare function saveResultsForMirror(_server: Server, clientFactory: Cli
             result_id: {
                 type: string;
                 format: string;
-                description: "The UUID returned by run_exam_for_mirror. When provided, the server retrieves the full result from the in-memory store — no need to pass the results array. This is the preferred approach.";
+                description: "The UUID returned by run_exam_for_mirror. When provided, the server retrieves the full result from the local file store — no need to pass the results array. This is the preferred approach.";
             };
             results: {
                 type: string;

package/shared/tools/save-results-for-mirror.js

@@ -1,9 +1,9 @@
 import { z } from 'zod';
-import { examResultStore } from '../exam-result-store.js';
+import { examResultStore, extractExamId, extractStatus } from '../exam-result-store.js';
 const PARAM_DESCRIPTIONS = {
     mirror_id: 'The ID of the unofficial mirror to save results for',
     runtime_id: 'The runtime ID that was used to run the exams',
-    result_id: 'The UUID returned by run_exam_for_mirror. When provided, the server retrieves the full result from the in-memory store — no need to pass the results array. This is the preferred approach.',
+    result_id: 'The UUID returned by run_exam_for_mirror. When provided, the server retrieves the full result from the local file store — no need to pass the results array. This is the preferred approach.',
     results: 'Array of exam results to save. Each result must include exam_id, status, and optional data. Only needed if result_id is not provided.',
     exam_id: 'The exam identifier (e.g., "auth-check", "init-tools-list")',
     status: 'The result status (e.g., "pass", "fail", "error", "skip")',
@@ -32,7 +32,7 @@ export function saveResultsForMirror(_server, clientFactory) {
         name: 'save_results_for_mirror',
         description: `Save proctor exam results for an unofficial mirror.
 
-**Preferred**: Pass the \`result_id\` returned by \`run_exam_for_mirror\`. The full result is retrieved from the in-memory store server-side — no need to pass the large results payload through the LLM context.
+**Preferred**: Pass the \`result_id\` returned by \`run_exam_for_mirror\`. The full result is retrieved from the local file store server-side — no need to pass the large results payload through the LLM context.
 
 **Fallback**: Pass results directly (as before) if result_id is not available.
 
@@ -88,20 +88,39 @@ Typical workflow:
                             content: [
                                 {
                                     type: 'text',
-                                    text: `No stored result found for result_id "${validatedArgs.result_id}". Results are stored in-memory and may have been lost if the server restarted. Pass the results array directly instead.`,
+                                    text: `No stored result found for result_id "${validatedArgs.result_id}". The result file may have been cleaned up or the /tmp directory cleared. Pass the results array directly instead.`,
                                 },
                             ],
                             isError: true,
                         };
                     }
-                    // Extract exam_result lines from stored data
+                    // Extract exam_result lines from stored data.
+                    // The exam_id may live at the top level of the stream line OR inside
+                    // line.data (the actual result payload). Prefer the data payload to
+                    // avoid reading from potentially incomplete display metadata.
+                    //
+                    // The real proctor API returns line.data as a metadata wrapper:
+                    //   { mirror_id, exam_id, status, result: { status, output: {...} } }
+                    // The actual output lives inside line.data.result. When we pass the
+                    // entire line.data as `data`, the output ends up nested too deeply
+                    // (result.data.result.output) and the backend saves empty output.
+                    // Use line.data.result when present so that `output` is at the
+                    // expected depth (result.data.output).
                     results = stored.lines
                         .filter((line) => line.type === 'exam_result')
-                        .map((line) => ({
-                        exam_id: (line.exam_id || line.exam_type || 'unknown'),
-                        status: (line.status || 'unknown'),
-                        ...(line.data ? { data: line.data } : {}),
-                    }));
+                        .map((line) => {
+                        const data = line.data;
+                        // Prefer the nested result object (contains output, input, etc.)
+                        // over the full data wrapper (contains metadata like mirror_id)
+                        const resultData = data?.result && typeof data.result === 'object' && !Array.isArray(data.result)
+                            ? data.result
+                            : data;
+                        return {
+                            exam_id: extractExamId(line),
+                            status: extractStatus(line),
+                            ...(resultData ? { data: resultData } : {}),
+                        };
+                    });
                     if (!runtimeId) {
                         runtimeId = stored.runtime_id;
                     }