onbuzz 4.4.0 → 4.6.0

package/src/tools/terminalDedup.js

@@ -0,0 +1,178 @@
+/**
+ * Terminal command deduplication.
+ *
+ * Goal: prevent the agent from accidentally launching the same command
+ * twice while the first invocation is still running. The user's intent
+ * is rarely "I really want two copies of `npm test` in flight at once" —
+ * it's more often a re-issued tool call after a perceived hang.
+ *
+ * Trigger: per-agent, deny only if an *identical* command (after
+ * whitespace normalization) is CURRENTLY RUNNING for the same agent.
+ * Once the prior run has exited (success OR failure), repeats are fine.
+ * Different agents running the same command in parallel are also fine.
+ *
+ * `force: true` override:
+ *   - Only honored if the SAME agent has just been denied for the SAME
+ *     command (within FORCE_TOKEN_TTL_MS).
+ *   - The first call of any command can NEVER bypass dedup with force.
+ *     This is the abuse-prevention property: `force` is a response to
+ *     a denial, not a preemptive permission.
+ *   - When force is honored, a parallel execution is allowed (option A
+ *     per design discussion). The token is cleared on use.
+ *
+ * The functions here are pure where possible and operate on plain
+ * arguments (a Map of running commands, a Map of denial tokens, a
+ * config). The TerminalTool instance owns those maps.
+ */
+
+import { TERMINAL_CONFIG } from '../utilities/constants.js';
+
+/** How long a "you were just denied; you may retry with force" token stays valid. */
+export const FORCE_TOKEN_TTL_MS = 5 * 60 * 1000;
+
+/**
+ * Normalize a command string for equality comparison. Trims surrounding
+ * whitespace and collapses CRLF to LF so a copy-pasted command from a
+ * Windows console doesn't fail to match its Unix-typed twin. Internal
+ * whitespace, env-var substitutions, etc. are NOT canonicalized — that
+ * way `echo $(date)` invocations naturally differ each call and the
+ * dedup never falsely fires.
+ *
+ * @param {string} command
+ * @returns {string}
+ */
+export function normalizeCommand(command) {
+  if (typeof command !== 'string') return '';
+  return command.replace(/\r\n/g, '\n').trim();
+}
+
+/**
+ * Locate a currently-running command for `agentId` that matches the
+ * given normalized command string. Returns the tracker entry (with
+ * commandId) or null.
+ *
+ * @param {Map<string, object>} commandTracker  Live tracker from TerminalTool.
+ * @param {string} agentId
+ * @param {string} normalizedCommand
+ * @returns {object|null}
+ */
+export function findRunningDuplicate(commandTracker, agentId, normalizedCommand) {
+  if (!commandTracker || !agentId || !normalizedCommand) return null;
+  for (const info of commandTracker.values()) {
+    if (info.agentId !== agentId) continue;
+    if (info.state !== TERMINAL_CONFIG.STATES.RUNNING) continue;
+    if (normalizeCommand(info.command) !== normalizedCommand) continue;
+    return info;
+  }
+  return null;
+}
+
+/**
+ * Decide whether to allow or deny the requested execution. Pure function:
+ * accepts the tool's state via parameters and returns a decision.
+ *
+ * @param {object} args
+ * @param {Map<string, object>} args.commandTracker
+ * @param {Map<string, object>} args.lastDeniedExec   Mutable; updated on deny / cleared on honored-force.
+ * @param {string} args.agentId
+ * @param {string} args.command                       Original (unnormalized) command.
+ * @param {boolean} [args.force=false]
+ * @param {object} [args.config]                      Effective per-agent terminal config.
+ * @param {number} [args.now=Date.now()]              Override for tests.
+ * @returns {{ allow: true } | { allow: false, reason: string, hint: string, status: object }}
+ */
+export function checkDedup({
+  commandTracker,
+  lastDeniedExec,
+  agentId,
+  command,
+  force = false,
+  config = {},
+  now = Date.now(),
+}) {
+  // Honor a per-agent kill switch — agent's tool config can opt out
+  // (e.g. test agents that need to spam identical commands). Default
+  // is enabled.
+  if (config.denyDuplicateConcurrentCommands === false) {
+    return { allow: true };
+  }
+  if (!agentId) return { allow: true };
+
+  const normalized = normalizeCommand(command);
+  if (!normalized) return { allow: true };
+
+  const duplicate = findRunningDuplicate(commandTracker, agentId, normalized);
+  if (!duplicate) {
+    // No conflicting in-flight command. `force` is silently ignored —
+    // there's nothing to override. Critically: this means an agent
+    // that ALWAYS passes force=true gains nothing by doing so. The
+    // flag is useless except as a response to a prior denial.
+    return { allow: true };
+  }
+
+  // There IS a duplicate in flight. Three sub-cases:
+  if (force === true) {
+    // Force is only honored if a matching denial token exists for THIS
+    // agent + THIS command, and isn't expired. That guarantees the
+    // agent has actually seen the denial first.
+    const token = lastDeniedExec.get(agentId);
+    const tokenValid = token
+      && token.command === normalized
+      && (now - token.deniedAt) <= FORCE_TOKEN_TTL_MS;
+    if (tokenValid) {
+      lastDeniedExec.delete(agentId);   // single-use token
+      return { allow: true };
+    }
+    // Force without a prior denial → fall through to denial path; record
+    // a new token, agent can then retry with force on its NEXT turn.
+  }
+
+  // Record the denial so the agent's next turn (with force=true) can
+  // legitimately retry. Single-use, agent-scoped.
+  lastDeniedExec.set(agentId, { command: normalized, deniedAt: now });
+
+  const elapsedMs = now - new Date(duplicate.startTime).getTime();
+  const elapsedSec = Math.max(0, Math.round(elapsedMs / 1000));
+  return {
+    allow: false,
+    reason: 'duplicate-running',
+    hint: 'An identical command is still running. Wait for it to finish — the result will arrive on a later turn. ' +
+          'If you are SURE you want to launch a parallel copy, retry the same call with force:true.',
+    status: {
+      commandId:        duplicate.commandId,
+      command:          duplicate.command,
+      state:            duplicate.state,
+      startedAt:        duplicate.startTime,
+      elapsedSec,
+      workingDirectory: duplicate.workingDirectory,
+    },
+  };
+}
+
+/**
+ * Build the "denied" tool result the terminal tool returns to the agent.
+ * Shape mirrors the standard run-command failure envelope so existing
+ * agent error-handling paths just work.
+ *
+ * @param {object} decision  Output of checkDedup with allow=false.
+ * @param {string} originalCommand
+ * @returns {object}
+ */
+export function denialResult(decision, originalCommand) {
+  return {
+    success: false,
+    deduped: true,
+    action: 'run-command',
+    command: originalCommand,
+    error: 'Duplicate command rejected: ' + decision.hint,
+    duplicateOf: decision.status,
+  };
+}
+
+export default {
+  normalizeCommand,
+  findRunningDuplicate,
+  checkDedup,
+  denialResult,
+  FORCE_TOKEN_TTL_MS,
+};

package/src/tools/terminalTool.js

@@ -20,6 +20,7 @@ import {
   SYSTEM_DEFAULTS,
   TERMINAL_CONFIG
 } from '../utilities/constants.js';
+import { checkDedup, denialResult } from './terminalDedup.js';
 
 /**
  * PromptDetector - Detects interactive prompts in command output
@@ -182,6 +183,15 @@ class TerminalTool extends BaseTool {
 
     // Phase 3 & 4: Background command tracking
     this.commandTracker = new Map(); // commandId -> { agentId, pid, process, state, buffers, timestamps }
+
+    // Per-agent denial token for the duplicate-command dedup feature.
+    // Map<agentId, { command, deniedAt }>. Set when the dedup check
+    // refuses a duplicate-while-running. Consumed (single-use) when the
+    // agent retries the SAME command with force:true. The token's
+    // existence is what makes force:true meaningful — without a prior
+    // denial, force is silently ignored, which prevents agents from
+    // preemptively passing force on every call to bypass dedup.
+    this.lastDeniedExec = new Map();
     this.commandIdCounter = 0;
 
     // Resource limits
@@ -488,6 +498,24 @@ After each invocation you will get a tool-result embedded in the following user
         
         switch (action.type) {
           case 'run-command':
+            // Dedup gate: if an identical command is already running for
+            // this agent, deny — unless the agent has been previously
+            // denied for this exact command and now passes force:true.
+            // See terminalDedup.js for the full contract.
+            {
+              const dedup = checkDedup({
+                commandTracker:  this.commandTracker,
+                lastDeniedExec:  this.lastDeniedExec,
+                agentId,
+                command:         action.command,
+                force:           action.force === true,
+                config:          effectiveConfig,
+              });
+              if (!dedup.allow) {
+                result = denialResult(dedup, action.command);
+                break;
+              }
+            }
             result = await this.executeCommand(action.command, currentWorkingDir, {
               timeout: customTimeout || this.timeout,
               async: forceAsync || false,

package/src/tools/videoTool.js

@@ -10,34 +10,40 @@ import { BaseTool } from './baseTool.js';
 import { getGalleryService } from '../services/galleryService.js';
 
 /**
- * Configuration constants for video generation
- * Updated for Sora 2 API specifications (supports both Sora 1 and Sora 2)
+ * Configuration constants for video generation.
+ *
+ * Per-model limits (allowed resolutions, allowed durations, whether
+ * variants are supported) are NOT hardcoded here — they live in the
+ * model catalog (`video_config` block on each video-generation row) and
+ * are enforced authoritatively by the backend route. The CLI sends the
+ * request and surfaces the backend's 400 verbatim if the inputs aren't
+ * compatible with the chosen model. This keeps the CLI provider-agnostic
+ * (Sora 1, Sora 2, future video models all just work) and avoids
+ * silently-stale validation tables here.
+ *
+ * The constants below are CLI-side controls only (queue/concurrency/
+ * polling/temp file lifetime/timeout). No model-specific values.
  */
 const VIDEO_CONFIG = {
-  DEFAULT_MODEL: null, // Resolved dynamically from modelsService via aiService
-  DEFAULT_WIDTH: 1920,
-  DEFAULT_HEIGHT: 1080,
-  DEFAULT_DURATION: 10, // seconds
-  DEFAULT_VARIANTS: 1,
-  // Sora 2 supported resolutions (superset of Sora 1 + new HD/4K options)
-  VALID_RESOLUTIONS: [
-    // Landscape
-    '1920x1080', '1792x1024', '1280x720', '854x480',
-    // Portrait
-    '1080x1920', '1024x1792', '720x1280', '480x854',
-    // Square
-    '1080x1080', '720x720', '480x480'
-  ],
-  MIN_DURATION: 1,
-  MAX_DURATION: 25,  // seconds (Sora 2 supports up to 25s)
-  MAX_VARIANTS: 4,
+  DEFAULT_MODEL: null,           // resolved dynamically by aiService
+  DEFAULT_WIDTH:  1280,          // sensible default; any video model in the
+  DEFAULT_HEIGHT: 720,           // catalog with `1280x720` accepts these
+  DEFAULT_DURATION: 4,           // Sora 2's smallest unit; safe across providers
+  DEFAULT_VARIANTS: 1,           // single video per job (Sora 2; Sora 1 supported up to 4)
+  MAX_PROMPT_LENGTH: 4000,
+  // Hard CLI safety bounds — let the backend/catalog be authoritative on
+  // *which specific values* are allowed for the chosen model, but reject
+  // obviously-bogus inputs locally so the user gets immediate feedback.
+  ABSOLUTE_MIN_DURATION: 1,
+  ABSOLUTE_MAX_DURATION: 60,
+  ABSOLUTE_MAX_VARIANTS: 8,
+  // Queue + scheduler controls (CLI-side only).
   MAX_CONCURRENT: 2,
   QUEUE_LIMIT: 5,
-  POLL_INTERVAL: 5000,   // 5 seconds between status checks
-  MAX_POLL_TIME: 600000, // 10 minutes max wait
-  TEMP_CLEANUP_MS: 86400000, // 24 hours (videos expire after 24h anyway)
-  MAX_PROMPT_LENGTH: 4000,
-  DOWNLOAD_TIMEOUT: 300000 // 5 minutes for video download
+  POLL_INTERVAL: 5000,           // ms between status checks
+  MAX_POLL_TIME: 600000,         // 10 minutes max wait per job
+  TEMP_CLEANUP_MS: 86400000,     // 24 h (Sora 2 expires content after 24 h anyway)
+  DOWNLOAD_TIMEOUT: 300000       // 5 minutes for video download
 };
 
 /**
@@ -118,8 +124,12 @@ export class VideoTool extends BaseTool {
       active:    Array.from(this.activeJobs.values()).map(summarize),
       completed: Array.from(this.completedJobs.values()).map(summarize),
       limits: {
-        concurrency: (typeof VIDEO_CONFIG !== 'undefined' && VIDEO_CONFIG?.CONCURRENCY_LIMIT) || this.activeJobs.size + this.queue.length,
-        queue:       (typeof VIDEO_CONFIG !== 'undefined' && VIDEO_CONFIG?.QUEUE_LIMIT) || null,
+        // Was reading the non-existent VIDEO_CONFIG.CONCURRENCY_LIMIT and
+        // falling through to `activeJobs.size + queue.length`, which is
+        // 0 when idle and produced a misleading `"concurrency": 0` in the
+        // /api/videos/jobs snapshot (the real cap is MAX_CONCURRENT=2).
+        concurrency: VIDEO_CONFIG?.MAX_CONCURRENT ?? null,
+        queue:       VIDEO_CONFIG?.QUEUE_LIMIT    ?? null,
       },
     };
   }
@@ -182,14 +192,14 @@ export class VideoTool extends BaseTool {
    * @returns {string} Formatted tool description
    */
   getDescription() {
-    return `Tool: Video Generator - Generate videos using Sora 2
+    return `Tool: Video Generator — generate videos from text using the platform's video model.
 
-**Purpose:** Generate videos from text descriptions using Sora 2. Videos are saved to files and displayed in chat.
+**Purpose:** Generate short videos from text descriptions. Videos are saved to disk and made available in chat. The active video model is resolved automatically from the model catalog (currently Sora 2 on this deployment); model-specific limits like allowed resolutions and durations come from the catalog and are enforced by the backend.
 
 **CRITICAL: Automatic Execution**
-- ANY \`\`\`json block with "toolId": "video-gen" will be EXECUTED IMMEDIATELY
-- Just output the command when you want to generate a video
-- Video generation takes several minutes - a job ID is returned immediately
+- ANY \`\`\`json block with "toolId": "video-gen" will be EXECUTED IMMEDIATELY.
+- Just output the command when you want to generate a video.
+- Video generation takes minutes — a job ID is returned immediately and the file lands when the job finishes.
 
 **USAGE:**
 \`\`\`json
@@ -198,22 +208,27 @@ export class VideoTool extends BaseTool {
   "parameters": {
     "prompt": "Detailed description of the video",
     "outputPath": "videos/filename.mp4",
-    "width": 1920,
-    "height": 1080,
-    "duration": 10
+    "width": 1280,
+    "height": 720,
+    "duration": 4
   }
 }
 \`\`\`
 
 **Parameters:**
-- **prompt** (required): Detailed description of video to generate
-- **outputPath** (optional): Path to save video (permanent). Omit for temp file.
-- **width** (optional): Video width - 480, 720, 854, 1024, 1080, 1280, 1792, 1920 (default: 1920)
-- **height** (optional): Video height - 480, 720, 854, 1024, 1080, 1280, 1792, 1920 (default: 1080)
-- **duration** (optional): Duration in seconds, 1-25 (default: 10)
-- **variants** (optional): Number of video variants to generate, 1-4 (default: 1)
+- **prompt** (required): Detailed description of the video.
+- **outputPath** (optional): Where to save the .mp4 (permanent). Omit for a temp file.
+- **width** / **height** (optional): Resolution. The catalog defines allowed pairs per model; the backend rejects unsupported pairs with a clear 400 listing valid values.
+- **duration** (optional): Duration in seconds. The catalog defines allowed values per model.
+- **variants** (optional): Number of video variants. Some models (e.g. Sora 2) only return one video per job; the backend will reject \`variants > 1\` with a clear message naming the model's limit.
+- **model** (optional): Override the auto-resolved model. Must be a video-generation model in the catalog.
+
+**Current model limits (Sora 2):**
+- Resolutions: \`720x1280\`, \`1280x720\`, \`1024x1792\`, \`1792x1024\` (16:9 / 9:16 only — no square or 480p)
+- Durations: \`4\`, \`8\`, or \`12\` seconds (those exact values, no others)
+- Variants: 1 per job — request multiple separately if you need more
 
-**Valid Resolutions:** 1920x1080, 1792x1024, 1280x720, 854x480, 1080x1920, 1024x1792, 720x1280, 480x854, 1080x1080, 720x720, 480x480
+These limits live in the catalog row, not in this tool. If the model rotates (e.g. Sora 3) or a new video model is added, the allowed values update automatically and you'll see them in the backend's 400 if you submit unsupported inputs.
 
 **EXAMPLE:**
 User: "create a video of a cat playing"
@@ -224,20 +239,20 @@ You output:
   "parameters": {
     "prompt": "A fluffy orange tabby cat with white paws crouches low on a sunlit hardwood floor, eyes locked on a red laser dot. The cat pounces forward, slides slightly on the polished wood, then quickly pivots to chase the dot as it darts away. Warm afternoon sunlight streams through sheer curtains, casting soft golden highlights on the cat's fur. Cozy living room with a beige sofa in the background.",
     "outputPath": "videos/cat-playing.mp4",
-    "width": 1080,
-    "height": 1080,
-    "duration": 5
+    "width": 1280,
+    "height": 720,
+    "duration": 4
   }
 }
 \`\`\`
 
 **Prompt Guidelines (IMPORTANT):**
-Sora simulates a physical world, so write prompts as narratives, not camera commands. Use the CAST method: describe the **Character** (appearance, clothing, posture), **Action** (break into beats like "takes three steps, pauses, looks back"), **Setting** (time of day, weather, specific objects), and **Tone/Atmosphere** (lighting quality, color palette, mood). Be specific and sensory—replace "beautiful street" with "rain-slick Tokyo asphalt reflecting neon signs." Keep prompts under 120 words, focus on ONE action per clip, and use simple camera cues only if needed ("wide shot," "close-up"). Anchor lighting explicitly ("warm golden hour sunlight with soft shadows") and name colors (teal, amber, magenta) for palette consistency. For character consistency across clips, repeat the same distinctive details (clothing colors, accessories, features) in each prompt. Avoid real people, copyrighted characters, and sensitive content.
+Sora simulates a physical world — write prompts as narratives, not camera commands. Use the CAST method: describe the **Character** (appearance, clothing, posture), **Action** (broken into beats: "takes three steps, pauses, looks back"), **Setting** (time of day, weather, specific objects), and **Tone/Atmosphere** (lighting quality, color palette, mood). Be specific and sensory — replace "beautiful street" with "rain-slick Tokyo asphalt reflecting neon signs." Keep prompts under 120 words, focus on ONE action per clip, and use simple camera cues only if needed ("wide shot," "close-up"). Anchor lighting explicitly ("warm golden hour sunlight with soft shadows") and name colors (teal, amber, magenta) for palette consistency. For character consistency across clips, repeat the same distinctive details (clothing colors, accessories, features) in each prompt. Avoid real people, copyrighted characters, and sensitive content.
 
 **Notes:**
-- Videos take 3-10 minutes to generate
-- Max ${VIDEO_CONFIG.QUEUE_LIMIT} videos in queue, max 2 concurrent jobs
-- Videos expire after 24 hours`;
+- Videos take 1-3 minutes to generate (Sora 2 short clips).
+- Max ${VIDEO_CONFIG.QUEUE_LIMIT} videos in queue, max ${VIDEO_CONFIG.MAX_CONCURRENT} concurrent jobs.
+- Generated videos expire on the provider after 24 hours; if you need to keep them, set an \`outputPath\` so the local copy persists.`;
   }
 
   /**
@@ -471,7 +486,17 @@ Sora simulates a physical world, so write prompts as narratives, not camera comm
   }
 
   /**
-   * Validate parameters
+   * Sanity-check inputs before submitting to the backend. Per-model limits
+   * (which exact resolutions and durations a given video model accepts)
+   * are NOT enforced here — the catalog declares them and the backend
+   * route validates against the catalog row. This keeps the CLI
+   * provider-agnostic; if the inputs aren't compatible with the chosen
+   * model, the backend's 400 surfaces back through the queue's `error`
+   * field with a clear message naming the allowed values.
+   *
+   * Local checks below cover only "obviously wrong" inputs that no video
+   * provider would accept — empty prompts, non-positive dimensions, etc.
+   *
    * @private
    */
   _validateParameters(params) {
@@ -483,25 +508,34 @@ Sora simulates a physical world, so write prompts as narratives, not camera comm
       if (!vid.prompt || vid.prompt.trim().length === 0) {
         throw new Error('Video prompt is required');
       }
-
       if (vid.prompt.length > VIDEO_CONFIG.MAX_PROMPT_LENGTH) {
         throw new Error(`Prompt too long (max ${VIDEO_CONFIG.MAX_PROMPT_LENGTH} characters)`);
       }
 
-      // Validate resolution
-      const resolution = `${vid.width}x${vid.height}`;
-      if (!VIDEO_CONFIG.VALID_RESOLUTIONS.includes(resolution)) {
-        throw new Error(`Invalid resolution: ${resolution}. Valid resolutions: ${VIDEO_CONFIG.VALID_RESOLUTIONS.join(', ')}`);
+      // Width/height must be positive integers — backend + provider
+      // decide which exact pairs are valid for the chosen model.
+      if (vid.width != null && (!Number.isInteger(vid.width) || vid.width <= 0)) {
+        throw new Error(`Invalid width: ${vid.width}. Must be a positive integer.`);
+      }
+      if (vid.height != null && (!Number.isInteger(vid.height) || vid.height <= 0)) {
+        throw new Error(`Invalid height: ${vid.height}. Must be a positive integer.`);
       }
 
-      // Validate duration
-      if (vid.duration < VIDEO_CONFIG.MIN_DURATION || vid.duration > VIDEO_CONFIG.MAX_DURATION) {
-        throw new Error(`Invalid duration: ${vid.duration}. Must be between ${VIDEO_CONFIG.MIN_DURATION} and ${VIDEO_CONFIG.MAX_DURATION} seconds`);
+      // Duration: only catch absurd values here; the catalog declares
+      // the per-model allowed list (e.g. Sora 2: {4, 8, 12} seconds).
+      if (vid.duration != null) {
+        if (typeof vid.duration !== 'number' || vid.duration < VIDEO_CONFIG.ABSOLUTE_MIN_DURATION || vid.duration > VIDEO_CONFIG.ABSOLUTE_MAX_DURATION) {
+          throw new Error(`Invalid duration: ${vid.duration}. Must be a number between ${VIDEO_CONFIG.ABSOLUTE_MIN_DURATION} and ${VIDEO_CONFIG.ABSOLUTE_MAX_DURATION} seconds. The chosen model may support a narrower set — see /llm/models.`);
+        }
       }
 
-      // Validate variants
-      if (vid.variants < 1 || vid.variants > VIDEO_CONFIG.MAX_VARIANTS) {
-        throw new Error(`Invalid variants: ${vid.variants}. Must be between 1 and ${VIDEO_CONFIG.MAX_VARIANTS}`);
+      // Variants: same approach — sanity-only. Some models (e.g. Sora 2)
+      // only return a single video per job; the backend will 400 if the
+      // model doesn't support multiple variants.
+      if (vid.variants != null) {
+        if (!Number.isInteger(vid.variants) || vid.variants < 1 || vid.variants > VIDEO_CONFIG.ABSOLUTE_MAX_VARIANTS) {
+          throw new Error(`Invalid variants: ${vid.variants}. Must be an integer between 1 and ${VIDEO_CONFIG.ABSOLUTE_MAX_VARIANTS}. The chosen model may only support 1.`);
+        }
       }
 
       if (vid.outputPath) {

package/src/tools/webTool.js

@@ -3073,6 +3073,15 @@ If blocked (CAPTCHA, access denied), use stealthLevel: "maximum" (visible browse
     this.cleanupTimer = setInterval(() => {
       this.cleanupIdleTabs();
     }, this.CLEANUP_INTERVAL);
+    // Don't let this background-maintenance timer hold the event loop
+    // open. When all real work is done (CLI exits, test suite finishes),
+    // the process should be free to exit on its own. Tests that create
+    // a WebTool without explicit cleanup() previously left jest workers
+    // hanging until --forceExit. unref() is the industry-standard fix
+    // for periodic-maintenance timers.
+    if (typeof this.cleanupTimer.unref === 'function') {
+      this.cleanupTimer.unref();
+    }
   }
 
   /**

package/src/utilities/__tests__/apiKeyPlaceholders.test.js

@@ -0,0 +1,120 @@
+/**
+ * Unit tests for the API-key placeholder detection helpers (shared by
+ * apiKeyManager + the web UI).
+ *
+ * Locks the heuristic that prevents UI display strings ("lx_•••• …
+ * (server-managed)" and the legacy "platform-managed" sentinel) from
+ * leaking into persistence and overwriting real keys with garbage.
+ */
+import { describe, test, expect } from '@jest/globals';
+import {
+  isPlaceholderApiKey,
+  sanitizeApiKeyValue,
+  sanitizeApiKeysObject,
+  LEGACY_BULLET_PLACEHOLDER,
+  LEGACY_PLATFORM_MANAGED_SENTINEL,
+} from '../apiKeyPlaceholders.js';
+
+describe('isPlaceholderApiKey', () => {
+  test('detects the canonical bullet placeholder', () => {
+    expect(isPlaceholderApiKey(LEGACY_BULLET_PLACEHOLDER)).toBe(true);
+    expect(isPlaceholderApiKey('lx_••••••••••••••••(server-managed)')).toBe(true);
+  });
+
+  test('detects any string containing bullets', () => {
+    expect(isPlaceholderApiKey('••••••••')).toBe(true);
+    expect(isPlaceholderApiKey('lx_•••')).toBe(true);
+  });
+
+  test('detects the platform-managed sentinel exactly', () => {
+    expect(isPlaceholderApiKey(LEGACY_PLATFORM_MANAGED_SENTINEL)).toBe(true);
+    expect(isPlaceholderApiKey('platform-managed')).toBe(true);
+  });
+
+  test('detects (server-managed) and [server-managed] tags anywhere in the string', () => {
+    expect(isPlaceholderApiKey('lx_abc(server-managed)')).toBe(true);
+    expect(isPlaceholderApiKey('foo[server-managed]bar')).toBe(true);
+    expect(isPlaceholderApiKey('(server-managed)')).toBe(true);
+  });
+
+  test('treats real-looking keys as not-placeholder', () => {
+    expect(isPlaceholderApiKey('lx_1234567890abcdef')).toBe(false);
+    expect(isPlaceholderApiKey('sk-proj-abcdefghijk')).toBe(false);
+    expect(isPlaceholderApiKey('FfX3H7zwazdmDG3wIslbiawTYLVxXyzGKa97IaXcH9nUjVprZ0HqJQQJ99BEACfhMk5XJ3w3AAAAACOG6efc')).toBe(false);
+  });
+
+  test('returns false for empty / null / undefined / non-string', () => {
+    expect(isPlaceholderApiKey('')).toBe(false);
+    expect(isPlaceholderApiKey(null)).toBe(false);
+    expect(isPlaceholderApiKey(undefined)).toBe(false);
+    expect(isPlaceholderApiKey(123)).toBe(false);
+    expect(isPlaceholderApiKey({})).toBe(false);
+    expect(isPlaceholderApiKey([])).toBe(false);
+  });
+
+  test('does NOT treat the substring "managed" alone as a placeholder', () => {
+    // Real keys should be allowed even if they happen to contain English
+    // words. Only the bracketed/parenthesised tag triggers detection.
+    expect(isPlaceholderApiKey('lx_managed_account_key_12345')).toBe(false);
+    expect(isPlaceholderApiKey('server')).toBe(false);
+  });
+});
+
+describe('sanitizeApiKeyValue', () => {
+  test('passes real keys through unchanged', () => {
+    expect(sanitizeApiKeyValue('lx_real_key_abc123')).toBe('lx_real_key_abc123');
+  });
+
+  test('strips placeholder values to empty string', () => {
+    expect(sanitizeApiKeyValue(LEGACY_BULLET_PLACEHOLDER)).toBe('');
+    expect(sanitizeApiKeyValue('platform-managed')).toBe('');
+    expect(sanitizeApiKeyValue('foo(server-managed)')).toBe('');
+  });
+
+  test('returns "" for non-string inputs (avoids leaking nulls into UI state)', () => {
+    expect(sanitizeApiKeyValue(null)).toBe('');
+    expect(sanitizeApiKeyValue(undefined)).toBe('');
+    expect(sanitizeApiKeyValue(0)).toBe('');
+    expect(sanitizeApiKeyValue({ value: 'real' })).toBe('');
+  });
+
+  test('preserves empty string', () => {
+    expect(sanitizeApiKeyValue('')).toBe('');
+  });
+});
+
+describe('sanitizeApiKeysObject', () => {
+  test('returns an empty object for null / undefined / non-object', () => {
+    expect(sanitizeApiKeysObject(null)).toEqual({});
+    expect(sanitizeApiKeysObject(undefined)).toEqual({});
+    expect(sanitizeApiKeysObject('not-an-object')).toEqual({});
+  });
+
+  test('strips placeholders, preserves real keys', () => {
+    const result = sanitizeApiKeysObject({
+      loxiaBackend: LEGACY_BULLET_PLACEHOLDER,
+      anthropic:    'sk-ant-real-key-1234567890',
+      openai:       'platform-managed',
+      azure:        'lx_real_azure_key',
+      deepseek:     'foo(server-managed)bar',
+    });
+    expect(result).toEqual({
+      loxiaBackend: '',
+      anthropic:    'sk-ant-real-key-1234567890',
+      openai:       '',
+      azure:        'lx_real_azure_key',
+      deepseek:     '',
+    });
+  });
+
+  test('does not mutate the input', () => {
+    const input = { loxiaBackend: LEGACY_BULLET_PLACEHOLDER };
+    sanitizeApiKeysObject(input);
+    expect(input.loxiaBackend).toBe(LEGACY_BULLET_PLACEHOLDER);
+  });
+
+  test('roundtrip: clean object passes through unchanged-shape', () => {
+    const clean = { loxiaBackend: 'lx_abc', anthropic: 'sk-ant-xyz' };
+    expect(sanitizeApiKeysObject(clean)).toEqual(clean);
+  });
+});