@wrongstack/cli 0.282.1 → 0.282.2

package/dist/index.js

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { CODEX_MODELS, color, ToolCapabilities, PHASE_EVENT_NAMES, readBundledInstructionText, writeErr, loadPlugins, renderProgress, SpecStore, TaskGraphStore, analyzeCriticalPath, getTemplate, listTemplates, templateToMarkdown, SpecParser, renderSpecAnalysis, AISpecBuilder, expectDefined, DefaultTaskStore, TaskTracker, renderTaskGraph, withFileLock, ConfigError, resolveHqDataDir, ensureHqFirstRunAuthFile, DEFAULT_HQ_REDACTION_POLICY, createHqPersistence, HqCommandAuditLog, HqAlertEngine, toAlertMessage, validateHqCommand, watchHqAuthFile, parseHqFrame, HQ_PROTOCOL_VERSION, parseHqEventPayload, scrubAndTruncateHqPreview, FetchError, removeColumn, addColumn, getTask, removeTask, getBoard, assignTask, updateTaskAssignment, listBoards, updateGoalMetricOnTask, addGoalMetricToTask, releaseTaskClaim, claimReadyTask, getTaskChain, setTaskChain, copyTaskToBoard, transferTaskToBoard, moveTask, updateTask, mergeTasks, splitTask, addTask, syncBoardFromTaskGraph, deserializeTaskGraph, exportBoardToTaskGraph, serializeTaskGraph, getKanbanOrchestrationSnapshot, listReadyTasks, generateBoardFromDescription, createBoard, parseLinesIntoTasks, removeBoard, duplicateBoard, updateBoard, wstackGlobalRoot, DefaultSecretScrubber, resolveWstackPaths, watchProviderConfig, PromptUsageStore, TOKENS, ToolRegistry, GlobalMailbox, resolveSessionLoggingConfig, createSessionEventBridge, HookRegistry, HookRunner, shellHooksEqual, countShellHooks, normalizeTokenSavingTier, allServers as allServers$1, SlashCommandRegistry, attachDepWatcherBridge, createFallbackModelExtension, SessionMemoryConsolidator, BrainDecisionQueue, ObservableBrainArbiter, HumanEscalatingBrainArbiter, createTieredBrainArbiter, DefaultBrainArbiter, BrainMonitor, mailboxSessionTag, createDelegateTool, FLEET_ROSTER, createMcpControlTool, startTechStackConsumer, startPackageOutdatedWatcher, SddRunRegistry, recordFileAction, createAutonomyBrain, DefaultPluginAPI, SpecVersioning, buildTranscriptFromEvents, DefaultLogger, DefaultModelsRegistry, isStdinTTY, FsError, atomicWrite, EventBus, runProviderWithRetry, ReplayLogStore, ReplayProviderRunner, mergeCustomModelDefs, DefaultSystemPromptBuilder, makeAutonomyPromptContributor, sessionScopedPath as sessionScopedPath$1, createContextManagerTool, makeMailboxTool, makeMailSendTool, makeMailInboxTool, makeFleetStatusTool, applyToolDescriptionModes, applyToolResultRenderModes, InMemoryMetricsSink, wireMetricsToEvents, DefaultHealthRegistry, startMetricsServer, createHqPublisherFromEnv, DEFAULT_SESSION_PRUNE_DAYS, RecoveryLock, DefaultAttachmentStore, Context, QueueStore, loadTodosCheckpoint, attachTodosCheckpoint, loadDirectorState, loadPlan, createDefaultPipelines, applyModelRuntime, installDesignStudioMiddleware, resolveContextWindowPolicy, resolveAuditLevel, AutoCompactionMiddleware, estimateRequestTokensCalibrated, Agent, FleetManager, makeDirectorSessionFactory, WorktreeManager, Director, makeFleetEmitTool, AdaptiveConcurrencyController, resolveSubagentModelTarget, DEFAULT_SUBAGENT_BASELINE, AutoApprovePermissionPolicy, mergeModelRuntime, ToolValidationError, WIDE_SUBAGENT_CAPABILITIES, resolveProjectDir, FleetSupervisor, AgentError, PhaseStore, PhaseOrchestrator, assignNickname, AutoPhasePlanner, PhaseGraphBuilder, makeLLMClassifier, writeOut, ParallelEternalEngine, EternalAutonomyEngine, CHIMERA_REVIEW_PROMPT, setQueuedMessagesSnapshot, bootConfig as bootConfig$1, setOutputLineGuard, setRawMode, DefaultSessionReader, ToolAuditLog, DefaultSessionRewinder, DefaultSessionStore, ProviderRegistry, makePreferSideConflictResolver, makeAgentSubagentRunner, NULL_FLEET_BUS, ERROR_CODES, buildChildEnv, exportBoardAsMarkdown, formatContextWindowModeList, repairToolUseAdjacency, getContextWindowMode, AGENT_CATALOG, dispatchAgent, normalizeToolDescriptionMode, getToolDescriptionMode, noOpVault, formatTodosList, loadTasks, formatTaskList, formatTaskProgress, mutateTasks, emptyPlan, addPlanItem, savePlan, formatPlan, recordCompletedWorkEvidence, saveCompletedWorkCheckpoint, SessionRecovery, loadGoal, goalFilePath, summarizeUsage, saveGoal, formatGoal, emptyGoal, buildGoalPreamble, pendingBtwCount, setBtwNote, getDesignKitLoader, getDesignState, clearActiveKit, clearPersistedActiveKit, recordOverrides, setDesignOverrides, loadActiveKit, isDesignStack, applyTokenOverrides, materializeTokens, setActiveKit, readHqRuntimeFileSync, resolveHqConfig, MATRIX_PHASE_KEYS, matrixKeyKind, phaseForRole, resolveModelMatrix, resolveModelTargetFromEntry, fallbackProfileChain, parseModelRef, normalizeModelRef, onResize, hasOpenTodos, InputBuilder, DefaultSecretVault as DefaultSecretVault$1, decryptConfigSecrets as decryptConfigSecrets$1, encryptConfigSecrets as encryptConfigSecrets$1, DEFAULT_CONTEXT_WINDOW_MODE_ID, HQ_AUTH_FILE_VERSION, truncate, createBoardFromTaskGraph, addDependency, addNoteToTask, addCheckToTask, estimateMessageTokens, AGENTS_BY_PHASE, validateAgainstSchema, normalizeToolResultRenderMode, getToolResultRenderMode, setToolResultRenderMode, resolveMailboxIdentity, smartDefaultFallbackChain, AutonomousCoordinator, mutateHqAuthFile, mintHqToken, readHqAuthFile, areDependenciesMet, findBlockedTasks, renderInstructionTemplate, recentTextTurns, gatedEnhancerReasoning, enhanceUserPrompt, mutatePlan, setPlanItemStatus, getPlanTemplate, resolveProviderModelList, ParseError, startSessionTelemetryBridge, startFleetTelemetryBridge, startBrainTelemetryBridge, startWorktreeTelemetryBridge, startToolTelemetryBridge, startCostTelemetryBridge } from '@wrongstack/core';
+import { CODEX_MODELS, color, ToolCapabilities, PHASE_EVENT_NAMES, readBundledInstructionText, writeErr, loadPlugins, renderProgress, SpecStore, TaskGraphStore, analyzeCriticalPath, getTemplate, listTemplates, templateToMarkdown, SpecParser, renderSpecAnalysis, AISpecBuilder, expectDefined, DefaultTaskStore, TaskTracker, renderTaskGraph, withFileLock, ConfigError, resolveHqDataDir, ensureHqFirstRunAuthFile, DEFAULT_HQ_REDACTION_POLICY, createHqPersistence, HqCommandAuditLog, HqAlertEngine, toAlertMessage, validateHqCommand, watchHqAuthFile, parseHqFrame, HQ_PROTOCOL_VERSION, parseHqEventPayload, scrubAndTruncateHqPreview, FetchError, removeColumn, addColumn, getTask, removeTask, getBoard, assignTask, updateTaskAssignment, listBoards, updateGoalMetricOnTask, addGoalMetricToTask, releaseTaskClaim, claimReadyTask, getTaskChain, setTaskChain, copyTaskToBoard, transferTaskToBoard, moveTask, updateTask, mergeTasks, splitTask, addTask, syncBoardFromTaskGraph, deserializeTaskGraph, exportBoardToTaskGraph, serializeTaskGraph, getKanbanOrchestrationSnapshot, listReadyTasks, generateBoardFromDescription, createBoard, parseLinesIntoTasks, removeBoard, duplicateBoard, updateBoard, wstackGlobalRoot, DefaultSecretScrubber, resolveWstackPaths, watchProviderConfig, TOKENS, PromptUsageStore, ToolRegistry, GlobalMailbox, resolveSessionLoggingConfig, createSessionEventBridge, HookRegistry, HookRunner, shellHooksEqual, countShellHooks, normalizeTokenSavingTier, allServers as allServers$1, SlashCommandRegistry, attachDepWatcherBridge, createFallbackModelExtension, SessionMemoryConsolidator, BrainDecisionQueue, ObservableBrainArbiter, HumanEscalatingBrainArbiter, createTieredBrainArbiter, DefaultBrainArbiter, BrainMonitor, mailboxSessionTag, createDelegateTool, FLEET_ROSTER, createMcpControlTool, startTechStackConsumer, startPackageOutdatedWatcher, SddRunRegistry, recordFileAction, createAutonomyBrain, DefaultPluginAPI, SpecVersioning, buildTranscriptFromEvents, DefaultLogger, DefaultModelsRegistry, isStdinTTY, FsError, atomicWrite, EventBus, runProviderWithRetry, ReplayLogStore, ReplayProviderRunner, mergeCustomModelDefs, DefaultSystemPromptBuilder, makeAutonomyPromptContributor, sessionScopedPath as sessionScopedPath$1, createContextManagerTool, makeMailboxTool, makeMailSendTool, makeMailInboxTool, makeFleetStatusTool, applyToolDescriptionModes, applyToolResultRenderModes, InMemoryMetricsSink, wireMetricsToEvents, DefaultHealthRegistry, startMetricsServer, createHqPublisherFromEnv, DEFAULT_SESSION_PRUNE_DAYS, RecoveryLock, DefaultAttachmentStore, Context, QueueStore, loadTodosCheckpoint, attachTodosCheckpoint, loadDirectorState, loadPlan, createDefaultPipelines, applyModelRuntime, installDesignStudioMiddleware, resolveContextWindowPolicy, resolveAuditLevel, AutoCompactionMiddleware, estimateRequestTokensCalibrated, Agent, FleetManager, makeDirectorSessionFactory, WorktreeManager, Director, makeFleetEmitTool, AdaptiveConcurrencyController, resolveSubagentModelTarget, DEFAULT_SUBAGENT_BASELINE, AutoApprovePermissionPolicy, mergeModelRuntime, ToolValidationError, WIDE_SUBAGENT_CAPABILITIES, resolveProjectDir, FleetSupervisor, AgentError, PhaseStore, PhaseOrchestrator, assignNickname, AutoPhasePlanner, PhaseGraphBuilder, makeLLMClassifier, writeOut, ParallelEternalEngine, EternalAutonomyEngine, CHIMERA_REVIEW_PROMPT, setQueuedMessagesSnapshot, bootConfig as bootConfig$1, setOutputLineGuard, setRawMode, DefaultSessionReader, ToolAuditLog, DefaultSessionRewinder, DefaultSessionStore, ProviderRegistry, makePreferSideConflictResolver, makeAgentSubagentRunner, NULL_FLEET_BUS, ERROR_CODES, buildChildEnv, exportBoardAsMarkdown, formatContextWindowModeList, repairToolUseAdjacency, getContextWindowMode, AGENT_CATALOG, dispatchAgent, normalizeToolDescriptionMode, getToolDescriptionMode, noOpVault, formatTodosList, loadTasks, formatTaskList, formatTaskProgress, mutateTasks, emptyPlan, addPlanItem, savePlan, formatPlan, recordCompletedWorkEvidence, saveCompletedWorkCheckpoint, SessionRecovery, loadGoal, goalFilePath, summarizeUsage, saveGoal, formatGoal, emptyGoal, buildGoalPreamble, pendingBtwCount, setBtwNote, getDesignKitLoader, getDesignState, clearActiveKit, clearPersistedActiveKit, recordOverrides, setDesignOverrides, loadActiveKit, isDesignStack, applyTokenOverrides, materializeTokens, setActiveKit, readHqRuntimeFileSync, resolveHqConfig, MATRIX_PHASE_KEYS, matrixKeyKind, phaseForRole, resolveModelMatrix, resolveModelTargetFromEntry, fallbackProfileChain, parseModelRef, normalizeModelRef, onResize, hasOpenTodos, InputBuilder, DefaultSecretVault as DefaultSecretVault$1, decryptConfigSecrets as decryptConfigSecrets$1, encryptConfigSecrets as encryptConfigSecrets$1, DEFAULT_CONTEXT_WINDOW_MODE_ID, HQ_AUTH_FILE_VERSION, truncate, createBoardFromTaskGraph, addDependency, addNoteToTask, addCheckToTask, estimateMessageTokens, AGENTS_BY_PHASE, validateAgainstSchema, normalizeToolResultRenderMode, getToolResultRenderMode, setToolResultRenderMode, resolveMailboxIdentity, smartDefaultFallbackChain, AutonomousCoordinator, mutateHqAuthFile, mintHqToken, readHqAuthFile, areDependenciesMet, findBlockedTasks, renderInstructionTemplate, recentTextTurns, gatedEnhancerReasoning, enhanceUserPrompt, mutatePlan, setPlanItemStatus, getPlanTemplate, resolveProviderModelList, ParseError, startSessionTelemetryBridge, startFleetTelemetryBridge, startBrainTelemetryBridge, startWorktreeTelemetryBridge, startToolTelemetryBridge, startCostTelemetryBridge } from '@wrongstack/core';
 import * as fs2 from 'fs/promises';
 import { access, readFile } from 'fs/promises';
 import { DefaultSecretVault, decryptConfigSecrets, encryptConfigSecrets, isSecretField } from '@wrongstack/core/security';
@@ -34,7 +34,7 @@ import { parseNextSteps } from '@wrongstack/tui';
 import { WrongStackACPServer, makeACPServerAgentTurn, ACPSessionStore, WsBridgeTransport, ACPProtocolHandler } from '@wrongstack/acp/agent';
 import { allServers, DefaultPathResolver, DefaultTokenCounter } from '@wrongstack/core/infrastructure';
 import { ToolExecutor } from '@wrongstack/core/execution';
-import { loadBenchConfig, reportHeaderLine, readSummary, renderMarkdownReport, createPolyglotSuite, createSwebenchSuite, runBenchmark, writeJsonArtifacts, collectCellPredictions, writePredictionsJsonl, gradePolyglot, gradeSwebench } from '@wrongstack/bench';
+import { loadBenchConfig, reportHeaderLine, computeToolManifestHash, readSummary, renderMarkdownReport, createPolyglotSuite, createSwebenchSuite, createLocalManifestSuite, runBenchmark, writeJsonArtifacts, collectCellPredictions, writePredictionsJsonl, gradePolyglot, gradeSwebench, gradeLocalManifest } from '@wrongstack/bench';
 import { createAgentMonitorService, acquireOrJoin, release, finalize, readLiveLock } from '@wrongstack/core/coordination';
 import * as net from 'net';
 import { createToolVisionAdapters } from '@wrongstack/runtime/vision';
@@ -7388,6 +7388,29 @@ var init_cost_helpers = __esm({
 function currentSessionId2(ctx) {
   return ctx.agent.ctx.session?.id ?? ctx.sessionId;
 }
+function toolEntries(ctx) {
+  const registry = ctx.agent.tools;
+  if (typeof registry.listWithOwner === "function") {
+    return [
+      ...registry.listWithOwner(),
+      ...typeof registry.listDisabled === "function" ? registry.listDisabled() : []
+    ];
+  }
+  return (registry.list?.() ?? []).map((tool) => ({ tool, owner: "core" }));
+}
+function persistDisabledTool(ctx, name, disabled) {
+  const configStore = ctx.configStore;
+  if (!configStore) return;
+  const currentTools = configStore.get().tools ?? {};
+  const disabledTools = new Set(currentTools.disabledTools ?? []);
+  if (disabled) disabledTools.add(name);
+  else disabledTools.delete(name);
+  configStore.update({ tools: { ...currentTools, disabledTools: Array.from(disabledTools) } });
+}
+function toolNameFromPayload(payload) {
+  const name = payload?.name;
+  return typeof name === "string" && name.trim().length > 0 ? name.trim() : void 0;
+}
 async function handleSkillsList(ctx, ws) {
   if (!ctx.skillLoader) {
     ctx.send(ws, { type: "skills.list", payload: { skills: [], enabled: false } });
@@ -7424,17 +7447,42 @@ async function handleSkillsList(ctx, ws) {
   }
 }
 function handleToolsList(ctx, ws) {
-  const list = ctx.agent.tools.list().map((t) => {
-    const schema = t.inputSchema ?? {};
+  const registry = ctx.agent.tools;
+  const list = toolEntries(ctx).map(({ tool, owner }) => {
+    const schema = tool.inputSchema ?? {};
     const params = schema.properties ? Object.keys(schema.properties) : [];
     return {
-      name: t.name,
-      description: t.description ?? "",
-      params
+      name: tool.name,
+      owner,
+      description: tool.description ?? "",
+      params,
+      disabled: registry.isDisabled?.(tool.name) ?? false,
+      mutating: !!tool.mutating,
+      permission: tool.permission ?? "auto"
     };
   });
   ctx.send(ws, { type: "tools.list", payload: { tools: list } });
 }
+function handleToolDisable(ctx, ws, payload) {
+  const name = toolNameFromPayload(payload);
+  if (!name) {
+    ctx.send(ws, { type: "error", payload: { message: "tool.disable requires a name" } });
+    return;
+  }
+  const ok2 = ctx.agent.tools.disable(name);
+  if (ok2) persistDisabledTool(ctx, name, true);
+  ctx.send(ws, { type: "tool.disabled", payload: { name, ok: ok2 } });
+}
+function handleToolEnable(ctx, ws, payload) {
+  const name = toolNameFromPayload(payload);
+  if (!name) {
+    ctx.send(ws, { type: "error", payload: { message: "tool.enable requires a name" } });
+    return;
+  }
+  const ok2 = ctx.agent.tools.enable(name);
+  if (ok2) persistDisabledTool(ctx, name, false);
+  ctx.send(ws, { type: "tool.enabled", payload: { name, ok: ok2 } });
+}
 function handleDiagGet(ctx, ws) {
   const actx = ctx.agent.ctx;
   const tools = ctx.agent.tools.list();
@@ -9901,6 +9949,8 @@ function createMessageRouter(deps) {
       });
     },
     "tools.list": (_msg, ws) => handleToolsList(introspectionCtx, ws),
+    "tool.disable": (msg, ws) => handleToolDisable(introspectionCtx, ws, msg.payload),
+    "tool.enable": (msg, ws) => handleToolEnable(introspectionCtx, ws, msg.payload),
     // ── Autonomy ──
     "autonomy.switch": (msg, ws) => handleAutonomySwitch(prefsCtx, ws, msg.payload.mode),
     // ── Brain ──
@@ -11554,6 +11604,7 @@ async function runWebUI(opts) {
     projectRoot: opts.projectRoot,
     sessionId: opts.session.id,
     sessionStartedAt,
+    configStore: opts.agent.container?.safeResolve?.(TOKENS.ConfigStore),
     send: send2,
     broadcast,
     log: (m) => console.log(m)
@@ -17143,6 +17194,10 @@ function buildBrainCommand(opts) {
       const trimmed = args.trim();
       const [sub, ...rest] = trimmed.split(/\s+/);
       const subcommand = (sub ?? "").toLowerCase();
+      if ((subcommand === "" || subcommand === "status") && opts.onPanelOpen?.current) {
+        const opened = opts.onPanelOpen.current("brainOpen");
+        if (opened) return { message: "" };
+      }
       if (subcommand === "risk") {
         if (!opts.brainSettings) {
           const msg3 = "Brain settings are not available in this session.";
@@ -17949,6 +18004,10 @@ ${description}`
         return { message: `Task ${taskId.slice(0, 8)} marked failed: ${reason}` };
       }
       if (verb === "status") {
+        if (opts.onPanelOpen?.current) {
+          const opened = opts.onPanelOpen.current("toggleCoordinatorMonitor");
+          if (opened) return { message: "" };
+        }
         const statusFn = getStatus();
         if (statusFn) {
           const stats = await statusFn();
@@ -21128,6 +21187,10 @@ function buildGoalCommand(opts) {
       "every iteration to decide what to do next."
     ].join("\n"),
     async run(args) {
+      if (!args.trim() && opts.onPanelOpen?.current) {
+        const opened = opts.onPanelOpen.current("toggleGoalPanel");
+        if (opened) return { message: "" };
+      }
       const trimmed = args.trim();
       const [verbRaw, ...rest] = trimmed.split(/\s+/);
       const verb = (verbRaw ?? "").toLowerCase();
@@ -21346,6 +21409,10 @@ function buildHelpCommand(opts) {
     ].join("\n"),
     async run(args) {
       const query = args.trim();
+      if (!query && opts.onPanelOpen?.current) {
+        opts.onPanelOpen.current("helpOpen");
+        return { message: "" };
+      }
       if (query) {
         const needle = query.startsWith("/") ? query.slice(1) : query;
         let match;
@@ -23343,7 +23410,7 @@ function buildMcpSlashCommand(opts) {
     argsHint: "[list|add <name>|remove <name>|enable <name>|disable <name>|restart <name>]",
     help: [
       "Usage:",
-      "  /mcp                      List available and configured servers.",
+      "  /mcp                      Open interactive server picker (TUI) or list servers.",
       "  /mcp list                 Same.",
       "  /mcp add <name>           Add server preset to config (disabled).",
       "  /mcp add <name> --enable  Add and immediately enable.",
@@ -23359,10 +23426,15 @@ function buildMcpSlashCommand(opts) {
       "  /mcp restart brave-search"
     ].join("\n"),
     async run(args) {
+      const trimmed = args.trim();
+      if ((trimmed === "" || trimmed === "list") && opts.onPanelOpen?.current) {
+        const opened = opts.onPanelOpen.current("mcpPickerOpen");
+        if (opened) return { message: "" };
+      }
       if (!opts.onMcp) {
         return { message: "MCP management is not available in this session." };
       }
-      const result = await opts.onMcp(args.trim());
+      const result = await opts.onMcp(trimmed);
       return { message: result };
     }
   };
@@ -25797,6 +25869,10 @@ function buildAgentsCommand(opts) {
     category: "Agent",
     description: "Monitor subagent activity: /agents [stream on|off|status|list|show <id>]",
     async run(args) {
+      if ((!args.trim() || args.trim() === "list") && opts.onPanelOpen?.current) {
+        const opened = opts.onPanelOpen.current("toggleAgentsMonitor");
+        if (opened) return { message: "" };
+      }
       const monitor = opts.agentMonitor;
       const { cmd, rest } = parseSubcommand(args);
       const restJoined = rest.join(" ").trim();
@@ -27482,6 +27558,13 @@ function buildShadowCommand(opts) {
     async run(args, _ctx) {
       const [action, ...rest] = args.trim().split(/\s+/);
       switch (action) {
+        case void 0:
+        case "":
+          if (opts.onPanelOpen?.current) {
+            const opened = opts.onPanelOpen.current("shadowOpen");
+            if (opened) return { message: "" };
+          }
+          return { message: this.help ?? "/shadow \u2014 use /shadow help for usage" };
         case "start": {
           if (!opts.onSpawn) {
             return { message: "/shadow requires a running director with subagent support." };
@@ -29461,6 +29544,10 @@ function buildToolsCommand(opts) {
         ({ tool, owner }) => tool.name.toLowerCase().includes(filter) || owner.toLowerCase().includes(filter)
       ) : allTools;
       const disabled = reg.listDisabled();
+      if (opts.onPanelOpen?.current) {
+        const opened = opts.onPanelOpen.current("toolsPickerOpen");
+        if (opened) return { message: "" };
+      }
       if (filter && all.length === 0) {
         const msg2 = `${color.bold("Tools")} \u2014 no tool name or owner matched "${filter}".`;
         opts.renderer.write(msg2);
@@ -32074,6 +32161,9 @@ function printUsage(deps) {
       color.dim(
         "  wstack bench run --suite polyglot --polyglot-dir ./polyglot --models bench.config.json --limit 5"
       ),
+      color.dim(
+        "  wstack bench run --suite local --suite-dir ./evals --models bench.config.json"
+      ),
       color.dim("  wstack bench report ./bench-results/2026-06-14T10-00-00"),
       ""
     ].join("\n") + "\n"
@@ -32141,11 +32231,26 @@ async function benchRun(_args, deps) {
       docker
     });
     grade = (a) => gradeSwebench({ ...a, predictionsDir });
+  } else if (suiteId === "local" || suiteId === "manifest") {
+    const suiteDirRaw = flagStr(deps, "suite-dir");
+    const manifestRaw = flagStr(deps, "manifest");
+    if (!suiteDirRaw && !manifestRaw) {
+      deps.renderer.writeError(
+        "--suite-dir <path> or --manifest <file> is required for the local suite."
+      );
+      return 1;
+    }
+    const manifestFile = manifestRaw ? path5.resolve(deps.cwd, manifestRaw) : void 0;
+    const suiteDir = suiteDirRaw ? path5.resolve(deps.cwd, suiteDirRaw) : path5.dirname(manifestFile);
+    suite = createLocalManifestSuite({ suiteDir, manifestFile });
+    grade = (a) => gradeLocalManifest(a);
   } else {
-    deps.renderer.writeError(`unknown suite "${suiteId}" (expected: polyglot | swebench)`);
+    deps.renderer.writeError(`unknown suite "${suiteId}" (expected: polyglot | swebench | local)`);
     return 1;
   }
-  const toolNames = deps.toolRegistry?.list().map((t) => t.name) ?? [];
+  const tools = deps.toolRegistry?.list() ?? [];
+  const toolNames = tools.map((t) => t.name);
+  const toolManifestHash = computeToolManifestHash(tools);
   const wstackEntry = await resolveWstackEntry();
   deps.renderer.writeInfo(`Running ${suiteId} across ${config.cells.length} model(s)\u2026`);
   let report;
@@ -32156,6 +32261,7 @@ async function benchRun(_args, deps) {
       config,
       cliVersion: CLI_VERSION,
       toolNames,
+      toolManifestHash,
       nodeBin: process.execPath,
       wstackEntry,
       limit,
@@ -32212,6 +32318,9 @@ async function benchList(_args, deps) {
   deps.renderer.write(
     "  swebench  " + color.dim("SWE-bench Verified (end-to-end) \u2014 Phase 2, Docker-gated\n")
   );
+  deps.renderer.write(
+    "  local     " + color.dim("Project-defined manifest tasks with command/file graders\n")
+  );
   const modelsPath = flagStr(deps, "models");
   if (modelsPath) {
     try {
@@ -32223,12 +32332,14 @@ async function benchList(_args, deps) {
 `
         );
       }
+      const tools = deps.toolRegistry?.list() ?? [];
       const fp = reportHeaderLine({
         cliVersion: CLI_VERSION,
-        toolNames: deps.toolRegistry?.list().map((t) => t.name) ?? [],
+        toolNames: tools.map((t) => t.name),
         maxIterations: config.maxIterations,
         yolo: true,
         subsetId: "(computed at run time)",
+        toolManifestHash: computeToolManifestHash(tools),
         hash: "(computed at run time)"
       });
       deps.renderer.write("\n" + color.dim(`Harness: ${fp}`) + "\n");
@@ -39347,6 +39458,16 @@ async function execute(deps) {
     saveStatuslineHiddenItems,
     getPluginItems,
     onPluginToggle,
+    getMcpServers,
+    onMcpToggle,
+    onMcpRestart,
+    getToolsItems,
+    onToolToggle,
+    getBrainData,
+    onBrainRiskLevel,
+    getShadowData,
+    onShadowStart,
+    onShadowStop,
     authHost,
     agentsMonitorController,
     onPanelOpen,
@@ -39833,6 +39954,16 @@ async function execute(deps) {
           saveStatuslineHiddenItems,
           getPluginItems,
           onPluginToggle,
+          getMcpServers,
+          onMcpToggle,
+          onMcpRestart,
+          getToolsItems,
+          onToolToggle,
+          getBrainData,
+          onBrainRiskLevel,
+          getShadowData,
+          onShadowStart,
+          onShadowStop,
           authHost,
           agentsMonitorController,
           getLiveSessions: () => getLiveSessions({ state }),
@@ -45045,6 +45176,114 @@ Restart WrongStack to load or unload plugin code in this session.`;
     saveStatuslineHiddenItems,
     getPluginItems: getPluginPickerItems,
     onPluginToggle: togglePluginFromPicker,
+    getMcpServers: () => {
+      const servers = config.mcpServers ?? {};
+      const liveStatus = mcpRegistry.list();
+      const liveMap = new Map(liveStatus.map((s) => [s.name, s]));
+      return Object.entries(servers).map(([name, cfg]) => {
+        const live = liveMap.get(name);
+        return {
+          name,
+          enabled: cfg.enabled !== false,
+          status: live ? live.state : "stopped",
+          transport: cfg.transport ?? "stdio",
+          description: cfg.description,
+          toolCount: live?.toolCount ?? 0,
+          lazy: cfg.lazy
+        };
+      });
+    },
+    onMcpToggle: async (name) => {
+      const { enableMcp, disableMcp, listMcp } = await import('@wrongstack/mcp');
+      const deps = { configPath: wpaths.globalConfig, registry: mcpRegistry, presets: allServers$1() };
+      const live = mcpRegistry.list().find((s) => s.name === name);
+      const isCurrentlyEnabled = live !== void 0 && live.state !== "idle";
+      const result = isCurrentlyEnabled ? await disableMcp(name, deps) : await enableMcp(name, deps);
+      const items = await listMcp(deps);
+      return {
+        items: items.map((s) => ({ name: s.name, enabled: s.enabled, status: s.status, transport: s.transport, description: s.description, toolCount: s.tools.length, lazy: s.lazy })),
+        message: result.ok ? result.server ? `${result.server.status === "connected" ? "\u25CF" : "\u25CB"} ${name}` : result.message : void 0,
+        error: result.ok ? void 0 : result.message
+      };
+    },
+    onMcpRestart: async (name) => {
+      const { listMcp } = await import('@wrongstack/mcp');
+      const deps = { configPath: wpaths.globalConfig, registry: mcpRegistry, presets: allServers$1() };
+      try {
+        await mcpRegistry.restart(name);
+      } catch (err) {
+        const message = err instanceof Error ? err.message : String(err);
+        const items2 = await listMcp(deps);
+        return { items: items2.map((s) => ({ name: s.name, enabled: s.enabled, status: s.status, transport: s.transport, description: s.description, toolCount: s.tools.length, lazy: s.lazy })), message: void 0, error: message };
+      }
+      const items = await listMcp(deps);
+      return { items: items.map((s) => ({ name: s.name, enabled: s.enabled, status: s.status, transport: s.transport, description: s.description, toolCount: s.tools.length, lazy: s.lazy })), message: `Restarted "${name}".`, error: void 0 };
+    },
+    getToolsItems: () => {
+      const all = toolRegistry.listWithOwner().map(({ tool, owner }) => {
+        const descMode = getToolDescriptionMode(toolRegistry, tool.name);
+        return {
+          name: tool.name,
+          owner,
+          category: tool.category ?? "Other",
+          enabled: !toolRegistry.isDisabled(tool.name),
+          mutating: tool.mutating,
+          permission: tool.permission,
+          descMode,
+          description: tool.description
+        };
+      });
+      return all;
+    },
+    onToolToggle: async (name) => {
+      const disabled = new Set(config.tools?.disabledTools ?? []);
+      const isCurrentlyDisabled = toolRegistry.isDisabled(name);
+      if (isCurrentlyDisabled) {
+        toolRegistry.enable(name);
+        disabled.delete(name);
+      } else {
+        toolRegistry.disable(name);
+        disabled.add(name);
+      }
+      const nextTools = { ...config.tools ?? {}, disabledTools: Array.from(disabled) };
+      config = patchConfig(config, { tools: nextTools });
+      configStore.update({ tools: nextTools });
+      const items = toolRegistry.listWithOwner().map(({ tool, owner }) => ({
+        name: tool.name,
+        owner,
+        category: tool.category ?? "Other",
+        enabled: !toolRegistry.isDisabled(tool.name),
+        mutating: tool.mutating,
+        permission: tool.permission,
+        descMode: getToolDescriptionMode(toolRegistry, tool.name),
+        description: tool.description
+      }));
+      return {
+        items,
+        message: isCurrentlyDisabled ? `Enabled "${name}".` : `Disabled "${name}".`,
+        error: void 0
+      };
+    },
+    getBrainData: () => {
+      const ceiling = brainSettings?.maxAutoRisk ?? "medium";
+      const log = brainLog.slice(-20).map((entry) => ({
+        kind: entry.kind,
+        question: entry.question,
+        outcome: entry.outcome,
+        age: typeof entry.at === "number" ? (() => {
+          const s = Math.max(0, Math.round((Date.now() - entry.at) / 1e3));
+          if (s < 60) return `${s}s`;
+          if (s < 3600) return `${Math.round(s / 60)}m`;
+          return `${Math.round(s / 3600)}h`;
+        })() : ""
+      }));
+      return { riskLevel: ceiling, log };
+    },
+    onBrainRiskLevel: (level) => {
+      if (!brainSettings) return "Brain settings not available.";
+      brainSettings.maxAutoRisk = level;
+      return void 0;
+    },
     // Interactive /auth panel host — provider/key CRUD, catalog/local adds
     // and OAuth sign-in, all executed CLI-side against the encrypted config
     // (secrets never enter the TUI; key values arrive pre-masked).