create-walle 0.9.19 → 0.9.21

package/template/wall-e/chat.js

@@ -31,6 +31,7 @@ const {
   inputForExternalActionEnvelope,
   reviewExternalAction,
 } = require('./external-action-controller');
+const { reviewExternalActionGateway } = require('./external-action-gateway');
 const { runShadow } = require('./eval/shadow');
 const {
   buildCodeReviewContextBlock,
@@ -984,7 +985,14 @@ function _externalActionLine(envelope = {}, result = {}) {
   const suffix = target ? `: ${target}` : '';
   if (result.alreadyExecuted) return `Already ${verb}${suffix}; I did not run it again.`;
   if (result.error) return `Failed to ${_externalActionVerb(envelope, 'present')}${suffix}: ${result.error}`;
-  return `Approved action ${verb}${suffix}.`;
+  if (_externalActionResultVerified(result)) return `Approved action ${verb} and verified${suffix}.`;
+  const verifyReason = result?.verification?.reason || result?.verification?.error || 'no read-after-write verification evidence was returned';
+  return `Approved action ${verb}${suffix}, but verification did not confirm it: ${verifyReason}`;
+}
+
+function _externalActionResultVerified(result = {}) {
+  if (!result || result.error || result.ok === false || result.success === false) return false;
+  return result.verified === true || result.verification?.ok === true || result.alreadyExecuted === true;
 }
 
 function _progressToolResultPayload(result, resultStr) {
@@ -1057,6 +1065,7 @@ async function _executeApprovedExternalActions({
     }
 
     const failed = !!(result && (result.error || result.ok === false || result.success === false));
+    const verified = _externalActionResultVerified(result || {});
     try {
       const confidence = require('./decision/confidence');
       if (envelope.domain && !result?.alreadyExecuted) confidence.recordAction(envelope.domain, !failed);
@@ -1071,7 +1080,9 @@ async function _executeApprovedExternalActions({
         ? 'Approved action was already executed'
         : failed
         ? 'Approved action failed'
-        : 'Approved action executed',
+        : verified
+        ? 'Approved action executed and verified'
+        : 'Approved action executed without verification',
       error: failed ? (result.error || 'External action failed') : null,
     });
     lines.push(_externalActionLine(envelope, result || {}));
@@ -2191,6 +2202,15 @@ async function chat(message, opts = {}) {
     name = normalizedCall.name;
     input = normalizedCall.input;
 
+    const gatewayReview = reviewExternalActionGateway({
+      toolName: name,
+      input,
+      userMessage: routingMessage || message,
+    });
+    if (!gatewayReview.admitted) {
+      return gatewayReview.result;
+    }
+
     // Eval hook: allow test harness to intercept tool calls with mock results
     if (typeof opts.toolInterceptor === 'function') {
       const intercepted = await opts.toolInterceptor(name, input);
@@ -2252,6 +2272,16 @@ async function chat(message, opts = {}) {
             args: name === 'run_shell' ? input.args : undefined,
             reason: perm.reason,
             source: perm.source,
+            approval_options: [
+              { id: 'allow_once', label: 'Yes', scope: 'this_tool_call' },
+              { id: 'allow_always', label: 'Always yes for this project/pattern', scope: 'project_permission_pattern' },
+              { id: 'deny', label: 'No', scope: 'this_tool_call' },
+            ],
+            approval_policy: {
+              kind: 'local_permission',
+              allow_always: true,
+              approval_scope: 'project_permission_pattern',
+            },
           });
           // Store resolver so server can call it when user responds
           if (!opts._permissionResolvers) opts._permissionResolvers = new Map();

package/template/wall-e/coding/stream-processor.js

@@ -12,6 +12,7 @@ const { recoverAllowedTextToolCalls } = require('../llm/text-tool-calls');
 const { ArtifactStore } = require('./artifact-store');
 const { normalizeResponse, transformRequest, providerId } = require('./provider-transform');
 const { Confidence, EventName, FailureClass, Provenance, createLaneEvent } = require('./lane-events');
+const { reviewExternalActionGateway } = require('../external-action-gateway');
 
 async function* streamFromChat(provider, request) {
   const response = normalizeResponse(recoverAllowedTextToolCalls(await provider.chat(request), request.tools), {
@@ -345,6 +346,41 @@ class StreamProcessor extends EventEmitter {
         name: call.name,
         data: { input: call.input },
       });
+      const gatewayReview = reviewExternalActionGateway({
+        toolName: call.name,
+        input: call.input,
+      });
+      if (!gatewayReview.admitted) {
+        const result = gatewayReview.result;
+        state.toolResults.push({ toolCallId: call.id, name: call.name, result });
+        await this._runtimeItem(sessionId, cwd, {
+          threadId: state.threadId,
+          turnId: state.turnId,
+          itemId: call.id,
+          role: 'tool',
+          status: 'completed',
+          name: call.name,
+          data: { result },
+        });
+        await this._record(sessionId, cwd, 'tool', {
+          state: 'blocked',
+          toolCallId: call.id,
+          name: call.name,
+          input: call.input,
+          result,
+        });
+        await this._laneEvent({
+          name: EventName.TOOL_FINISHED,
+          sessionId,
+          cwd,
+          provider: state.provider,
+          model: state.model,
+          runId: state.messageId,
+          confidence: Confidence.HIGH,
+          data: { toolCallId: call.id, name: call.name, blocked: true, reason: result.reason },
+        });
+        return;
+      }
       if (this.permissionService?.authorize) {
         await this._record(sessionId, cwd, 'tool', {
           state: 'permission_check',

package/template/wall-e/coding-orchestrator.js

@@ -565,6 +565,27 @@ function hasToolCall(toolCallHistory = [], names = new Set()) {
   return (toolCallHistory || []).some((call) => names.has(call.name));
 }
 
+function isVerificationToolCall(call = {}) {
+  const name = String(call.name || '');
+  const input = String(call.inputHash || JSON.stringify(call.input || {}));
+  if (name === 'browser_screenshot') return true;
+  if (name === 'run_shell') {
+    return /\b(?:test|spec|lint|build|typecheck|tsc|pytest|jest|mocha|vitest|playwright|node\s+--(?:test|check)|npm\s+(?:test|run)|pnpm\s+(?:test|run)|yarn\s+(?:test|run)|git\s+diff\s+--check)\b/i.test(input);
+  }
+  return /(?:test|verify|screenshot|diagnostic|lint|build)/i.test(name);
+}
+
+function hasVerificationEvidence(toolCallHistory = []) {
+  return (toolCallHistory || []).some(isVerificationToolCall);
+}
+
+function isVerificationBlockerResponse(content) {
+  const text = contentToText(content);
+  if (!text.trim()) return false;
+  return /\b(?:could not|couldn'?t|unable to|not able to|cannot)\b[\s\S]{0,120}\b(?:test|verify|build|run|execute)\b/i.test(text)
+    || /\b(?:tests?|verification|build)\b[\s\S]{0,120}\b(?:not run|not available|blocked|unavailable|missing)\b/i.test(text);
+}
+
 function stripPathLikeTokens(text) {
   return String(text || '').replace(/(?:^|[\s`'"(])((?:\.?[A-Za-z0-9_.@-]+\/)+[A-Za-z0-9_.@-]+)(?=[\s`'",):;.\]]|$)/g, ' ');
 }
@@ -610,6 +631,29 @@ function getNoActionContinuation({ prompt, content, toolCallHistory = [], mode,
   if (!isActionRequiredPrompt(prompt, { mode })) return null;
 
   const madeEdits = hasToolCall(toolCallHistory, EDIT_TOOL_NAMES);
+  if (madeEdits && !hasVerificationEvidence(toolCallHistory) && !isVerificationBlockerResponse(content)) {
+    const reason = 'The assistant made file changes but ended before running verification.';
+    if (!toolsAvailable) {
+      return {
+        action: 'fail',
+        reason: `${reason} No tool turns remain.`,
+      };
+    }
+    if (nudges >= maxNudges) {
+      return {
+        action: 'fail',
+        reason: `${reason} Verification continuation limit reached.`,
+      };
+    }
+    return {
+      action: 'continue',
+      reason,
+      message: `[SYSTEM] ${reason} This is not complete.\n` +
+        `Run the relevant verification now: tests, lint, build, typecheck, browser screenshot, or at minimum git diff --check when no project test exists.\n` +
+        `Only summarize success after a tool result proves the work. If verification is genuinely impossible, state the blocker with tool-backed evidence.\n` +
+        `Working directory: ${cwd}`,
+    };
+  }
   if (madeEdits) return null;
   if (isLegitimateNoEditResponse(content, toolCallHistory)) return null;
 
@@ -3294,6 +3338,7 @@ module.exports = {
   isActionRequiredPrompt,
   isPrematureActionResponse,
   getNoActionContinuation,
+  hasVerificationEvidence,
   subtaskRequiresFileChanges,
   screenshotTrackerHook,
   collectEmptyChangedFiles,

package/template/wall-e/deploy.sh

@@ -14,7 +14,7 @@ fi
 if ! fly apps list 2>/dev/null | grep -q my-wall-e; then
   echo "Creating Fly.io app..."
   fly apps create my-wall-e
-  fly volumes create wall_e_data --region sjc --size 1
+  fly volumes create wall_e_data --region sjc --size 5
 fi
 
 # Set secrets

package/template/wall-e/docs/external-action-controller.md

@@ -32,6 +32,38 @@ user confirmation, and the exact approved envelope is replayed back to Wall-E.
 Wall-E then executes the original payload directly rather than asking the model
 to recreate it.
 
+## Approval Tiers
+
+Wall-E uses two approval tiers:
+
+1. **Local permission approval** is for reversible project-local work such as
+   shell test commands, builds, and file edits. The UI may offer `Yes`,
+   `Always yes for this project/pattern`, and `No`. The persisted rule is scoped
+   to the project plus the permission pattern, never to vague model wording.
+2. **External action approval** is for real-world side effects: email, calendar,
+   Slack, SMS, reminders, and notifications. These approvals are exact-payload
+   approvals. The default choices are `Approve once` or `Cancel`; broad
+   `always yes` is intentionally disabled because the next payload may target a
+   different person, account, calendar, or time.
+
+This keeps the fast Claude Code-style flow for local coding work without giving
+models an ambient ability to send messages or schedule events.
+
+## Side-Effect Gateway
+
+All side-effecting routes pass through a host-side gateway before normal
+permission checks. The gateway blocks shell, AppleScript, and generic MCP calls
+that try to dispatch external actions directly, for example:
+
+- `osascript`/JXA creating Calendar events.
+- `gws calendar events insert`.
+- Gmail `messages.send` through shell, curl, or MCP.
+- Slack `chat.postMessage` through shell or MCP.
+
+The gateway returns a structured `external_action_gateway` tool result that
+tells the model which dedicated tool to use. The action is not executed, and the
+model is instructed not to claim success.
+
 Sequence:
 
 1. Wall-E blocks `mail_send`, `mail_reply`, `calendar_create`, and other external actions and
@@ -46,8 +78,9 @@ Sequence:
    to `chat()`.
 5. `chat()` validates that each approval id/hash still matches the reconstructed
    tool input, checks validation issues, applies a session-scoped idempotency
-   guard, executes the local tool, and emits normal `tool_call` / `tool_result`
-   progress events.
+   guard, executes the local tool, runs read-after-write verification where the
+   connector supports it, and emits normal `tool_call` / `tool_result` progress
+   events.
 
 This means provider wording is not part of the safety decision. DeepSeek,
 Anthropic, OpenAI, and other providers all use the same envelope replay path.
@@ -81,14 +114,39 @@ Anthropic, OpenAI, and other providers all use the same envelope replay path.
   approve mail/calendar side effects.
 - Approved envelopes are idempotent per Wall-E session and payload hash to avoid
   accidental duplicate sends from double-submit or retry.
+- Calendar approval envelopes preserve `account`, `source`, `calendarId`,
+  `calendar`, `location`, and time fields so replay cannot silently fall back to
+  a different provider or calendar.
+- Google Calendar creates verify with `calendar.events.get` against the same
+  account, calendar id, and event id before Wall-E can summarize the action as
+  verified.
+- Gmail sends and replies verify with `gmail.messages.get`. Replies keep the
+  original message id separately from the sent reply id so thread evidence is
+  not confused with source evidence.
+- Final summary guards treat `sent` or `created` without `verified` evidence as
+  incomplete. Shell stdout, AppleScript UIDs, or model prose do not count as
+  external-action completion evidence.
+
+## Coding-Agent Completion Contract
+
+Wall-E coding sessions use the same evidence rule. If an action-oriented coding
+prompt caused file edits, the agent must run a relevant verification tool before
+ending with a success summary. Accepted evidence includes tests, lint, build,
+typecheck, Playwright/browser screenshots for UI work, or `git diff --check`
+when no project-specific verifier exists. If verification is impossible, the
+agent must say so with tool-backed evidence instead of claiming success.
 
 ## Tests
 
 Focused regressions:
 
 - `claude-task-manager/tests/walle-permission-policy.test.js`
+- `claude-task-manager/tests/walle-external-actions.test.js`
 - `wall-e/tests/external-action-controller.test.js`
+- `wall-e/tests/external-action-gateway.test.js`
 - `wall-e/tests/local-tools-gws-live-files.test.js`
+- `wall-e/tests/coding-orchestrator.test.js`
+- `wall-e/tests/coding-stream-processor.test.js`
 - `wall-e/tests/execution-trace.test.js`
 - `wall-e/tests/chat.test.js` with `stages a draft email`
 

package/template/wall-e/external-action-controller.js

@@ -133,7 +133,13 @@ function targetForTool(toolName, input = {}) {
     return { channel: input.channel_name || input.channel || null };
   }
   if (toolName === 'calendar_create') {
-    return { calendar: input.calendar || null, attendees: normalizeAddressList(input.attendees) };
+    return {
+      calendar: input.calendar || input.calendar_name || input.calendarId || input.calendar_id || null,
+      calendarId: input.calendarId || input.calendar_id || null,
+      account: input.account || null,
+      source: input.source || input.provider || null,
+      attendees: normalizeAddressList(input.attendees),
+    };
   }
   if (toolName === 'reminder_create') {
     return { list: input.list || null };
@@ -163,6 +169,7 @@ function payloadForTool(toolName, input = {}) {
       title: input.title || '',
       start: calendarStart(input),
       end: calendarEnd(input),
+      location: input.location || '',
       notes: input.notes || '',
     };
   }
@@ -208,10 +215,14 @@ function inputForExternalActionEnvelope(envelope = {}) {
   if (toolName === 'calendar_create') {
     return {
       calendar: target.calendar || null,
+      calendarId: target.calendarId || null,
+      account: actor.account || target.account || null,
+      source: target.source || null,
       attendees: normalizeAddressList(target.attendees),
       title: payload.title || '',
       start_date: payload.start || null,
       end_date: payload.end || null,
+      location: payload.location || '',
       notes: payload.notes || '',
     };
   }
@@ -315,6 +326,17 @@ function buildBlockedToolResult(envelope, decision) {
       ? 'External action was staged because the user asked for prepared content, not dispatch.'
       : 'External action requires validation and explicit confirmation before execution.',
     action: envelope,
+    approval_options: isPreview
+      ? []
+      : [
+        { id: 'approve_once', label: 'Approve once', scope: 'exact_action_payload' },
+        { id: 'cancel', label: 'Cancel', scope: 'exact_action_payload' },
+      ],
+    approval_policy: {
+      kind: 'external_action',
+      allow_always: false,
+      approval_scope: 'exact_action_payload',
+    },
     draft: envelope.domain === 'email'
       ? {
         to: envelope.toolName === 'mail_reply' ? 'derived_from_original_message' : envelope.target.to,

package/template/wall-e/external-action-gateway.js

@@ -0,0 +1,163 @@
+'use strict';
+
+const SHELL_SIDE_EFFECT_RULES = [
+  {
+    domain: 'calendar',
+    operation: 'create_event',
+    recommendedTool: 'calendar_create',
+    reason: 'Calendar event creation must use calendar_create so account, calendar, approval, and verification are tracked.',
+    patterns: [
+      /\bgws\b[\s\S]*\bcalendar\b[\s\S]*\bevents?\b[\s\S]*\binsert\b/i,
+      /calendar\.googleapis\.com[\s\S]*\/events/i,
+      /osascript[\s\S]*(?:tell\s+application\s+"Calendar"|Application\(["']Calendar["']\))[\s\S]*(?:make\s+new\s+event|events\.push|new\s+event\s+with\s+properties)/i,
+      /(?:tell\s+application\s+"Calendar"|Application\(["']Calendar["']\))[\s\S]*(?:make\s+new\s+event|events\.push|new\s+event\s+with\s+properties)/i,
+    ],
+  },
+  {
+    domain: 'email',
+    operation: 'send_email',
+    recommendedTool: 'mail_send or mail_reply',
+    reason: 'Email dispatch must use mail_send or mail_reply so recipients, threading, approval, and sent-mail verification are tracked.',
+    patterns: [
+      /\bgws\b[\s\S]*\bgmail\b[\s\S]*(?:\+send|messages\s+send|users\s+messages\s+send)\b/i,
+      /gmail\.googleapis\.com[\s\S]*\/messages\/send\b/i,
+      /osascript[\s\S]*(?:tell\s+application\s+"Mail"|Application\(["']Mail["']\))[\s\S]*(?:make\s+new\s+outgoing\s+message|send\b)/i,
+      /(?:tell\s+application\s+"Mail"|Application\(["']Mail["']\))[\s\S]*(?:make\s+new\s+outgoing\s+message|send\b)/i,
+    ],
+  },
+  {
+    domain: 'slack',
+    operation: 'send_message',
+    recommendedTool: 'slack_send_message',
+    reason: 'Slack message dispatch must use slack_send_message so channel, approval, and execution evidence are tracked.',
+    patterns: [
+      /slack\.com\/api\/chat\.postMessage/i,
+      /\bslack\b[\s\S]*\bchat\.postMessage\b/i,
+    ],
+  },
+  {
+    domain: 'reminder',
+    operation: 'create_reminder',
+    recommendedTool: 'reminder_create',
+    reason: 'Reminder creation must use reminder_create so approval and execution evidence are tracked.',
+    patterns: [
+      /osascript[\s\S]*(?:tell\s+application\s+"Reminders"|Application\(["']Reminders["']\))[\s\S]*(?:make\s+new\s+reminder|new\s+reminder)/i,
+      /(?:tell\s+application\s+"Reminders"|Application\(["']Reminders["']\))[\s\S]*(?:make\s+new\s+reminder|new\s+reminder)/i,
+    ],
+  },
+];
+
+const MCP_SIDE_EFFECT_RULES = [
+  {
+    domain: 'calendar',
+    operation: 'create_event',
+    recommendedTool: 'calendar_create',
+    pattern: /(?:calendar|google[-_ ]?calendar).*?(?:create|insert|add|schedule)|(?:create|insert|add|schedule).*?(?:calendar|event)/i,
+  },
+  {
+    domain: 'email',
+    operation: 'send_email',
+    recommendedTool: 'mail_send or mail_reply',
+    pattern: /(?:gmail|mail|email).*?(?:send|reply|respond)|(?:send|reply|respond).*?(?:gmail|mail|email)/i,
+  },
+  {
+    domain: 'slack',
+    operation: 'send_message',
+    recommendedTool: 'slack_send_message',
+    pattern: /(?:slack).*?(?:send|post|message)|(?:send|post).*?(?:slack|channel)/i,
+  },
+];
+
+function commandText(input = {}) {
+  if (!input || typeof input !== 'object') return '';
+  if (typeof input.command === 'string') {
+    const args = Array.isArray(input.args) && input.args.length ? ` ${input.args.join(' ')}` : '';
+    return `${input.command}${args}`;
+  }
+  if (typeof input.script === 'string') return input.script;
+  return '';
+}
+
+function classifyExternalActionBypass(toolName, input = {}) {
+  const name = String(toolName || '').trim();
+  if (name === 'run_shell' || name === 'shell' || name === 'bash' || name === 'terminal' || name === 'applescript') {
+    const text = commandText(input);
+    if (!text) return null;
+    for (const rule of SHELL_SIDE_EFFECT_RULES) {
+      if (rule.patterns.some((pattern) => pattern.test(text))) {
+        return {
+          toolName: name,
+          domain: rule.domain,
+          operation: rule.operation,
+          recommendedTool: rule.recommendedTool,
+          reason: rule.reason,
+          evidence: text.slice(0, 500),
+        };
+      }
+    }
+  }
+
+  if (name === 'mcp_call' || name.startsWith('mcp__')) {
+    const callName = name === 'mcp_call'
+      ? `${input.server || ''}.${input.tool || ''}`
+      : name;
+    const text = `${callName} ${JSON.stringify(input.arguments || input.args || {})}`.slice(0, 1000);
+    for (const rule of MCP_SIDE_EFFECT_RULES) {
+      if (rule.pattern.test(text)) {
+        return {
+          toolName: name,
+          domain: rule.domain,
+          operation: rule.operation,
+          recommendedTool: rule.recommendedTool,
+          reason: `MCP side-effect tools must be routed through ${rule.recommendedTool} so approval and verification are tracked.`,
+          evidence: text.slice(0, 500),
+        };
+      }
+    }
+  }
+
+  return null;
+}
+
+function buildGatewayBlockedResult(finding) {
+  return {
+    external_action: true,
+    external_action_gateway: true,
+    blocked: true,
+    executed: false,
+    verified: false,
+    decision: 'dedicated_tool_required',
+    domain: finding.domain,
+    operation: finding.operation,
+    original_tool: finding.toolName,
+    recommended_tool: finding.recommendedTool,
+    reason: finding.reason,
+    evidence: finding.evidence,
+    approval_policy: {
+      kind: 'external_action',
+      allow_always: false,
+      approval_scope: 'exact_action_payload',
+    },
+    model_instruction: [
+      `Do not retry this ${finding.domain} side effect through ${finding.toolName}.`,
+      `Use ${finding.recommendedTool} instead so Wall-E can stage an exact approval envelope, execute the approved payload, and verify the result before summarizing success.`,
+      'Do not claim this action was executed.',
+    ].join(' '),
+  };
+}
+
+function reviewExternalActionGateway({ toolName, input = {} } = {}) {
+  const finding = classifyExternalActionBypass(toolName, input);
+  if (!finding) return { admitted: true };
+  return {
+    admitted: false,
+    finding,
+    result: buildGatewayBlockedResult(finding),
+  };
+}
+
+module.exports = {
+  buildGatewayBlockedResult,
+  classifyExternalActionBypass,
+  reviewExternalActionGateway,
+};

package/template/wall-e/fly.toml

@@ -20,6 +20,7 @@ primary_region = "sjc"  # Change to your nearest region: https://fly.io/docs/ref
   WALL_E_PORT = "3457"
   NODE_ENV = "production"
   WALLE_TELEMETRY_SERVER = "1"
+  WALLE_TELEMETRY_CLEANUP_ENABLED = "1"
 
 [[vm]]
   size = "shared-cpu-1x"