feat: add pre/post verification to task execution pipeline

- Before executing: verify if work is already done on server
- If already done: skip execution, comment evidence, mark complete
- After executing: run verification again to confirm work was applied
- Verification agent uses Pi with read-only investigation prompt
- Adds verification status (PASSED/COULD NOT CONFIRM) to Asana comments
- Prevents redundant work and catches false completions

pi-worker/src/pi/executor.ts

@@ -39,15 +39,16 @@ export class PiExecutor {
     });
 
     try {
+      // Don't pass ANTHROPIC_API_KEY — Pi uses OAuth from ~/.pi/agent/auth.json
+      // Passing the API key overrides OAuth and may hit a depleted prepaid balance
+      const { ANTHROPIC_API_KEY: _removed, ...cleanEnv } = process.env;
+
       const proc = spawn({
         cmd: args,
         cwd: this.config.targetProjectPath,
         stdout: "pipe",
         stderr: "pipe",
-        env: {
+        env: cleanEnv,
-          ...process.env,
-          ANTHROPIC_API_KEY: this.config.anthropicApiKey,
-        },
       });
 
       this.runningProcess = proc;
@@ -172,15 +173,15 @@ export class PiExecutor {
     logger.info("pi-executor", `Executing with agent-team: ${task.substring(0, 100)}...`);
 
     try {
+      // Don't pass ANTHROPIC_API_KEY — Pi uses OAuth from ~/.pi/agent/auth.json
+      const { ANTHROPIC_API_KEY: _removed, ...cleanEnv } = process.env;
+
       const proc = spawn({
         cmd: args,
         cwd: this.config.targetProjectPath,
         stdout: "pipe",
         stderr: "pipe",
-        env: {
+        env: cleanEnv,
-          ...process.env,
-          ANTHROPIC_API_KEY: this.config.anthropicApiKey,
-        },
       });
 
       this.runningProcess = proc;

pi-worker/src/scheduler/task-loop.ts

@@ -129,25 +129,62 @@ export class TaskLoop {
 
   private async executeTask(task: AsanaTask) {
     try {
-      // Move to "In Progress"
+      // STEP 1: Check if the work is already done
+      const alreadyDone = await this.verifyTaskAlreadyDone(task);
+      
+      if (alreadyDone.done) {
+        logger.info("task-loop", `Task already done: ${task.name}`, { evidence: alreadyDone.evidence });
+        
+        await this.asana.addComment(
+          task.gid,
+          `✅ Pi Worker verified this task is **already completed** on the server.\n\nEvidence:\n${alreadyDone.evidence}`
+        );
+
+        // Move to Done and mark complete
+        const doneSection = await this.findFirstSection(SECTIONS.DONE);
+        if (doneSection) {
+          await this.asana.moveTaskToSection(task.gid, doneSection.gid);
+        }
+        await this.asana.updateTask(task.gid, { completed: true });
+        logger.info("task-loop", `Task marked done (was already complete): ${task.name}`);
+        return;
+      }
+
+      // STEP 2: Move to "In Progress"
       const inProgressSection = await this.findFirstSection(SECTIONS.IN_PROGRESS);
       if (inProgressSection) {
         await this.asana.moveTaskToSection(task.gid, inProgressSection.gid);
       }
 
       // Add starting comment
-      await this.asana.addComment(task.gid, `🤖 Pi Worker picked up this task.\nStarting execution at ${new Date().toISOString()}`);
+      await this.asana.addComment(task.gid, `🤖 Pi Worker picked up this task.\nVerification: Not yet done on server.\nStarting execution at ${new Date().toISOString()}`);
 
       // Build context from task
       const context = this.buildTaskContext(task);
 
-      // Execute with Pi agent
+      // STEP 3: Execute with Pi agent
       const result = await this.pi.execute(task.name + (task.notes ? `\n\nDetails:\n${task.notes}` : ""), context);
 
+      // STEP 4: After execution, verify the work was actually done
+      let verified = false;
+      if (result.success && result.toolCalls > 0) {
+        const postCheck = await this.verifyTaskAlreadyDone(task);
+        verified = postCheck.done;
+        
+        if (!verified) {
+          logger.warn("task-loop", `Task executed but verification failed: ${task.name}`);
+        }
+      }
+
       // Add completion comment
-      const statusEmoji = result.success ? "✅" : "❌";
+      const statusEmoji = result.success ? (verified ? "✅" : "⚠️") : "❌";
+      const verificationNote = result.success 
+        ? (verified ? "Post-execution verification: PASSED" : "Post-execution verification: COULD NOT CONFIRM — please review manually")
+        : "";
+      
       const comment = [
         `${statusEmoji} Pi Worker ${result.success ? "completed" : "failed"} this task.`,
+        verificationNote,
         `Duration: ${(result.durationMs / 1000).toFixed(1)}s`,
         `Tool calls: ${result.toolCalls}`,
         result.error ? `Error: ${result.error}` : "",
@@ -161,13 +198,12 @@ export class TaskLoop {
       await this.asana.addComment(task.gid, comment);
 
       if (result.success) {
-        // Move to "Done" and mark complete
         const doneSection = await this.findFirstSection(SECTIONS.DONE);
         if (doneSection) {
           await this.asana.moveTaskToSection(task.gid, doneSection.gid);
         }
         await this.asana.updateTask(task.gid, { completed: true });
-        logger.info("task-loop", `Task completed: ${task.name}`);
+        logger.info("task-loop", `Task completed: ${task.name}`, { verified });
       } else {
         // Move back to To Do on failure
         const todoSection = await this.findFirstSection(SECTIONS.TODO);
@@ -184,6 +220,48 @@ export class TaskLoop {
     }
   }
 
+  private async verifyTaskAlreadyDone(task: AsanaTask): Promise<{ done: boolean; evidence: string }> {
+    try {
+      const verifyPrompt = `You are a verification agent. Your ONLY job is to check if the following task/bug/feature is ALREADY done on this server/codebase. Do NOT fix anything. Do NOT make any changes. ONLY investigate and report.
+
+Task: ${task.name}
+${task.notes ? `Details: ${task.notes}` : ""}
+
+INSTRUCTIONS:
+1. Use your tools (bash, read, grep, find, ls) to investigate if this work is already completed
+2. Check the codebase, running services, server state, logs — whatever is relevant
+3. Be thorough but quick
+
+RESPOND WITH EXACTLY THIS FORMAT (nothing else):
+DONE: true
+EVIDENCE: <one paragraph explaining what you found that proves it's done>
+
+OR:
+DONE: false
+EVIDENCE: <one paragraph explaining what's missing or broken>`;
+
+      const result = await this.pi.execute(verifyPrompt);
+
+      if (!result.success) {
+        return { done: false, evidence: "Verification check failed to run" };
+      }
+
+      const output = result.output.trim();
+      const doneMatch = output.match(/DONE:\s*(true|false)/i);
+      const evidenceMatch = output.match(/EVIDENCE:\s*(.+)/is);
+
+      const isDone = doneMatch ? doneMatch[1].toLowerCase() === "true" : false;
+      const evidence = evidenceMatch ? evidenceMatch[1].trim().substring(0, 1000) : output.substring(0, 1000);
+
+      logger.info("task-loop", `Verification result for "${task.name}": done=${isDone}`, { evidence: evidence.substring(0, 200) });
+
+      return { done: isDone, evidence };
+    } catch (error: any) {
+      logger.error("task-loop", `Verification error: ${error.message}`);
+      return { done: false, evidence: `Verification error: ${error.message}` };
+    }
+  }
+
   private buildTaskContext(task: AsanaTask): string {
     const parts = [
       `Asana Task: ${task.name}`,