@elizaos/plugin-agent-orchestrator 0.3.0 → 0.3.2

package/dist/index.js

@@ -115,7 +115,7 @@ function buildTurnCompletePrompt(taskCtx, turnOutput, decisionHistory) {
 ` + `Guidelines:\n` + `- BEFORE choosing "complete", enumerate each objective from the original task and verify ` + `evidence in the output. If ANY objective lacks evidence, use "respond" with the missing work.\n` + `- A PR being created does NOT mean the task is done \u2014 check that the PR covers ALL requested changes.
 ` + `- If the task mentions multiple features/fixes, verify EACH one is addressed, not just the first.\n` + `- If the agent only analyzed code or read files, it hasn't done the actual work yet \u2014 send a follow-up.
 ` + `- If the agent wrote code but didn't test it and testing seems appropriate, ask it to run tests.\n` + `- If the output shows errors or failed tests, send a follow-up to fix them.\n` + `- IMPORTANT: If the working directory is a git repository clone (not a scratch dir), the agent ` + `MUST commit its changes, push them, and create a pull request before the task can be "complete". ` + `If the output only shows code edits with no git commit or PR, respond with "Now commit your changes, push, and create a pull request".\n` + `- IMPORTANT: Creating a PR is NOT the final step. If this is the turn where the PR was created ` + `(i.e. "Created pull request" or a PR URL appears for the FIRST time and no previous decision ` + `already sent a review follow-up), respond with "Review your PR, run each test plan item to verify ` + `it works, update the PR to check off each item, then confirm all items pass".\n` + `- If a previous decision ALREADY sent a review/verification follow-up (check the decision history), ` + `and the agent has now responded with its review results, you MAY mark "complete" if the agent ` + `indicates the work is done (e.g. "Done", "verified", "all checks pass", "Here's what I did", ` + `or a clear summary of completed work). Do NOT require exact phrases \u2014 use judgment.
-` + `- Keep follow-up instructions concise and specific.\n` + `- Default to "respond" \u2014 only use "complete" when you're certain ALL work is done.
+` + `- Keep follow-up instructions concise and specific.\n` + `- When asking agents to verify work, prefer CLI tools (gh, curl, cat, git diff, etc.) over ` + `browser automation. Browser tools may not be available in headless environments and can cause delays.\n` + `- Default to "respond" \u2014 only use "complete" when you're certain ALL work is done.
 
 ` + `Respond with ONLY a JSON object:\n` + `{"action": "respond|complete|escalate|ignore", "response": "...", "useKeys": false, "keys": [], "reasoning": "..."}`;
 }
@@ -135,7 +135,7 @@ function buildTurnCompleteEventMessage(taskCtx, turnOutput, decisionHistory) {
 ` + `- "escalate" \u2014 something looks wrong, ask the user
 ` + `- "ignore" \u2014 should not normally be used here
 
-` + `Guidelines:\n` + `- Verify evidence for EVERY objective before using "complete".\n` + `- If code was written but not committed/pushed/PR'd, respond with next step.\n` + `- If a PR was just created, respond to review & verify test plan items.\n` + `- Default to "respond" \u2014 only "complete" when certain ALL work is done.
+` + `Guidelines:\n` + `- Verify evidence for EVERY objective before using "complete".\n` + `- If code was written but not committed/pushed/PR'd, respond with next step.\n` + `- If a PR was just created, respond to review & verify test plan items.\n` + `- When asking agents to verify work, prefer CLI tools (gh, curl, cat, etc.) over browser automation.\n` + `- Default to "respond" \u2014 only "complete" when certain ALL work is done.
 
 ` + `Include a JSON action block at the end of your response:\n` + "```json\n" + `{"action": "respond|complete|escalate|ignore", "response": "...", "useKeys": false, "keys": [], "reasoning": "..."}\n` + "```";
 }
@@ -425,7 +425,7 @@ async function executeDecision(ctx, sessionId, decision) {
       } catch {
       }
       ctx.sendChatMessage(summary ? `Finished "${taskCtx?.label ?? sessionId}".\n\n${summary}` : `Finished "${taskCtx?.label ?? sessionId}".`, "coding-agent");
-      ctx.ptyService.stopSession(sessionId).catch((err) => {
+      ctx.ptyService.stopSession(sessionId, true).catch((err) => {
         ctx.log(`Failed to stop session after LLM-detected completion: ${err}`);
       });
       checkAllTasksComplete(ctx);
@@ -469,7 +469,7 @@ async function handleBlocked(ctx, sessionId, taskCtx, data) {
       });
       ctx.sendChatMessage(`[${taskCtx.label}] WARNING: Auto-approved access to path outside workspace (${taskCtx.workdir}). ` + `Prompt: "${promptText.slice(0, 150)}". Stopping session for safety.`, "coding-agent");
       taskCtx.status = "error";
-      ctx.ptyService?.stopSession(sessionId).catch((err) => {
+      ctx.ptyService?.stopSession(sessionId, true).catch((err) => {
         ctx.log(`Failed to stop session after out-of-scope auto-approval: ${err}`);
       });
       return;
@@ -1980,6 +1980,14 @@ isBun,
 PTYManager,
 ShellAdapter
 } from "pty-manager";
+function forwardReadyAsTaskComplete(ctx, session) {
+  if (!ctx.hasActiveTask?.(session.id) || !ctx.hasTaskActivity?.(session.id)) {
+    return;
+  }
+  const response = ctx.taskResponseMarkers.has(session.id) ? captureTaskResponse(session.id, ctx.sessionOutputBuffers, ctx.taskResponseMarkers) : "";
+  ctx.log(`session_ready for active task ${session.id} \u2014 forwarding as task_complete (stall classifier path, response: ${response.length} chars)`);
+  ctx.emitEvent(session.id, "task_complete", { session, response });
+}
 async function initializePTYManager(ctx) {
   const usingBunWorker = isBun();
   if (usingBunWorker) {
@@ -1996,11 +2004,7 @@ async function initializePTYManager(ctx) {
     bunManager.on("session_ready", (session) => {
       ctx.log(`session_ready event received for ${session.id} (type: ${session.type}, status: ${session.status})`);
       ctx.emitEvent(session.id, "ready", { session });
-      if (ctx.hasActiveTask?.(session.id) && ctx.taskResponseMarkers.has(session.id)) {
-        const response = captureTaskResponse(session.id, ctx.sessionOutputBuffers, ctx.taskResponseMarkers);
-        ctx.log(`session_ready for active task ${session.id} \u2014 forwarding as task_complete (stall classifier path)`);
-        ctx.emitEvent(session.id, "task_complete", { session, response });
-      }
+      forwardReadyAsTaskComplete(ctx, session);
     });
     bunManager.on("session_exit", (id, code) => {
       ctx.emitEvent(id, "stopped", { reason: `exit code ${code}` });
@@ -2079,11 +2083,7 @@ async function initializePTYManager(ctx) {
   }
   nodeManager.on("session_ready", (session) => {
     ctx.emitEvent(session.id, "ready", { session });
-    if (ctx.hasActiveTask?.(session.id) && ctx.taskResponseMarkers.has(session.id)) {
-      const response = captureTaskResponse(session.id, ctx.sessionOutputBuffers, ctx.taskResponseMarkers);
-      ctx.log(`session_ready for active task ${session.id} \u2014 forwarding as task_complete (stall classifier path)`);
-      ctx.emitEvent(session.id, "task_complete", { session, response });
-    }
+    forwardReadyAsTaskComplete(ctx, session);
   });
   nodeManager.on("blocking_prompt", (session, promptInfo, autoResponded) => {
     ctx.emitEvent(session.id, "blocked", { promptInfo, autoResponded });
@@ -2151,26 +2151,40 @@ async function sendKeysToSession(ctx, sessionId, keys) {
     ptySession.sendKeys(keys);
   }
 }
-async function stopSession(ctx, sessionId, sessionMetadata, sessionWorkdirs, log) {
-  const session = ctx.manager.get(sessionId);
-  if (!session) {
-    throw new Error(`Session ${sessionId} not found`);
-  }
-  if (ctx.usingBunWorker) {
-    await ctx.manager.kill(sessionId);
-  } else {
-    await ctx.manager.stop(sessionId);
-  }
-  const unsubscribe = ctx.outputUnsubscribers.get(sessionId);
-  if (unsubscribe) {
-    unsubscribe();
+async function stopSession(ctx, sessionId, sessionMetadata, sessionWorkdirs, log, force = false) {
+  try {
+    const session = ctx.manager.get(sessionId);
+    if (!session) {
+      throw new Error(`Session ${sessionId} not found`);
+    }
+    if (ctx.usingBunWorker) {
+      if (force) {
+        await ctx.manager.kill(sessionId, "SIGKILL");
+      } else {
+        await ctx.manager.kill(sessionId);
+      }
+    } else {
+      if (force) {
+        await ctx.manager.stop(sessionId, { force: true });
+      } else {
+        await ctx.manager.stop(sessionId);
+      }
+    }
+  } finally {
+    try {
+      const unsubscribe = ctx.outputUnsubscribers.get(sessionId);
+      if (unsubscribe) {
+        unsubscribe();
+      }
+    } catch {
+    }
     ctx.outputUnsubscribers.delete(sessionId);
+    sessionMetadata.delete(sessionId);
+    sessionWorkdirs.delete(sessionId);
+    ctx.sessionOutputBuffers.delete(sessionId);
+    ctx.taskResponseMarkers.delete(sessionId);
+    log(`Stopped session ${sessionId}`);
   }
-  sessionMetadata.delete(sessionId);
-  sessionWorkdirs.delete(sessionId);
-  ctx.sessionOutputBuffers.delete(sessionId);
-  ctx.taskResponseMarkers.delete(sessionId);
-  log(`Stopped session ${sessionId}`);
 }
 function subscribeToOutput(ctx, sessionId, callback) {
   if (ctx.usingBunWorker) {
@@ -2579,7 +2593,7 @@ async function scanIdleSessions(ctx) {
       ctx.sendChatMessage(`[${taskCtx.label}] Session stopped \u2014 idle for ${idleMinutes} minutes with no progress.`, "coding-agent");
       if (ctx.ptyService) {
         try {
-          await ctx.ptyService.stopSession(taskCtx.sessionId);
+          await ctx.ptyService.stopSession(taskCtx.sessionId, true);
         } catch (err) {
           ctx.log(`Idle watchdog: failed to stop session ${taskCtx.sessionId}: ${err}`);
           taskCtx.status = "error";
@@ -3193,6 +3207,13 @@ class PTYService {
           return false;
         const taskCtx = coordinator.getTaskContext(sessionId);
         return taskCtx?.status === "active";
+      },
+      hasTaskActivity: (sessionId) => {
+        const coordinator = this.coordinator;
+        if (!coordinator)
+          return false;
+        const taskCtx = coordinator.getTaskContext(sessionId);
+        return (taskCtx?.decisions.length ?? 0) > 0;
       }
     });
     this.manager = result.manager;
@@ -3361,10 +3382,10 @@ class PTYService {
       throw new Error("PTYService not initialized");
     return sendKeysToSession(this.ioContext(), sessionId, keys);
   }
-  async stopSession(sessionId) {
+  async stopSession(sessionId, force = false) {
     if (!this.manager)
       throw new Error("PTYService not initialized");
-    return stopSession(this.ioContext(), sessionId, this.sessionMetadata, this.sessionWorkdirs, (msg) => this.log(msg));
+    return stopSession(this.ioContext(), sessionId, this.sessionMetadata, this.sessionWorkdirs, (msg) => this.log(msg), force);
   }
   get defaultApprovalPreset() {
     const fromEnv = this.runtime.getSetting("PARALLAX_DEFAULT_APPROVAL_PRESET");
@@ -3834,7 +3855,7 @@ function registerSessionEvents(ptyService, runtime, sessionId, label, scratchDir
             text: preview ? `Agent "${label}" completed the task.\n\n${preview}` : `Agent "${label}" completed the task.`
           });
         }
-        ptyService.stopSession(sessionId).catch((err) => {
+        ptyService.stopSession(sessionId, true).catch((err) => {
           logger4.warn(`[START_CODING_TASK] Failed to stop session for "${label}" after task complete: ${err}`);
         });
       }
@@ -6059,5 +6080,5 @@ export {
   CodingWorkspaceService
 };
 
-//# debugId=0100F6538A0A9D6264756E2164756E21
+//# debugId=BA5A6C5F800D868764756E2164756E21
 //# sourceMappingURL=index.js.map