Lieutenant/swarm output streaming drops all tool execution events — only text_delta captured

[wwaIII]

Bug: Lieutenant Output Shows Only LLM Text, Drops Tool Execution Progress

Component: vers-lieutenant.ts extension (also affects vers-swarm.ts)
Severity: High
Affects: All lieutenant and swarm orchestration workflows

Summary

vers_lt_read returns stale output while a lieutenant is actively executing tool calls because the event handler in installEventHandler() only captures text_delta events. All tool execution events (tool_execution_start/update/end) and tool call events (toolcall_start/delta/end) are silently dropped. This makes lieutenants appear "stuck" during tool-heavy work, causing the orchestrator to poll aggressively and waste tokens.

Root Cause

File: pi-v/extensions/vers-lieutenant.ts, lines ~717-718
Function: installEventHandler()

// Current code — only captures text_delta
handle.onEvent((event) => {
    if (event.type === "agent_start") {
        lt.status = "working";
        lt.lastOutput = "";
    } else if (event.type === "agent_end") {
        lt.status = "idle";
    } else if (event.type === "message_update" && event.assistantMessageEvent?.type === "text_delta") {
        lt.lastOutput += event.assistantMessageEvent.delta;  // ← ONLY this is captured
    }
    // ALL OTHER EVENTS SILENTLY DROPPED
});

Pi's RPC mode (rpc-mode.js) emits ALL events via session.subscribe((event) => output(event)) — they are streamed over the SSH tail -f connection and arrive at the extension. They're just ignored.

Same bug exists in vers-swarm.ts at lines ~830-831 and ~1292-1293.

Events Dropped

Event Type	Description	Captured?
agent_start	Agent begins	✅
agent_end	Agent finishes	✅
message_update + text_delta	LLM text streaming	✅
message_update + toolcall_start	LLM requests tool	❌ DROPPED
message_update + toolcall_delta	Tool call args streaming	❌ DROPPED
message_update + toolcall_end	Tool call complete	❌ DROPPED
tool_execution_start	Tool begins running	❌ DROPPED
tool_execution_update	Partial tool output	❌ DROPPED
tool_execution_end	Tool finished with result	❌ DROPPED
turn_start / turn_end	Turn lifecycle	❌ DROPPED

Incident Data

• Marketing lieutenant sent a task to create 7 files (ad copy, mockups)
• Lieutenant responded: "Let me bang all of this out right now. Creating everything in parallel." (71 chars)
• Orchestrator polled vers_lt_read 22 times over ~2 minutes
• Every read returned the same 71-char string
• Lieutenant was actually working — writing files via tool calls
• Output only updated ~3 minutes later when the LLM emitted more text in a new turn

Reproduction Steps

1. Create a lieutenant: vers_lt_create({ name: "test", ... })
2. Send a tool-heavy task: vers_lt_send({ name: "test", message: "Create 5 files with different content" })
3. Immediately poll: vers_lt_read({ name: "test" }) repeatedly
4. Observe: Output freezes after initial text, shows no tool execution progress

Proposed Fix

Extend the event handler to capture tool execution events:

handle.onEvent((event) => {
    if (event.type === "agent_start") {
        lt.status = "working";
        lt.lastOutput = "";
    } else if (event.type === "agent_end") {
        lt.status = "idle";
    } else if (event.type === "message_update") {
        const ame = event.assistantMessageEvent;
        if (ame?.type === "text_delta") {
            lt.lastOutput += ame.delta;
        } else if (ame?.type === "toolcall_end") {
            lt.lastOutput += `\n[calling ${ame.toolCall?.name}...]`;
        }
    } else if (event.type === "tool_execution_start") {
        lt.lastOutput += `\n[executing ${event.toolName}...]`;
        lt.lastActivityAt = new Date().toISOString();
    } else if (event.type === "tool_execution_end") {
        const status = event.isError ? "ERROR" : "done";
        lt.lastOutput += `\n[${event.toolName}: ${status}]`;
        lt.lastActivityAt = new Date().toISOString();
    }
});

Apply the same fix to vers-swarm.ts at both locations.

Impact

• Token waste: ~2,640+ tokens per incident from redundant polling
• Observability: Zero visibility into tool execution progress
• Reliability: Orchestrator may incorrectly assume lieutenant is stuck and take recovery actions
• UX: Operators watching fleet status see frozen output with no indication of progress

Metadata

• Discovered by: orchestrator + investigator lieutenant
• Date: 2026-02-24
• Confirmed via: Source code analysis of vers-lieutenant.ts, vers-swarm.ts, rpc-mode.js, pi-agent-core types
• Full investigation report: Available on request (12KB with architecture diagrams and code refs)