bug: Tool call activities created but never completed, accumulating as orphans

[vybe]

Summary

Tool call activities are created with STARTED state but never completed, causing them to accumulate as orphaned records. On a moderately active instance, this creates hundreds of orphan records requiring periodic cleanup. The issue is in the activity tracking logic, not backend restarts.

Component

Backend / Activity Service / Chat Router

Priority

P2 - Feature impaired (workaround: periodic cleanup script), data accumulates but no functionality loss

Error

Not an error per se, but database shows accumulation:

=== TOOL_CALL ACTIVITIES BY STATE ===
completed : 246
failed : 932  # These are all orphans marked by cleanup

=== CHAT_START ACTIVITIES BY STATE ===
completed : 410
failed : 45   # Normal failure rate

Tool calls have 21% completion vs chat_start at 90%.

Location

• File: src/backend/routers/chat.py
• Lines: ~200-220 (tool call tracking loop)
• Function: Chat message handler

Root Cause

In the chat endpoint, individual tool_call activities are created for each tool used:

# Each tool creates a STARTED activity
for tool_call in execution_log_simplified:
    await activity_service.track_activity(
        activity_type=ActivityType.TOOL_CALL,  # Created as STARTED
        parent_activity_id=chat_activity_id,
        ...
    )

# Only parent activity is completed
await activity_service.complete_activity(
    activity_id=chat_activity_id,  # NOT the tool_call activities!
    ...
)

track_activity() creates activities in STARTED state. The code only calls complete_activity() on the parent chat_activity_id, leaving all tool_call activities in STARTED state forever.

Reproduction Steps

1. Send a chat message to any agent
2. Agent responds using tools (e.g., Read, Write, Bash)
3. Check database: SELECT activity_state, COUNT(*) FROM agent_activities WHERE activity_type = 'tool_call' GROUP BY activity_state
4. Observe tool_call activities remain in 'started' state
5. Repeat - orphans accumulate with each response

Suggested Fix

Option A - Complete tool_call activities immediately (they're historical records):

for tool_call in execution_log_simplified:
    tool_activity_id = await activity_service.track_activity(
        agent_name=name,
        activity_type=ActivityType.TOOL_CALL,
        ...
    )
    # Complete immediately since this is a historical record
    await activity_service.complete_activity(
        activity_id=tool_activity_id,
        status="completed",
        details={"tool_name": tool_call.get("tool", "unknown")}
    )

Option B - Create with completed state directly:

# Add state parameter to track_activity() or use a new method
await activity_service.record_completed_activity(
    activity_type=ActivityType.TOOL_CALL,
    ...
)

Option C - Don't create individual tool_call activities at all (data already stored in chat message tool_calls JSON column).

Environment

• Trinity version: d34ca28
• Instance: Production with multiple active agents

Impact

• Database bloat: ~1000 orphaned records per week on active instance
• Requires periodic cleanup via /cleanup-orphans skill
• Cleanup marks them as failed with error message, which is misleading
• UI may show stale "in progress" activities

Related

• Activity service: src/backend/services/activity_service.py
• Activity tracking: track_activity() always creates with STARTED state
• Chat message already stores tool calls in tool_calls JSON column (possible duplicate data)

[vybe]

Fixed in commit - removed tool_call activity tracking from chat.py.

Solution: Removed the granular tool_call activity tracking loop (lines 311-327) since:

1. Tool calls are already stored in chat_messages.tool_calls JSON column
2. No frontend component queried agent_activities for tool_call type
3. The data was duplicate and creating orphan accumulation

The TOOL_CALL enum is preserved for querying historical data.