Files

shokollm 54aa6419eb fix(kugetsu): prevent excess agent spawning with flock + sequential processing

- count_active_dev_sessions() now excludes pm-agent.json from count
- process_queue() now calls kugetsu start directly (not opencode run)
- process_queue() uses dynamic batch size = available_slots
- process_queue() has retry logic (max 3 attempts) on failure
- cmd_start() now uses flock around critical section
- Added notification types: task_queued, task_dequeued, task_started, task_completed, task_error
- Removed QUEUE_DAEMON_BATCH_SIZE config (no longer needed)

Fixes issue #146

2026-04-05 08:44:45 +00:00

2.6 KiB

Raw Permalink Blame History

Fix: Queue daemon spawning excess agents due to race condition

Problem

When enqueueing multiple tasks (e.g., 6 tasks), the queue daemon was spawning many more subagents than expected, eventually exhausting container memory.

Root Cause: The combination of:

process_queue() calling opencode run directly instead of kugetsu start, bypassing all concurrency logic
count_active_dev_sessions() counting pm-agent.json toward MAX_CONCURRENT_AGENTS, reducing effective dev agent slots
No atomic locking around session count check + session file creation (TOCTOU race condition)
Background spawning of multiple concurrent processes in process_queue()

Expected behavior: With MAX_CONCURRENT_AGENTS=3 and 6 tasks:

Tasks should be processed sequentially via kugetsu start
Only 3 dev agents should run at a time
Tasks should queue and wait for slots to free up

Solution

1. `count_active_dev_sessions()` - Exclude pm-agent

Only count actual dev agent session files (exclude pm-agent.json).

2. `process_queue()` - Call `kugetsu start` directly + retry logic

Call kugetsu start directly (foreground, sequential) instead of spawning opencode run background process
Dynamic batch size = available slots (removes need for QUEUE_DAEMON_BATCH_SIZE)
Retry logic (max 3 attempts) on failure
On failure: cleanup worktree/session and revert to pending state
Save fork_pid to queue item for timeout handling

3. `cmd_start()` - Add flock

Add flock around critical section (count check + fork)
Track fork_pid for queue item timeout handling

4. Notification System

New notification types:

Event	Type
Task enqueued	`task_queued`
Task dequeued	`task_dequeued`
Task started	`task_started`
Task completed	`task_completed`
Task error	`task_error`

5. Config

Remove QUEUE_DAEMON_BATCH_SIZE (no longer needed - batch size is now dynamic)

Notification Flow

Event	Location	Type
Task enqueued	`enqueue_task()`	`task_queued`
Task dequeued	`process_queue()` after state change to `notified`	`task_dequeued`
Task started	`cmd_start()` after session file created	`task_started`
Task completed	`update_queue_item_state()`	`task_completed`
Task error	`update_queue_item_state()`	`task_error`

Out of Scope

Re-check loop in cmd_start (checking if session DB is reliable) - deferred to separate research issue
Buffer mechanism for excess forking (safety failsafe only)

Status

Issue created
Implementation
PR created (#147)
Merged

2.6 KiB Raw Permalink Blame History