Reference task queue

Task Queue

Location: forge/tasks/ - pending/ — new tasks waiting (JSON files matching tasks/schema.json) - active/ — currently being worked on by a Claude Code worker - completed/ — done with results - failed/ — tasks that errored or timed out (after max 2 retries)

Schema: forge/tasks/schema.json — required fields: name, model, machine, prompt, priority, budget_minutes, created_by, created_at

Dispatcher Script

Location: forge/scripts/dispatcher.sh

Modes: - ./scripts/dispatcher.sh — foreground (default) - ./scripts/dispatcher.sh --daemon — background via nohup - ./scripts/dispatcher.sh --once — single cycle, good for testing

What it does: - Polls tasks/pending/ every 30s - Sorts by priority (critical > high > normal > low) - Spawns tmux sessions named worker-{taskname}_{Model} - Writes prompts to /tmp/forge-tasks/task-{name}.txt (avoids tmux paste issues) - Sends /remote-control then instructs worker to read the prompt file - Monitors budget — warns at 80%, kills at limit - Retry logic: failed tasks re-queue up to MAX_RETRIES=2 before permanent failure - Max 5 concurrent workers - Logs to forge/logs/dispatcher.log - Handles SIGTERM/SIGINT gracefully

Dependencies: jq, tmux, claude CLI