worker-recovery

What problem does it solve?

In multi-agent Claude Code orchestration, stalled, killed, or misbehaving worker sessions are indistinguishable from successfully completed tasks until you verify their output artifacts, leading to lost work, broken workflows, and wasted compute when workers go idle without delivering required results.

Core Features & Use Cases

• Failure Taxonomy & Triage: Identifies common production-observed worker failure signatures from tmux pane scrollback, including rate limit cuts, usage policy false positives, ghost text hazards, stranded dispatches, and context window exhaustion.
• Targeted Recovery Playbooks: Provides step-by-step, production-tested recovery procedures for each failure type, from simple re-triggers for rate-limited workers to full TCC permission revival for macOS permission loss without losing session context.
• Prevention Guardrails: Includes proactive rules to avoid common failure modes, such as staggered dispatch to prevent first-token rate limit cuts, debounce settings for idle watchers, and resource-aware verification to avoid host OOM kills during concurrent builds.
• Use Case: Use this skill when a worker pane shows idle status without producing its expected report, PR, or build artifact, or when an orchestration watcher fires unexpectedly during a multi-worker workflow.

Quick Start

Invoke the worker-recovery skill when a Claude Code worker pane appears idle without producing its expected output to identify the failure type from the pane scrollback and execute the matching recovery steps.