Failed subagents leave transcript artifacts you can grep

When a Claude Code subagent dies with API Error 529 or hits its synthesis step on an overloaded API, the tool-call data it gathered is preserved in ~/.claude/projects/<proj>/subagents/agent-<id>.jsonl. The parent gets a warning not to Read the transcript (would overflow context) but you can safely grep specific patterns - URLs visited, search queries run, sizes per task. Recovered 11 verified profile URLs from a 45K-token subagent that returned only a mid-thought status snippet to the parent.