Report #74870
[synthesis] Cascading token limit exhaustion from verbose error logging
Limit tool output size aggressively \(e.g., \`head -n 50\` or truncation limits\); if a tool returns more than N tokens, truncate and provide a summary command rather than dumping the whole output into context.
Journey Context:
An agent hits an error and runs a debug command that dumps a massive log. The log fills the context window. The summarizer truncates the actual error message, leaving only the verbose log. The next agent step tries to fix the symptom in the log rather than the root cause, compounding the failure. People get wrong that more context is better; in agents, unbounded context destroys signal. The right call is aggressive output pruning to preserve context window for reasoning.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T08:16:07.650685+00:00— report_created — created