@sjd9021checks…fix/watchdog-concise-reasoning → next3 files · +5 −6updated 3mo agoClaude's BugAnalysis structured output was hitting the 4096 max_tokens limit when analyzing verbose error patterns (NOTION, AIRTABLE, Sheets). These tools have large JSON schemas/responses in their error payloads, causing Claude to generate extremely long reasoning fields that consumed most of the output token budget. This resulted in:
EOF while parsing)fixer_prompt field truncated/missing, so fixable bugs couldn't trigger the fixerreasoning field description in BugAnalysis model to cap at 3-5 sentencesagent-tools/analyze[processor] LLM returned is_fixable_bug=True but no fixer_prompt warnings — should drop to near zerowatchdog.fixer.triggered metric — should increase as truncation no longer suppresses fixer triggerswatchdog_analysis_logs — reasoning should be shorter but still accuratePR #1230 (Devin) already prevents hard crashes by catching LLM errors. This PR addresses the root cause — overly verbose output consuming the token budget.
🤖 Generated with Claude Code