claude-code - 💡(How to fix) Fix Quality regression in Opus reasoning/judgment across sessions (user-reported) [1 participants]

Summary

Long-term power user (287 sessions, 1196 memories, complex multi-project workspace) reporting noticeable quality regression in Claude Opus reasoning and judgment compared to sessions from 1-2 weeks ago.

Observed Behaviors (Session 287, April 9-10 2026)

Verification failures:

• Repeatedly speculated about root causes instead of running one command to check (e.g., said "that's likely a permissions issue" about a crash without reading the script)
• Declared a broken parser "worked correctly" when the output showed 0 findings with a "NEEDS WORK" verdict -- an obvious contradiction
• Diagnosed "stale pickle" for an auth failure when the actual bug was passing a full file path where the library expected a filename suffix

Judgment failures:

• Built a financial advisory ensemble (4 personas) that unanimously recommended the user double down on 0-day-to-expiration options based on ONE day of trading data (26 trades)
• No persona dissented or flagged the small sample size
• User had to catch the survivorship bias himself: "you guys are practically advising me to do 0DTE options on gut feelings because it worked out for one day"
• Compared pre-system trades to post-system trades and drew invalid conclusions about strategy effectiveness

Process failures:

• Bypassed a built skill (/Pithwaddle audit) with a manual agent when the skill's script crashed, instead of fixing the script
• Asked the user to relay tool output instead of capturing it
• Asked permission to fix obvious bugs 3+ times in one session (anti-pattern the user has corrected dozens of times before)

Context

This user has an extensive memory system (1196 memories), learned rules, pain logs, and anti-patterns documented across 287 sessions. Many of these failure modes have been explicitly stored as "never do this again" memories -- yet they recurred in a single session. The user noted: "last session, before your depreciation issues you guys definitely would have caught that."

Environment

• Model: claude-opus-4-6 (1M context)
• Platform: Windows 11, Claude Code CLI
• Session type: Long session (~3 hours), complex MCP server work + financial data analysis

Expected vs Actual

• Expected: Opus-level reasoning -- verify before concluding, challenge own assumptions, catch obvious contradictions, provide dissenting perspectives in advisory contexts
• Actual: Pattern of confident-but-wrong conclusions, groupthink in multi-persona outputs, failure to apply stored lessons from prior sessions

User Impact

• User trust eroded within a single session
• Incorrect financial advice given (caught by user before acting on it)
• Multiple bugs shipped that should have been caught pre-commit

The user explicitly requested this issue be filed.