Repository navigation
[Bug] Claude Opus 4.6 reasoning regression: speculation loops instead of systematic debugging #33774
Closed as not planned
Description
Bug Description
Model quality regression — Claude Opus 4.6 in Claude Code (2026-03-12)
Experienced significant degradation in reasoning quality during a multi-hour debugging session. The task was straightforward: a V2 Excel parser wasn't extracting FDSO values despite the data being in the expected cell (C13).
Symptoms:
- Speculation loops instead of testing: Rather than downloading an actual production file and running the parser against it to see what the xlsx library returns, the model spent hours speculating about possible causes (wrong
columns, unicode characters, merged cells) with zero evidence. - Over-engineered fixes without evidence: Expanded the parser to scan additional columns (B/F) and added dozens of label aliases — when I told the model the data is consistently in C13/E13 across all files. Changes were made based
on imagination, not data. - Ignored project instructions repeatedly: The project's CLAUDE.md explicitly mandates test-driven development — "build validation tools FIRST, then iterate." The model never wrote a test that reproduces the failure, never
downloaded a real file to inspect, and never ran the parser against actual production data. - Required multiple user corrections: I had to intervene several times to point out the model was going in the wrong direction. Each correction was acknowledged but the underlying pattern (speculate → code → fail → speculate again)
continued. - Net output after hours of work: One commit adding diagnostic logging to warning messages. Zero progress on the actual fix.
This is a notable regression from typical Opus performance on this codebase. The model felt like it was operating with reduced reasoning depth — pattern-matching on surface-level hypotheses rather than methodically investigating the
problem.
Environment Info
- Platform: darwin
- Terminal: WarpTerminal
- Version: 2.1.47
- Feedback ID: 195af669-be9c-4c82-922d-6c8681564d81
Errors
[{"error":"Error: ENOENT: no such file or directory, open '/Users/adammay/.claude/remote-settings.json'\n at openSync (unknown)\n at readSync (/$bunfs/root/claude:12:1441)\n at K2 (/$bunfs/root/claude:6132:113)\n at k5 (/$bunfs/root/claude:6134:226)\n at r9A (/$bunfs/root/claude:178:8371)\n at tZA (/$bunfs/root/claude:2166:20976)\n at Ac7 (/$bunfs/root/claude:2166:22238)\n at <anonymous> (/$bunfs/root/claude:2166:22451)","timestamp":"2026-03-12T08:34:21.930Z"},{"error":"Error: ENOENT: no such file or directory, open '/Users/adammay/.claude/remote-settings.json'\n at openSync (unknown)\n at readSync (/$bunfs/root/claude:12:1441)\n at K2 (/$bunfs/root/claude:6132:113)\n at k5 (/$bunfs/root/claude:6134:226)\n at r9A (/$bunfs/root/claude:178:8371)\n at tZA (/$bunfs/root/claude:2166:20976)\n at Ac7 (/$bunfs/root/claude:2166:22238)\n at <anonymous> (/$bunfs/root/claude:2166:22451)","timestamp":"2026-03-12T09:34:21.955Z"},{"error":"Error: ENOENT: no such file or directory, open '/Users/adammay/.claude/remote-settings.json'\n at openSync (unknown)\n at readSync (/$bunfs/root/claude:12:1441)\n at K2 (/$bunfs/root/claude:6132:113)\n at k5 (/$bunfs/root/claude:6134:226)\n at r9A (/$bunfs/root/claude:178:8371)\n at tZA (/$bunfs/root/claude:2166:20976)\n at Ac7 (/$bunfs/root/claude:2166:22238)\n at <anonymous> (/$bunfs/root/claude:2166:22451)","timestamp":"2026-03-12T10:47:35.978Z"},{"error":"Error: 1P event logging: 1 events failed to export (code=ECONNREFUSED, ECONNREFUSED)\n at queueFailedEvents (/$bunfs/root/claude:1900:2350)\n at async doExport (/$bunfs/root/claude:1900:1257)\n at processTicksAndRejections (native:7:39)","timestamp":"2026-03-12T10:47:41.003Z"},{"error":"Error: {\"message\":\"Failed to export 1 events (code=ECONNREFUSED, ECONNREFUSED)\",\"originalLine\":\"1900\",\"originalColumn\":\"1409\",\"line\":\"1900\",\"column\":\"1409\"…Note: Content was truncated.
Activity
github-actions commentedon Mar 12, 2026
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
mehulparmariitr commentedon Mar 13, 2026
Same happening with me as well, it has become quite dumb. I had to switch to Codex with GPT 5.4 to quickly solve the issue!
SeanTAllen commentedon Mar 18, 2026
Very very very dumb
cjohnstonr commentedon Apr 12, 2026
Model if failing at just about every complex tasks. I have been using claude code exclusively for the last 9 months. I just did side by side comparisons of codex and opus 4.6...codex >>. What is going on? Anthropic, please respond publicly, you are losing our trust.
github-actions commentedon May 23, 2026
Closing for now — inactive for too long. Please open a new issue if this is still relevant.
github-actions commentedon Jul 14, 2026
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.