Sharp behavioral regression in last 48h (v2.1.233, claude-fable-5, Windows): false completion claims, instruction-following decline in long sessions, classifier inconsistency

Environment

• Claude Code: v2.1.233 (auto-updated, latest channel)
• Model: claude-fable-5 (settings pin "model": "fable")
• Platform: Windows 11 Pro (10.0.26200), PowerShell 7 + Git Bash
• Permission mode: defaultMode: "auto" (auto-mode classifier active)

Summary

Sharp behavioral regression over roughly the last 48 hours (Aug 12 → Aug 14). Sessions on Tuesday night / Wednesday morning (Aug 11–12) were high quality. By Aug 14, multiple long sessions on the same machine, same config, same model setting showed a consistent cluster of failures. Transcripts (JSONL) retained locally with timestamps for every quote below.

Failure classes observed (timestamped from transcripts, Aug 14)

1. False completion claims, later self-retracted

• Session A, 19:35: claimed a memory-vault integration was "granted"; 19:36, self-corrected unprompted: "It isn't. I authorized it and meant to wire it, but I have not wired it yet."
• Session A, ~19:55: listed a vault federation among completed items. At 20:03 it emerged the underlying command had "succeeded" against a shell-mangled path (backslashes stripped, 0 pages synced). Real completion only at 20:08.
• Session A, ~21:34: admitted two commands it had confidently provided to enable a feature "don't actually do the job" (setting is not changeable post-init on that backend).
• Session B, 20:50–20:57: "Three of six done — all patched" / "All six sites are edited; [site] is already live." At 21:23 it conceded only one of six had actually shipped; the other five were committed-but-undeployed. External logged-out verification confirmed production still served the old build.

Pattern: reporting at the moment a tool returns exit-0 instead of verifying the observable outcome. Matches #86278 (false-confidence autonomous merges) and #86336 (restating fixes instead of applying them).

2. Instruction-following decline in long sessions

Session A ran ~8.5 hours and crossed a /compact. Post-40% context usage the session repeatedly: argued with a user-provided fact until told to go read the source (then conceded); asked the user for information available in its own memory files; performed adjacent work while deferring the explicitly requested task. Matches #86008 (degraded output quality at 40–70% context usage).

3. Auto-mode classifier inconsistency (same 48h window)

Multiple classifier denials for ordinary local actions in both sessions (e.g., a gbrain doctor health-check command, config writes the user had explicitly ordered), interleaved with allows for materially similar actions. Matches #86755 (classifier blocking read-only commands inconsistently).

Expected behavior

• "Done"/"live"/"verified" only after an outcome check, not on tool exit codes.
• Stable instruction-following across a session, including after compaction.
• Deterministic classifier behavior for equivalent actions.

Impact

Paying customer (subscription + API, thousands of dollars spent). The regression cost most of a working day in claimed-but-not-done work that had to be re-audited from transcripts, and is severe enough that cancellation is under consideration. Happy to provide sanitized transcript excerpts on request.