[Bug] claude-opus-4-6: Model behavior regression - hallucination and lost self-correction in code auditing

Bug Description
Bug Report: Claude Code v2.1.105 — Model Behavior Regression (Increased Hallucination / Loss of Self-Correction)
Version: Claude Code 2.1.105
Model: claude-opus-4-6 (max reasoning depth)
Regression Start: After v2.1.100

Description:
Since v2.1.100, using claude-opus-4-6 with max reasoning depth, the model has exhibited a significant behavioral regression in the following areas:

1. Superficial Code Auditing

The model no longer performs genuine code-level verification (file-by-file, function-by-function cross-checking). It relies on documents or diff surfaces to draw conclusions, resulting in outputs that contradict the actual state of the codebase.

1. Guessing Instead of Verifying

When faced with uncertain context, the model fills gaps with assumptions rather than actively retrieving ground truth via tool calls (Read File, Bash, etc.). This introduces bugs instead of fixing them.

1. Loss of Self-Correction

In prior versions, the model proactively caught its own errors before outputting. In v2.1.105, errors are only corrected after the user explicitly points them out — and even then, only the surface issue is patched without re-evaluating broader constraints.

1. Degraded Instruction Adherence

User-defined constraints set within a session (e.g., "produce only one file per turn", "align on format rules before writing") are repeatedly violated within the same session. After correction, the model re-violates the same constraint in the next step.

Steps to Reproduce:

Model: claude-opus-4-6
Reasoning depth: max
Task type: multi-file codebase refactoring + structured document generation
Issue is most pronounced in long-context sessions (500+ exchange work sessions)

Expected Behavior:

Audit steps should actively read source files rather than infer from memory or documentation
Session-level constraints should remain enforced throughout the conversation without repeated reminders
Model should self-identify when its conclusions conflict with actual file contents and pause to re-verify

Actual Behavior:

Audit conclusions are based on inference and guesswork rather than actual file reads
The same constraint requires repeated re-assertion before being respected
No proactive self-correction; every error requires the user to manually discover and explicitly flag it

Impact:
Multiple bugs have been introduced into the project codebase as a direct result of this regression. The reliability of the model as a coding agent has dropped substantially, shifting the burden of verification back onto the user and defeating the purpose of autonomous code review.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.104
• Feedback ID: 6901fa31-2108-4221-93a8-61920314cdd1

Errors

[{"error":"Error: NON-FATAL: Lock acquisition failed for /Users/fushihua/.local/share/claude/versions/2.1.104 (expected in multi-process scenarios)\n    at Tc_ (/$bunfs/root/src/entrypoints/cli.js:2836:2153)\n    at Se6 (/$bunfs/root/src/entrypoints/cli.js:2836:1233)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-13T00:33:14.722Z"},{"error":"Error: File does not exist. Note: your current working directory is /Users/fushihua/Desktop/CrabClaw.\n    at call (/$bunfs/root/src/entrypoints/cli.js:4458:7632)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-13T00:34:26.885Z"},{"error":"MaxFileReadTokenExceededError: File content (23893 tokens) exceeds maximum allowed tokens (10000). Use offset and limit parameters to read specific portions of the file, or search for specific content instead of reading the whole file.\n    at kY7 (/$bunfs/root/src/entrypoints/cli.js:4449:12671)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-13T00:34:53.982Z"},{"error":"MaxFileReadTokenExceededError: File content (12557 tokens) exceeds maximum allowed tokens (10000). Use offset and limit parameters to read specific portions of the file, or search for specific content instead of reading the whole file.\n    at kY7 (/$bunfs/root/src/entrypoints/cli.js:4449:12671)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-13T00:36:19.859Z"},{"error":"Error: EISDIR: illegal operation on a directory, read '/Users/fushihua/Desktop/CrabClaw/CrabCode/crates/acosmi-supervisor/src'\n    at G1H (/$bunfs/root/src/entrypoints/cli.js:3942:15399)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-13T00:36:29.205Z"},{"error":"MaxFileReadTokenExceededError: File content (31694 tokens) exceeds maximum allowed tokens (10000). Use offset and limit parameters to read specific portions of the file, or search for specific content instead of reading the whole file.\n    at kY7 (/$bunfs/root/src/entrypoints/cli.js:4449:12671)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-13T00:36:34.275Z"},{"error":"MaxFileReadTokenExceededError: File content (11287 tokens) exceeds maximum allowed tokens (10000). Use offset and limit parameters to read specific portions of the file, or search for specific content instead of reading the whole file.\n  …

Note: Content was truncated.