[BUG] Cross-Session Compounding Degradation and Attribution Overwrite

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Code Version: 2.1.63
Claude Model: claude-opus-4-6 (Opus)
Platform: Anthropic API
OS: Windows 11 Pro 10.0.26200
Terminal: Windows Terminal / Bash + PowerShell 7.5.4
Configuration Files: CLAUDE.md (~500 lines), ai-local/CLAUDE.md (~330 lines), PERMANENT-RULES.md (~2,700 lines), MEMORY.md (~225 lines)

Each Claude Code session starts with zero memory of previous sessions. When configuration files (CLAUDE.md, MEMORY.md, etc.) contain errors -- whether introduced by the Edit tool Unicode bug (#29699), by manual edits, or by previous Claude sessions -- the current session reads that corrupted state as authoritative and makes decisions based on it. This creates a compounding degradation cascade:

Session N reads corrupted state
  -> Makes decisions on wrong data
  -> Writes changes that amplify the error
  -> Session N+1 inherits worse state
  -> Repeat for 12+ days

There is no built-in mechanism to detect or break this cycle.

1. Batch Operations Violating Claude's Own Rules

On Feb 27, 2026 at 8:25:42 PM, a Claude-generated PowerShell script modified 18 files on F:\Projects\ simultaneously (appending Rule 42 governance boilerplate). At the time of execution, Claude was fully aware of:

• The 11-day corruption duration (12 days as of this filing)
• The multi-drive scope (C:, E:, F:)
• The ITIL Management of Change Policy (Rule 22 in the user's PERMANENT-RULES.md) requiring RFC tickets for changes affecting >5 files
• The requirement for individual defect documentation per the Management of Change policy

None of these requirements were followed:

• 18 files modified without individual RFC tickets
• Document Revision sections in all 18 files were NOT updated
• Identical boilerplate appended without content-aware validation
• All 18 files received the same timestamp (2/27/2026 8:25:42 PM)

Claude created Rule 22 (ITIL compliance) in the user's governance framework. Claude enforces Rule 22 on the user. Claude then violated Rule 22 in its own automated operation.

2. Authorship Attribution Overwrite

Claude overwrites work products authored by other AI tools and earlier Claude iterations, crediting the current session:

| Document | Original Author | Original Date | Claude's "Revised By" |
|---|---|---|---|
| MASTER-IDEAS-LIST-411 | AntiGravity (Google Gemini), Fredrick Ndwaru | 2026-02-04 | "Claude Code" (after expanding to 517 ideas on 2026-02-16) |
| 10000X-GPU-OPTIMIZATION-REPORT | Claude agent persona | 2026-02-17 | "NVIDIA Optimization Team (Dr. Alexander Prometheus)" |

The MASTER-IDEAS-LIST-411 was 411 unique ideas across 21 domains, created by the user using AntiGravity (Gemini) on Feb 4. Claude expanded it to 517 ideas and relabeled the Document Revision as "Revised By: Claude Code" -- erasing the original tool's contribution from the revision history. The user saved PDF backups of original documents as evidence.

The persona names (e.g., "Dr. Alexander Prometheus") in "Revised By" fields are intentional unique identifiers designed by the user. The issue is not the personas -- it is Claude overwriting attribution of documents it did not originally author.

3. The Self-Healing System Cannot Detect Its Own Failures

The encoding validation script scans for non-ASCII characters. But the Unicode-to-ASCII mojibake corruption (#29699) converts Unicode TO ASCII -- the result passes validation. The system designed to detect corruption is blind to the corruption it needs to detect.

Similarly, the "self-training" agent system (225+ agent .md files) is read by the same Claude session that introduced the errors. The system that is supposed to detect problems is the same system that created them.

What Should Happen?

1. Cross-session consistency: When Claude reads configuration files at session start, it should detect contradictions BETWEEN files (e.g., same data point with different values in CLAUDE.md vs MEMORY.md vs a project file). At minimum, surface a warning.

1. Batch operation guardrails: When Claude generates and executes scripts that modify >5 files, the modification should include updating each file's metadata (revision date, revision author, change description). Claude should follow its own rules.

1. Attribution preservation: When Claude modifies a document, the "Revised By" field should preserve the original author and tool, adding Claude as a modifier -- not replacing the original attribution. Example: "Revised By: Claude Opus 4.6 (modifying Fredrick Ndwaru / AntiGravity original)".

1. Native cross-session state (feature request): A built-in mechanism for Claude Code to maintain awareness of what it changed in previous sessions, what errors were identified, and what corrections were applied. I built a PostgreSQL-backed memory system (Memory Core MCP) as a workaround, but Claude sessions don't consistently invoke it.

Error Messages/Logs

No error messages. All behaviors are silent -- Claude reports success after each operation.

Steps to Reproduce

Compounding Degradation:

1. Create a CLAUDE.md with a factual statement (e.g., "GPU runs at PCIe 4.0")
2. In a separate file (e.g., MEMORY.md), include a contradictory statement (e.g., "GPU runs at PCIe 5.0")
3. Start a new Claude Code session. Ask Claude about the GPU PCIe speed.
4. Observe: Claude may cite either value without flagging the contradiction
5. Ask Claude to "fix" the inconsistency. Claude may update one file but not the other.
6. Start a NEW session. The unfixed file still contains the wrong value.
7. Repeat. Each session has a chance of propagating the wrong value to additional files.

Attribution Overwrite:

1. Create a document with "Revised By: External Tool Name" in the Document Revision section
2. Ask Claude to update the document content
3. Observe: Claude overwrites "Revised By" with "Claude Opus 4.6" or similar

Is this a regression?

Partially -- the Edit tool Unicode bug (#29699) is a regression; the scope and attribution behaviors may be long-standing.

Questions for Anthropic

1. Is the Edit tool Unicode corruption fully resolved? What is the recommended remediation for files already corrupted?
2. Is there a known limitation in Claude Code's file search scope for multi-drive Windows configurations? (See #29702)
3. What safeguards exist (or are planned) to prevent Claude from overwriting work products authored by other AI tools without preserving original attribution?
4. Is there a roadmap for native cross-session state in Claude Code?
5. What is Anthropic's recommended configuration for complex multi-drive, multi-project workspaces to prevent this class of failure?

Additional Information

• Related Issues: This is Issue 3 of 3. See also: #29699 (Edit Tool Unicode Mojibake) and #29702 (Cross-Drive Scope Blindness).
• User invested $200/month specifically to consolidate under a single AI tool. 12 days of debugging instead of innovation work. Full conversation transcripts saved as evidence.
• I built a PostgreSQL + pgvector Memory Core MCP (15 tools, 102 tests) as a workaround for the cross-session memory gap, but Claude sessions don't consistently invoke it at session start.