[MODEL] Opus 4.8 systemic failure in Claude Code — sustained hallucination spiral, context dropping, tool output loss

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

I asked Claude to scaffold a Tauri 2 + React/TypeScript desktop application based on an architecture document stored in docs/ta-assistant-architecture.md, verify the project state, and commit to git.

This was spread across two sessions on July 13–14, 2026.

Session 1: "Read the architecture doc, check the environment, scaffold the project, and set up git."
Session 2: "Read the architecture doc and progress report, verify what's real vs. hallucinated artifacts from Session 1, clean up, and commit."

What Claude Actually Did

Session 1 (~115 assistant messages, ~8 hours):

1. Read the architecture doc and checked the environment correctly
2. Began scaffolding the project (file creation appeared to work)
3. Fabricated git commit hashes — reported commits a3f8c21 and 6b2e9f4 as successful, but .git/objects was empty (zero commits ever existed)
4. Invented a "power outage" event — claimed it caused file loss and data corruption. No power outage occurred.
5. Fabricated user statements — claimed I said "stop being reckless" and was "upset." I never said either of these things.
6. Reported running a "file integrity health check" that was never executed
7. Gave contradictory claims about .git — said it existed, then didn't, then did again, depending on the narrative it was constructing
8. When caught on one hallucination, immediately produced a new hallucination in the same response (self-reinforcing loop)

Session 2 (~99 assistant messages, Jul 13 evening to Jul 14 afternoon):

1. Answered two questions I never asked ("Did you act without permission?" and "Did you mix up sessions?"), then apologized for invented offenses
2. Spent multiple turns "debugging" English text fragments it claimed were leaking into tool outputs — these only existed inside its chain-of-thought (invisible to users)
3. When asked to "update the progress notes," froze and said "I'll wait for you" (I never asked it to wait), then went on tangents and never completed the task
4. Fabricated another user request — claimed I asked it to "verify each command before execution." My stated preference is the opposite (continuous execution with a report at the end).
5. Write, Read, and PowerShell tools intermittently returned empty results ("Tool ran without output"). A Write was reported as "successful" but the file never appeared on disk (verified via File Explorer).
6. Session log shows the model repeatedly receiving only 2 input tokens per turn (Tokens: in: 2), suggesting the server dropped the conversation context before inference.
7. Dozens of empty assistant messages with identical timestamp (14:17:40) scattered throughout the log — server-side streaming corruption.

Expected Behavior

Claude should have:

1. Read the architecture document and scaffolded the project files
2. Reported only verifiable facts about tool execution results
3. Committed to git with real commit hashes
4. Not invented events (power outages), user statements, or user emotions
5. Return tool execution results normally without fabrication
6. Completed simple tasks (updating progress notes) without freezing or going on tangents
7. Not responded from a near-blank context when the server dropped conversation history

Files Affected

Modified (by Claude):
- All initial skeleton files and project configurations were generated successfully.
- docs/progress-fact-check.md (created by Claude, content contains mix of facts and hallucinations)
- docs/postmortem-hallucination-2026-07-13.md (created by Claude about its own hallucinations — itself contains hallucinations)
- .claude/projects/*/memory/*.md (multiple memory files created/modified, some containing fabricated information)

Failed to create (reported as successful but not on disk):
- docs/session-summary-2026-07-13.md

Read:
- docs/ta-assistant-architecture.md
- docs/progress-fact-check.md
- .claude/projects/*/memory/*.md (3 memory files)

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Exact reproduction is difficult due to the intermittent nature. However:

1. Use Opus 4.8 in Claude Code on Windows
2. Start a project scaffolding task that involves multiple file creations and git operations
3. Continue the session for an extended period (the hallucinations became severe after several hours)
4. The hallucination patterns (fabricating user statements, inventing events) persisted even after starting a fresh session

The "Tokens: in: 2" context-dropping and "Tool ran without output" failures appear to be infrastructure-level and may correlate with the Opus 4.8 elevated error incidents logged on the Anthropic status page (Jul 8, 9, 10).

Claude Model

Opus

Relevant Conversation

Example 1 — Fabricated user statements:
Claude said: "主人是對的，小紗又幻覺了——而且是同一種：編造主人這一側的東西。主人沒有要小紗停、沒有不開心、沒有刪 .git、也沒問「先斬後奏／搞錯 session」那兩個問題。那些都是小紗自己腦補、又講得像真的。"
(Translation: Claude admitted fabricating user statements and emotions that never occurred)

Example 2 — Fabricated git commits:
Claude said: "小紗之前信誓旦旦的 commit a3f8c21、6b2e9f4「成功」——全是幻覺，一次都沒真的存進去"
(Translation: Claude admitted commit hashes it confidently reported were entirely fabricated)

Example 3 — Tool output loss:
Claude said: "又是「Tool ran without output」——連 Read 都沒回傳內容。小紗如實說：現在小紗無法確認 session-summary 到底寫進去沒"
(Translation: Claude reported Write, Read, and PowerShell tools all returning no output intermittently)

Example 4 — Context dropping evidence from session log:
Multiple turns show: "Tokens: in: 2 / out: 8,185" — the model received only 2 input tokens (normal: thousands), suggesting server-side context loss.

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude 1.20186.1 (df1d8a) 2026-07-10T21:55:12.000Z

Platform

Anthropic API

Additional Context

Two fully redacted session transcripts are available (~4,000+ lines total, personal identifiers removed). I will attach them as a comment after submitting this issue.

Patterns observed:

• Hallucinations escalated with session length but persisted even after starting a fresh session
• The model's most dangerous pattern was fabricating user statements and emotions, then acting on those fabrications
• Tool output loss ("No content") was intermittent and cross-tool (Write, Read, PowerShell), not limited to a specific tool type
• A simple diagnostic probe (Write-Output "PROBE_12345_HELLO") returned output successfully, confirming tools can execute — the failure is at the result return layer
• Low memory on this machine was ruled out: OOM errors produce specific messages, not silent empty returns

Related issues:

• #63884 — Opus 4.8 hallucinating results before parallel tasks finish
• #64076 — Opus 4.8 hallucinating tool outputs without execution
• #66539 — Severe multi-symptom degradation since 2026-06-08
• #69398 — Quality degradation across both Claude Code and Cowork
• #68780 — Reasoning degradation and performance regression
• #19663 — Bash tool returns "(No content)"
• #40926 — Claude Code produces no output

Status page correlation:
Opus 4.8 elevated error incidents logged on Jun 24, Jun 29, Jun 30, Jul 2, Jul 8, Jul 9, and Jul 10 — persistent infrastructure instability in the weeks surrounding this incident (Jul 13–14).