Claude Code (Opus 4.8) Silently Corrupts All Tool Calls — An Unfixed Bug Easy to Trigger in Japanese Environments

After running Claude Code on long, always-on sessions, you may suddenly find every tool call is broken. I hit this while running multiple sessions for hours in a Japanese environment. Here's what I found.

What happens

• Function call opening tags get corrupted into nonsense tokens like count / court
• Arguments leak into chat as raw text (filenames appear inline)
• MCP connection is fine (/mcp shows Connected) — not a tool/server issue
• It does not self-heal once it starts. The model imitates the corrupted output already in history as if it were the correct format (in-context self-reinforcing loop, not learning)
• The model itself does not notice the failure — it acts as if everything succeeded. Only a human watching from the outside catches it.

Trigger (observed)

GitHub issue anthropics/claude-code #63875 (OPEN, unfixed) reports the same symptoms, and the conditions match:

• Model: Opus 4.8 (does not occur on 4.7; the analogous 4.7 bug #61133 is closed/fixed)
• OS: Windows (issue carries platform:windows label)
• Long turns / accumulated context
• Non-ASCII (Japanese) characters in tool arguments

→ In short, "Opus 4.8 + Windows + Japanese + long-running" is the easiest path to hitting this.
Despite that, the English-language GitHub issue has almost no reports from Japanese users. Probably under-reported through the English channel rather than rare. Writing this for the next person stuck on the same thing.

Recovery (after corruption)

• /rewind (Esc twice also works) — roll back to a checkpoint before the corrupted turn. Lighter than /compact.
• /compact — replaces history with a summary, removing the corrupted examples (breaks the imitation loop).
• --continue makes it worse — restores raw history, which brings back the corrupted examples.
• For deep corruption (orphaned tool_result etc.) where the above don't reach, /clear (new session).

Prevention (root)

1. Downgrade to Opus 4.7

The bug is specific to 4.8. Switching to 4.7 (fixed) avoids it. Default in /config, per-session in /model. Capability gap is practically negligible for the version difference.

2. Decompose tasks (the real fix)

4.8's advantage shows up when you throw a single huge task at it. But that throw itself has a low success rate, and bumping the model isn't the solution.
If you decompose, 4.7 handles it reliably, and turns stay short and context stays light — exactly the conditions that don't trigger this bug.
= higher success + less corruption, two birds with one stone. Isolating heavy work into subagents keeps the main session's context clean.

3. Compact early + CLAUDE.md

CLAUDE.md is always loaded and survives summarization. Put critical constraints and decision rules in CLAUDE.md (not chat) and you can compact frequently without losing them.

Takeaway

"Brute-force with a stronger model" loses to "decompose and run on a stable model."
Until Anthropic ships a fix for the 4.8-specific bug, Opus 4.7 + task decomposition is enough. If you're running long sessions in Japanese and something feels off lately, this is the first thing to check.

---

Originally published in Japanese on Zenn.