Bash tool output with ANSI escape sequences causes context overflow and unrecoverable API errors

Description

When a Bash tool call captures large output containing ANSI escape sequences (e.g., colored log output from a subprocess), the raw escape sequences are stored in the context window without stripping. This can quickly consume ~35KB+ of context, trigger repeated API errors, and leave Claude Code completely unresponsive to any commands except /exit.

Steps to Reproduce

1. Run a command via the Bash tool that produces verbose colored stderr output (e.g., a Rust service with tracing producing DEBUG-level logs with ANSI color codes)
2. The command output contains hundreds of lines of ANSI-encoded log lines like:

``
[2m2026-02-17T18:25:15.784395Z[0m [34mDEBUG[0m [1;34menoch::job_runner[0m...
``

1. The Bash tool captures all of this (stdout + stderr) into the context
2. Context window fills up rapidly due to the verbose escape sequences

Expected Behavior

One or more of:

• The Bash tool should strip ANSI escape sequences from captured output before storing in context
• The Bash tool's 30,000 character truncation should be applied more aggressively when escape sequences are detected
• When context overflow triggers API errors, Claude Code should remain responsive (allow /compact, /clear, or other recovery commands — not just /exit)

Actual Behavior

• ANSI escape sequences are stored verbatim in the context window
• ~35KB of colored log output triggers context overflow
• Repeated API errors occur with no way to recover
• Claude Code becomes completely unresponsive to all commands except /exit
• The entire session's work is effectively lost (must start a new session)

Workaround

Redirect output to a file and read it separately:

./run.sh > /tmp/output.txt 2>&1; echo "Exit: $?"

Then use the Read tool to examine the file. This avoids the Bash tool capturing the large output directly.

Environment

• macOS (Darwin 24.6.0)
• Claude Code CLI
• Model: Claude Opus 4.6
• The subprocess producing the output was enoch (Rust binary using tracing with Bunyan-style colored JSON output)

Impact

This is particularly problematic for long-running sessions where significant work has been done. The context overflow makes the session unrecoverable, forcing a fresh start. The ANSI sequences add significant byte overhead compared to plain text — stripping them before storing in context would substantially reduce the risk.