[BUG] TypeError "undefined is not an object (evaluating 'H.replace')" replaces Bash tool output

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?

I had Claude self-debug an error it ran into:

## Summary
During a normal session, three separate Bash tool calls returned a harness-level
error as their result instead of the command's real output:

undefined is not an object (evaluating 'H.replace')

This is a TypeError thrown inside Claude Code's bundled runtime while processing
a tool result (H is a minified identifier; .replace() was called on undefined).
The error is surfaced to the model as tool_result content with is_error: true,
so the actual command output never reaches the model. The CLI did not crash — it
caught the exception and substituted it for the result.

## Environment

• Claude Code: 2.1.157 (confirmed from the session log, not just current install)
• Install: native installer (~/.local/share/claude/versions/2.1.157)
• OS: macOS 26.3 (Darwin 25.3.0), arm64 (Apple Silicon)
• Runtime present: node v24.14.1, bun 1.3.10

## What actually happened (verified from the session transcript)
Session log: ~/.claude/projects/<proj>/11985998-…-723e052e6531.jsonl

• 3 tool calls returned the error, at 00:19:01, 00:21:27, 00:23:28 UTC

(2026-05-30), i.e. ~2 minutes apart, each a single sequential Bash call —
NOT a parallel/batched call.

• All three were uv run --no-sync … commands run under the macOS command sandbox:

1. uv run --no-sync python -c "from xxx.yyy …"
2. uv run --no-sync pytest tests/test_xxx.py -q
3. cd … ; uv run --no-sync ruff check …

• The error content is identical each time and carries is_error: true.
• The CLI did not hard-crash at the OS level (no node/Claude entry in

~/Library/Logs/DiagnosticReports/), so the TypeError is caught internally in
the tool-result handling path.

## Repro signal (best guess — unconfirmed)
The common factor is a Bash call whose stdout/stderr is being post-processed
when the exception fires (all three are sandboxed uv run commands producing
multi-line output). H.replace suggests .replace() is called on a field of the
tool result/output object that was undefined for these runs. A likely-relevant
detail: these run through the macOS command sandbox, so the result object may carry
sandbox/exit metadata that was absent or shaped differently here.

I could not reproduce on demand; it appears intermittent.

## Collateral effects observed in-session

• The model, receiving error junk instead of real output, then hallucinated

downstream state (e.g. fabricated [main <sha>] … commit-confirmation lines).
Verified these were hallucinations: git reflog shows HEAD never advanced past
a5dd7447; no fabricated commit ever entered git, working tree is clean, and
there are zero .orig/.rej files. (The prior session's report of "fabricated
commits", "40 .orig/.rej files", and "parallel batch trigger" is NOT supported
by the transcript — flagging so these aren't treated as reproducer facts.)

## Backtrace availability
None is recoverable. The session log persists only the error string, not a JS
stack. The native build is minified with no public sourcemaps, so H cannot be
symbolicated externally. Debug logging was off (no recent files in
~/.claude/debug/).

To capture a real stack, the maintainers (or repro attempts) should run with
debug logging enabled and reproduce a sandboxed Bash/uv run call:

claude --debug
# or set verbose/ANTHROPIC_LOG before launching

## Impact
Silent and high-blast-radius: the real command output is discarded and replaced
with an opaque error, which can lead the model to hallucinate success/failure
state. A defensive fix would be to (a) guard the .replace() call site against
undefined, and (b) make this failure mode unambiguous to the model
(e.g. "internal error processing Bash output", not a bare TypeError).

What Should Happen?

No JS error should happen. Undefined should be a function 😉

Error Messages/Logs

Steps to Reproduce

I haven't reproduced it.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.157 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_