fix(redact): preserve code-like tool output

[zapabob]

Summary

• pass code_file=True when redacting execute_code and terminal output so code-like lines such as MAX_TOKENS=100 are not rewritten as secrets
• keep prefix-based secret masking active for actual leaked tokens
• add regression coverage for both terminal output and execute_code output

Context

This addresses the remaining code-output corruption reported in #33801. It is related to #33840, but includes behavior tests covering the user-visible regression.

Tests

• .\.venv\Scripts\python.exe -m ruff check tools\code_execution_tool.py tools\terminal_tool.py tests\tools\test_code_execution.py tests\tools\test_terminal_output_transform_hook.py
• .\.venv\Scripts\python.exe scripts\run_tests_parallel.py tests\tools\test_terminal_output_transform_hook.py tests\tools\test_code_execution.py tests\agent\test_redact.py -- --tb=short --timeout-method=thread -k redaction_preserves_code_like_env_names
• .\.venv\Scripts\python.exe -m ruff check .

Closes #33801

[alt-glitch]

Duplicate of #33840 — identical fix (pass code_file=True to redact_sensitive_text in execute_code and terminal). Both close #33801. This PR adds regression tests that #33840 lacks; consider merging the test coverage into #33840 or vice versa.

[zapabob]

Thanks, agreed that #34666 overlaps with #33840 on the implementation: both pass code_file=True to redact_sensitive_text for execute_code and terminal output.

I rechecked the current diff and the practical options are:

• merge fix(redact): preserve code-like tool output #34666 as the superset, since it has the same code change plus regression coverage; or
• if maintainers prefer fix: pass code_file=True to redact_sensitive_text in execute_code and terminal #33840, cherry-pick the two regression tests from fix(redact): preserve code-like tool output #34666 into fix: pass code_file=True to redact_sensitive_text in execute_code and terminal #33840:
  • tests/tools/test_code_execution.py::TestExecuteCodeEdgeCases::test_output_redaction_preserves_code_like_env_names
  • tests/tools/test_terminal_output_transform_hook.py::test_terminal_output_redaction_preserves_code_like_env_names

Fresh local validation on Windows:

• .\.venv\Scripts\python.exe -m ruff check tools\code_execution_tool.py tools\terminal_tool.py tests\tools\test_code_execution.py tests\tools\test_terminal_output_transform_hook.py -> passed
• .\.venv\Scripts\python.exe scripts\run_tests_parallel.py tests\tools\test_terminal_output_transform_hook.py tests\tools\test_code_execution.py tests\agent\test_redact.py -- --tb=short --timeout-method=thread -> 121 passed, 0 failed

So my recommendation is to keep/merge #34666 unless #33840 absorbs those tests.

[teknium1]

Closing as part of the #33801 secret-redaction-corrupts-code cluster, fixed at the root in #54061 (merged: 674e16e).

The real defect was _DB_CONNSTR_RE's password group matching greedily across newlines, corrupting displayed tool output (read_file/terminal/execute_code) for any source containing a postgresql:// f-string template. The fix forbids whitespace in the userinfo/password groups so the match can't span a line, preserves brace-only template passwords under code_file=True, and passes code_file=True at the terminal/execute_code output sites. Thanks for the contribution to this area.