`Path.read_text()`/`write_text()` called without `encoding=` — 4 test failures on non-UTF-8 locales

Severity: Medium Type: Portability / Correctness Files: agent/integrations/local.py:32, agent/prompt.py:28, and 12 call sites under tests/

Description

Python's Path.read_text() and open() default to locale.getencoding(), not UTF-8. On Windows that is typically cp1252. The repository's source and Markdown files are UTF-8 and use non-ASCII characters liberally (em dashes, curly quotes, arrows), so any encoding-less read of a repo file fails on a non-UTF-8 locale.

Call sites missing encoding=:

agent/integrations/local.py:32    scoped.write_text(...)
agent/prompt.py:28                Path(DEFAULT_PROMPT_PATH).read_text().strip()   # see the separate DEFAULT_PROMPT_PATH issue
tests/agent/test_desktop.py:33,35,53
tests/agent/test_structured_input_hygiene.py:37
tests/e2e/fakes.py:165
tests/e2e/harness.py:608,613
tests/sandbox/test_langsmith_sandbox_config.py:55
tests/sandbox/test_local_integration.py:47,55,56
tests/tools/test_background_execute.py:76

Impact

Four tests fail on Windows purely because of this:

• tests/agent/test_structured_input_hygiene.py::test_no_module_hand_rolls_a_user_message
• tests/sandbox/test_langsmith_sandbox_config.py::test_nothing_deletes_sandboxes
• tests/sandbox/test_local_integration.py::test_create_local_sandbox_scopes_global_git_config
• (plus the test_desktop.py sites, which are latent)

Observed failure:

lines = _role_user_literals(ast.parse(path.read_text()))
UnicodeDecodeError: 'charmap' codec can't decode byte 0x8d in position 204:
character maps to <undefined>

The three failing tests are all repo-wide invariant guards — they walk agent/**/*.py and assert properties across every module (no hand-rolled user messages; nothing deletes sandboxes). They are exactly the checks you least want silently unavailable to a contributor, and they are the first thing to break on a non-UTF-8 developer machine.

Suggested fix

Pass encoding="utf-8" at every site. The codebase already does this correctly elsewhere (agent/prompt.py:32 uses .read_text(encoding="utf-8")), so this is an inconsistency rather than a new convention. A ruff rule (PLW1514 / flake8-encodings) would prevent regressions.

Found during a full-repo audit (ruff + basedpyright + pytest on a Windows host, plus manual review). Filing each finding separately so they can be triaged independently.