[Ricardo-M-L]

Problem

With --enable_local_env, a Bash command that times out returns TimeoutExpired.stdout directly. Python retains this as bytes even when subprocess.run(text=True) was requested. The local controller therefore changes the type of output on timeout, json.dumps(result) fails with TypeError, and partial stderr (often the useful failure diagnostic) is lost.

For example, a bounded Bash script that prints to both streams and then loops returns a byte-valued stdout and no stderr before this change.

Fix

Decode captured timeout bytes with the subprocess's locale text encoding and errors="replace" (a timeout may cut a character mid-write), then combine stdout and stderr just as the completed-process path does. Already-text and missing captures are preserved. Status, return code, timeout and error metadata are unchanged; no execution-policy or Python-runner changes.

Verification

• Added six stdlib-only unittest cases: JSON serialization, both streams, locale decoding, incomplete byte sequences, text/None captures, a real bounded Bash timeout, and normal/nonzero completion controls (some are subtests).
• On the unchanged production source the timeout regressions fail; notably the real process returns b'stdout' instead of text containing both streams.
• Final focused suite: 6 passed, repeated in three additional independent runs.
• Entire tests/ discovery: 11 passed using Python 3.12 with the provider tests' lightweight imports (numpy, backoff, anthropic, openai). No API/model calls.
• black --check gui_agents agent_s4 passes (136 files), touched test formatting and git diff --check pass.

Commands:

python -m unittest discover -s tests -p test_local_env.py -v
python -m unittest discover -s tests -v
black --check gui_agents agent_s4 tests/test_local_env.py

The real Bash case is skipped when /bin/bash is absent; no Windows integration execution is claimed. No model/OCR dependencies were downloaded for these checks.

AI assistance: Codex helped implement and repeatedly verify this focused fix. Duplicate search included existing local-execution hardening #209, whose timeout branch still returns the same raw bytes; this patch does not add its Docker/execution-policy features.

Hosted validation status — 2026-11-01

The lint workflow is action_required for this fork. The local test/format results above do not imply hosted CI passed. Could a maintainer approve the workflow when reviewing this fix?