fix: isolate per-execution log buffers for parallel eval runs

[smflorentino]

Summary

Fixes two bugs that caused jumbled and truncated log output when UIPATH_JOB_KEY is set and eval runs execute in parallel.

Bug 1: Interleaved output

When eval runs N items in parallel, each item creates a child UiPathRuntimeLogsInterceptor via UiPathExecutionRuntime.execute(). The old code called _redirect_stdout_stderr() in child mode, which did:

sys.stdout = LoggerWriter(stdout_logger, ...)
sys.stderr = LoggerWriter(stderr_logger, ...)

sys.stdout is a single global. Each child overwrites it with a new LoggerWriter instance. If 4 eval items run concurrently:

• Child 1 sets sys.stdout to LoggerWriter_1
• Child 2 sets sys.stdout to LoggerWriter_2 (LoggerWriter_1 is now orphaned)
• Child 3 sets sys.stdout to LoggerWriter_3 (LoggerWriter_2 is now orphaned)
• Child 4 sets sys.stdout to LoggerWriter_4 (LoggerWriter_3 is now orphaned)

All 4 tasks call print(), which writes to sys.stdout — but that's LoggerWriter_4. One buffer, four writers. Lines merge. When earlier children tear down and restore sys.stdout, the later children's LoggerWriter instances are orphaned along with any buffered content.

Bug 2: Truncated output

LoggerWriter.buffer was never flushed before teardown. Partial lines (no trailing \n) were silently discarded when the handler was closed.

Changes

• LoggerWriter now maintains per-context buffers keyed by current_execution_id — concurrent tasks never share a buffer
• Child interceptors no longer replace sys.stdout/sys.stderr — only the master owns global streams. Children register their handler on the stdout/stderr loggers; the existing filter system routes records correctly.
• Teardown flushes buffers before clearing context and removing handlers

Test plan

• Existing tests pass (86/86)
• New unit tests for per-context buffer isolation
• New test verifying child interceptors don't replace global streams
• New async end-to-end test: 4 concurrent executions with interleaved logging.info() + print() — verified zero cross-contamination

🤖 Generated with Claude Code