fix(oneshot): write response to stdout instead of silently dropping it

[luyao618]

Summary

Fixes #22975 — hermes -z (oneshot mode) returns empty stdout despite a successful API response.

Root Cause

run_oneshot() wraps the entire agent call tree in redirect_stdout(open(os.devnull, 'w')) to suppress diagnostic output. However, AIAgent.__init__ calls _install_safe_stdio(), which wraps sys.stdout in a _SafeWriter proxy. The proxy's __getattr__ forwards attribute lookups to the inner stream — when that stream is a devnull file handle, it lacks text-stream attributes like .encoding (returns None on some platforms) and .buffer (absent entirely). Downstream SDK code and logging formatters that inspect these attributes silently malfunction, causing agent.chat() to return an empty string.

Fix

1. Replace open(os.devnull) with io.StringIO() as the redirect target. StringIO is a well-behaved in-memory text stream that exposes all standard file-object attributes, avoiding the edge cases that devnull triggers through _SafeWriter's proxy.

2. Set agent._print_fn = lambda: None to route all agent-internal print calls (_safe_print, _vprint) through a no-op, eliminating a second vector for stdout interference during response capture.

Testing

• All existing oneshot tests pass (11 passed, 7 skipped)
• The fix is minimal (1 file, +22/-8 lines) and confined to hermes_cli/oneshot.py

Checklist

• Scope: 1 file changed
• Tests pass
• SSH-signed commit

[Bartok9]

Reviewed this against the other two competing fixes for #22975 (#23033, and mine which I just closed).

This is the most complete fix. The StringIO over devnull change removes the underlying proxy edge case rather than papering over the symptom — luyao618 correctly identified that devnull's missing .encoding/.buffer attributes are what causes _install_safe_stdio() to malfunction downstream.

The _print_fn = lambda: None addition is a good belt-and-suspenders move that eliminates the second stdout interference vector.

Compared to my approach (_disable_streaming = True) which only fixed the streaming path, and KhanCold's fallback chain which adds complexity to work around the root cause rather than fix it — this one is the right call.

One minor suggestion: the test coverage in #23007 could be strengthened with a test that specifically verifies _run_agent returns non-empty when _disable_streaming is NOT set (to guard against regression back to the streaming path dropping content). But the fix itself is solid. 👍

[luyao618]

Rebased onto current main — all blocking checks now green (lint, e2e, Windows footguns, nix, builds). The original Windows footgun + ty-diff failures were baseline drift, not regressions from this change.

Per @Bartok9's review above, this is the most complete of the three competing fixes for #22975: it uses io.StringIO() instead of open(os.devnull) so _SafeWriter proxy attribute lookups (e.g. .encoding, .buffer) keep working, and it neutralizes the agent's internal print path via _print_fn. Ready for a maintainer pass when there's time 🙏

[luyao618]

Closing — open too long, no longer relevant.