fix: shut down Honcho memory threads in oneshot mode

[pajoric]

Summary

This fixes a process shutdown issue affecting Hermes oneshot mode when the Honcho memory provider is active.

The observed behavior was:

• The oneshot request completed successfully.
• The expected response was printed.
• The process then terminated with SIGABRT (exit code 134).

The issue occurred during process teardown rather than request execution.

Reproduction

Using current upstream:

time hermes -z "Reply with exactly: Hermes update validation complete."
echo $?

[alt-glitch]

This was generated by AI during triage.

Duplicate of #31664 — same mechanism: a try/finally in hermes_cli/oneshot.py that calls agent.shutdown_memory_provider(...), plus Honcho-side thread-join hardening in session.shutdown(), to stop the oneshot SIGABRT (exit 134) when the Honcho memory provider is active. Related: #49498 (atexit-hook + httpx-close variant) and #50217 (skip-prefetch-enqueue variant) attack the same SIGABRT cluster via different mechanisms. Flagging for a maintainer to pick the canonical fix.

[hugoalvespereira]

This was generated by AI during triage.

Duplicate of #31664 — same mechanism: a try/finally in hermes_cli/oneshot.py that calls agent.shutdown_memory_provider(...), plus Honcho-side thread-join hardening in session.shutdown(), to stop the oneshot SIGABRT (exit 134) when the Honcho memory provider is active. Related: #49498 (atexit-hook + httpx-close variant) and #50217 (skip-prefetch-enqueue variant) attack the same SIGABRT cluster via different mechanisms. Flagging for a maintainer to pick the canonical fix.

Ok. I’ll wait for y'alls call on which fix you guys prefer

[teknium1]

Thanks for targeting the oneshot memory-worker lifecycle gap; current main still returns from the oneshot agent path without memory-provider teardown.

Problems

• The PR changes an older agent.chat() call, but current main calls agent.run_conversation() at hermes_cli/oneshot.py:425-426 (introduced by 093f567f0d). The finalizer must be transplanted onto that current tuple-return path.
• The new prefetch-thread list can still race shutdown: session.shutdown() snapshots it, but no closed-state guard prevents a concurrent prefetch_context() from appending a new daemon afterward. Current provider paths call it at plugins/memory/honcho/__init__.py:677 and :819.
• The PR adds no regression test. The existing oneshot integration coverage is in tests/hermes_cli/test_tui_resume_flow.py:823-881.

Suggested changes

• Finalize in a try/finally around current run_conversation(), forwarding _session_messages using the guarded pattern in cli.py:1128-1147.
• Synchronize shutdown with prefetch registration using a terminal closed state, and add tests for transcript forwarding, exceptional exit, and no post-shutdown prefetch creation.

This is an automated hermes-sweeper review.