fix(cli): suppress memory prefetch for single-query runs

[hyper-factor]

Summary

• Suppresses post-turn memory prefetch for non-interactive single-query CLI runs (hermes chat -q / quiet one-shot paths).
• Prevents Honcho/background memory prefetch work from racing Python interpreter shutdown after the response has already printed.
• Keeps normal interactive/gateway prefetch behavior intact.

Why

We reproduced a legitimate upstream bug: hermes chat -q can print the correct answer, then exit with SIGABRT / rc=134 when Honcho hybrid memory is enabled. The crash happens after successful output, which breaks automation/worker callers that rely on exit code.

This overlaps with existing shutdown-related reports, but the open PRs I checked focus on hermes -z, provider shutdown, or forced one-shot exit. This PR targets the chat -q path by avoiding unnecessary prefetch work when there is no next turn to warm.

Related

• Fixes Intermittent SIGABRT (core dump) on process exit under concurrent fresh hermes chat -q subprocesses #43186
• Refs hermes -z one-shot crashes with SIGABRT (exit 134) on exit: Honcho memory daemon threads blocked in httpx I/O during interpreter finalization #37632
• Refs [Bug]: Honcho hybrid memory can leave daemon dialectic threads alive during CLI shutdown, causing intermittent CPython abort #33485
• Refs [Bug]: Hindsight sync can race interpreter shutdown after successful one-shot CLI exit #15073
• Related PRs reviewed: fix(oneshot): join memory daemon threads before process exit (#37632) #37635, fix: shut down Honcho memory provider after oneshot #31664, fix(cli): prevent SIGABRT after successful one-shot runs #34963, fix(cli): prevent SIGABRT on hermes -z process exit #33543, fix(memory/honcho): align shutdown join timeout with Honcho HTTP timeout #33701

Test plan

• uv run --with pytest --with pytest-asyncio python -m pytest -o addopts='' tests/cli/test_cli_active_agent_ref.py tests/agent/test_memory_post_turn_prefetch.py -q
  • Result: 4 passed
• Live smoke test from patched branch against Honcho-enabled HERMES_HOME:
  • HERMES_HOME=/tmp/hermes-home PYTHONFAULTHANDLER=1 .venv/bin/hermes chat -Q --toolsets safe -q 'Reply exactly: UPSTREAM_PR_OK'
  • Result: printed UPSTREAM_PR_OK, exit code 0

[liuhao1024]

Positive verification: Suppress post-turn memory prefetch in single-query runs

Reviewed: cli.py signal handler, cli_agent_setup_mixin.py atexit cleanup.

The fix is well-scoped: single-query non-interactive runs (-q flag) set _suppress_post_turn_prefetch on both the CLI and agent objects, preventing Honcho background threads from racing interpreter shutdown. The _active_agent_ref fix in the setup mixin correctly assigns to the owning module (cli or __main__) instead of using a bare global that would create a dead reference in the wrong module namespace.

No findings. LGTM.

[liuhao1024]

Reviewed this PR — the _active_agent_ref module-owner fix is a real bug. The old code used global _active_agent_ref inside the mixin module (hermes_cli.cli_agent_setup_mixin), creating a shadow variable invisible to the cli module's atexit cleanup. The sys.modules lookup correctly targets the concrete owner module. The _suppress_post_turn_prefetch flag for single-query runs is clean and prevents background thread races at shutdown. Both regressions have dedicated test files. No issues found.

[teknium1]

Thanks for the focused shutdown investigation. This is now implemented on main through a more general lifecycle fix, so this PR's suppression flag is redundant.

• cli.py:1230-1236 finalizes every single-query invocation through _run_cleanup() before releasing its session.
• cli.py:1114-1147 flushes pending memory work and calls shutdown_memory_provider() for the active CLI agent.
• agent/memory_manager.py:654-662 uses a daemon-backed memory worker so wedged provider work cannot block interpreter exit.
• Commit 64b21e50fb637a9445cc83ed12cf12a7109b8e34 fixed the same _active_agent_ref publication issue at hermes_cli/cli_agent_setup_mixin.py:404-405; tests/cli/test_cli_active_agent_ref_wiring.py:49-69 covers it. That commit is contained in v2026.7.1.

Automated hermes-sweeper review.