fix(honcho): cleanly shutdown one-shot memory

[motoblurr]

Summary

• flush/shutdown Honcho memory provider from oneshot paths before process exit
• gracefully stop Honcho session async writer during shutdown
• expose safer honcho_conclude cleanup helpers (list_recent, delete_text, delete_tag, dry-run, IDs)
• add regression coverage for one-shot active-session release before hard exit
• add Honcho sync hygiene guard: exact-response probes and configured temporary verification patterns are not persisted

Verification

• source .venv/bin/activate && python -m pytest tests/honcho_plugin/test_session.py::TestConcludeToolDispatch tests/test_honcho_client_config.py tests/test_honcho_session_context.py tests/test_honcho_startup_fail_open.py tests/agent/test_oneshot.py tests/cli/test_cli_shutdown_memory_messages.py tests/cli/test_single_query_session_finalize.py -q -o 'addopts=' --tb=short → 65 passed in 1.06s
• source .venv/bin/activate && ruff check plugins/memory/honcho/__init__.py plugins/memory/honcho/client.py tests/honcho_plugin/test_session.py tests/test_honcho_client_config.py → passed
• python scripts/check-windows-footguns.py --all → passed
• git diff --check → passed
• Live smoke: hermes chat -q 'Reply exactly: SKIPMEM-20260704' --toolsets safe -Q → SKIPMEM-20260704, EXIT_CODE=0
• hermes honcho status → Enabled True, workspace fernando-ai-memory, recall hybrid, write freq async, connection OK

Notes

• Upstream GitHub Actions currently requires maintainer approval for the fork PR (action_required / no jobs started).

[motoblurr]

Local verification is green for the affected Python/Honcho paths. GitHub Actions did not start jobs because this fork PR is in action_required state and needs maintainer approval. Latest run: https://github.com/NousResearch/hermes-agent/actions/runs/28708541617

[tonydwb]

Code Review Summary

Verdict: Comment

Cleans up one-shot memory after a Honcho one-shot task completes. The one-shot memory was not being released, causing potential memory buildup on long-running agents.

Observations

• 6 files, 308 additions — this is a reasonable scope for a memory lifecycle fix.
• The finally block pattern is correct for guaranteed cleanup.

Looks Good

• No hardcoded secrets
• No debug artifacts
• Memory cleanup is well-placed

---

Reviewed by Hermes Agent

[teknium1]

Thanks for addressing the missing one-shot memory boundary; the premise remains valid on current main: hermes_cli/oneshot.py:425-426 returns directly after run_conversation() without calling the agent lifecycle shutdown.

Problems

• The new os._exit(0) in cli.py's _finalize_single_query() masks failures. Current main invokes that finalizer in the finally around cli.chat() (cli.py:16456-16459), so a failed human single-query run with Honcho configured would terminate with status 0 instead of propagating its error.
• The one-shot lifecycle change needs direct coverage. The existing -z tests at tests/hermes_cli/test_tui_resume_flow.py:633-718 stub _run_agent(), so they cannot verify that _run_agent() calls shutdown_memory_provider() after success and failure.

Suggested changes

• Remove the unconditional hard exit and retain the existing bounded cleanup watchdog (cli.py:994-1049) as the forced-exit backstop.
• Add focused _run_agent() lifecycle tests for both normal and exception paths.

Automated hermes-sweeper review.