fix(cli): shut down oneshot memory providers

[madeat3am]

Summary

• mirror the interactive CLI memory-provider lifecycle in hermes -z
• close the oneshot-owned SessionDB after agent shutdown
• extend the existing oneshot integration test to assert both resources close

Reproduction

With Honcho active, a successful hermes -z tool call printed its final response and then exited 134 with SIGABRT. A native backtrace showed Python finalizing while a daemon Honcho HTTP worker was blocked in socket receive. hermes_cli/oneshot.py bypasses cli.py and returned without calling AIAgent.shutdown_memory_provider().

After this patch, real Honcho write and read oneshots both returned the expected marker, empty stderr, and exit 0.

Test

scripts/run_tests.sh tests/hermes_cli/test_tui_resume_flow.py

Result: 48 passed.

[alt-glitch]

This was generated by AI during triage.

Related: #31664 (earliest-open oneshot memory-provider shutdown), #52186 (join Honcho prefetch threads), #61875 (atexit hook), #50217 (skip oneshot prefetch). Same oneshot-SIGABRT class but this PR is a superset (also closes the oneshot-owned SessionDB); not a duplicate. A maintainer should pick one canonical approach for the cluster.

[teknium1]

Thanks for the focused lifecycle fix. Current main still creates the oneshot SessionDB and AIAgent in hermes_cli/oneshot.py:388-417, then returns immediately after run_conversation() at hermes_cli/oneshot.py:425-426; it does not perform the equivalent of interactive cleanup. The proposed finally mirrors the existing interactive CLI behavior at cli.py:1113-1147, and AIAgent.shutdown_memory_provider() performs the required provider end-session and shutdown calls at run_agent.py:3388-3404.

The proposed test also verifies both resources are released from the oneshot owner path. The GitHub comparison reports the PR as 14 commits behind its base, but none of those base changes touch either PR-modified path, so this should be mechanically salvageable.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the oneshot SIGABRT/lifecycle failure: #63931 adds owner-path cleanup for the memory provider and SessionDB, while #31172 attempted a broader shutdown strategy that also skipped oneshot prefetch and stopped Honcho’s async writer.

Related pull requests

• #31172 [closed] related — (+101/-12) — closed reference implementation, not a strict duplicate: it overlaps #63931 by shutting down the oneshot memory provider and closing the SessionDB, but additionally disables next-turn prefetch, changes Honcho manager shutdown, and includes unrelated gateway enablement changes; it remains relevant as evidence for the broader worker-teardown approach.
• #63931 related — (+67/-34) — keep open with a salvage path: its focused finally mirrors the interactive lifecycle by calling shutdown_memory_provider() and closing the oneshot-owned SessionDB, with a test covering both releases. This agrees with the maintainer-bot keep_open review, while the contributor’s cluster note should be addressed by treating #63931 as the focused owner-lifecycle candidate rather than conflating it with the extra prefetch and Honcho changes in #31172.

Duplicates

#31172 and #63931 overlap on oneshot memory-provider and SessionDB cleanup, but they are not strict duplicates: #31172 also changes prefetch behavior, Honcho shutdown, and gateway configuration.

Suggested consolidation

Keep #63931 open with the concrete salvage path of rebasing its focused oneshot finally cleanup and its dual-resource lifecycle test onto current main. Retain closed #31172 only as a reference for separately evaluating the additional prefetch and Honcho worker-shutdown changes; no PR in this pair should be closed as a duplicate of the other.

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 17 kB of PR diffs, 2 kB of issue/PR text, <1 kB of discussion (1 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.