fix: close SessionDB FDs on timeout and lazy recall paths (#72782)

[webtecnica]

Summary

Fixes #72782: Two ownership gaps in SessionDB lifecycle were leaking SQLite file descriptors (.db, WAL, SHM) on cron-heavy gateways.

Changes

1. cron/scheduler.py — Timeout path FD leak

run_job() constructs SessionDB in a one-worker executor with future.result(timeout=N). When initialization exceeds the timeout, the worker was abandoned and its eventual successful result was never retrieved or closed — the returned SessionDB's FDs leaked for the process lifetime.

Fix: Keep the future reference. On TimeoutError, register a done callback (_close_late_session_db_fds) that retrieves and closes the result when the constructor eventually completes.

2. agent/agent_init.py + run_agent.py — Lazy recall ownership

When AIAgent receives session_db=None, _get_session_db_for_recall() opens its own SessionDB but AIAgent.close() never closed it. Caller-injected gateway/CLI DBs must remain shared.

Fix:

• agent_init.py: Initialize _owns_session_db = False for injected stores.
• run_agent.py: Set _owns_session_db = True when lazily creating the fallback.
• run_agent.py: In close(), step 8, close _session_db only when the agent owns it.

Testing

All existing cron tests pass (verified against a clean base):

• test_sessiondb_init_hang.py (4 tests)
• test_scheduler_mcp_init.py
• test_cron_inactivity_timeout.py (11 tests)

[teknium1]

Thanks for identifying the two SessionDB ownership gaps. The production direction is sound: current cron/scheduler.py:2935 abandons a timed-out SessionDB future without retaining its eventual result, and current run_agent.py:610 lazily opens a SessionDB that AIAgent.close() does not close (run_agent.py:4048-4053).

Problems

• This PR adds no regression tests. Existing tests/cron/test_sessiondb_init_hang.py:75-183 covers timeout continuation, but not closing a late SessionDB result; there is also no lazy-owned versus injected-store teardown test.
• The diff combines Cron SessionDB timeout and lazy recall paths leak file descriptors #72782 with unrelated CLI, Photon, and Himalaya changes. The Photon change has independently landed on main as cd158466d1 (plugins/platforms/photon/auth.py:183), so the full stack is not a focused salvage candidate.

Suggested changes

• Salvage the focused Cron SessionDB timeout and lazy recall paths leak file descriptors #72782 production commit 20fb735f02cc6064f46578a344fe0fc5ebe92547 rather than the complete stacked diff.
• Add delayed-future closure and lazy/injected ownership regression tests; the open fix(cron): close leaked SessionDB connections on timeout-abandon and lazy recall (#72782) #72822 provides focused examples.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Two open PRs address #72782. Both close late SessionDB results after cron initialization timeouts and track lazy-versus-injected database ownership, but #72822 isolates those fixes and adds targeted regression coverage while #72804 includes unrelated changes.

Related pull requests

• #72804 fixes — (+210/-158) — fixes, but duplicated and unfocused: The SessionDB future callback and ownership-aware teardown address both reported leak causes, but the diff also changes model aliases, Photon setup/authentication, and Himalaya documentation and adds no regression tests for #72782. Despite the MAINTAINER-BOT keep_open verdict on #72804, its suggested salvage is the focused production commit 20fb735, while #72822 already contains the equivalent production fix with dedicated tests.
• #72822 best fix — (+211/-1) — focused fix with a concrete salvage path: It addresses both leak paths in agent/agent_init.py, cron/scheduler.py, and run_agent.py, and adds tests for late-result closure, constructor failure, lazy-versus-injected ownership, and idempotent teardown.

Duplicates

#72804 and #72822 implement essentially the same production changes for both SessionDB ownership gaps; #72822 is the focused, test-backed duplicate target, while #72804 carries unrelated changes.

Suggested consolidation

Keep #72822 open with a salvage path: retain its focused three-file production fix and targeted regression tests for both causes in #72782. Close #72804 as a duplicate of #72822 despite its MAINTAINER-BOT keep_open verdict, because the complete diff is unfocused and untested for this issue, while the verdict's salvageable production commit 20fb735 is represented by the equivalent tested implementation in #72822; preserve attribution as appropriate.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I72782(["issue #72782 (open)"])
    subgraph Dup72804 ["PRs duplicating each other"]
        P72804["PR #72804 (open)"]
        P72822["PR #72822 (open)"]
    end
    P72804 -->|fixes| I72782
    class I72782 open
    class P72804 open
    class P72822 open
    class P72822 best
    class P72804 target
    click I72782 "https://github.com/NousResearch/hermes-agent/issues/72782"
    click P72804 "https://github.com/NousResearch/hermes-agent/pull/72804"
    click P72822 "https://github.com/NousResearch/hermes-agent/pull/72822"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 36 kB of PR diffs, 7 kB of issue/PR text, 5 kB of discussion (1 comments), 3 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.