[Bug] Background subagent completion notifications dropped with "database is not open" — parent turn never re-invoked (3.10.2 macOS)

提交前确认 · Pre-submission checklist

• I searched existing issues and confirmed this isn't a duplicate. (Closest: #397 — same symptom class "no turn starts after background completion", but different root cause: remote compact vs local subagent DB admission failure.)
• I've read CONTRIBUTING.md.

问题类别 · Category

后台任务 / 子代理 · Background tasks / subagents

涉及的 Agent 框架 · Agent framework

ZCode Agent（自研）

严重程度 · Severity

影响功能 · Major (work is completed but the orchestrating session is never notified)

复现频率 · Reproducibility

偶发，但一次会话内三个并行后台 agent 全部命中（3/3）

问题描述 · Description

When background subagents (spawned via the Agent tool with run_in_background: true) complete, the completion notification is generated and enqueued, but admitting it into the session ledger fails with "database is not open". The background turn that should re-invoke the parent session is then cancelled, and the task registry drops the agent handles entirely — TaskOutput returns No task found with ID, SendMessage returns No active local_agent task found. The user must manually prompt the session to continue.

Log events from ~/.zcode/cli/log/zcode-2026-08-31.jsonl (all within one session, timestamps 15:33–15:46 local):

runtime.background_task_notifications.sealed  (reason: subagent_terminal)
background_task.notification.runtime_enqueued (queueSize: 1)
session_input.admit_failed  "Failed to admit background notification to ledger" — errorMessage: "database is not open"
background_task.notification.batch_failed     — errorMessage: "database is not open"
subagent.background.failed                    — errorMessage: "database is not open"
bootstrap.zcode_protocol "v4 background turn failed" — error: "Turn was cancelled."

5 occurrences of "database is not open" in total. The session DB (~/.zcode/cli/db/db.sqlite + WAL) is present and writable before/after the failure window, which suggests the ledger connection handle was closed or not yet (re)opened when the notification arrived — a race between subagent terminal notification and session DB lifecycle. For reference, the OpenCode lineage fixed a similar signature ("database is not open in ephemeral CLI flows", openclaw #20556/#7464).

复现步骤 · Steps to reproduce

1. ZCode 3.10.2 desktop app on macOS 27.0 arm64, local project session.
2. Spawn 3 subagents concurrently with the Agent tool, run_in_background: true (each doing multi-minute multi-file work).
3. End the turn while they run (wait for automatic completion notifications).
4. All three complete (their edits land in the working tree), but no notification turn starts; task registry has no record of the agents.

期望行为 · Expected behavior

The completion notification is durably admitted to the session ledger (retry/reopen the DB handle if closed) and the parent turn is re-invoked, or at minimum the task handles remain queryable via TaskOutput.

环境信息 · Environment

• ZCode desktop 3.10.2 (latest at time of writing; changelog has no entry covering this)
• macOS 27.0.0 arm64
• Local (non-remote) session, default SQLite session storage

临时规避 · Workaround

Dispatch parallel subagents as multiple synchronous Agent calls in one message (they run concurrently and return in-turn), avoiding the background notification path entirely.

Start with the events in ~/.zcode/cli/log/zcode-2026-08-31.jsonl, especially session_input.admit_failed and background_task.notification.batch_failed, then trace the session database lifecycle around subagent completion. Done means completion notifications are admitted, the parent turn is re-invoked, and task handles remain queryable when admission encounters a closed database.