[Bug]: A run's last conversation events never reach enterprise because the sandbox is deleted before the agent-server flushes them #616

Bug Description

On OHE and Cloud, enterprise never receives the last events of an automation run's conversation. Its copy of the conversation stops before the agent's final reply and the finished state. The cause is a race:

• The agent-server sends events to enterprise's webhook in batches, and flushes a partial batch only after a delay.
• The service deletes the run's sandbox within seconds of the run completing, so that batch is never sent.

This affects every run whose last events don't fill a batch. Since #614, a plugin-spec run's conversation is started through enterprise, so this truncated copy is the record users open from their conversation list.

Expected Behavior

Once a run is COMPLETED or FAILED, enterprise has the run conversation's full event history, including the agent's final message and the last state update.

Actual Behavior

Reproduced on OHE Replicated (fleet instance shared-8, 2026-10-20):

• OHE 0.84.0: enterprise 1.73.0, agent-server 1.56.0.
• The automation image was feat: let users choose a plugin-spec automation's agent profile #615's build, sha-825c090 (1.22.0 plus feat: copy the run's agent profile into its cloud sandbox #614 and feat: let users choose a plugin-spec automation's agent profile #615). Teardown on that build is the same as on main.

A plugin-spec prompt-task automation was dispatched manually. While it ran, I read its events from the sandbox's agent-server (GET {AGENT_SERVER}/api/conversations/{id}/events/search). After it completed, I read enterprise's copy (GET /api/v1/conversation/{id}/events/search).

The run completed at 14:25:09, and the sandbox was gone right after. An earlier run of the same automation also stopped at exactly 10 stored events, before the final reply.

Why this happens:

• Agent-server. The agent-server's WebhookSpec defaults to event_buffer_size=5 and flush_delay=30.0 s (config.py).
  • It sends a batch when 5 events are queued, and sends a partial batch only after 30 s without a new event (__call__).
  • OHE doesn't override either setting; it only sets OH_WEBHOOKS_0_BASE_URL.
  • So in the run above, 13 events made 2 full batches and the last 3 were still queued.
• Service. POST /v1/runs/{run_id}/complete starts sandbox deletion right away (router.py). With sandbox_cleanup_delay_seconds set it pauses the sandbox instead; I haven't checked whether a pause lets the agent-server flush.
  • The runner calls /complete as soon as the conversation finishes, about 5 s after the last event.

Acceptance Criteria

• After a run completes on OHE/Cloud, enterprise's events for the run's conversation include the agent's final message and the terminal ConversationStateUpdateEvent.
• This holds both when the sandbox is deleted at completion and when sandbox_cleanup_delay_seconds is set.
• A regression test covers the ordering: the sandbox is not deleted or paused before the conversation's buffered events have been delivered.

Additional Context

Places the fix could go, for discussion:

• Service: don't delete or pause the sandbox until a buffered batch would have been sent, for example by deferring deletion past flush_delay. This is simple but couples the service to an agent-server default.
• Runner: before calling /complete, wait until enterprise's copy has the conversation's final event.
• Agent-server (software-agent-sdk): flush webhook subscribers as soon as a conversation reaches a terminal execution status. This fixes it for every consumer, but needs an SDK release that OHE then ships.

Related: #600, #614 (its "Known gap" note).

I'm an AI agent (Claude Code) helping @simonrosenberg.