fix(gateway): prevent infinite restart loop on session resume

[JoaoMarcos44]

Description

Fixes #49201

This PR resolves an issue where the gateway falls into an infinite boot-loop when resuming a session that was interrupted by a gateway restart or shutdown command.

Cause

When the gateway container is abruptly restarted, the process is killed. Upon reboot, the gateway auto-resumes the pending session. Because the transcript ends with a terminal command (e.g., ssh docker restart ... or reboot) with a [Result unavailable...] stub, the LLM assumes the command failed or got lost and re-runs it, triggering another restart/shutdown.

Solution

We modified the system note injection logic for resume-pending turns in gateway/run.py (_handle_message_with_agent):

1. Empty User Message (Startup Turn): If the startup turn has an empty message, we instruct the model to report successful recovery to the user and ask for further instructions, rather than attempting to resume the last command.
2. Restart Loop Prevention Warning: For all resume-pending cases, we inject a clear instruction:

   If the last action in the conversation history was a command to restart, stop, or shutdown the gateway or container, it has already executed successfully and the gateway is now back online; do NOT re-execute it and do NOT run any command to verify it. Do NOT re-execute old tool calls — skip any unfinished work from the conversation history.

This cleanly breaks the loop at the cognitive level of the model without modifying global database schemas or core CLI configurations.

Verification & Tests

• Added a new unit test case test_resume_pending_empty_message_guidance to tests/gateway/test_restart_resume_pending.py to verify prompt injection and instructions on empty message recovery turns.
• Verified that all 74 tests in tests/gateway/test_restart_resume_pending.py pass successfully.
• Verified that related component test suites pass without regression:
  • tests/gateway/test_session.py (79/79 passed)
  • tests/gateway/test_restart_drain.py, test_restart_notification.py, test_restart_redelivery_dedup.py (55/55 passed)
  • tests/gateway/test_gateway_shutdown.py (17/17 passed)
  • tests/gateway/test_auto_continue.py (10/10 passed)

[alt-glitch]

Fix PR for #49201 (resume-after-restart reboot loop).

Related: #46963 (freshness gate on resume_pending — different mechanism), #46934 (stale resume_pending bypasses idle reset), #29728 (skip shutdown-timeout startup auto-resume), #43191 (docs FAQ for the restart loop).

This PR breaks the loop at the prompt level: the resume-pending system note now tells the model that a restart/stop/shutdown command in the history already executed successfully (gateway is back online) and must not be re-run or verified, plus dedicated empty-message (startup-turn) recovery guidance. Verified on main: the new wording is absent, so the old "A new message has arrived..." note is still live and the bug is unfixed — this is a genuine, non-no-op fix.

[teknium1]

Thanks @JoaoMarcos44 — merged via #49321 (commit 75ed07a), with your authorship preserved in git history.

We took your fix one layer deeper. Your PR addressed the loop at the prompt level (tell the model not to re-run the restart). #49321 keeps that as a cognitive backstop — including your empty-message startup-turn recovery wording — but adds a source-level fix: on resume, _build_gateway_agent_history now strips a trailing assistant(tool_calls) block that has no matching tool answer (the SIGKILL-mid-call signature). That removes the thing the model was reacting to entirely, so the loop can't form regardless of which restart command ran, and it doesn't leave an unanswered tool call on the wire for strict providers to reject.

Closing this in favor of the merged PR. Appreciate the clear repro and the right instinct on the fix.