Summary

When an agent run times out (embedded run timeout, default 600s), background exec processes continue running independently. When those exec processes later complete, their results are injected as System: messages into a new session that has zero context from the previous session. If the user then sends a message (e.g. "continue"), the agent has no idea what was being worked on and may start an entirely unrelated task.

Reproduction Steps

1. Start a task that involves long-running background exec commands (e.g. npm install in a slow environment, CI pipeline polling)
2. The agent run hits the 600s timeout → run is killed
3. Background exec processes keep running (they are independent OS processes)
4. Wait for exec to complete (could be minutes to hours later)
5. User sends a message in the same topic/chat (e.g. "continue")
6. A new session is created for that topic
7. The accumulated System: Exec completed/failed messages + the user message are bundled into a single first user message in the new session

What Happens

The new session receives something like:

System: [01:51] Exec completed (session-a, code 0) :: apt done configure done
System: [01:57] Exec completed (session-b, code 0) :: 1min
System: [01:59] Exec failed (session-c, signal SIGKILL) :: proot warning...

[User]: continue

The agent has no prior conversation context. It:

1. Sees cryptic exec output snippets with no semantic meaning
2. Interprets "continue" by searching memory files
3. Finds the most prominent recent TODO in memory (which may belong to a different topic/session)
4. Starts working on the wrong task in the wrong topic

Root Causes

1. Agent run timeout kills the run but not background exec processes — This creates orphaned exec results that arrive later with no consumer context
2. Session rotation drops all conversation history — The new session has zero knowledge of what the old session was doing
3. Accumulated System messages are batched with user messages — The signal-to-noise ratio is very low; the agent cannot distinguish which exec results are relevant
4. No per-topic/per-session state persistence — There is no mechanism to carry forward "this session was working on task X" across session boundaries

Observed Impact

• Agent started working on Project A in a topic that was dedicated to Project B
• Cross-topic contamination: work intended for one topic bled into another
• User had to re-explain context that was already established in the previous session
• Wasted tokens and time on duplicate/wrong work

Suggested Improvements

1. Carry forward session summary on rotation — When a session times out and a new one is created for the same topic, inject a compact summary of the previous session's last N messages as context
2. Kill or re-parent orphaned exec processes — When an agent run times out, either kill associated exec processes or tag their results with the original session context
3. Separate System exec results from user messages — Don't batch them into a single user turn; deliver them as distinct system events that the agent can triage
4. Per-topic persistent state — Allow sessions to write a small state blob (e.g. "current task", "last working context") that survives session rotation

Environment

• OpenClaw latest (npm)
• Telegram channel with topic-based group chat
• Agent run timeout: 600000ms (default)
• Multiple background exec processes running concurrently