Agent loses context and halts after first session compaction

[maxcleme]

Description

After the first session compaction, the agent stops mid-task and responds as if it has no conversation history, outputting something like:

This breaks the agent's current thought process entirely. The only workaround is to re-enter the original prompt manually, after which it manages to continue — but all in-progress state is lost.

Steps to Reproduce

1. Start a docker-agent session with a multi-agent setup
2. Let the session run until the first compaction triggers
3. Observe the agent halt and output the "no conversation history" message
4. Re-enter the original prompt to confirm the agent can resume

Expected Behavior

The agent should handle compaction transparently and continue its work without interruption or loss of context.

Actual Behavior

• Agent halts immediately after the first compaction
• Outputs a confused "no history" message
• Breaks the thought process of the agent team
• Requires manual intervention to resume

Additional Context

• Happens on the first compaction, not subsequent ones
• Appeared recently with no changes to the main .yaml config — this is a regression
• Re-prompting does allow the agent to continue, suggesting the compacted context is not fully lost — the agent just fails to resume autonomously