openclaw - 💡(How to fix) Fix `openclaw agent` defaults --channel to last, silently resumes most recent session [1 participants]

Summary

openclaw agent --agent <id> -m "..." invoked without --session-id defaults --channel to last, which silently resumes whatever session was most recently active. There's no warning, no banner, and no easy way for an operator running a fresh test to discover this without reading the resulting agentMeta.sessionId and noticing it's old.

Reproduction

1. Run any openclaw agent --agent foo -m "first message".
2. Wait an arbitrary amount of time (could be days).
3. Run openclaw agent --agent foo -m "second message" (no --session-id, no --channel).
4. Inspect the run report's agentMeta.sessionId and usage.input_tokens.

Observed behaviour

The second run is appended to the first session. input_tokens includes the entire prior conversation. The agent may produce confusing replies if the original context was unrelated (for us: a stale 17k-token investigation got "ping" appended and the model tried to continue the investigation).

Expected behaviour

One of:

• Default --channel to new (always start a fresh session), or
• Add a --new-session / --fresh flag and document that not passing it keeps the previous session, or
• Print a one-line banner: Resuming session <id> from <timestamp>; pass --new-session for a fresh run.

Why this matters

Combined with [related issue on --session-id being silently remapped], operators have no obvious knob to force fresh sessions for testing. We resorted to moving agents/<id>/sessions/ aside between runs.

Environment

OpenClaw 2026.4.12 (1c0672b).