Summary

I've observed an intermittent issue where the openclaw status --json command (and consequently the session_status tool) reports a negative value for inputTokens for the main agent's current session.

Steps to reproduce

1. Initiate a long-running OpenClaw session with continuous interaction.
2. Periodically run openclaw status --json from a separate process (e.g., via a Python script calling subprocess.run(["openclaw", "status", "--json"], ...) or directly via exec tool).
   • Note: The stdout of openclaw status --json often includes non-JSON lines (e.g., [plugins] ... registration messages, gateway connect failed) before the actual JSON object. This non-JSON prefix needs to be handled if parsing the output programmatically.
3. Observe the inputTokens value in the sessions.recent[0].inputTokens field within the parsed JSON output.

Expected behavior

inputTokens should always be a non-negative (zero or positive) integer, representing the cumulative tokens consumed by the session.

Actual behavior

When running openclaw status --json (or session_status within an agent's context), the inputTokens field in the JSON output occasionally shows a negative number (e.g., -284957, -296318). outputTokens appears to remain positive.

This occurs seemingly randomly and affects the accuracy of token usage tracking.

OpenClaw version

2026.2.19-2 (45d9b20) (as per openclaw status output)

Operating system

Linux 6.17.0-1008-gcp (x64)

Install method

package (pnpm)

Logs, screenshots, and evidence

Please attach the complete log output from running `openclaw status --json` when a negative `inputTokens` value is observed. (e.g., copy-paste terminal output or take a screenshot).

Impact and severity

Impact:

• Leads to inaccurate reporting of token usage.
• Breaks downstream scripts designed to track and summarize token consumption, as negative values are unexpected and logically incorrect for cumulative usage.

Severity: High (as it affects the accuracy of a core metric)

Additional information

Temporary Mitigation (Agent-side):
As a workaround, my agent-side script (record_tokens.py) that logs token usage now clamps any negative inputTokens values to 0 before recording, but this does not fix the root cause in the OpenClaw core.