Summary

Description
Context usage is severely undercounted when using Ollama provider (ollama/minimax-m2.5:cloud, etc.). The displayed context (e.g., "2.7k/205k (1%)") does not reflect actual token usage.

Steps to reproduce

1. Use Ollama model (e.g., ollama/minimax-m2.5:cloud)
2. Run multiple prompts/conversations
3. Check context via session_status or //status
4. Compare with actual Ollama usage at https://ollama.com/usage

Expected behavior

Context should accumulate correctly and reflect actual token usage, similar to other providers (Zhipu, Google, etc.)

Actual behavior

Context display stays very low (0-3%) even after extensive use
"Tokens: X in / Y out" also appears undercounted
Watchdog triggers compact based on inaccurate context data
Risk: User runs out of quota without knowing (quotas may deplete silently)

OpenClaw version

2026.2.14

Operating system

Model: ollama/minimax-m2.5:cloud OS: WSL2 (Linux)

Install method

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

This affects all Ollama cloud models
Context tracking works correctly for glm-4.7, gpt52, etc.
Need external monitor (ollama.com/usage) to track actual usage