Summary

When a llamacpp provider model is active, the Sessions table displays the model's training context window (n_ctx_train) as the denominator rather than the actual runtime context window (n_ctx). This makes token usage percentages misleading.

Environment

• OpenClaw version: 2026.4.25
• Provider: llamacpp (OpenAI-compatible server via llama.cpp)
• Model: qwen3.6-mxfp4-moe
• llama.cpp build: b8951-665abc609

Observed behavior

Sessions table shows:

agent:main:main  |  qwen3.6-mxfp4-moe  |  56k/262k (21%)

The denominator 262k is n_ctx_train — the model's trained context ceiling — not the server's actual loaded context.

Expected behavior

The denominator should reflect the actual runtime context window the server was started with (--ctx-size 65536), showing 56k/65k and a correct usage percentage (~86%).

Root cause

The /v1/models OpenAI-compatible endpoint only exposes n_ctx_train in the meta field:

"meta": {
  "n_ctx_train": 262144,
  ...
}

It does not include n_ctx (the actual loaded context). However, llama.cpp's /props endpoint does expose the correct value:

"n_ctx": 65536

Additionally, the agent's models.json file has contextWindow: 65536 set correctly, but this value appears to be overridden by the live server query result.

Suggested fix

For llamacpp provider types, OpenClaw should either:

1. Prefer models.json — trust the contextWindow value configured in the agent's models.json over the live server metadata, or
2. Query /props — for llama.cpp servers, fetch n_ctx from the /props endpoint instead of (or in addition to) n_ctx_train from /v1/models

Option 1 is simpler and consistent with how other providers work. Option 2 is more accurate but requires provider-specific logic.