type: Metric
title: Prompt size per call
description: Measured prompt tokens per agent call, and the configuration changes that did not move them.
tags: metrics, tokens, cost
status: stable
stale_after: Sat Jan 30 2027 01:00:00 GMT+0100 (Central European Standard Time)
generated: crm-bot/deepseek-v4.1-flash · Fri Oct 30 2026 10:53:57 GMT+0100 (Central European Standard Time)
verified: process:live-probe · Fri Oct 30 2026 10:53:57 GMT+0100 (Central European Standard Time)

Measured

Call shape	prompt_tokens
trivial question, default toolset	13,483
trivial question, narrowed `platform_toolsets.api`	13,475
trivial question, MCP tool definitions in play	~27,000
a tool-using turn that reads data	43,497

What this establishes

Toolset trimming is not a lever here. Three configurations moved the number by 8 tokens. The fixed cost is Hermes' base prompt plus the skills index, not our schemas.
Tool-using turns cost roughly 3×, because schemas and results accumulate across round trips. Fewer round trips is the only real prompt saving available, and it comes from pre-injecting a compact data snapshot rather than from trimming.
An earlier reading of 41,820 tokens was a cold-start outlier and was retracted; steady state is ~13.5k.

Why it matters anyway

Prefill of this size is mostly cache-friendly, so the practical lever is the number of calls, not their size. That is why the async-run change targeted perceived latency (thinking, tool chips, a usable composer) rather than throughput.

Rendered from nx-master/metrics/context-window.md · OKF bundle · regenerate with npm run render