TUI Status Bar

The bottom status bar in Arcana's terminal UI shows live session metrics. This page documents what each field means and how the numbers are calculated.

<duration>  ·  <tokens-in>↓  <tokens-out>↑  ·  <total> total  ·  <ttft> ttft  ·  <cached>↺

Field Reference

How Each Metric Works

<duration> — Session Duration

Wall-clock time from session creation to now. Updates in real-time. Format uses the largest two units: days and hours, hours and minutes, minutes and seconds.

A long duration does not necessarily mean high cost. A session left idle (for example, overnight) accumulates duration without consuming tokens.

<tokens-in>↓ — Tokens In

Every token sent to the model provider. This includes:

• System prompt — Instructions, tool definitions, skills loaded for this session
• Conversation history — All previous user messages and assistant responses
• Tool results — File contents, search results, command output, web pages
• File context — Any files the agent read and included in context

In short: everything the model reads. This is the primary driver of prompt-cacheable tokens — repeated sends of the same system prompt and conversation history can be served from cache.

<tokens-out>↑ — Tokens Out

Every token the model generates. This includes:

• Text responses — The agent's replies to you
• Tool calls — Function calls the agent makes (file edits, shell commands, web requests)
• Reasoning tokens — Internal chain-of-thought on models that expose them (Claude thinking, o1/o3 reasoning)
• Structured output — JSON or other structured data the agent produces

Output tokens are never cached. Every new generation is a fresh computation.

<total> total — Total Tokens

The sum of Tokens In and Tokens Out. This is the number that matters for cost calculations, since providers bill for both input and output (output is typically 3–5× more expensive per token).

<ttft> ttft — Time to First Token

How long you waited before the agent started responding. Measured from the moment the HTTP request was sent to the provider to the moment the first token of the streaming response arrived.

Factors that affect ttft:

• Model size — Larger models (Opus, o1) are slower to start than smaller ones (Haiku, Flash)
• Prompt length — Longer prompts take more time to process before generation begins
• Provider load — High traffic at the provider increases queue time
• Prompt cache — A cache hit reduces ttft because the provider skips re-processing cached tokens
• Network latency — Distance between Arcana and the provider endpoint

A ttft above 5s usually means a long prompt or a cache miss on a large model. Below 1s is typical for small models with short prompts.

<cached>↺ — Cached Tokens

Prompt tokens served from the provider's prompt cache instead of being re-processed. This is a cost and latency saver — cached tokens are billed at a fraction of the normal input price (typically 10–25% of the standard rate, depending on provider).

The ↺ symbol represents reuse. When you see a high cached-token count relative to Tokens In, it means Arcana is efficiently reusing context across turns. A typical well-cached session has 50–80% of input tokens cached after the first few turns.

Providers that support prompt caching:

Reading the Bar

The separators group related metrics:

<duration>        ·  <tokens-in>↓  <tokens-out>↑     ·  <total> total     ·  <ttft> ttft  ·  <cached>↺
^^^^^^^^          ^^^^^^^^^^  ^^^^^^^^^^     ^^^^^^^^^^^^^^     ^^^^^^^^^^^^  ^^^^^^^^^
Session           Input       Output         Combined           Latency       Cache
Duration          tokens      tokens         token count        metric        reuse

Left to right: how long you have been working → how much the model read → how much it wrote → total volume → how fast it started → how much was reused.

Tips

• High cached ratio (↺ near ↓) — Your sessions are well-warmed. The model is reusing context efficiently. This saves money.
• Low ttft (< 1s) — Fast start. You are likely on a small model, short prompt, or cache hit.
• High ttft (> 5s) — Slow start. Check if your prompt is very long or if the provider is under load.
• ↑ much larger than ↓ — The agent is generating a lot of output. Common during large code changes or analysis tasks.
• ↑ much smaller than ↓ — Short responses. The agent is doing targeted work (small edits, quick answers).

Related