Context and cost | GBS SE Agent

Every model call carries the whole conversation. Keeping that small and stable is most of what makes an agent fast and cheap, especially on a local model with a modest context window.

A request is a bar of segments: the system prompt with rules and the tool list, the environment block with OS and AGENTS.md, and the skills and MCP catalogue lines. Those three are byte-identical on every step, so the server can cache them. Then the conversation, your requests and tool results, which grows every step, and finally space reserved for the answer, up to maxOutputTokens. Compaction starts at about 85 percent of the space left after the reserve. When the conversation outgrows the budget, the cheapest thing goes first: one, each new request turns big tool output from earlier requests into short notes, which is free; two, when over budget, the six newest tool results stay and older ones are dropped, also free; three, summarise, where a model call condenses older history, costing one request, which /compact also triggers. In normal numbers, the default window is 32,768 tokens, about 24,000 words, and with an 8,192-token answer reserve compaction starts near 20,900 tokens. — Growth happens at the end and cuts come from the middle. The opening is never touched, because changing it would throw away the server's cache.

Growth happens at the end and cuts come from the middle. The opening is never touched, because changing it would throw away the server's cache.

A stable opening

The system prompt, the tool schemas and the environment block are byte-identical on every step of a session; the environment block is computed once and cached. The conversation is append-only. That lets Ollama's KV cache and cloud providers' prompt caches reuse everything up to the newest message, so each step after the first only pays for what is new.

This is also why skills and MCP tools are fixed when a run starts: adding one mid-run would change the opening and lose the cache.

The fixed overhead is about 2,600 tokens for the main agent and about 1,200 for an explore subagent.

Keeping tool output bounded

Reads are ranged (1,000 lines by default), and searches are capped.
Command output keeps the first 6,000 and last 14,000 characters, where errors usually are.
Every result is trimmed to gbsAgent.maxToolOutputChars (20,000 by default) before the model sees it.
Broad searches happen in subagents, whose output never enters the main conversation.

The compaction ladder

Trimming happens in a fixed order, cheapest first:

Earlier requests' output. Each time you send a new request, tool results over 1,500 characters from earlier requests become a short note saying what the call was. The agent can read the file again if it needs to. Free.
Old tool output in this request. The budget is the context window minus the answer reserve (the smaller of gbsAgent.maxOutputTokens and a quarter of the window), times 0.85. When the estimated prompt passes it, all but the six most recent tool results over 600 characters become notes. Free.
A summary. If that is still not enough, one model call condenses the older part of the conversation into a summary, cutting only at a point that does not separate a tool call from its result. /compact does this on demand.

In normal numbers

The default window is 32,768 tokens, about 24,000 words or 50 pages. A bigger window is not automatically better: models get noticeably worse at finding things in the middle of a very full context, which is why the agent trims even when there is room.

Setting	Default	Raise it when
`gbsAgent.contextWindow`	32,768	Your model and hardware support more; set it to the real value
`gbsAgent.maxOutputTokens`	8,192	Long files get cut off mid-write
`gbsAgent.maxSteps`	60	Big tasks stop to check in too early
`gbsAgent.maxToolOutputChars`	20,000	Test failures get trimmed right where it matters

Counting requests, not tokens

One instruction becomes many model calls: three to eight for a small edit, twenty to sixty for a refactor with subagents. Request limits, not token allowances, are what run out first on a hosted model. See How free tiers run out.