Context and cost

Every call to the model sends the whole context: the system prompt (the CLI's, yours, and what the kit appends), the tool definitions, the skills listing, memory files, and the conversation so far. Most of it is cached between calls, but it still counts against the context window and the usage limits.

What a session sends

Measured on Captain Whiskers answering "hola" (input tokens of one API call):

Configuration	Input tokens
Every setting source (the SDK's default)	~16.1k
settingSources: ["project"] (the kit's default)	~15.7k
settingSources: [], its own two skills only, no auto-memory	~11.4k

Most of the fixed part is the tools' definitions (about 6k tokens) and the skills listing (3.5k with every skill found, 0.1k with two). Memory files (CLAUDE.md files found up the directory tree) add the rest.

Those figures predate the kit's later built-in tools. Today the same configuration's first call takes about 21.8k tokens, mostly tool definitions. What each part costs, measured by taking it out (input tokens of the first call):

Part	Tokens	Goes away with
Agent (with the list of subagents)	4.1k	no subagents
TodoWrite (the task list)	3.4k	disallowedTools: ["TodoWrite"]
The knowledge base: ten knowledge_* tools, its prompt section and skills	3.4k	without the knowledge extension
Read, Glob, Grep (Grep alone 1.4k)	2.7k	no sourcesDir nor extraWritableDirs (with the knowledge base on its tools)
The sources tools	2.2k	without the sources extension
Bash	1.9k	no subagent listing it (the kit leaves it out then)
The approval and question tools	1.5k	autonomous mode
WebFetch, WebSearch	1.4k	disallowedTools
Skill and the skills listing	1.1k	no plugins nor file tools

The kit's MCP tools (the knowledge base's, the sources', the approvals', the date tools) only go away with their feature: disallowedTools blocks them, but their definitions are still sent (see SDK behaviors).

Making it smaller

1. Offer only your skills: skills: "plugins" (those of the plugins the agent loads), or a list, skills: ["my-agent:a", "my-agent:b"]. The SDK's own and unrelated skills are no longer listed. Slash commands keep working. See Skills and plugins.
2. Load no settings files when your plugin brings everything: settingSources: []. The CLAUDE.md files of parent folders (a repository's developer instructions, say) stop being sent.
3. Grant only the tools you need: no file tools without a knowledge or sources folder, no Agent without subagents, no Bash unless a subagent lists it. disallowedTools removes a built-in tool the configuration would otherwise get; an MCP tool it only blocks, still sending its definition.
4. Keep the prompt short and move rarely needed detail into skills, which are only loaded when they apply.
5. Use subagents for noisy work: a subagent's searches and page reads stay in its own context; only its answer reaches the main one. Use a smaller model for them.

Watching it

The Ink chat's status bar shows the session's tokens and how full the context window is after each turn:

⏵⏵ guided (shift+tab) · 6 turns · 26.6k in (+126.7k cached) / 1.4k out · context 3%

• context 3% is what the conversation takes up now. It's the number that tells you when compaction is near.
• 26.6k in is the new input over the whole session: the first call's fixed part plus what each call added.
• (+126.7k cached) is the context sent again from the cache, once per call. It grows fast because every call sends everything again, but it doesn't mean the context is growing, and cache reads cost a tenth of new input. In a terminal too narrow for it, it's left out.

From code:

const run = runQuery(input, options);
// …after a turn:
const usage = await run.contextUsage(); // { percentage, totalTokens, maxTokens } or null

Each turn-end event carries the session's cumulative usage (inputTokens including cache reads and writes, cacheReadTokens, the part of it read from the cache, outputTokens, and an estimated costUsd). See Events.

Long conversations

autoCompactEnabled is on by default: when the context fills up, the CLI compacts the conversation into a summary and goes on. Turn it off with buildSessionOptions(config, runDir, spec, { autoCompactEnabled: false }) if you'd rather have the session fail than lose detail.