Skip to content

Context and model loop

This chapter follows a Claude Code turn from input collection to provider-facing request and streamed response handling. It covers what becomes model-visible context, how model/provider/auth state is selected, and how headless/SDK streaming frames drive a scriptable run.

Read this chapter when the question is: what did the model see, which provider/model handled it, and how did the runtime frame the turn?

Source-anchor policy

This page is a chapter guide. Linked implementation pages carry concrete cli.js anchors.

Semantic alias	Minified anchor	Scope
Context/model chapter	N/A — navigation page	Groups prompt/context assembly, memory, settings, provider/auth selection, and headless streaming.
Context/model implementation pages	See linked source-anchor tables	Concrete bundle anchors live in destination pages.

Model-turn map

flowchart TD
    Input[Prompt / stdin / files / settings] --> Context[Prompt and context sources]
    Context --> Memory[CLAUDE.md / rules / output styles / slash commands]
    Context --> Tools[Tool schemas and agent definitions]
    Memory --> Request[Provider request]
    Tools --> Request
    Auth[Auth and provider selection] --> Request
    Request --> Stream[Streaming response]
    Stream --> Headless[Headless stream-JSON frames]
    Stream --> TUI[Interactive session loop]
    Headless --> Events[Session events and result]
    TUI --> Events

Primary reading order

Order	Page	Question answered
1	Prompt, context, and memory	Which `CLAUDE.md`, settings, system-prompt, output-style, slash-command, memory, `system-reminder`, and dynamic-injection sources feed context?
2	Prompt assembly scenarios	How do major runtime paths assemble default, custom, appended, bare, agent, subagent, WebSearch/WebFetch, compaction, hook, and helper prompts?
3	Context, memory, compaction, checkpoints, and rewind	How are memory scopes selected, how does auto/manual compaction summarize context, and how do file checkpoints plus rewind work?
4	Prompt template catalog	Which long prompt/template-like literals are embedded in `cli.js`, grouped by runtime surface with anchors, previews, and hashes?
5	Models, providers, and auth	How do API keys, OAuth tokens, provider env vars, model flags, and provider selection connect?
6	Model selection, calls, usage, quota, and billing	How are model roles resolved, how are provider calls made, and how do rate limits, usage, quota, budget, and billing surfaces work?
7	Headless streaming and resilience	How does `-p`/SDK mode validate input/output formats, stream messages, multiplex frames, and handle permission/control/result framing inside `HeadlessRunner`/`HeadlessControlLoop`?
8	Context and model loop architecture	How is the context assembler + streaming multiplexer decomposed, what shapes the static/dynamic prompt boundary, and how do interactive and headless reuse one pipeline?

Handoffs

Tool schemas and permissions are owned by Tools, integrations, and security.
Durable transcript state is owned by Sessions, persistence, and remote.
Agents and task-specific prompts are owned by Agents and automation.

Created and maintained by Yingting Huang.