Skip to content

l10-query-engine.en.md

Latest commit

 

History

History
49 lines (32 loc) · 1.59 KB

l10-query-engine.en.md

File metadata and controls

49 lines (32 loc) · 1.59 KB
Raw

Layer 10: QueryEngine And System Prompt Assembly

中文 | English

Core Question

Why is QueryEngine.ts more than a wrapper around query? Why is it the session-level orchestration center?

Review First

  • l1-startup
  • l2-agent-loop
  • l9-context

Read First

  • QueryEngine.ts
  • bootstrap/state.ts
  • constants/prompts.ts

Search Anchors

  • class QueryEngine
  • before_getSystemPrompt
  • after_getSystemPrompt
  • system_message_yielded
  • fetchSystemPromptParts

What To Notice

  • how the default system prompt, append prompt, and memory mechanics prompt get assembled
  • why slash-command processing, transcript persistence, and skill/plugin loading happen before the main query
  • why QueryEngine acts like both a session controller and a staging area before each turn

Source Evidence

  • QueryEngine.ts:184: export class QueryEngine
  • QueryEngine.ts:284-301: profiler checkpoints around system prompt assembly
  • QueryEngine.ts:554: system_message_yielded

What You Should Actually Learn

query.ts answers “how does a turn run?” QueryEngine.ts answers “what must be prepared before a turn can run at all?” If you want to understand how Claude Code unifies slash commands, memory, skills, plugins, and structured output under one entry path, this layer matters more than the loop itself.

Questions To Answer

  1. Why is system-prompt assembly centralized in QueryEngine instead of scattered across tools?
  2. Why is transcript persistence handled before entering the main loop?
  3. Where is the boundary between QueryEngine and the REPL?