中文 | English
This is not a loose backlog. It is the bilingual article track tied directly to the repository.
-
query.tsand the Agent Loop: the heart of Claude Code- repo companion:
l2-agent-loop - focus:
query,queryLoop,while (true), tool-result feedback
- repo companion:
-
Why Claude Code uses async generators
- repo companion:
l8-streaming - focus: event flow, cancellation,
yield*composition
- repo companion:
-
Claude Code’s three-layer permission model
- repo companion:
l7-permissions - focus: semantic denial, rule matching, user confirmation
- repo companion:
-
How prompt cache requirements shape Claude Code architecture
- repo companion:
l9-context - focus: dynamic boundary, snapshotting, sticky latches
- repo companion:
-
Why Claude Code memory is an index system
- repo companion:
l6-advanced - focus:
MEMORY.md, typed memory, on-demand loading
- repo companion:
-
The boundary between SyntheticOutputTool and the coordinator
- repo companion:
l6-advanced - focus: structured output, worker isolation, multi-agent composition
- repo companion:
-
Why QueryEngine is the session-orchestration center
- repo companion:
l10-query-engine - focus: system-prompt assembly, message staging, session orchestration
- repo companion:
-
Claude Code API streaming is more than “typing as it goes”
- repo companion:
l11-api-streaming - focus: event model, TTFB, stall detection, watchdogs
- repo companion:
-
Why
REPL.tsxis an interaction-orchestration layer, not just a screen- repo companion:
l12-repl-ui - focus: initial messages,
onSubmitvsonQuery, local state boundaries
- repo companion:
-
Why Claude Code has MCP, hooks, and plugins instead of one extension API
- repo companion:
l13-mcp-hooks-plugins - focus: capability ingress, lifecycle interception, feature packaging
- The real boundary between memory extraction and team memory
- repo companion:
l14-memory-system - focus: typed memory, extractMemories, team-memory safety
- Why Claude Code needs print, serve, and bridge runtime modes
- repo companion:
l15-runtime-modes - focus: entrypoints, multiple runtime surfaces, long-lived bridge sessions
- every article must link back to one layer note and one source-map node
- Chinese and English versions must share the same evidence points
- do not write generic AI-agent advice; write only what the source can support