English catalog.
🔍

Learning station · English · source-traced

The same corpus, cited to source.

Eighteen agent codebases, from-scratch tutorials, and interactive benches—translated as lectures, not marketing. Identifiers, commit SHAs, and path:line citations are copied from the Chinese source-traced originals. English does not add claims.

18
codebases dissected
60+
lectures
30+
interactive benches
file:line
kept verbatim
Public: kaiyuan-xiangmu-jiexi-zixi.pages.dev/en/ · Chinese catalog: 总目录.html · How to publish: online-access.html
Interactive Vite stations and handwritten labs keep their executable JavaScript from the Chinese station so lab logic stays faithful to source. English lectures cover the same mechanisms with citations. Local preview: ./start-html-previews.sh, then open 127.0.0.1 (not localhost).

I want to…

Learn official courses / papers (not source)
🆕 95 items · Anthropic · Andrew Ng · Kaggle whitepapers · 36 papers · MLSysBook
Know where to start
Four-stage path · acceptance checks, not vibes
Understand Claude Code end to end
14 chapters · agent loop is chapter 6
Build an agent from scratch
Coding assistant / Manus / host / harness tracks
See the same problem, thirteen answers
Cache, stop, permissions, sandbox, compaction…
Compare architectures on one table
Ten dimensions · product position decides shape
Know what the boards actually measure
HumanEval → SWE-bench → GAIA / OSWorld · cost must be reported
See “everything is a plugin,” including the loop
241 packages · seam triangle · monotonic guards
See a host that never writes the loop
OpenClaw vertical host · content_guard exit 7 · commit 765f5a6
Open this on a phone
Cloudflare Pages · site branch · /en/

Learning maps

AI learning path complete list new
AI学习路径完整清单.html
95 official courses / papers / textbook chapters. Path A (why it works) · Path B (what it can do). Complements the graded source reading path. Chinese lecture; course titles already English.
Graded reading path
en/reading-roadmap.html
Stage 0–4. After each stage you should be able to produce something, not merely “feel that you get it.”
Agent evaluation and benchmarks
en/evaluation-and-benchmarks.html
Four layers. You get what you measure. Uncontrolled cost is not a ranking.
Further projects worth dissecting
en/further-projects.html
Twelve repos chosen because they contradict a decision you have already seen. Each names a source reading entry.
Catalog of agent engineering patterns
en/agent-engineering-patterns.html
Fifteen chapters. Same problem, different answers, with path:line back to the thirteen trees.
Architecture-diagram guide
en/architecture-diagram-guide.html
Six figure types. Every box and edge exists to support one load-bearing claim. Not a directory tree.
Online access
en/online-access.html
Merge to main → Actions publishes the site branch → Cloudflare hosts. English lives at /en/.

Interactive stations

Lab JavaScript is the Chinese station’s, so executable logic stays aligned with source. Use the English lecture beside each lab for the cited walkthrough.

Claude Code station v1 7300
/learn/claude-v1/
Three labs: prompt journey · agent loop · permission gate. Lecture: claude-code-deep-dive.
Claude Code station v2 7301
/learn/claude-v2/
Same content as v1, 18 chapters / 6 parts.
OpenManus × Suna station 7200
/learn/manus/
Two general-task agents. Lectures: analyses/openmanus, suna, openmanus-vs-suna.
Build-an-agent guide 7397
/learn/agent-guide/
Ten-chapter knowledge tree. Tutorial: tutorials/build-coding-assistant.
Open Design labs
open-design-解析.html
Four benches: adapter · prompt bands · anti-AI-taste linter · five-juror scoring. Analysis: analyses/open-design.
MiMo Code labs
MiMo-Code-解析.html
Goal judge · four loop gates · prompt router · BM25 floor · attribution vs OpenCode. Analysis: analyses/mimo-code.
OpenAI Codex labs
openai-codex-解析.html
Approval × sandbox truth table · AGENTS.md 32 KiB budget · run_turn stop protocol. Analysis: analyses/openai-codex.
MiroFish labs
MiroFish-解析.html
Ontology normalize · persona split · schedule curve · dual-world parameters. Analysis: analyses/mirofish.
Easel labs
Easel-解析.html
Four-layer prompt stack · Skill routing · outputs gate · outbound BLOCK/WARN. Frozen at 765f5a6. Analysis: analyses/easel.
CodePilot labs
CodePilot-解析.html
Three runtimes · session lock · Native while · Bridge formats · Rewind. Analysis: analyses/codepilot.
Pi labs
Pi-解析.html
Double while · leases · lazyStream · TUI diff · compaction retainedTail. Analysis: analyses/pi.
Architecture gallery
架构图库.html
Seven projects · nine figures · six types. Each figure states the question and the load-bearing claim first.

Claude Code track

How Claude Code works · deep dive flagship
en/claude-code-deep-dive.html
14 chapters. Chapter 6 is the loop. Tools, slash commands, MCP, effort, compaction, sub-agents—cited to Claude Code source.
How Claude Code works · primer archive
en/claude-code-primer.html
v1 text, superseded by the deep dive, kept as archive.
Tool inventory
en/reports/tool-inventory.html
43 built-in tools, one by one, with source paths.
Component inventory
en/reports/component-inventory.html
Plugins, skills, MCP, auth, effort.
Slash-command inventory
en/reports/slash-command-inventory.html
60+ commands, 12 dissected.

From-scratch tutorials

Build an AI coding assistant
MiniClaude in TypeScript. Parts list before you write the loop.
Build a general-task agent (Manus lineage)
Follow one real task. Each wall adds one part, including terminate.
Build an AI coworker (OpenWorker lineage)
Morning briefing to Slack. Out-of-band approval. Persist-and-resume.
Build an AI design agent (Open Design lineage)
You write zero lines of the agent loop. Socket + script + gates. 30 acceptance assertions.
Build a social-content workbench (Easel lineage)
15 steps against ZJU-REAL/Easel @ 765f5a6. Persona prefix, not global USER.md. content_guard exit 7.
Build a plug-in agent harness (DeepSeek Harness lineage)
20 steps, ~700-line minidsh. Do not write main. Service repo → inject → reversible effect → seam.
Build CodePilot
PRD → three swappable runtimes. Electron + Next.js.
Build Cline
Stateless AgentRuntime. 17 modules in build order. After the user hits enter.
Build Kanban
One card, one worktree. Hooks inject state. After the play button.
Build OpenAI4S
Code-as-Action scientific agent. Dual plane. Prompt from composer to bubble.

Source analyses

Each lecture names its frozen commit. path:line refers to that tree. English does not relocate citations.

OpenCode
Server-as-ontology. Effect stack. Teachable refusal.
hermes-agent
Prefix cache as sacred. Persist-before-act. 74 tools.
Raven
Spine scheduler. Curator lossless context. EverOS dual-track memory.
CodeWhale
Rust single binary. Safety by construction. Provenance downgrade.
goose
Everything is MCP. Engine-as-server (ACP).
nanobot
Explicit state machine. MessageBus. Dream. bwrap sandbox.
grok-build
xAI official CLI. Goal Mode kit. LLM classifier permissions.
OpenWorker
Inbox + durable resume. Out-of-band approval. Provider-agnostic on aisuite.
Open Design
Host, not engine. Adapter-as-data. Programmatic anti-AI-taste linter.
MiMo Code
Deep fork of opencode. Independent Goal judge. Chapter 20 is a layer-by-layer diff with byte counts.
OpenAI Codex
run_turn in turn.rs. Approval × OS sandbox as orthogonal walls. apply_patch as a first-class crate.
MiroFish
The series’ only non-coding agent. Graph ↔ simulation closed loop. Interview living agents.
CodePilot
Three swappable runtimes. Electron 40 + Next.js 16. IM bridges.
Pi
Nine-package monorepo. Double while. Runtime-extensible Extensions. CBOR wire protocol.
Cline
Five-layer monorepo. Stateless AgentRuntime. 28 tools. 179 provider IDs.
Kanban
One card = one detached worktree + one pty. Hooks reverse-inject state.
CodexMonitor
Official codex app-server JSON-RPC. Not PTY scraping. Four cross-checks against Codex source.
DeepSeek-Reasonix
Cache-first constitution. Compose turn tail. PrefixShape miss diagnosis. use_capability isolates MCP.
DeepSeek Harness
Everything is a plugin—including the loop. 241 packages. Capability-seam triangle. Monotonic guards. dsh-invariants.
DeepTutor
Switch the goal, keep the engine. L1/L2/L3 auditable memory. Subagent consults a coding CLI.
Easel
ZJU-REAL/Easel @ 765f5a6. Persona inlined in the message. 112 SKILL. SSE disconnect does not kill. Dual lock + uuid5.
OpenAI4S
JSON control plane + Python/R science plane. Ledger. 34 skills.
OpenManus
state_context rollback trap. terminate as a first-class exit. PlanningFlow.
Suna
session_id == sandbox_id == branch. Executor key boundary.
OpenManus vs Suna
Product axis, where the loop lives, MIT vs Elastic 2.0, which to pick.
Claude Code vs OpenCode vs Codex
Product sovereignty, permission lineage, tool ABI, instruction loading, session assets.
Comparative review
Ten dimensions. Eight duels. When the host sits at the table. When the agent stops doing work and starts evolving.

Reports

Liang Wenfeng investor meeting · views
Already English. Close to the transcript. No embellishment.
Liang Wenfeng technical ladder · industry research
LM → CoT → Agent → continual learning → gradual self-iteration → embodiment. P1–P5 evidence grades. Three falsifiers.
DeepSeek hiring · skill map
34 JDs from talent.deepseek.com. Ten tracks. Skill frequencies.
DeepSeek hiring page · design notes
Google Web quality-line diagnosis. P0–P2. The interactive reader is the Chinese page.
OpenAI4S analysis · verification notes
Cross-check of the OpenAI4S lecture against source. Hedges kept.
OpenWorker enterprise plan
How a traditional firm productizes the open-source coworker. Citations stay on the openworker tree.
No match for “”.
Try another term, or clear search.