Documentation Index
Fetch the complete documentation index at: https://base.bangwu.me/llms.txt
Use this file to discover all available pages before exploring further.
Cursor
go-cursor-helpCursor 3: What Users Actually Want
Based on 431 community feedback responses (April 2026), the core insight is: users don’t just want “AI that writes code” — they want a stable, controllable AI development workbench. The top priorities are:- Agent + IDE seamless fusion — Agent Window is promising, but must preserve full IDE capabilities (LSP, debugging, task running, extensions, keybindings, code navigation, diff accept/reject). Users shouldn’t have to switch back to the old IDE for small operations.
- Worktree / Git / PR workflow productization — Developers want a low-friction, auditable Git console, not natural-language Git commands. Key needs: branch switching, PR detection, selective staging, multi-repo diff, submodule support, CI status, and review integration.
- Stability over features — Startup speed, memory usage, context loss, LSP failures, and large-repo indexing are trust killers. If the basics don’t work, users will switch to alternatives.
- Model cost transparency — Users want to know “which model is cheapest for this task?” not just more models. Good direction: Cursor suggests “use a cheap model here, upgrade to a strong model there.”
- Keyboard-first — The entire product must be operable without a mouse. Custom keybindings, quick panel switching, and inherited VS Code muscle memory are essential for heavy users.
- Task memory system — When agent work grows, “chat history list” must evolve into “task memory system”: auto-rename chats, pin messages, fork sessions, cross-project context.
Synthesized from Eric Zakariasson’s Cursor 3 feedback thread — 431 community responses, April 2026.
Workflow for studying unfamiliar LLMs
When you use an AI IDE to understand an open-weight model rather than just write app code, this order works well:- Read the official report first
- Use it to understand the model family, goals, and top-level architecture.
- Treat it as the overview, not the final source of implementation truth.
- Inspect the Hugging Face config
- Check layers, hidden size, attention heads, RoPE settings, MoE options, KV heads, and other structural fields.
- Config files often expose details that summaries skip.
- Read the reference implementation
- If the model is supported by
transformersor an official repo, inspect the real code path. - Working code is usually the most reliable way to confirm module order and implementation details.
- If the model is supported by
- Only then summarize or diagram it
- Once you have your own mental model, use the AI IDE to generate diagrams, compare related architectures, or explain local modules.
Best fit
- Open-weight models
- Models with public configs and reference implementations
- Less useful for closed models that only expose product-level descriptions
Reference
GitHub Copilot: Usage-Based Billing (June 2026)
Starting June 1, 2026, GitHub Copilot transitions from “premium request units” to AI Credits based on token consumption. This reflects Copilot’s evolution from a code-completion assistant into an agentic platform running long multi-step coding sessions.What stays the same
- Base prices unchanged — Pro 39/month, Business 39/user/month
- Code completions and Next Edit suggestions — Still included, do not consume AI Credits
- Monthly credits equal subscription price — Pro gets 39 credits
What changes
- Premium request units replaced by AI Credits — Usage calculated by input, output, and cached tokens at each model’s API rate
- No fallback to cheaper models — Previously, exhausted PRUs could fall back. Now credits spent = credits spent, no downgrade path
- Copilot Code Review — Consumes both AI Credits and GitHub Actions minutes
- Pooled credits for organizations — Business/Enterprise credits can pool across the team, reducing stranded capacity
Transition buffer (June–August 2026)
| Plan | Normal monthly credits | Transitional credits (June–Aug) |
|---|---|---|
| Business | $19/user | $30/user |
| Enterprise | $39/user | $70/user |
Practical implications
- Light users — Nearly unaffected; completions stay free
- Heavy Agent users — Monitor bills after June; long autonomous sessions consume credits fast
- Annual subscribers — Stay on old PRU model until plan expires, then transition to monthly with credits
- Preview billing — Early May: preview bill page shows projected costs before the switch
Coding Agent Safety: How OpenAI Controls Codex
When coding agents can read/write repos, run commands, and call development tools, you need both efficiency and control. OpenAI’s answer is a four-layer framework for Codex: sandbox + approval + network policy + identity governance. The guiding principle: zero friction for low-risk daily operations, mandatory human review for high-risk ones.1. Sandbox + Approval
- Sandbox defines technical execution boundaries: where to write, whether to allow networking, which paths are read-only
- Approval policy defines when to stop and ask a human — typically triggered when crossing sandbox boundaries
- Auto-review mode: A separate sub-agent reviews Codex’s pending actions and context, auto-approving low-risk requests and only interrupting the user when risk escalates. This is “AI reviewing AI” — turning approval itself into an intelligent layer
2. Network Access
Default-deny, explicit-allow model:- Allow known compliant destinations
- Block explicitly unwanted domains (e.g., pastebin.com — typical data exfiltration channel)
- Require approval for unknown domains
3. Identity & Credentials
- OAuth credentials for CLI and MCP are forced into OS keyring (macOS Keychain)
- Authentication via ChatGPT login — no bypass
- Locked to specific enterprise workspace UUID
4. Command Rules
Commands are classified by semantic risk level:- Read-only commands (
gh pr view,kubectl get) → auto-approve - Dangerous commands → explicit block or require approval
5. Configuration Distribution
Three-layer distribution ensures org-wide consistency:- Cloud managed requirements + macOS managed preferences + local requirements file
requirementsare admin-mandatory — users cannot override
Agent-Native Telemetry
Traditional EDR and audit logs tell you “what happened” — but for AI agents, security teams need “why”:- User’s original prompt
- Agent’s reasoning path
- Approval decisions
- Tool call chains
- MCP server usage
- Network proxy allow/deny events
- Expected agent behavior
- Benign mistake
- genuinely needs escalation
- OpenAI: Running Codex Safely — official blog post
- Twitter @dotey Chinese summary — detailed translation of the four-layer framework