todai
Wednesday, 6 May
5 min read · 802 words

ACTION NEEDED

  • MCP server named "workspace" will be skipped in v2.1.128+ — If any of your MCP config entries use workspace as the server name, they'll be silently dropped with a warning after updating. Rename the server before running npm update -g @anthropic-ai/claude-code. https://github.com/anthropics/claude-code/releases/tag/v2.1.128

TODAY'S ITEMS

1. "AI Didn't Delete Your Database, You Did" — Accountability Essay Trending on HN

  • Ibrahim Diallo's essay reframes the viral Cursor/Claude production DB wipe: the problem wasn't the agent, it was a production delete endpoint with no auth friction, no confirmation step, and no staging restriction — then handing an agent access to it. The essay draws a direct parallel to a 2010 SVN trunk deletion and argues AI agent accidents follow the same human-infrastructure failure pattern as pre-AI mistakes.
  • Source: Hacker News (409pts, 218 comments)
  • Why it matters: Before your next Claude Code session, audit which APIs and shell commands your agent can reach — anything with production-delete access should be sandboxed, revoked, or gated behind a manual confirmation step.
  • Verified

2. context-mode — MCP Server That Claims 98% Tool Output Reduction

  • context-mode is an MCP server that sandboxes raw tool output to keep it out of the context window — a 315KB Playwright snapshot becomes 5.4KB, 59KB of GitHub issues becomes an indexed summary. It also tracks every file edit, git operation, and decision in SQLite with BM25 retrieval, so sessions pick up exactly where they left off after /compact without re-dumping history. Works with Claude Code via the plugin marketplace, plus Cursor, Windsurf, and 11 other platforms.
  • Source: GitHub
  • Why it matters: Install with npm install -g context-mode and run it on a project that regularly hits context limits — the tool-output sandboxing means significantly more work happens before /compact triggers.
  • Verified

3. Claude Code v2.1.128 — Worktree Bug Was Dropping Unpushed Commits

  • v2.1.128 (released May 4) fixes a significant bug where EnterWorktree branched from origin/<default-branch> instead of local HEAD, silently discarding any uncommitted or unpushed local work each time you entered a new worktree. Also: sub-agent prompt cache tracking reduces cache_creation tokens ~3×, parallel read-only tool calls no longer cancel on a single failure, and large stdin (10MB+) no longer crashes claude -p.
  • Source: GitHub Releases
  • Why it matters: If you use git worktrees with Claude Code for parallel sessions, update now — previous versions were silently discarding local uncommitted work every worktree branch.
  • Verified

NEW TOOL / PRODUCT SPOTLIGHT

  • context-mode — MCP server that sandboxes tool output and tracks session state in SQLite, solving both sides of the Claude Code context problem (input bloat + forgetting state after /compact). Install: npm install -g context-mode | https://github.com/mksglu/context-mode

PROMPT OF THE DAY

Add to your CLAUDE.md — Karpathy's four principles
for better LLM coding behaviour:

## How to Approach Tasks

**Think Before Coding**
- State assumptions explicitly; don't silently pick an
  interpretation when ambiguous
- If uncertain, ask rather than guess
- Push back if a simpler approach exists

**Simplicity First**
- Minimum code that solves the problem — nothing speculative
- No abstractions for single-use code
- If 200 lines could be 50, rewrite it

**Surgical Changes**
- Touch only what you must; don't improve adjacent code
- Match existing style, even if you'd do it differently
- If you notice unrelated dead code, mention it — don't
  delete it

**Goal-Driven Execution**
- Define success criteria before writing a line
- Write a test that reproduces the bug, then make it pass
- For multi-step tasks, state a brief plan with verify steps

Drop directly into your CLAUDE.md — addresses Karpathy's specific criticisms: silent wrong assumptions, overcomplicated code, orthogonal edits. Source: https://github.com/forrestchang/andrej-karpathy-skills (110K+ stars, from @karpathy's observations on LLM coding pitfalls)

LANDSCAPE NOTES