Tuesday, April 28, 2026
Daily picks
20
articles scored
#1 GOLDReleaseClaude Code Releases
v2.1.121
- New `alwaysLoad` option on MCP server config: set it to `true` and all tools from that server skip tool-search deferral and are always immediately available — no more hunting for tools that should just be there
- `claude plugin prune` removes orphaned auto-installed plugin dependencies; `plugin uninstall --prune` cascades the cleanup automatically
- **PostToolUse hooks can now replace tool output for every tool**, not just MCP tools — `hookSpecificOutput.updatedToolOutput` is now universal, which opens up a huge range of output-transformation use cases
- Two major memory leak fixes: unbounded multi-GB RSS growth when processing many images, and up to ~2GB leak on `/usage` with large transcript histories — both are now gone
- Fullscreen mode gets several quality-of-life fixes: typing no longer jumps scroll position, long dialogs are scrollable, wrapping URLs are now fully clickable
#2 SILVERGuideDev.to Claude
Claude Opus 4.7: Effort Controls and Migration Guide
- **Breaking change**: Opus 4.7 rejects `temperature`, `top_p`, and `top_k` with a 400 error — if you swap the model string in an existing Opus 4.6 integration without removing these params, it silently breaks at runtime
- `thinking.budget_tokens` is also gone: replace with `thinking: {type: "adaptive"}` and the model manages its own reasoning budget based on task difficulty
- New `effort` levels (`low`, `medium`, `high`, `xhigh`, `max`) replace manual sampling tuning — `xhigh` is now the Claude Code default across all tiers
- New task budgets beta (`task-budgets-2026-03-13` header) lets you cap total tokens across an entire agentic loop so the model finishes gracefully instead of running out mid-task
#3 BRONZETutorialDev.to Claude
How to Build a Self-Verification Loop in Claude Code (3 Layers, 20 Minutes)
- Claude Code's Stop hook blocks the agent from declaring "done" until verification passes — pair it with PostToolUse feedback injection to build a loop the agent can't escape until the code is actually correct
- Three layers: syntax check on every file change (PostToolUse command), intent verification against the original request (Stop prompt/agent), and regression test suite (Stop command)
- The third layer is often the only one teams have; Layer 2 catches the case where tests pass but Claude built the wrong thing entirely