Wednesday, April 1, 2026

Daily picks

articles scored

#1 GOLDReleaseClaude Code Releases (GitHub)

v2.1.89

Claude Code v2.1.89 lands with 20+ changes — the headline is a new `"defer"` value for PreToolUse hooks: headless sessions can now pause mid-task at any tool call and resume later with `-p --resume`, so long automation pipelines can checkpoint instead of running all-at-once or stopping cold
`CLAUDE_CODE_NO_FLICKER=1` is a new env var for flicker-free alt-screen rendering — if you're running Claude Code in tmux or over SSH and seeing visual glitches, set this and they go away
Auto mode now shows denied commands in a `/permissions` → Recent tab where you can retry with `r` — previously blocked actions were invisible; now you can see and re-evaluate them
Windows users get two critical fixes: Edit/Write were doubling CRLF line endings and stripping Markdown hard line breaks (two trailing spaces), silently corrupting files — update now if you're on Windows
`/stats` now correctly counts subagent token usage and preserves data past 30 days — your token history was likely undercounting the whole time

#2 SILVERReleaseGitHub anthropics/claude-code-action

The Claude Code GitHub Action hits v1.0 GA — the old `mode`, `direct_prompt`, and `override_prompt` inputs are replaced with a single `prompt` input, and all CLI options now flow through `claude_args`, making setup dramatically simpler
The action now auto-detects interactive vs automation mode (no more manual configuration), and adds native AWS Bedrock and GCP Vertex AI support if you want to route away from the Anthropic API directly
Breaking changes from v0.x — review the migration guide before upgrading production workflows

#3 BRONZEReleaseGitHub anthropics/claude-agent-sdk-typescript

New `startup()` function pre-warms the CLI subprocess before your first `query()` call — pay the cold-start cost upfront and get ~20x faster first responses when you control the timing
New `listSubagents()` and `getSubagentMessages()` functions expose what subagents did inside a session — finally lets you inspect multi-agent execution history programmatically