Saturday, March 21, 2026
Daily picks
35
articles scored
#1 GOLDReleaseClaude Code Releases
Claude Code v2.1.81
- New `--bare` flag makes Claude Code safe for scripted/automation pipelines: it skips hooks, LSP, and plugin sync so `-p` calls run cleanly without interactive interruptions — requires an API key (no OAuth/keychain in bare mode)
- Multi-session OAuth fix: if you're running several Claude Code windows at once, one session refreshing its token no longer forces all the others to re-authenticate
- `--channels` permission relay expanded: channel servers that declare the permission capability can now forward tool approval prompts to your phone (approve/deny file writes, bash commands, etc., without being in the terminal)
- Voice mode gets two reliability fixes: silent retry failures now show the real error instead of "check your network", and audio recovers when the server silently drops the WebSocket connection
- MCP read/search tool calls now collapse into a single "Queried {server}" line (expand with Ctrl+O), cutting transcript clutter; plugin hooks no longer block prompt submission if the plugin directory is deleted mid-session
#2 SILVERReleaseGitHub anthropics/claude-code-action
Claude Code GitHub Action v1.0 — Now Generally Available
- The Claude Code GitHub Action officially hits v1.0 GA with a cleaned-up API: the old `mode`, `direct_prompt`, `override_prompt`, `custom_instructions`, `allowed_tools`, and `mcp_config` inputs are gone — everything now flows through a single `prompt` input and `claude_args` (matching the CLI format)
- Automatic mode detection means you no longer configure interactive vs. automation mode manually — the action figures it out from your setup
- Breaking changes from v0.x require migration — check the migration guide before upgrading. New example workflows cover automated PR reviews, CI failure auto-fix, issue triage, documentation generation, and security scanning
#3 BRONZEReleaseGitHub anthropics/claude-agent-sdk-python
claude-agent-sdk-python v0.1.49
- `AgentDefinition` now accepts `skills`, `memory`, and `mcpServers` fields — you can define agents with full Claude Code infrastructure (skill files, memory, MCP servers) directly in Python code rather than wiring it separately
- Per-turn token usage now preserved on `AssistantMessage`, new typed `RateLimitEvent`, and programmatic `tag_session()` / `rename_session()` — better observability and organization for multi-session agent runs
