Claude Code Digest — 2026-05-08 00:20:53
Version updates
2.1.133 introduces the worktree.baseRef setting, giving you control over where --worktree and isolated agents branch from. Anthropic reverted the default behavior to fresh (branching from origin/<default>) rather than local HEAD.
This rollback points to a common developer friction point: context pollution. When an agent branches from a local HEAD, it inherits unstaged experiments and broken states. Forcing agents to branch from a pristine origin guarantees repeatable execution. If you rely on local uncommitted code to inform your subagents, you must explicitly configure this setting back to head.
What the docs reveal
Taking back the terminal
Anthropic pushes its fullscreen terminal user interface; developers demand their standard scrollback buffers. The new CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN environment variable bridges this gap.
Setting this variable strips out the alternate screen renderer. You regain native terminal text search (Cmd+f), tmux copy mode, and standard scrollback history. For developers who live inside multiplexers, this removes the frustrating barrier between Claude Code and the surrounding shell environment. Furthermore, JetBrains IDE users receive explicit scroll-wheel mitigations, confirming that Anthropic is patching editor-specific TUI edge cases as they arise rather than waiting for IDE vendors to fix terminal emulation.
Context window economics
The status line finally tells the truth about your context window. Previously, token counts displayed cumulative session totals. Developers saw massive token counts and assumed they were hitting hard limits, prompting unnecessary context flushes. Now, total_input_tokens accurately reflects active context window usage—aggregating fresh input, cache creation, and cache read tokens. If the number spikes, your immediate context is bloated.
Tool development receives a corresponding token-saving upgrade. You can use the new structuredContent field in custom MCP tools to feed machine-readable JSON directly to the model. This keeps JSON payloads out of standard text input blocks, preventing context duplication and saving tokens. Note that the Python @tool decorator lacks this feature; you must build a standalone MCP server to use it.
Agent isolation and longevity
Background agents require guardrails to prevent runaway API costs. The new CLAUDE_ASYNC_AGENT_STALL_TIMEOUT_MS environment variable defaults to 10 minutes. If a subagent makes no progress within that window, the process aborts and surfaces partial results. You stop burning tokens on infinite command-line loops.
The Python and TypeScript SDKs indicate a shift toward persistent, stateful agent architectures. Developers gain precise control over reasoning cycles via the new effort configuration. You get deep session state management through fork_session and session_store. Instead of funneling every task through a single linear history, you can now branch conversations mid-stream, write transcripts to external databases, and update flag settings (like model selection) mid-session without dropping the current task.
Untangling enterprise workflows
Corporate network topography routinely breaks default LLM configurations. The documentation reveals several new escape hatches for enterprise environments.
You no longer need to rely on the official marketplace for plugins. You can install tools directly from internal GitLab or Bitbucket servers by appending .git to the repository URL. To secure these custom workflows, you can enforce the new httpHooksAllowedDomains setting. This wildcard-supported allowlist prevents compromised scripts or hallucinated bash commands from exfiltrating data to external domains via HTTP hooks.
Finally, Anthropic simplified the ultimate fallback: collaborative debugging. When configuration breaks down, running /doctor provides system diagnostics. Pressing f instantly pipes that raw diagnostic output into Claude. You let the model fix its own environment.