Claude Code Digest — 2026-03-26 00:38:12
What the docs reveal
Collapsing the boundary between web and terminal
Anthropic is linking the browser to the terminal. The new /remote-control command (alias /rc) lets you drive local Claude Code sessions directly from the Claude.ai web or desktop application.
Previously, developers faced a hard choice: use the CLI for direct file system access, or use the web app for a richer UI and better artifact handling. This update merges those workflows. You can configure environments via the new /remote-env and /chrome commands. Anthropic views the CLI not merely as a standalone executable, but as a local backend execution engine for the broader Claude ecosystem.
Native PowerShell replaces Git Bash workarounds
Windows developers no longer need to rely on Git Bash. By setting CLAUDE_CODE_USE_POWERSHELL_TOOL=1, users enable native PowerShell execution. Claude Code automatically detects and prioritizes pwsh.exe (PowerShell 7+) over legacy system defaults. You can configure the shell preference globally in settings.json or target it specifically within command hooks and skills.
This remains an opt-in preview. Native PowerShell currently lacks Auto mode, sandboxing, and profile loading. Anthropic shipped an incomplete feature because forcing Windows users through Git Bash creates persistent path mapping and syntax errors. They traded safety and automated tooling for immediate execution accuracy on Windows and .NET stacks.
Context hygiene limits hallucination
Context management dictates model performance. Anthropic now explicitly instructs developers to keep CLAUDE.md under 200 lines. Monolithic configuration files bloat the context window, consume tokens, and degrade Claude’s reasoning capabilities.
To encourage modularity, Anthropic introduced the .claude/rules/ directory. You can scope specific instructions to targeted file types or subdirectories. If you manage multiple agents via AGENTS.md, you can now import it directly into CLAUDE.md using the @AGENTS.md syntax.
Anthropic also added a token-saving mechanism. The CLI parser now strips block-level HTML comments (<!-- comment -->) from CLAUDE.md. You can document your agent configuration for human contributors without burning context window space.
The shift toward a plugin ecosystem
Anthropic deprecated the native /review command. Developers must now install the code-review plugin from the marketplace.
This deprecation signals a strategic shift. Anthropic is moving workflow-specific functionality out of the core binary and into the community ecosystem. Core Claude Code will remain a lean execution framework. Coinciding with this shift, developers gain more control over plugin project structure. Plugin configuration now supports custom paths for commands, agents, skills, and output styles by overriding default directories.
Granular model routing for subagents
Subagents now accept a model parameter during invocation. The CLI resolves model selection in a strict cascade: environment variables override invocation parameters, invocation parameters override definition frontmatter, and frontmatter overrides the main conversation model.
This solves a major cost and latency issue. You can build multi-agent workflows that route simple documentation or linting subtasks to Haiku, reserving complex architectural refactoring for Opus or Sonnet.
Deep linking in VS Code
The new vscode://anthropic.claude-code/open URI handler opens a primary Claude Code tab in VS Code. It accepts prompt and session query parameters. You can programmatically launch VS Code, resume a specific session, and inject a pre-filled prompt via simple bash scripts or external dashboards. This functionality integrates Claude Code directly into automated developer workflows.