Claude Code Digest — 2026-03-22 16:09:38
What the docs reveal
Automating CI/CD Pipelines
Anthropic rebuilt Claude Code for automated pipelines. The new --bare flag and CLAUDE_CODE_SIMPLE=1 environment variable strip the agent of local dependencies. Bare mode aggressively disables hooks, skills, plugins, MCP servers, and auto-memory. It bypasses keychain OAuth storage entirely, requiring standard API key ingestion.
Anthropic targets the unpredictability of local development setups. Local workspaces accumulate uncommitted settings, causing scripts to fail in CI/CD. Bare mode ensures Claude operates neutrally. Anthropic plans to make --bare the default behavior for -p prompt commands. This shifts the fundamental CLI design: single-shot operations become pure functions, immune to workspace drift.
Remote Approvals for Asynchronous Tasks
Claude Code now asks for permission through remote interfaces. The update introduces remote permission relays via external channels, backed by race-condition handling between the terminal and the remote device. Headless execution gains the --permission-prompt-tool flag and a brute-force --dangerously-skip-permissions absolute bypass.
Developers routinely abandon their desks during large refactors. Finding Claude paused on step two, waiting for permission to run a simple script, ruins productivity. Anthropic eliminates this friction. Remote relays allow developers to approve actions from a phone or chat bridge. The race-condition logic—accepting the first response and closing the pending prompt—demonstrates rigorous, real-world testing. The unattended mode flag confirms Anthropic finally trusts developers to run fully autonomous agents in sandboxed environments.
Removing Friction from Interactive Sessions
The interactive session gained keyboard shortcuts for mid-flight adjustments. Developers press Option+T (macOS) to toggle extended thinking, and Option+P to cycle the active model without clearing the current prompt.
Slash commands break flow state. When an agent stumbles on complex logic, developers need a stronger model immediately. Allowing users to toggle extended thinking mid-prompt transforms the CLI from a simple chat interface into a professional editor. Anthropic understands developers adjust compute resources based on immediate, granular needs.
Hardening Context Management
Anthropic formalized its context compaction strategy. The documentation instructs users to migrate persistent rules to the CLAUDE.md file to survive automated memory background purges. A new showClearContextOnPlanAccept setting enables users to flush the context window immediately after accepting a proposed plan.
Context pollution degrades agent accuracy. As task history accumulates, models lose focus and consume unnecessary tokens. The new clearing option forces a workflow shift: plan, approve, and reset. Moving project rules to CLAUDE.md secures constraints against aggressive background pruning. Protect your invariants in static files; treat dynamic memory as ephemeral.
Integrating Legacy Infrastructure The Model Context Protocol (MCP) integration now supports legacy servers. Claude Code accommodates MCP servers lacking Dynamic Client Registration through manual credential structures. Additionally, it auto-discovers endpoints utilizing the new Client ID Metadata Document (CIMD) standard.
Enterprise adoption requires integrating rigid, internal tools. Older corporate microservices ignore dynamic registration standards. Passing pre-configured credentials lowers the integration barrier for secure networks. CIMD discovery automates connection configuration everywhere else. Anthropic builds for the enterprise backend.