Claude Code Digest — 2026-03-18 04:43:50
Version updates
Anthropic released version 2.1.78 to address state persistence and runaway execution. The addition of the ${CLAUDE_PLUGIN_DATA} variable fundamentally changes the plugin architecture. Previously, updating a plugin risked wiping local configuration. Now, plugins store persistent state—like cached syntax trees or authentication tokens—in a durable directory. The CLI prompts the user for confirmation before deleting this data during an uninstall. Plugins now behave less like utility scripts and more like isolated applications.
The release introduces a StopFailure hook. This event fires when API errors, such as rate limits or invalid authentication headers, kill a turn. Developers running Claude Code in continuous integration or long-running background tasks previously lacked visibility into silent headless crashes. You can now catch StopFailure to execute alert webhooks, initiate exponential backoffs, or revert partial file changes immediately.
Finally, Anthropic added effort and maxTurns parameters. Autonomous agents occasionally lock into infinite retry loops when a test fails repeatedly, burning through context tokens. These new parameters establish hard boundaries on execution. Developers dictate the exact limits of a task before the model racks up the bill.
What the docs reveal
Documentation changes in settings.md and sub-agents.md reveal an aggressive pivot toward a routed, multi-agent architecture. Anthropic is deprecating the monolithic system prompt. Claude Code now operates as a dispatcher for specialized subagents. You configure each subagent with unique system prompts, distinct AI models, and tightly restricted toolsets.
Users invoke these subagents through two distinct workflows: ad-hoc delegation or persistent session configuration. For isolated tasks, developers use the new @-mention syntax to hand off work dynamically mid-thread. For dedicated workflows, you assign an agent to an entire session via the --agent CLI flag or .claude/settings.json. Session assignment replaces the default context entirely. A backend engineer can configure their terminal to exclusively use a database-oriented agent, and this explicit configuration persists securely across session resumes.
Anthropic linked this multi-agent architecture directly to the plugin ecosystem. Developers access third-party agents using scoped syntax: @<plugin-name>:<agent-name>. This feature lays the groundwork for an agent integration marketplace. Instead of loading every installed tool into a generic Claude context, developers will install expert plugins. You will ping @docker:builder to arrange your containers, then hand the resulting environment to @playwright:tester to run end-to-end verifications.
This architectural shift resolves severe context bottlenecks. Squashing dozens of tools into a single prompt degrades model reasoning, spikes latency, and wastes API output tokens. By isolating tools and context into strict subagent boundaries, Anthropic forces prompt economy. You provide the model exactly what it needs for the assigned task, and absolutely nothing else.