Claude Code Digest — 2026-03-20 04:38:30
What the docs reveal
The End of Hanging Exits
Closing Claude Code should return you to the prompt instantly. When developers wrote complex SessionEnd hooks to clean up resources, export logs, or upload telemetry, sessions hung. Anthropic fixed this by imposing a strict 1.5-second execution ceiling via CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS. This hard limit applies to manual exits, /clear commands, and session switching. If your teardown takes longer, the process dies. You must optimize your cleanup scripts or move heavy tasks to asynchronous background processes.
The update also exposes a new termination reason: bypass_permissions_disabled. This signals a hard abort when an agent attempts to evade safety constraints or lacks required execution rights. Anthropic prioritizes secure termination over graceful execution.
Surgical Compute Allocation
Running every command at peak system intelligence burns tokens and time. Anthropic now allows you to assign specific effort levels directly in skill and subagent frontmatter. You can instruct Claude Code to use Opus 4.6 at max effort for writing complex algorithms, while throttling a simple linter subagent down to low.
The system enforces a strict hierarchy for this compute: environment variables veto frontmatter, frontmatter overrides workspace configurations, and everything falls back to model defaults. Assigning proportional intelligence to specific tools puts you in control of your API burn rate. You should audit your subagents and downgrade the effort level on deterministic tasks.
Hot-Reloading the Developer Loop
Restarting the CLI to test a minor plugin tweak kills momentum. The new /reload-plugins command and automatic refresh system eliminate this friction. Claude Code now actively watches file modifications and updates plugins, skills, hooks, and MCP/LSP servers on the fly.
To support this sprawling ecosystem, the interface now displays exact counts for loaded capabilities. Anthropic understands that developers are treating Claude Code as a platform extending into their whole environment. This update transforms plugin authoring from a tedious restart loop into a fluid, live-reload experience.
Securing the Subagent Ecosystem
The plugin ecosystem is maturing, bringing inevitable supply-chain threats. Anthropic now enforces a strict naming policy for third-party marketplaces. Operators can no longer use names like claude-code-marketplace or mimic official Anthropic services. This anti-impersonation rule prevents phishing and protects developers from installing malicious code disguised as native integration. Allow-listing for extensions now also supports regex for hostPattern, giving enterprise teams fine-grained control over accepted plugin sources.
Local execution security tightens concurrently. The custom statusLine feature now demands explicit workspace trust. Since status lines often execute shell commands to fetch git branches or container states, they represented a silent remote code execution vector in untrusted repositories. If you decline workspace trust, the statusline disables itself and displays a "skipped" notification. Anthropic accepts visible execution friction to close invisible vulnerabilities.
Persistent Subagent Memory
Background agents now hold state. The new memory field for JSON-configured --agents allows you to set up persistent memory across invocations. The documentation also standardizes the terminology, dictating that the JSON prompt field functions exactly like the markdown body in file-based agents: it serves as the absolute system prompt. Equipping subagents with enduring memory pushes Claude Code away from simple scripts toward autonomous, long-running project assistants that retain context without saturating your primary session window.