← Claude Docs

Week 19 · May 4–8, 2026 - Claude Code Docs

Claude Docs · May 8, 2026
Claude Code v2.1.128 through v2.1.136 introduce the --plugin-url flag, allowing plugins to be loaded directly from URLs for testing before marketplace distribution. Ctrl+R reverse-search now defaults to all prompts across projects rather than limiting to the current session, restoring the ability to find commands run in other repositories. The update includes new configuration options for worktree handling, enhanced credential reliability for OAuth and MCP servers, performance optimizations for sub-agent caching, and multiple infrastructure improvements.

Detailed Analysis

Claude Code's May 4–8, 2026 release cycle (versions v2.1.128 through v2.1.136) introduces a cluster of enhancements centered on extensibility, developer ergonomics, and operational reliability. The most prominent additions involve the plugin system: the `--plugin-dir` flag now accepts `.zip` archives in addition to directories, and a new `--plugin-url` flag allows developers to fetch a plugin archive directly from a URL for the duration of a session. This architectural shift meaningfully lowers the friction for plugin experimentation and internal distribution, enabling teams to ship plugins from artifact stores without requiring marketplace publication. Alongside this, the reverse-search behavior (`Ctrl+R`) has been reverted to its pre-v2.1.124 default, restoring cross-project history lookup while still allowing users to narrow scope to the current project or session via `Ctrl+S` — a pragmatic resolution to a regression that had disrupted muscle memory for long-term users.

Several additions this week reflect a broader push to give developers and administrators finer-grained control over Claude Code's behavior in automated and enterprise contexts. The new `settings.autoMode.hard_deny` rules introduce an unconditional block mechanism that overrides even explicit allow exceptions, addressing a meaningful gap in the permission model for auto mode workflows where certain actions must be categorically prohibited regardless of other policy configurations. Separately, the new `worktree.baseRef` setting (`fresh | head`) gives teams control over whether new worktrees branch from the remote default branch or local HEAD, with the default `fresh` behavior deliberately isolating worktrees from unpushed commits — a design choice that promotes cleaner agent-isolation patterns. A new admin key, `parentSettingsBehavior`, further extends enterprise configurability by allowing SDK `managedSettings` to be opted into policy merges, signaling continued investment in deploying Claude Code within governed organizational environments.

Performance and observability improvements round out the release. Sub-agent progress summaries now hit the prompt cache, reducing `cache_creation` token costs by approximately threefold — a significant efficiency gain for workflows that rely heavily on parallel or nested agent execution. Hooks now receive the active effort level through both a JSON input field (`effort.level`) and the `$CLAUDE_EFFORT` environment variable, and `CLAUDE_CODE_SESSION_ID` is now consistently available in Bash tool subprocess environments, enabling more robust logging and tracing integrations. The deliberate isolation of `OTEL_*` environment variables from subprocesses such as Bash, hooks, MCP, and LSP prevents the CLI's own telemetry endpoint from leaking into user-run applications — a subtle but consequential fix for developers running OTEL-instrumented services through the Bash tool.

OAuth and credential reliability receives notable attention this week, with fixes targeting three distinct race conditions: parallel sessions deadlocking at 401 after a refresh-token race, concurrent MCP OAuth token refreshes causing token loss across multiple servers, and a rare login loop triggered by simultaneous credential writes. These fixes suggest Claude Code is being used at increasing scale in multi-session and multi-server configurations where concurrency edge cases have become practically significant. The `/mcp` interface improvement — now displaying tool counts per connected server and flagging servers that connected with zero tools — further supports operational visibility in complex MCP deployments. Taken together, the week's updates trace a consistent trajectory: Claude Code is maturing from a powerful individual developer tool into infrastructure capable of supporting sophisticated, policy-governed, multi-agent workflows at organizational scale.

Read original article →