May 15, 2026Samarth at CLSkillsclaude code updatesclaude code 2.1claude code agent view

Claude Code This Week (May 8-15, 2026): Agent View, Fast Mode on Opus 4.7, and the terminalSequence Hook

Claude Code shipped seven releases between May 8 and May 15, 2026 (v2.1.136 through v2.1.142). The headline changes: a unified agent dashboard, Fast mode now defaults to Opus 4.7, hooks can emit desktop notifications, and Agent tool dispatch gained eight new flags. Here is what each one means in daily use.

Seven Releases in Eight Days

Between May 8 and May 15, 2026, Claude Code shipped versions 2.1.136 through 2.1.142. That is a release every ~1.1 days, which is the fastest sustained cadence the product has run since launch. The pattern from April held: a mix of headline features, quieter quality-of-life wins, and one or two changes that look small in the changelog but change daily behavior more than you expect.

This post covers what actually changed, what each release means if you use Claude Code every day, and the two updates I think are genuinely worth a workflow change.

TL;DR

v2.1.139 (May 11): claude agents command lands. Unified dashboard for every Claude Code session across projects (running, blocked, done). New /goal and /scroll-speed slash commands.
v2.1.140 (May 12): Agent tool subagent_type matching is now case- and separator-insensitive. The agent color palette got refreshed.
v2.1.141 (May 13): Hook JSON output gained a terminalSequence field, so hooks can fire desktop notifications, window titles, and bell sounds without owning a terminal. CLAUDE_CODE_PLUGIN_PREFER_HTTPS env var added for plugin cloning behind firewalls.
v2.1.142 (May 14): claude agents dispatch gained eight configuration flags. Fast mode now defaults to Opus 4.7 on every plan that has Fast access.
v2.1.136 (May 8): settings.autoMode.hard_deny for unconditional permission blocks, plus a fix for MCP servers vanishing after /clear.
v2.1.137 / 2.1.138 (May 9): Windows extension activation fix, internal fixes.
Net impact: If you only update for one feature, update for Agent View. If you use hooks, the terminalSequence field is the second reason. Everything else compounds in the background.

What Changed, Release by Release

v2.1.136 (May 8, 2026): Unconditional permission blocks and an MCP fix

The quiet-but-important change here is settings.autoMode.hard_deny. Auto mode in Claude Code is the routing logic that decides which tools and shell commands run without explicit user approval. The existing settings already let you allow or deny specific patterns, but the denials were soft: in some edge cases (especially when a tool was wrapped inside another tool, or when a plugin re-dispatched a sub-action), denial rules could be overridden by upstream auto-allows.

hard_deny is the override. Anything matching a hard_deny pattern is blocked unconditionally, regardless of what the rest of the auto mode logic decides. If you run Claude Code in a setup where some commands must never run automatically (production deploys, destructive git operations, anything that writes to a shared system), this is the right setting to use.

The MCP fix in this release is mundane but real: MCP servers were sometimes vanishing from the active server list after a /clear command. They were still connected at the process level but invisible to the model. Fixed.

Also shipped: CLAUDE_CODE_ENABLE_FEEDBACK_SURVEY_FOR_OTEL, which lets organizations using OpenTelemetry pipe Claude Code feedback prompts into their telemetry stack instead of the default in-CLI flow. Niche for individual users, useful for enterprise rollouts.

v2.1.137 and v2.1.138 (May 9, 2026): Windows fix and internal cleanup

Two small releases on the same day. The Windows VS Code extension was failing to activate on certain machine configurations, especially ones running with non-default user profile paths. Fixed in 2.1.137. The 2.1.138 release contains internal fixes that were not enumerated in the public changelog, which usually means dependency bumps, test infrastructure, or behavior changes that did not warrant user-facing notes.

Nothing in either release changes how you use Claude Code day to day.

v2.1.139 (May 11, 2026): Agent View lands

This is the big one.

Run claude agents and you get a single live list of every Claude Code session you have running, across every project on the machine. The list shows which sessions are actively working, which are blocked waiting on your input, and which finished. You can attach to any of them from this view.

Why this matters: a lot of Claude Code users (especially anyone running long agent workflows, multi-project setups, or background-dispatched sub-sessions) ended up with five or six terminal windows open just to keep track of what was running where. The cognitive cost of switching between them, remembering which directory each was rooted in, and figuring out which one was blocked on a permission prompt, was real. Agent View collapses that into one command.

The other two additions in this release are smaller but worth knowing:

/goal lets you set a completion condition for the current session. Instead of "work until I tell you to stop," you can declare "work until tests pass," "work until this file compiles," or any other condition expressible in plain text. Useful for unattended runs.
/scroll-speed changes how fast text scrolls in fullscreen mode. Trivial-looking, but if you are reading long Claude output and the default scroll feels too fast (or too slow), this is the lever.

Agent View is flagged as a Research Preview, which Anthropic uses for features that are stable enough to ship but may change shape based on usage. Expect refinements in the next two or three releases.

v2.1.140 (May 12, 2026): Agent tool matching, refreshed palette

If you have ever called the Agent tool with subagent_type: "Explore" and gotten an error because the tool definition spelled it explore or Explore-Agent or some other near-match, this release fixes that. Matching is now case-insensitive and separator-insensitive, so Explore, explore, Explore-Agent, and explore_agent all resolve to the same subagent if the underlying agent registry intends them to.

This is a small but real source of bugs in custom agent configurations. The error messages when the match failed were rarely clear about why, and most users solved it by trial-and-error.

The color palette refresh is purely cosmetic. The colors agents render in (used for differentiating which subagent is talking in multi-agent flows) are now slightly more distinguishable in dark mode and more accessible to users with red-green color vision differences. No behavior change.

v2.1.141 (May 13, 2026): terminalSequence hooks and HTTPS plugin clones

Two additions in this release that matter for different audiences.

The terminalSequence field on hook JSON output is the bigger one for power users. Hooks in Claude Code are scripts that run at specific lifecycle points (pre-tool-call, post-tool-call, on session end, etc.). Until this release, the only way a hook could communicate with the user was by writing to stdout, which assumes the hook is running in a terminal that the user is actively watching. That assumption breaks the moment you dispatch a Claude Code agent in the background.

terminalSequence lets hooks emit raw terminal escape sequences from their JSON output. The practical uses are:

Desktop notifications. A hook can now ring the user with a system-level notification when a long-running session completes, without owning the terminal.
Window title updates. A hook can update the terminal window or tab title with current session status, so you can glance at your taskbar to see which sessions are doing what.
Bell sounds. Any hook can emit the BEL character via this field to audibly signal a state change.

If you have been running multiple Claude Code sessions in the background, this is the feature you wanted. Combined with Agent View from 2.1.139, the entire experience of managing background Claude Code work just got dramatically less stressful.

The second addition, CLAUDE_CODE_PLUGIN_PREFER_HTTPS, is narrower. By default, Claude Code clones plugin repositories from GitHub using SSH. If your network blocks SSH outbound (common in corporate environments), plugin install would fail in a way that was not always obviously about SSH. Set this env var to 1 and plugin cloning falls back to HTTPS. Useful in exactly the situations where you would otherwise have spent an hour debugging it.

The /feedback command was also extended to optionally include recent session context when you send feedback. This makes feedback reports more useful to the Claude Code team and is opt-in per submission.

v2.1.142 (May 14, 2026): Eight new agents flags and Fast mode on Opus 4.7

The headline change here is operational: Fast mode now uses Opus 4.7 by default on every Claude Code plan that has Fast access.

Fast mode is the toggle that prioritizes lower-latency output over deeper reasoning for the current session. Until this release, Fast mode routed to Sonnet by default on most plans. With Fast mode on Opus 4.7, the latency tradeoff is now between a less-aggressive Opus and a more-aggressive Opus, instead of between Opus and Sonnet. The practical result is that Fast mode no longer feels like a quality drop. If you toggled Fast off in the past because output quality dipped too much, give it another try now.

The claude agents dispatch flags are the second story. When you dispatch a background agent (the agents tooling that landed with Agent View), the new flags let you configure that session's environment without writing a wrapper script. The full list:

--add-dir to add additional working directories to the session
--settings to override settings.json for that one session
--mcp-config to override MCP server config
--plugin-dir to load a specific plugin directory
--permission-mode to set the permission mode (default, accept-edits, bypass-permissions, plan)
--model to pin a specific model for that session
--effort to set reasoning effort (default through xhigh)
--dangerously-skip-permissions for sessions that should run without prompting at all

This is the kind of release that does not affect casual users but unlocks new workflows for anyone running Claude Code as part of automation. You can now write CI/CD jobs that dispatch a Claude Code session with exactly the permission and tool surface you want, without baking it into a settings file the rest of your team has to inherit.

The Two Updates That Are Actually Worth Changing Your Workflow For

Most releases in this cluster are incremental. Two are not.

Agent View (v2.1.139) changes how you keep track of multiple sessions. If you have been running one Claude Code window because juggling multiple felt like more overhead than value, that calculation just flipped. Open four or five sessions, dispatch some to background work, and use claude agents as your dashboard. The cognitive overhead is real but it is now manageable.

Fast mode on Opus 4.7 (v2.1.142) is the one I expect most users to feel without realizing why. Fast mode used to be a sacrifice of quality for speed. Now it is a sacrifice of reasoning depth for speed within the same model class. The output for most daily tasks (file edits, debugging, writing test cases) is materially better in Fast mode now than it was in early May.

What This Cadence Tells Us

A release every 27 hours, on average, for eight days straight. That is not the cadence of a product on autopilot.

Three things stand out:

The agent surface area is hardening into a real product. Agent View, the claude agents flags, the subagent matching improvements, and the terminalSequence hooks are all converging on the same goal: making it possible to run Claude Code as a fleet, not a single REPL. Up through April, the agent features felt like a research direction. From May 11 onward, they look like the core product roadmap.

Hooks are getting first-class treatment. The terminalSequence field is a small change with big implications. It signals that hooks are now expected to be a primary way teams customize Claude Code behavior, not just an escape hatch.

Enterprise readiness is moving fast. hard_deny, CLAUDE_CODE_ENABLE_FEEDBACK_SURVEY_FOR_OTEL, the HTTPS plugin fetch, the per-session permission mode flag. None of those features matter to a solo developer. Every one of them removes a blocker for a team that needs to roll Claude Code out across an org with security review.

Practical Upgrade Guide

Upgrade is a one-liner:

npm install -g @anthropic-ai/claude-code@latest

Or on macOS:

brew upgrade claude-code

There are no breaking changes between 2.1.135 and 2.1.142. Your settings.json, CLAUDE.md files, plugin configs, and hooks all continue to work.

Two recommendations once you are on 2.1.142:

Run claude agents once. Even if you only ever run one session at a time, the agents view is now where the product is heading. Familiarize yourself with the command before you need it.

Try Fast mode again. If you turned Fast off in the past, give it three days on the default. The Opus 4.7 routing is materially different from the Sonnet routing it replaced.

What Is Not In These Releases

A few items that have been rumored or partially shipped but are not in this cluster:

Native plugin marketplace. Still no first-party plugin directory. Plugins are installed from URLs, GitHub repos, or local paths.
GitHub Actions integration. Still third-party only.
Claude Cowork integration with Claude Code. Discussed since Q1, still not shipping.
Native Apple Silicon optimization pass. The native binary from April helped Windows and Linux more than ARM64 macOS. The deeper macOS optimization is still pending.

Honest Take

If you have to skip a release cycle (because you are on a regulated machine and updates are slower than monthly), 2.1.142 is the one to land on. It is the cleanest snapshot of the entire May cluster, and it contains the two updates I think actually move the needle.

The larger story is that Claude Code's release cadence is accelerating, not slowing. Products that lose direction usually slow their releases first. Claude Code is speeding up while shipping increasingly substantive features. That is the pattern of a product where the team has a clear roadmap and is being given the resources to execute on it.

The one thing I would watch is permission complexity. Between auto mode, hard_deny, the new per-session permission flags, and the existing permissions config, the surface area for getting your security posture wrong is growing. The defaults are sensible, but if you are configuring Claude Code for a team, take an extra hour to write down what your permission posture is supposed to be and verify the config matches.

Sources

Questions about Agent View, the new hooks, or how to actually use the per-session permission flags in CI? Reply on the newsletter and I answer every email.

Want all 160+ tested prompt codes?

Lifetime updates, before/after output for every code, indexed for quick ctrl-F.

PayPal (cards, Apple Pay, Google Pay) · Lifetime updates · 30-day refund