Claude Code releases 2.1.149 with `/usage` breakdown and PowerShell cwd fix
Claude Code 2.1.149 added `/usage` cost breakdowns and fixed a PowerShell working-directory bypass, sandbox issues in git worktrees, and macOS file-table exhaustion from `find`. Anthropic also expanded auto mode to Pro plans and Sonnet 4.6 in the same update window, so users should check their available modes.

TL;DR
- Claude Code 2.1.149 shipped 26 CLI changes, and ClaudeCodeLog's release post centered three of them: a new
/usagebreakdown, a PowerShell working-directory fix, and a Bashfindfix for macOS host crashes. - According to the 2.1.149 changelog thread, the most serious fixes were in permission handling: built-in PowerShell
cdforms could move the cwd without detection, git worktree sandbox rules were too broad, and stale directory variables could confuse permission analysis. - Boris Cherny's preview post teed up the new
/usageview a day earlier, and the shipped changelog says it now breaks limits usage out by skills, subagents, plugins, and per-MCP-server cost. - In the same update window, ClaudeDevs' auto mode update expanded auto mode to Pro accounts and added Sonnet 4.6 support alongside Opus 4.7, with the permission modes docs linked from the thread.
- Less than 10 hours after 2.1.149, ClaudeCodeLog's 2.1.150 post said Claude Code switched its search guidance to ripgrep-based
grep, while the follow-up inventory also noted removal of aclaude-empty-model slug.
You can jump straight to the official 2.1.149 changelog, the auto mode docs, and the 2.1.150 changelog. The weirdly useful detail is that the new cost view is not just token totals, it is split by skills, agents, plugins, and MCP servers per Boris Cherny's preview and the shipped notes. The security-adjacent detail is that two of the 2.1.149 fixes were really about the agent's idea of where it was allowed to be in the filesystem, per the changelog thread.
Usage gets a real cost breakdown
The new /usage screen is the most immediately legible change in 2.1.149. Boris Cherny's preview post described it as a breakdown of which skills, agents, MCPs, and plugins are using tokens, and the release changelog says the shipped version surfaces per-category limits usage, including per-MCP-server cost.
That turns Claude Code's limits view into a cost-driver view. Instead of one aggregate number, 2.1.149 breaks usage into:
- Skills
- Subagents or agents
- Plugins
- Per-MCP-server cost
LLMpsycho's summary post picked the same item as the headline feature, which matches the emphasis in both the release thread and the official changelog.
Permissions and sandboxing
The densest part of 2.1.149 is the permission cleanup. According to the changelog thread, Anthropic fixed four separate cases where filesystem boundaries or approvals could be interpreted too loosely.
- Built-in PowerShell
cdforms likecd..,cd\,cd~, and drive switches could change the working directory without detection, which let a later command read outside the workspace. - The sandbox write allowlist for git worktrees covered too much of the main repository root, instead of only the shared
.gitdirectory. - PowerShell prefix and wildcard allow rules did not always pre-approve native executables and scripts correctly.
- Permission analysis could trust stale
PWD,OLDPWD, andDIRSTACKvalues acrosscd,pushd, andpopd.
The same release also fixed a host-stability bug that the release post described in plainer terms: Bash find could exhaust the macOS file and vnode table while scanning large trees. One nearby report from koltregaskes' Windows complaint shows users were already hitting update-related stability pain on desktop, although that post predates 2.1.149 and does not claim the same root cause.
Auto mode expands to Pro and Sonnet 4.6
The other shipping change in this window was not inside the 2.1.149 changelog at all. ClaudeDevs' auto mode update says auto mode is now available on the Pro plan, and that Sonnet 4.6 joined Opus 4.7 as a supported model.
The thread points to Anthropic's permission modes docs, and the linked follow-up post told users to update either the CLI or desktop app to see it. That matters because Boris Cherny's /usage preview explicitly said the new breakdown was in CLI first, with desktop following later in his preview post. In practice, this update window split across surfaces:
/usageappeared in Claude Code CLI first, with desktop still catching up, per Boris Cherny's preview.- Auto mode was opened to Pro and updated for Sonnet 4.6, per ClaudeDevs' product post.
- Anthropic told users to update the CLI or desktop app to get the new behavior, per the follow-up instruction.
2.1.150 arrives less than 10 hours later
Claude Code's release cadence stayed twitchy. the follow-up inventory says 2.1.150 landed 9 hours, 56 minutes after 2.1.149, and the release post framed the visible change as a search-tool update: tool descriptions now use ripgrep-based grep for search tasks, with "more consistent and faster results."
The official 2.1.150 changelog is terse and calls the release "internal infrastructure improvements," but the additional changes post adds two concrete details that are easy to miss:
- The CLI surface removed a
claude-empty-model entry. - Prompt tokens increased again, rising another 1,946 tokens, with tool-token share moving from 72.7% to 73.3%.
That makes 2.1.149 look like the user-facing release, and 2.1.150 the immediate cleanup pass that quietly changed search guidance on top.