Claude Code 2.1.147 ships /code-review, then 2.1.148 squashes a Bash-exit-127 regression within 24 hours

TL;DR

Claude Code v2.1.147 (released 2026-05-21) introduced two big things: a renamed /code-review command that finds correctness bugs and can post inline GitHub PR comments, and pinned background sessions that survive idle time and survive Claude Code updates. Then early on 2026-05-22, a regression was discovered where the Bash tool returned exit code 127 on every command, and Anthropic shipped v2.1.148 as a hotfix roughly five hours later. This post bundles "the useful new features" and "broken-and-fixed-in-five-hours" into one read, framed for anyone who depends on AI coding tools daily.

Bottom line first: upgrade claude to v2.1.148 or later right now. If you stay on v2.1.147 you risk hitting a state where the Bash tool fails on every command (details below). The good news is that the /code-review command and pinned background sessions introduced in v2.1.147 are still active in v2.1.148. The upgrade is pure upside. Anyone caught on v2.1.147 can move to v2.1.148 in about five minutes.

"Claude Code shipped again, but is it true that Bash is broken?" and "Why did /simplify disappear?" —— in this post we answer both questions using Anthropic's official GitHub Releases for v2.1.147 / v2.1.148 and the Claude Code Docs Changelog as primary sources, from a Sales Claw maintainer's perspective.

Inside a 24-hour window, Anthropic shipped v2.1.147 at 5/21 20:39 UTC followed by v2.1.148 at 5/22 01:16 UTC, about 4 hours 37 minutes later. The first is a large release containing five new features plus 30+ bug fixes. The second is a hotfix that fixes only one thing: the Bash regression. ("Regression" here means a feature that worked in the previous version stopped working in the new one.)

The takeaway from these two releases is that "AI development infrastructure is now operated on a 'ship fast, fix fast' assumption". The cadence — feature release at night UTC, hotfix five hours later — is the new normal, not an exception. This post unpacks the new features and adds an operational survival kit for the regression-injection era, so that teams running daily AI tooling can build a posture rather than reacting to each incident as a surprise. For series context, see also: Claude Code 2.1.145 / Codex Mobile / Anthropic postmortem roundup and Claude Code v2.1.144 explained.

This post uses GitHub Releases v2.1.147 / v2.1.148 / the Claude Code Docs Changelog as primary sources. Sales Claw is available as a free download from the download page.

1. What are v2.1.147 and v2.1.148 — why two releases shipped in 24 hours

[Official] Per Anthropic's GitHub Releases page, v2.1.147 went out at 2026-05-21 20:39 UTC (5:39 AM JST on 5/22), and v2.1.148 followed at 2026-05-22 01:16 UTC (10:16 AM JST on 5/22). The gap is about 4 hours 37 minutes.

First, the prerequisites. For non-engineer readers, here's the jargon de-jargoned:

• A regression is when something that worked in the previous version stops working in the new one. In Japanese it's sometimes called a "throwback bug"
• A hotfix is a release that ships a fix for a critical production bug along the shortest path, with a shorter validation window than a normal release
• Exit 127 is the shell exit code that means "command not found". You'd get it if you typed foo in your terminal. In v2.1.147, even git status and ls were being told "not found"

[Author's view] The fact that a hotfix can ship in 5 hours means Anthropic has the in-house plumbing to cut a hotfix release on demand when production hits a wall. That includes a CI pipeline that runs in minutes rather than hours, on-call engineers who can verify a one-line patch out of hours, and an npm publish path that doesn't require a coordinated release window. Operators of OSS business agents like Sales Claw need a matching operational mindset — versioned releases, rollback drills, and a clear "who decides when to ship" line of accountability (more in §6 on risks).

2. /simplify → /code-review — what actually changed

The headline change in v2.1.147 is that the /simplify command was renamed to /code-review and the underlying behavior was rewritten.

First, the prerequisites. /simplify originally meant "clean up this code and rewrite it so it reads better" — a one-shot way to ask "Can you make this simpler?".

The new /code-review changes direction. Translating Anthropic's release notes verbatim:

"Renamed /simplify to /code-review. It now reports correctness bugs at a chosen effort level (e.g., /code-review high); pass --comment to post findings as inline GitHub PR comments. The old cleanup-and-fix behavior has been removed."

[Official] Boiled down, /code-review now does three things:

Usage examples

Pattern 1: quick local invocation

# Lightweight review of the currently open code
/code-review

# Take more time and dig harder
/code-review high

Pattern 2: post inline comments on a GitHub PR

# Ask Claude to review your own PR and write findings directly into GitHub
/code-review high --comment

[Author's view] "Let Claude review the PR and persist the findings as GitHub comments" quietly solves a very real problem: "the reviewer is on vacation, so merges are blocked". Even for small OSS projects like Sales Claw, the workflow of "ask Claude to review your own PR with --comment, leave the findings in GitHub, then have a human make the call" becomes feasible. It also creates a paper trail that didn't exist before: the inline comments are searchable, link back to the line of code, and stay attached to the PR even after merge. That makes it easier for downstream readers (auditors, new hires, the author six months later) to reconstruct why the code looks the way it does.

3. Pinned background sessions — Ctrl+T removes review-wait pain

First, the prerequisites. Claude Code has had "background sessions" for a while — you offload a long-running task (a test run, a large refactor) to a separate session and continue your own work while it runs. Roughly: "ask Claude to do something, then let it run while you switch to something else."

Until v2.1.146, background sessions had a few painful properties: they would close on idle, vanish when Claude Code updated itself, and were aggressively closed when the system ran low on memory. Anyone who's left a session running and come back to find it gone has felt this.

[Official] v2.1.147 introduced "pinning" to opt out of those behaviors. Quoting the release notes verbatim:

"Pinned background sessions (Ctrl+T in claude agents) now stay alive when idle, are restarted in place to apply Claude Code updates, and are shed under memory pressure only after non-pinned sessions"

How to use it

1. While in claude, type claude agents
2. In the background-session list, select the session you want to pin
3. Press Ctrl+T (pin toggle)
4. Press again to unpin

[Author's view] The single most important behavior here is "pinned sessions are restarted in place when Claude Code updates itself". Until now, "I updated Claude and lost yesterday's session" happened often enough to be annoying. The psychological safety of offloading 1+ hour tasks (long searches, test runs, large refactors) is meaningfully different now.

Now that parallel work in Claude Code is normal, the form-submission side deserves the same automation.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub

4. The 30+ Windows / PowerShell / GNOME Terminal fixes

v2.1.147 ships more than 30 bug fixes. Many of them target the Windows + PowerShell environments that dominate Japanese enterprise dev shops. Six standouts:

Windows / PowerShell fixes worth highlighting

1. Fixed: PowerShell (pwsh) installed via winget or Microsoft Store causing Claude to exit 1
   → Many corporate PCs install pwsh through winget, so this fix is a big deal
2. Fixed: PowerShell tool dropping output for commands that rely on the default formatter
   → PowerShell-native commands like Get-Process were not flowing back into Claude
3. Fixed: "Yes, and don't ask again" for PowerShell script invocations now writes a rule that actually matches on subsequent runs
   → Removes the bug where the approval prompt kept reappearing
4. Fixed: hook if conditions like PowerShell(git push*) not matching — only PowerShell(*) worked
   → Crucial for teams writing fine-grained hooks in CLAUDE.md / settings.json
5. Fixed: full-screen strobing in attached background sessions on Windows Terminal during streaming
   → Direct perceived-quality win for older Windows Terminal builds
6. Fixed: on Windows, removing a background-job worktree no longer follows NTFS junctions into the main repo
   → Closes the worst-case "junction makes Claude delete other folders" scenario

Other fixes that matter (non-Windows)

• Unknown slash commands in headless / SDK mode were silently doing nothing → they now report an error
• MCP server resources / templates / prompts past page 1 → pagination fix
• Backgrounded sessions re-asking for tool permissions you'd already approved with "don't ask again" → fixed
• CLAUDE_CODE_SUBAGENT_MODEL not propagating to teammate processes spawned by agent teams → fixed
• Plugin agents with multiple Agent(...) entries in tools: frontmatter dropping all but the last → fixed
• GNOME Terminal right-click / middle-click paste not inserting text → fixed

[Author's view] The one-liner summary of this fix list: this is a particularly big release for Japanese enterprise Windows environments. Sales Claw is also developed on Windows 11 + PowerShell 7 (winget-installed), so the pwsh exit-1 issue was painful. Fixing that alone justifies the upgrade. The hook-condition fix is the second hidden gem — many teams had quietly accepted that fine-grained policies "didn't work" and reverted to PowerShell(*) blanket rules, which kills the whole point of having approval gates. With v2.1.147 the granular rules finally match, which is the difference between "we have a policy" and "the policy is enforced".

5. Five v2.1.147 improvements that hit AI sales workflows

Claude Code is primarily for software engineers, but v2.1.147's changes also matter for AI sales automation (including Sales Claw). Five concrete examples:

1. Automate copy-prompt PR review with /code-review --comment
   Sales Claw's outreach copy templates live in preferences.yaml and TypeScript defaults. When you update these via a GitHub PR, running /code-review high --comment lets Claude flag any tone that might come across as rude to the reader (prospect). It's a quality gate for outreach copy.
2. Protect overnight automation with pinned sessions
   Business agents like Sales Claw realistically run a "prepare tomorrow's send candidates overnight" pattern. With Claude Code updates no longer wiping work, you can safely delegate overnight jobs to Claude as well.
3. MCP pagination fix unblocks large connectors
   When Sales Claw integrates with a customer CRM (Salesforce / HubSpot etc.) via MCP, previously only the first page of resources / templates was returned. Now you actually get records past page 1.
4. The hook if fix lets you block only dangerous PowerShell commands
   "Make Remove-Item -Recurse -Force require human approval but auto-approve git push"-style policies depend on hook conditions like PowerShell(Remove-Item *) actually matching. That works correctly now.
5. NTFS junction safety fix prevents repo-deletion accidents
   Production Sales Claw setups frequently spin up dataset worktrees. Removing them no longer risks following a junction into the main repo and nuking it.

What it looks like in Sales Claw

Sales Claw is an OSS tool designed to lower mis-send and ToS-violation risk through policy control, pre-send automated inspection, sales-NG detection, CAPTCHA-stop behavior, send-rate limiting, audit logging, and auto-stop conditions. The concrete steps to take advantage of v2.1.147 / v2.1.148 are in the Quick start guide.

# 1. Upgrade Claude Code to v2.1.148 or later
npm i -g @anthropic-ai/claude-code@latest

# 2. On the Sales Claw side, open a PR for the copy prompts and let Claude review
git checkout -b copy-update
# (edit preferences.yaml copy)
git commit -m "feat: refresh subject and greeting copy"
gh pr create --title "Refresh subject and greeting copy" --body "..."

# 3. Launch claude and request a PR review
claude
/code-review high --comment

# 4. Put long-running overnight jobs in the background and pin them
# (in the claude agents view, press Ctrl+T on the target session)

6. Risks and safety design — surviving the regression-injection era

What v2.1.147 → v2.1.148 shows over 24 hours is the reality that "regressions ship alongside new features as a matter of course". That's true beyond Claude Code: the same applies to Codex (OpenAI), Cursor, and Gemini CLI. If you treat these as daily-use infrastructure, the operator side needs self-defense practices.

Mechanical risks and safety design at upgrade time

For operating a business agent like Sales Claw through Claude Code, the realistic approach is to lower these risks via automated checks and operational design. Don't claim absolute safety — claim "structurally lower risk".

• Canary the upgrade: codify a rule that "only one team member runs the new version for 24–48 hours; if no issues, everyone follows"
• Never disable the audit log (action-log.json): Sales Claw's auto-stop conditions depend on it
• Keep send-rate limiting permanently on: throttle repeat sends to the same domain to leave recovery headroom on a misfire
• Keep sales-NG detection on (preferences.respectNoSales: true): auto-skip pages that decline sales pitches
• On CAPTCHA, stop with awaiting_approval and save an audit log: the system does not try to bypass CAPTCHA
• Always AND together a count cap, an elapsed-time cap, and a turn cap: prevents runaway loops

Rollback procedure when a new version breaks production

Residual risks

Some risks can't be fully eliminated by automated checks:

• New CAPTCHA variants that Sales Claw hasn't accounted for yet (possible misfires before we adapt)
• Immediate adaptation to ToS changes (legal info still needs manual updates)
• Industry-specific regulations (BFSI and similar — verify separately)
• [Unverified] Whether other regressions beyond Bash exit 127 remain in v2.1.148 cannot be fully verified as of 5/22. Anthropic may ship additional fixes within the next 24–48 hours

7. Pre-deployment checklist — upgrade now, or wait?

8. Conclusion — design ops around "fix-fast" rhythms

Claude Code v2.1.147 lands two improvements that pay off for daily users: /code-review and pinned background sessions. It also demonstrates the operational reflex of killing the Bash exit 127 regression with v2.1.148 five hours later.

These 24 hours show that AI coding tools have two faces: "they evolve through feature additions" and "they operate through regression-injection followed by quick patching". If you intend to depend on them as daily-use infrastructure, the operator side needs canarying, rollback procedures, and audit logging as table-stakes practices. The same logic applies to anything you build on top of these tools — Sales Claw included — because your own users will hold you to "infrastructure" reliability expectations even when the underlying stack still ships at "developer-tool" pace.

Next actions: (1) run npm i -g @anthropic-ai/claude-code@latest right now to land on v2.1.148; (2) pin any background sessions you care about with Ctrl+T; (3) try /code-review high --comment on one of your own PRs. Start with the Sales Claw Quick Start guide if you want the operational template.

If Claude Code is becoming 'infrastructure', the form-submission side should climb the same staircase too.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub

Japanese-language original: Claude Code 2.1.147 で /code-review 登場、翌日 2.1.148 で Bash が全部死ぬバグを 24 時間で潰した話