TL;DR — Shopify closed two multi-year plumbing holes this week (POS cash management and non-Plus B2B), and Claude Code shipped four back-to-back releases (2.1.94, 2.1.97, 2.1.98, 2.1.101) plus a Cowork GA that together change what a serious agent session looks like. The common thread is infrastructure that used to require a spreadsheet or a shadow app now ships in the box.
The theme
Every item this week is the same story told in two stacks: Shopify and Anthropic both spent the week deleting work you used to have to do yourself. Shopify POS finally has a register-session model with drawer audits, and Shopify B2B finally works on Basic. Claude Code bumped its default effort to high, added a focus view, shipped a Monitor tool for background scripts, and generated ramp-up docs from your own usage. Cowork went GA with an Analytics API, OpenTelemetry, and RBAC.
None of these are headline features in isolation. Together they are the quiet shift from "platform that gives you primitives" to "platform that closes the loop." The week's mental model: if you have a spreadsheet, a cron job, a polling subagent, or a hand-rolled onboarding doc filling a gap one of these vendors sells you, this is the week to delete it and let the platform own it. The engineering work for the next quarter is not building more glue. It is figuring out which glue to throw away.
1. Shopify POS Cash Management Foundations (original)
Overview
If your retail location handles any meaningful cash, you have probably built a shadow ledger — a spreadsheet, a Google Form, or an end-of-day ritual — to reconcile the drawer against what Shopify POS thought happened. The old flow was three verbs: open day, close day, hope. The new Cash Management Foundations replace that with named register sessions, opening and closing counts by denomination, mid-shift cash-in and cash-out with reason codes, staff-role drawer access control, and an audit trail that survives turnover. This is unsexy plumbing and the single most-requested POS feature since 2022.
Technical
The core unit is now a register session — a named open-to-close window tied to one staff member and one drawer, with multiple sessions per day supported. Opening a session stores cash counts per denomination ($1s, $5s, $10s), so end-of-day variance traces to a specific bill type. Mid-shift movements each require a reason code; Shopify ships Refund, Expense, Deposit, Tip Out, and Other, and you can add custom codes in Settings → Apps and channels → POS → Cash management. Match those codes to your bookkeeper's chart of accounts and monthly close gets faster.
The real fraud control is the new cash_management_admin staff permission. Without it, staff can ring sales but cannot do a no-sale drawer pop — historically the main theft vector on Shopify POS. Closing a session computes variance against opening plus cash sales plus cash-in minus cash-out, flags anything over a configurable $5 threshold for manager review, and writes to an audit log exportable under Analytics → Reports → Cash management. Included with Shopify POS Pro at no extra cost; Lite gets only basic open/close. One breakage risk: if you have a POS UI extension app that touches the drawer, it may need to move to the new register extension targets — test on one device before fleet rollout.
Takeaway
This week, turn the new flow on at one register at one location. Have your closing manager run it for three days and compare variance numbers to your existing spreadsheet. Define five custom reason codes that match your bookkeeper's accounts. Next week, train every store lead and migrate the fleet. Kill the spreadsheet on day one of month two — if you leave it alive, half your team will keep filling it in.
2. Claude Code Bumps Default Effort to High (original)
Overview
Claude Code 2.1.94 flipped the default effort level from medium to high for API-key, Bedrock, Vertex, Foundry, Team, and Enterprise users. If you have not run /effort to set it explicitly since upgrading, every session you open is spending more thinking tokens per turn. For interactive agentic work that is the right call. For batch workers, quick edits, and anything where you already trust the model to ship, it is the wrong call — and you will feel it on the bill.
Technical
Effort controls how much extended thinking the model does per turn. More effort means more thinking tokens, longer responses, and more tool calls per answer. You override the new default three ways: in-session with /effort (status bar updates immediately); per skill via the effort frontmatter field added in 2.1.80; and per agent, supported in subagent frontmatter since 2.1.78.
yaml
---
name: quick-edit
description: One-shot edit, no deep thinking
effort: low
---
The pattern that works: leave interactive sessions on the new high default, set effort: low on every skill whose job is "do one thing fast," and benchmark for a week before touching the global. If you run a nightly polish worker or a PR worker that handles dozens of tasks an hour, that worker belongs on low or medium — the marginal value of extended thinking collapses when the task is "polish this draft" instead of "design this system." Two smaller fixes in the same release matter: 429 retries now respect exponential backoff as a minimum instead of burning attempts in 13 seconds, and plugin skills declared via "skills": ["./"] now take their invocation name from frontmatter, stabilizing names across marketplace and local installs.
Takeaway
Run /effort at the top of your next session and check what the default resolved to. Audit your batch workers this week and add effort: low to anything that does not need to think hard. Leave interactive sessions on high for at least two weeks before deciding it is too expensive — in practice it almost never is.
3. Claude Code Focus View and Ctrl+O (original)
Overview
A long Claude Code session is a wall of tool calls, tool results, fifteen-line diffs, and search output with the actual answer buried somewhere in the middle. Claude Code 2.1.97 added a focus view — toggle Ctrl+O in NO_FLICKER mode — that collapses every tool call to a one-line summary with diffstats and shows only your prompt and the model's final response. It sounds cosmetic. It is not. It changes how I prompt, because I stop trying to read the model's work and start trusting it.
Technical
Ctrl+O toggles focus view, which renders tool calls as one-liners like +12 -3 for an Edit or Listed 4 directories for an ls, then shows the final response. Alongside it, the status line got a refreshInterval setting that re-runs a custom status command every N seconds — useful if you pull CI URLs, deploy status, or token usage from external state. The status line JSON now also exposes workspace.git_worktree whenever you are inside a linked worktree, so your status line can finally say which worktree you are in without hand-wiring it. And /agents gained a ● N running indicator next to agent types with live subagent instances.
Smaller but real: pasted images now compress to the same token budget as Read-tool images (so long sessions with screenshots stop silently burning tokens), markdown blockquotes render a continuous left bar across wrapped lines, and slash command completion now triggers after CJK sentence punctuation. The tradeoff with focus view is that when something goes wrong you toggle back out to debug — the first time the model edits the wrong file you will lose 30 seconds finding the Edit call. After a week of using /freeze or a scope-fence skill, you stop worrying about it.
Takeaway
Turn on NO_FLICKER and build Ctrl+O muscle memory by Friday. Set a refreshInterval on your status line. If you do not have a custom status line yet, write one that prints your branch, your worktree, and your unstaged file count — it will be the most useful 30 lines of code you write all month.
4. Claude Code Monitor Tool (original)
Overview
Most of the noise around 2.1.98 was the interactive Vertex AI setup wizard. Important if you are on GCP, but the change that matters long-term is the new Monitor tool. It gives Claude Code a first-class way to watch a long-running background process and react as events stream in, instead of polling a log file and pretending that is a control loop. Pair it with the same release's CLAUDE_CODE_SUBPROCESS_ENV_SCRUB and you have a defensible story for letting Claude Code drive a long build, a test run, or a dev server without leaking API keys to the child process.
Technical
Monitor is documented as "a tool for streaming events from background scripts." Spawn a long-running process, get a handle, and Claude Code attaches to its stdout and reacts as new lines arrive — no sleep loops, no manual tail -f shims, no dropped output between polls. The killer use case is subagents: before this, a "watch the dev server and tell me when the homepage is ready" agent polled a log every five seconds and burned tokens summarizing nothing. With Monitor, the agent attaches once, sleeps until signal, and only spends tokens on something interesting.
Four related switches from the same release change how you use it: CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 strips Anthropic and cloud-provider credentials from any subprocess environment; CLAUDE_CODE_SCRIPT_CAPS caps script invocations per session as a runaway seatbelt; when env scrub is on, the Linux sandbox uses PID namespace isolation so the child literally cannot see the rest of your machine; and the W3C TRACEPARENT env var is now forwarded to Bash subprocesses when OTEL tracing is on, so child-process spans parent correctly in distributed traces. Decision rule: if you would tail -f it as a human, use Monitor. If you would make && echo done, use the older TaskCreate + TaskGet pattern — Monitor complements it, does not replace it.
Takeaway
Pick one background process you currently watch with a polling subagent — dev server, test watcher, build — and rewrite the prompt to use Monitor this week. Set CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 in the same session so you can show your security person without flinching. The token savings on any non-trivial watch loop pay for the upgrade in a day.
5. Claude Code /team-onboarding (original)
Overview
Claude Code 2.1.101 added /team-onboarding, which scans your local Claude Code usage on a project and generates a teammate ramp-up guide from the patterns it sees — commands you run, subagents you reach for, directories you live in. It is the closest thing I have seen to "tribal knowledge as a command," and it ships in the same release that fixed around two dozen --resume bugs and enabled OS CA certificate trust by default. If you run a small team and have watched a new engineer spend two days figuring out which package manager you use and where the dev scripts live, this is for you.
Technical
Call /team-onboarding from a session in the project root. It introspects recent prompts, common tool calls, frequently edited paths, and enabled skills and plugins, then writes a project-scoped ramp-up doc — how this project gets worked on, not how the CLI works in general. A few practical notes: the guide is only as thick as your history, so run it on a project you have been in for at least a few weeks. There is no documented way yet to pass a custom template or pin which paths to emphasize, so treat the output as a draft a senior engineer edits before handing it to a new hire.
Three other fixes in 2.1.101 deserve a line. OS CA certificate trust is now default — this is finally the version that works behind an enterprise TLS proxy without NODE_EXTRA_CA_CERTS gymnastics (opt back into the bundled list with CLAUDE_CODE_CERT_STORE=bundled). A hardcoded 5-minute request timeout that was killing slow backends regardless of API_TIMEOUT_MS is fixed, which matters if you run local LLMs or extended thinking through a slow gateway. And the --resume/--continue chain recovery now handles dead-end branches cleanly, so long sessions over large repos stop occasionally losing context.
Takeaway
Run /team-onboarding in your most active project this week, read the output critically, edit it down, and commit it as ONBOARDING.md linked from your README. Twenty minutes of work that saves the next engineer two days. If you are behind a corporate proxy or pointing Claude Code at a slow gateway, the certificate and timeout fixes make 2.1.101 a no-brainer upgrade regardless.
6. Claude Cowork Hits GA (original)
Overview
Cowork went GA on April 9, 2026 on macOS and Windows through Claude Desktop. On the surface it reads like a polish announcement — research preview to GA, the usual arc. The interesting detail is in the bullets: an Analytics API, OpenTelemetry support, and role-based access controls that let Enterprise admins group users with custom capability sets. For founders running small ops teams on Claude, this is the moment Cowork stops being a toy and starts looking like something you can defend in a security review.
Technical
The Analytics API is the load-bearing change. Cowork usage was a black box — you could feel someone burning tokens but could not put a number on which workflows consumed which budgets. Now you pull usage and engagement data programmatically and dump it into whatever observability stack you already run. OpenTelemetry support is the second half: Cowork emits OTel traces you wire into Datadog, Honeycomb, or Grafana Tempo without a custom shim, so you can root-cause why a long agent task drifted, which tool calls cost the most latency, and where the model paused for clarification. Same debugging loop you run on a production microservice.
RBAC is the third pillar and it is really a blast-radius feature. Let customer support use Cowork with the filesystem tool disabled, give engineering the full computer use surface, and lock skills and connectors per group. A few limits to watch: Analytics API is Enterprise-only (Team still sees only the in-app dashboard), OTel is opt-in per workspace via admin settings, and RBAC groups are flat — no nested groups, so deep org charts end up duplicating capability sets.
bash
curl -H "Authorization: Bearer $ANTHROPIC_ENT_TOKEN" \
"https://api.anthropic.com/v1/organizations/$ORG_ID/usage?range=30d" \
| jq '.data[] | {user, model, input_tokens, output_tokens, cost_usd}'
Takeaway
If you have been treating Cowork as a research toy, GA week is the right week to flip it. Wire the Analytics API into whatever finance dashboard you run, point OTel traces at your existing backend, and draft RBAC groups before you onboard non-engineers. The capability is Monday-morning ready; the discipline is on you.
7. Shopify B2B on Basic, Grow, and Advanced (original)
Overview
Shopify opened the native B2B feature set — company profiles, custom catalogs, customer-specific price lists, and net 30/60/90 payment terms — to merchants on Basic, Grow, and Advanced. Until this week, the only way to get any of this without a third-party app was to be on Plus at $2,300/month. For brands running a few wholesale orders a month off a Google Form or a hidden customer tag, this is the upgrade to make. The catch is volume: Plus still gets unlimited B2B catalogs and expansion stores, while Basic tops out fast. For most merchants under $1M in wholesale GMV, those ceilings will not bite.
Technical
Enable B2B from Settings → Customers → Companies. A company holds one or more locations, each with its own shipping address, payment terms, and assigned catalog. A buyer logs in with their email, sees their company's catalog, and pays under the configured terms. Custom catalogs use Shopify's price list feature, so you can do percentage-off-base or fixed prices per SKU — but not volume break tiers ("buy 50, get 20%"), which still need a Shopify Function or app. Payment terms cover net 7 / 15 / 30 / 60 / 90 / on-fulfillment; orders move to "Payment due" status until the buyer settles via the customer portal. Shopify does not advance you cash, and invoice-paid-by-bank-transfer is the realistic flow for net 30 — track those in your bank recon, not Admin.
Plan ceilings: Basic caps at 25 active companies and 5 catalogs; Grow at 100 and 25; Advanced at 500 and 100; Plus unlimited. What still requires an app: tiered volume pricing, MOQ enforcement at the line item level, approval workflows (orders over $5,000 require manager sign-off), and ERP sync. Migrating from a tag-based wholesale setup is non-trivial — plan an afternoon to manually create a Company per existing wholesale customer, assign the location, and build or pick a catalog. Test with one customer first; the customer-facing login UX on Online Store 2.0 themes still has rough edges, especially the "log in to see prices" empty state.
Takeaway
If you run wholesale on Basic or Grow with a tag-and-discount-code stack, schedule a half-day next week to migrate your top five accounts to native Companies. Keep the tag system alive for two billing cycles before you deprecate it. If you are on Plus for B2B reasons alone and running under 25 companies, get on the phone with your account manager and renegotiate — the B2B-only premium just shrank.
Original sources
- Shopify POS Cash Management Foundations: What Changed — originally published 2026-04-06
- Claude Code Bumps Default Effort From Medium to High — originally published 2026-04-07
- Claude Code Focus View Cuts the Noise With Ctrl+O — originally published 2026-04-08
- Claude Code Monitor Tool Streams Background Scripts — originally published 2026-04-09
- Claude Code 2.1.101 Ships /team-onboarding for Ramp-Up — originally published 2026-04-10
- Claude Cowork Hits GA: What Operators Should Actually Care About — originally published 2026-04-11
- Shopify B2B Lands on Basic, Grow, and Advanced Plans — originally published 2026-04-12
