If you use Claude Code every day, you don't need eight changelog entries — you need the two that change how you work. Here they are.

The 30-second version: As of v2.1.199 you can stack skills — /skill-a /skill-b do XYZ loads every leading skill in the line, up to five — so skills are now composable in a single invocation. And as of v2.1.200, AskUserQuestion decision dialogs no longer auto-continue by default: the agent pauses at a real decision instead of picking a default and rolling on. Everything else in the July batch is reliability and reach — background subagents, progress heartbeats, an abuse cutoff, screen-reader mode.

Claude Code shipped roughly eight releases in the first two weeks of July 2026, up through v2.1.211. Most were bug fixes and hardening. Two are worth changing a habit over.

1. Skills stack now — up to five in one line#

Before this month, a skill was a single invocation: you called one, it ran, you called the next. v2.1.199 changed that. A stacked slash-skill line like:

/research-a-topic /format-as-brief summarize the MCP stateless spec

now loads all the leading skills — research-a-topic and format-as-brief — before Claude acts, with a ceiling of five stacked skills. That turns skills from one-shot macros into a small composable pipeline you assemble at the prompt. The natural pattern is a gather skill plus a shape skill: one pulls and verifies, the next enforces your house format, and you invoke both in one turn instead of chaining two round-trips.

There's a quiet companion fix in v2.1.202: re-invoking an already-loaded skill used to append a duplicate copy of it to context, silently eating your token budget. That's fixed. If you lean on skills heavily — and stacking makes that more likely — the context math is now honest.

New to skills entirely? Start with how to build a Claude agent skill, then how to structure one for progressive disclosure. Stacking rewards skills that each do one thing well.

2. The agent pauses at decisions now#

v2.1.200 changed AskUserQuestion dialogs to no longer auto-continue by default. This is the sleeper change.

Previously, a session could reach a genuine fork — "which of these approaches do you want?" — surface the question, and, if nothing answered, proceed on a default. Convenient for a human watching the terminal; quietly dangerous for anything running unattended, because the agent would commit to a guess and keep going. Now it stops and waits at that fork.

If you run Claude Code interactively, you'll barely notice. If you run it headless, in CI, or on a schedule, this is the one thing to test before you upgrade: a loop that depended on auto-continue will now block at the first decision unless you explicitly configure it to continue. That's almost certainly the safer default — but it's a behavior change, and silent behavior changes in automation are how 2am pages happen.

3. Everything else: longer runs, wider reach#

The rest of the batch makes long and fan-out work more legible:

None of those change a habit the way stacking and pause-by-default do, but together they're the tell of where Claude Code is heading: unattended, multi-agent, longer-horizon runs that report themselves clearly and fail safe.

What to actually do before you upgrade#

  1. Audit your automation for auto-continue. Any headless or scheduled Claude Code run that leaned on AskUserQuestion proceeding on its own will now pause. Decide per-loop whether to answer, configure continue, or restructure the decision out.
  2. Reorganize skills to stack. Split fat skills into single-purpose ones you can chain /a /b. The five-skill ceiling is generous; small skills compose better than big ones.
  3. Re-baseline your context budget. With the duplicate-skill fix, sessions that re-invoke skills should show lower context growth — good, but it means your old numbers were inflated.

If you're choosing between coding agents rather than tuning one, this batch is also a data point: see Claude Code vs Cursor vs Cline on stopping a runaway subagent, and how to cap runaway Claude Code subagents if the background-by-default change makes you nervous about fan-out cost.

The headline isn't a marquee feature. It's that skills got composable and autonomous loops got a safer default — two small changes that quietly move Claude Code toward the unattended, multi-agent workflows everyone keeps saying they want.