Claude Code 2.1.287: The ‘You Should Know’ Safety Agent and 60+ Fixes That Make Claude Safer to Use
Claude Code 2.1.287 introduces Claude Mods and the 'You Should Know' background safety agent, which flags overlooked risks during sessions. The release also closes 60-plus bugs covering file handling, screen reader accessibility, remote session stability, and a dangerous shell command safeguard bypass.
Claude Code Just Got a Built-In Safety Watcher — Here’s What Changed
When an AI tool starts handling real work — drafting client reports, writing code, managing files — the cost of a missed detail goes up fast. The Claude Code 2.1.287 changelog is a dense release: one meaningful new capability, a new plugin architecture, and more than sixty bug fixes. Most of those fixes close edge cases that could quietly cause data loss or confusing errors. If you use Claude Code at work, this update is worth understanding.
What Are Claude Mods?
The headline addition in this release is Claude Mods, a plugin architecture that lets third-party and built-in extensions modify Claude’s deeper behavior — not just add tools, but change how Claude reasons, monitors, and responds during a session.
According to the Claude Code changelog, version 2.1.287 states: “Added Claude Mods: plugins may now modify deeper behavior.” This is a structural shift. Earlier, plugins could only extend what Claude could do (fetch a URL, run a script). Now they can shape how Claude thinks and what it flags during a session.
The first built-in Mod to ship with this architecture is called You Should Know.
‘You Should Know’: A Side Agent Watching Your Back
The changelog describes You Should Know as “a built-in mod where a side agent watches your back and flags things you or Claude might miss.”
In practical terms: while you work in a Claude Code session, a background agent runs alongside the main conversation. It watches for oversights — a file about to be overwritten, an assumption Claude made that hasn’t been verified, a step that seems correct but carries a risk you might not notice unless you were looking for it.
You enable it with the command /plugin enable cc-plugin-you-should-know@builtin. The changelog notes this works “for first-party sessions with telemetry on” — an important restriction covered below.
Think of it as a second pair of eyes in the background. You don’t have to ask it questions. It volunteers observations when something seems off.
A Concrete Scenario: A Finance Team in Pune
Imagine a small finance team at a logistics company in Pune. They use Claude Code to automate parts of their month-end close — reconciling vendor payment files, generating summary reports from raw Excel exports, and occasionally running scripts that rename or move files in bulk.
Without You Should Know, a junior analyst could ask Claude to clean up a folder of vendor invoices and Claude would proceed confidently, perhaps deleting duplicates based on filename alone — missing that two files with similar names were actually for different vendors.
With You Should Know enabled, the background agent could flag: “These two files have similar names but different timestamps and sizes — confirm which to keep before proceeding.” The main session keeps moving; the safety agent interjects only when something is genuinely worth pausing for.
This isn’t hypothetical caution. The 2.1.287 changelog also documents a specific dangerous-command fix: “Fixed a dangerous rm (such as one on / or the home directory) losing its always-ask safeguard when the same command also redirected output to a ~ or wildcard path.” The safeguard existed before, but a specific command pattern could bypass it. That gap is now closed.
The Bug Fix List Is the Real News
The You Should Know mod gets the headline, but the 60-plus fixes in this release are where most users will feel the difference. A few stand out for non-technical users who rely on Claude Code for day-to-day professional work.
File Handling Fixes
- Messages sent from Claude apps with 17 to 20 attached files were only delivering the first 16. The changelog confirms this is now fixed — all attached files reach Claude.
- PNG, JPEG, and WebP images over 8,000 pixels on a side were failing to send from remote sessions. Claude now sends a scaled-down copy automatically.
- Cloud sessions could sometimes lose the earlier conversation when the session restarted during compaction. This is fixed.
For a marketing team in Hyderabad sharing design mockups or campaign screenshots with Claude, the image size fix alone removes a frustrating silent failure that offered no clear error message.
Screen Reader and Accessibility Fixes
This release includes an unusually thorough set of screen reader fixes — more than eight separate issues addressed. These include: the cursor moving away from typed text in search boxes, the Enter key being blocked on optional fields, misleading hints about Tab key behavior, and diff content being left out of approval prompts.
For visually impaired professionals using Claude Code in India — where assistive technology adoption in knowledge work is growing — this represents a meaningful jump in usability.
Remote and Cloud Session Stability
Several fixes address situations where remote sessions would drop messages, stall, or fail to reconnect:
- Remote Control was not receiving messages for minutes at a time when a reconnect request got no response. It now gives up after 30 seconds and retries.
- On macOS, Remote Control sessions were stopping mid-turn when the Mac went to idle sleep.
- Files uploaded from cloud sessions were failing when the upload finished just after a 30-second timeout. The window is now extended to 35 seconds, with one automatic retry.
The 1M Context Window Change
The changelog notes that “Opus 4.7+ and Fable now use a 1M context window by default on Bedrock, Vertex, Foundry and the Claude apps gateway, with no [1m] suffix.” A setting called CLAUDE_CODE_DISABLE_1M_CONTEXT=1 keeps the older 200K limit if needed.
For teams running Claude on Amazon Bedrock or Google Vertex — common in enterprises that need data residency guarantees — this means substantially longer conversations or document sets can now be processed in a single session by default.
Limitations and Tradeoffs to Know
Before you rush to enable You Should Know, the constraints matter.
What to Watch For Next
The Claude Mods architecture is the thread worth following. According to the Claude Code changelog, plugins can now “modify deeper behavior” — which means future Mods could add things like organization-specific compliance checks, automatic citation requirements, or domain-specific reasoning guardrails. The You Should Know mod is the proof-of-concept for what that looks like in practice.
If you are evaluating Claude Code for your team, the accessibility fixes in this release make it meaningfully more usable for a broader range of people. The file-handling fixes reduce silent failures that were hard to diagnose. And if your organization is on Bedrock or Vertex, the default 1M context window upgrade arrives without any configuration change on your part.
Check whether your current Claude Code installation has updated to 2.1.287 — and if You Should Know is available in your session type, it costs nothing to turn on.
