Claude Code v2.1.287 Adds Claude Mods, 'You Should Know' Side Agent, and MCP URL Prompts

Claude Code v2.1.287 landed with two headline additions: Claude Mods, which let plugins modify deeper behavior than before, and a built-in mod called You Should Know that runs a side agent to flag things you or Claude might miss.

Enable the built-in mod with:

/plugin enable cc-plugin-you-should-know@builtin

That mod is limited to first-party sessions with telemetry on.

Agents view and MCP protocol changes

• New n:<text> filter in the agents view matches session names and tasks. Filters now show matches inside collapsed sections, and Enter opens the first match.
• URL prompts from MCP servers are now supported on the 2025-11-25 protocol (for things like sign-in). If a server stops connecting after this update, add "bareElicitationCapability": true to its MCP config entry.
• OpenTelemetry's user_prompt event gained a prompt_text field — a copy of prompt for backends that nest dotted keys. Drop or mask it wherever you drop or mask prompt (#70763).

Platform and runner changes

• Windows: startup warning when denying the Bash tool also disables the PowerShell tool, meaning Claude has no shell.
• Self-hosted runner: built-in gh api (REST only) for Anthropic-managed git sessions on macOS and Linux where the GitHub CLI isn't installed.

Notable fixes

• Dangerous rm commands (e.g. on / or the home dir) lost their always-ask safeguard when the same command redirected output to a ~ or wildcard path — patched.
• Remote Control could stall for minutes waiting on a reconnect; it now gives up after 30 seconds and retries.
• Hooks with asyncRewake no longer spam "found issues" when the hook script is missing; the broken hook is reported once.
• Bedrock and Vertex startup model checks now respect an enforced availableModels list, which previously could collapse /model to a single Opus row.
• Amazon Bedrock Guardrails blocks arriving mid-response no longer end the turn with an API error when the reply began with thinking.
• claude -p and SDK sessions no longer repeat a model fallback on every later message after a mid-reply model switch.
• A folder's CLAUDE.md is no longer attached twice after resuming a session or a compaction.
• Switching between Opus 5.5 and Sonnet 5.5 (/model, opusplan) no longer rewrites earlier MCP tool announcements, which could drop earlier extended thinking.
• /advisor pairing now allows Sonnet 5.5 to advise Opus 4.7 and 4.8, and flags API-refused advisors up front.
• Bash permission prompts no longer show internal parser names like "Contains simple_expansion".
• Organization per-tool permission ceilings are no longer silently dropped for an MCP tool named __proto__.
• Claude is no longer told to page large MCP JSON results with Read's offset and limit, which can't split one long line.
• Tool heartbeats now reach SDK hosts when the model's response stream stalls with no data.
• Fable picked in /model on a claude.ai login now saves the newest version's id, matching Opus and Sonnet behavior.
• Background sessions can be reopened from claude agents after the agent exits and removes its worktree.
• Fullscreen sessions on busy machines no longer exit with "unrecoverable interface error" while a scroll key is held in long conversations.
• Several screen reader mode fixes: cursor placement in search boxes and sign-in code fields, Enter on optional /rewind summarize options, removed "Tab to amend" hint where Tab does nothing, and cleanup of arrow-key/number-select hints in /permissions and /mcp.

If you're on a first-party session with telemetry and want the side agent watching for gaps, the enable command above is the fastest path. The MCP config note is the one to flag if you maintain custom servers.