← Blog · 2026-10-19

This Week in Claude Code: Oct 13-20, 2026 (6 releases, the code review playbook, 14 posts)

Six Claude Code releases shipped this week (2.1.299 to 2.1.304), Anthropic published how the Claude Code team reviews code, and the community wrote a lot about hooks, CLAUDE.md and keeping parallel sessions under control. Below are the changes that matter if you use Claude Code for your own or client work, then the posts I'd send to a friend. Every item links to its source. This covers Oct 13 through Monday evening, Oct 19.

Releases: the changes worth knowing

Full notes are on the official changelog and the GitHub releases page. These are the ones that change day-to-day use.

2.1.299 (published Oct 13, late evening ET), the big one:

• Idle compaction. In the terminal, a conversation over 200K tokens that sits idle for about 54 minutes is now compacted before its one-hour prompt cache expires, so your next message doesn't reprocess everything. "idleCompaction": false in settings.json turns it off, and CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS changes the 200K minimum.
• MCP allow rules got stricter. mcp__a no longer also allows a server named a__b or a_. Write mcp__<server>__* with the server's whole normalized name. Agent tools: lists changed the same way. If you allow MCP tools by prefix, check your rules (our free permissions builder writes permissions blocks you can compare against).
• Deny and ask rules for an MCP server whose normalized name contains a double underscore now apply (they didn't before).
• UserPromptSubmit hooks now run before a skill or custom slash command expands, so a hook that blocks the prompt also stops the command's embedded shell commands.
• Skills and commands with odd file names (hidden, direction-changing or blank-like characters, < or >) are no longer offered until renamed. If a skill vanished after updating, check its folder name (more checks in skill not triggering?).
• Fixes you may have hit: an empty .mcp.json reported as invalid by claude mcp list, claude doctor and /status; a settings file ignored when enabledMcpjsonServers or disabledMcpjsonServers held a single server name as a plain string; /usage, the limit warning and the status line sometimes showing more than 100% used.
• New: /usage warnings for max effort, workflows, always-on context and large tool results, and CLAUDE_CODE_DISABLE_FILE_MENTIONS so @ file mentions pass paths instead of file contents.

2.1.300 (Oct 14): updatedInput from PreToolUse and PermissionRequest hooks is now validated like Claude's own input. Unknown keys are dropped, and input the tool rejects denies the call. If your hooks rewrite tool input, test them again (the hooks generator and hooks not working? help).

2.1.301 (Oct 15): switching to a model with a smaller context window now has the previous model summarize a conversation that's too long for the new one first, and a long-running subagent's progress trail now updates incrementally, so rendering is faster.

2.1.302 (Oct 16): an Idle compaction row in /config turns the feature off or back on. Fork subagents are now on by default in claude -p, the Agent SDK, the desktop app and VS Code (CLAUDE_CODE_FORK_SUBAGENT=0 turns them off). The desktop app and VS Code now always run the main conversation's subagents in the background. At most 12 prompt or agent hooks of one event run at a time.

2.1.303 (Oct 18): bug fixes and reliability improvements.

2.1.304 (Oct 19):

• On Pro and Max plans, Claude Code now warns when resuming an inactive conversation would use a noticeable share of your 5-hour usage limit, and offers to start a new conversation instead.
• After compaction, Claude gets the skill list again, and /reload-skills now sends only new or changed skills.
• /usage says when a rate limit on usage data ends instead of "try again in a moment".
• CLAUDE_CODE_AUTO_COMPACT_WINDOW=200k no longer silently sets a 100K window; the variable accepts k and M suffixes.
• claude --worktree no longer removes the worktree and its branch without asking at exit when HEAD was detached or on another branch.
• /auto-mode-setup, a guided setup that reads your project and recent sessions to teach auto mode about your environment, is listed in this release, including on Bedrock, Vertex, Foundry and telemetry-disabled sessions.

From Anthropic: how the Claude Code team does code review

How the Claude Code team does code review (claude.dev blog, Oct 15; HN discussion). From June to September, merged pull requests to Claude Code's codebase more than doubled while the share that had to be reverted fell by more than 40 percent. Most reading moved to machines: Claude verifies its own work and a second agent reviews the diff in a fresh context before a PR, then CI, AI code review, security review and risk-based human approval run on every PR.

The "How you can adopt this process" section is practical even for a one-person shop: list the checks your agent can run in a short CLAUDE.md (our CLAUDE.md generator gives you a starting file), add hooks that run formatters and linters, run /code-review and /security-review locally, keep repeated review comments in a REVIEW.md, and add "Testing", "Risk" and "How to revert" to your PR template.

Community: hooks and guardrails

• AI Hooks for Claude Code, Cursor, and Codex: Block Secrets Before Agents Can Use Them by @dwayne_mcdaniel (GitGuardian). Using ggshield hooks to scan prompts, commands, file reads, MCP calls and outputs for secrets.
• How I stopped Claude Code from lying about finishing tasks by @98dkcc488bglitch. Four bash hooks built on the exit-2 contract that block claimed-but-undone work, pushes to the wrong branch and edits outside scope.
• Your First Claude Code Mod: Block Edits Outside the Task's Worktree in 30 Lines by @gabrielanhaia. A small, tested mod that refuses Edit and Write calls outside the task's git worktree.
• How I work with Claude Code: memory, gates and the mistakes that built them by @muchassak. CLAUDE.md routes into a wiki, hooks restore rules after compaction, and corrections become rules and then tests.

Community: CLAUDE.md and permissions

• State of AI Instruction Quality: /init Files Score Higher by @cleverhoods. A scoring study of 8,220 CLAUDE.md files from top GitHub repos: files started with /init scored higher at every size (median 8.4 vs 7.7).
• Your CLAUDE.md Isn't Memory. It's a Sticky Note That Keeps Growing by @icegaming. Why instruction files fail as team memory, and six fixes from rules-only files to decision logs and pruning.
• Let Claude Code Read Your Backend Repo, Not Edit It by @gabrielanhaia. --add-dir plus one Edit deny rule gives Claude a second repo as read-only reference.
• An Allow Rule for MCP Server 'a' Also Matched 'a__b' by @gabrielanhaia. The background to the 2.1.299 MCP rule change above, with a test and a grep to check your own settings.
• Run Two Claude Code Accounts Side by Side Without Logging Out by @gabrielanhaia. A separate config directory per account, routing by folder, and a macOS Keychain pitfall. Handy if you keep client and personal work apart.

Community: sessions, plans and orchestration

• Watching Claude Code subagents live by @bargan. ccdeck, a local dashboard that draws each session and its subagents as a live tree of cards.
• session-archive: back up Claude Code session logs before the 30-day auto-delete by @haoli. A stdlib-only Python CLI that gzips, indexes, searches and restores session logs.
• I made Claude Code plans reviewable by @yahav10. plan-live opens a markdown plan as a local review page where comments and an Apply button get Claude to revise it.
• Show HN: Pi-cowboy – use Codex, Claude Code, OpenCode as subagents. Splits big jobs into subtasks for parallel background agents in isolated copy-on-write git worktrees.

Community: building with it

• How I Fixed 19 Timezone Bugs in My Time-Tracking App With Claude Code by @yureki_lab. A harness replaying the app across 40 timezones and DST boundaries, with a failing test required before each fix.

My takeaway this week

Two habits pay off right away. First, if you allow MCP tools in settings.json, write the full mcp__<server>__* form now that 2.1.299 matches server names exactly. Second, borrow the smallest piece of Anthropic's review process: a short list of checks in CLAUDE.md plus one formatter or test hook, so Claude checks its own work before you read it.

Release facts on this page come from Anthropic's Claude Code changelog, the GitHub releases and the claude.dev blog. Community summaries are mine; all credit to the authors. This page isn't affiliated with Anthropic.

The full, searchable list of Claude Code write-ups (287 posts, sorted by category, refreshed daily) is at solostackkits.netlify.app, and the free browser tools are here.