Claude Code release notes: what to change after an update

Where to read the official Claude Code release notes, how to parse an entry, and which 2.1.294 to 2.1.296 changes may need action in your setup.

The Claude Code release notes are the primary source for knowing what an update changed in your settings, hooks, subagents and MCP servers. This guide shows where the official Claude Code release notes live, how to read an entry, and which recent entries may need a change on your side. Everything below was read in the official changelog and its RSS feed on 2026-10-10; the latest version at that time was 2.1.296.

Where the official Claude Code release notes live

Three official places that list the same entries, in different formats:

Third-party trackers exist, but when a detail matters, cite the official page and the version.

How to read a changelog entry

Entries are grouped under headings such as Added, Changed and Fixed. Some lines start with a bracketed tag such as [VSCode], [Cloud sessions] or [Claude Tag]; these name the surface the change applies to. If you only use the terminal CLI, you can skip entries tagged for other surfaces.

Three habits help:

  1. Note the version and the publication date from the RSS item, not only the heading.
  2. Scan Changed first: changed defaults can alter behaviour without any edit on your side.
  3. Search the entry for the names you use: hook events, frontmatter fields, environment variables, MCP.

Recent releases that can affect your setup

Checked on 2026-10-10. Entries are paraphrased from the changelog; follow the links for exact wording.

2.1.296 (published 2026-10-09)

Per the 2.1.296 entry:

  • Added autoCompactWindow to subagent frontmatter and --agents definitions, so a subagent can auto-compact earlier than the main conversation’s window.
  • Added the CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL environment variable to run every workflow agent on one model while other subagents keep theirs.
  • Added CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS to set a longer maximum backoff delay when retrying an overloaded (529) request.
  • Added a note in /plugin for a plugin whose hooks are left out because another enabled plugin has the same name.
  • Added an allow_large option to the Read tool, so Claude can read a text file past the usual size limits in one call when it needs the whole file and the context has room.
  • Changed the default limit on MCP tool descriptions sent up front and on MCP server instructions from 2,048 to 4,096 characters.

2.1.295 (published 2026-10-08)

Per the 2.1.295 entry:

  • Added onFailure: "block" for command and HTTP hooks: a hook that cannot start, times out, or exits with an unexpected code blocks the action instead of letting it through.
  • Changed subagents to preload at most 32 skills from the skills field, each once; a subagent with the Skill tool can still invoke the rest.
  • Changed MCP tool descriptions loaded through tool search to be cut at 16,384 characters instead of 2,048.

2.1.294 (published 2026-10-08)

Per the 2.1.294 entry: fixed prompt and agent hooks written as instructions (such as “Block commands that…”) allowing what they should block, and improved how prompt hooks on Stop and SubagentStop written as instructions are judged.

If you use hooks, read the hooks examples alongside these entries.

Check your installed version and update safely

Run claude --version (or claude -v, listed in the CLI reference as “Output the version number”, checked 2026-10-10) and compare it with the newest entry in the changelog. Then:

  1. Read every entry between your version and the latest, not just the newest one.
  2. Update on a branch or in a throwaway project first if your team shares .claude/ settings.
  3. After updating, re-run the workflows that depend on hooks, subagents and MCP servers; those are where the entries above change behaviour.

This guide does not cover update commands per install method; see the official Claude Code docs for yours.

What to change in your setup

  • Hooks: if you rely on a hook as a guardrail, review it against the 2.1.295 onFailure: "block" entry and the 2.1.294 fix for instruction-style prompt and agent hooks. Retest hooks that were previously letting actions through.
  • Plugins: open /plugin and look for the new note about hooks left out because two enabled plugins share a name.
  • Subagents: consider autoCompactWindow for long-running subagents, and check that none preloads more than 32 skills (see what Claude Code skills are for how skills load).
  • MCP: if a server’s tool descriptions or instructions were near 2,048 characters, re-check what the model now receives under the 4,096 default, and the 16,384 cut for tool search.
  • Retries and models: set CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS or CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL only if you actually hit those cases.
  • Subscribe: add the RSS feed to your reader so you see entries on the day they are published.

FAQ

Where are the official Claude Code release notes?
In the Claude Code docs changelog (code.claude.com/docs/en/changelog), which also has an RSS feed, and in CHANGELOG.md in the anthropics/claude-code GitHub repository.
Did the MCP description limit change recently?
The 2.1.296 entry (published 2026-10-09) says the default limit on MCP tool descriptions sent up front and on MCP server instructions went from 2,048 to 4,096 characters. The 2.1.295 entry (2026-10-08) separately lists a 16,384-character cut for descriptions loaded through tool search.
Can a subagent compact its context earlier than the main conversation?
Per the 2.1.296 entry, yes. autoCompactWindow was added to subagent frontmatter and to --agents definitions.