Claude docs changes for October 8th, 2026 [diff]

Executive Summary

  • Claude Haiku 5.5 (claude-haiku-5-5) launches: 1M context, $0.10/$0.50 per Mtok, new model docs, migration guide and prompting guide. It is now the default haiku alias on the Anthropic API in Claude Code 2.1.293.
  • New Managed Agents docs cover live response previews via event deltas, and session inspection and usage tracking in the Console.
  • New browser and computer use SDK toolset docs, plus a page on claiming monthly API credits with Max and Team plans.
  • Claude Code docs add keybindings for the above-prompt band and mod panes, a full subagentStatusLine task field reference, and --json output for plugin marketplace commands.
  • Claude Code 2.1.294 fixes prompt and agent hooks written as instructions (such as "Block commands that...") so they actually block.

New Claude Code versions

2.1.293

New features

  • Added Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API (1M context, $0.10/$0.50 per Mtok)
  • Added agentType to the subagentStatusLine payload
  • Added isDeferred to $.tool.register for mods, so a tool's schema can be listed in the prompt from the start instead of behind tool search
  • mock.session added to claude plugin test so tests can read rows appended with $.session.append

Existing feature improvements

  • Faster startup for Team and Enterprise orgs: policy and managed settings are fetched earlier, with a retry after 3 seconds
  • Improved Claude in Chrome reliability and its cloud-session error message for multi-organization users
  • Artifacts now pin libraries to exact versions that are at least two weeks old
  • claude.ai skill syncing now checks about every 40 minutes instead of every 10 while idle
  • Agent lists and MCP server announcements now sort non-ASCII names after ASCII names
  • claude purge now deletes what it can, lists what it couldn't, and exits 1
  • Claude Tag: channel rule limit raised from 20 to 50; Claude joins announcement channels without posting its intro
  • Code Review's Add a repository dialog now explains why each repository couldn't be added

Major bug fixes

  • Fixed Claude treating its own last actions before a context compaction as done, and retracting or redoing finished work
  • Fixed a memory leak in HTTP MCP connections
  • Fixed messages being lost when ← moved a session to the background, and several related backgrounding issues
  • Fixed /model effort selection wrapping around and accidentally saving Low as a model's default
  • Fixed path-scoped rules and nested CLAUDE.md files not loading when Claude views files through Bash cat, head, tail, sed -n or grep
  • Fixed claude logs, stop, kill, rm and daemon commands sometimes signing you out
  • Fixed claude plugin eval refusing Bash-granting runs on Macs with Docker Desktop
  • Fixed /ultrareview upload refusing some repositories on Linux
  • Reverted the 2.1.281 auto mode denial message change and the 2.1.290 cloud session wakeup fix

2.1.294

Existing feature improvements

  • Improved how prompt hooks on Stop and SubagentStop written as instructions are judged, so Claude is less likely to stop early

Major bug fixes

  • Fixed prompt and agent hooks written as instructions (such as "Block commands that...") allowing what they should block

Claude Code changes

Changed documents

headless [Source]

  • Stream output docs now cover forked skills as well as subagents, and explain parent_tool_use_id values, including the forked-command- prefix for a skill started with /<skill-name> as the prompt. [lines 190-199] [Source]
  • Added a table of how a run starts, which parent_tool_use_id it carries and when its messages arrive, plus minimum Claude Code versions for each forwarding behavior. [lines 199-215] [Source]

hooks [Source]

  • WorktreeRemove now documents when it runs on exiting an interactive worktree session, including automatic removal of unnamed worktrees with no changes. It warns that non-git directories always look clean, so the hook should check for uncommitted work. [lines 2977-2985] [Source]

keybindings [Source]

  • New contexts AbovePrompt, AbovePromptInput and AbovePromptSelect. [lines 63-65] [Source]
  • New "Above-prompt actions" (abovePrompt:toggle, focus, next, press and more) and "Pane actions" (pane:scrollUp, grow, close and more) for mod-drawn UI, with default keys. [lines 396-437] [Source]

model-config [Source]

  • Alias table now includes haiku: Haiku 5.5 on the Anthropic API, Haiku 4.5 on other providers. Haiku 5.5 needs v2.1.293 or later. [lines 36-57] [Source]
  • Resumed sessions on a Haiku model move to whatever haiku resolves to now. [lines 139-141] [Source]
  • Haiku 5.5 supports all five effort levels and defaults to medium. /effort status shows the active level. Fallback effort text was simplified. [lines 525-593] [Source]

plugins/cli-reference [Source]

  • plugin marketplace add, remove and update accept --json and print a result object, documented with command, outcome, marketplace and failureCode fields (v2.1.287+). [lines 120-134] [Source]
  • plugin eval JSON report adds gatingHooks, which shows whether refusing mod hooks have a .catch handler. [line 636] [Source]

plugins/mods/api [Source]

  • Tools deferred by MCP tool search hide their description; set isDeferred: false to always show it. [line 64] [Source]
  • Documented what happens when a mod refuses a $.process.spawn call after the command has already started or finished. [lines 189-193] [Source]

plugins/mods/gallery [Source]

  • Explained when Image elements fall back to dimmed alt text (unsupported terminals, tmux/screen, background sessions) and the CLAUDE_CODE_FORCE_TERMINAL_IMAGES override. [lines 374-381] [Source]

quickstart [Source]

  • Noted that the installer shows no progress while downloading, and reorganized the example prompts and command lists into clearer sections. [lines 46-250] [Source]

statusline [Source]

  • New "Task fields" table for subagentStatusLine listing every field, including the new agentType (v2.1.293). [lines 1095-1121] [Source]
  • Troubleshooting restructured into headed sections. [lines 1132-1151] [Source]

env-vars [Source]

  • Values in the settings env key are copied as written with no shell expansion, so ~ and $HOME stay literal. Use absolute paths. [lines 82-83] [Source]

Many other Claude Code pages received smaller edits, including amazon-bedrock, microsoft-foundry, claude-apps-gateway-config, cloud-environments, desktop, ide-integrations, vs-code, skills and slash-commands.


API changes

New Documents

api-credits-for-subscribers [Source]

How Claude Max and Team plans include monthly Claude API credits. It explains how to claim them by linking a Console organization, which plans are eligible, and what the credits cover. They apply only to the Claude API in the Console, not to other platforms.

browser-use-sdk [Source]

How to run the browser use tool or computer use tool from the Python or TypeScript SDK. You subclass a toolset class and implement one method per member tool. The SDK handles routing, policies, approval callbacks and tool_result building. It covers the computer toolset too.

event-deltas [Source]

Managed Agents beta feature. Opt in with the event_deltas[] query parameter on the event stream to render agent.message text as a live preview while the model generates it. The buffered agent.message stays the authoritative record.

session-observability [Source]

Inspect Managed Agents sessions in the Console session viewer (timeline minimap, per-thread lanes), and read token usage and list cost to debug unexpected agent behavior.

haiku-5-5/migration-guide [Source]

Code changes needed to move from Claude Haiku 4.5 to Haiku 5.5.

haiku-5-5/overview [Source]

Model ID, pricing and limits for Claude Haiku 5.5.

haiku-5-5/whats-new-haiku-5-5 [Source]

Haiku 5.5 targets high-volume, latency-sensitive work such as classification, routing, extraction and subagents. It supports adaptive thinking with effort and a 1M context window with up to 128k output. It uses the newer tokenizer, which counts about 30% more tokens than Haiku 4.5, and its thinking blocks work only in the account that produced them or a linked one. Includes a table of changes from Haiku 4.5 and the action each needs.

prompting-claude-haiku-5-5 [Source]

Prompting guidance specific to Claude Haiku 5.5.

claude-haiku-5-5 [Source]

Published system prompt for Claude Haiku 5.5.

Changed documents

SDK reference pages

The generated per-language api/*/beta/messages reference pages (CLI, Ruby, Java, C#, Go and others) were regenerated with very large diffs, and the Haiku 5.5 model is likely reflected in them. They have no hand-written changes to summarize.