ClaudeCodeChanges

New version Docs only
S M T W T F S
1 2 3 4 5 6 7
8 9 10 11 12 13 14
15 16 17 18 19 20 21
22 23 24 25 26 27 28

Claude docs changes for February 24th, 2026 [diff]

Executive Summary

  • Claude Code 2.1.51 adds a new claude remote-control subcommand for external builds, BashTool performance improvements, and a security fix for hooks executing without workspace trust
  • The prompt engineering best practices guide was significantly overhauled and consolidated into a single comprehensive reference, replacing individual topic pages
  • New extensive guidance on multi-session software development patterns using memory files, git checkpointing, and context recovery for long-running agent workflows
  • Introduced the concept of "context rot" — how accuracy degrades as context grows — with new recommendations on context curation for agentic systems
  • New C# SDK endpoint for the Files API: Upload File (beta)

New Claude Code versions

2.1.51

New features

  • Added claude remote-control subcommand for external builds, enabling local environment serving for all users
  • Added support for custom npm registries and specific version pinning when installing plugins from npm sources
  • Added CLAUDE_CODE_ACCOUNT_UUID, CLAUDE_CODE_USER_EMAIL, and CLAUDE_CODE_ORGANIZATION_UUID environment variables for SDK callers to provide account info synchronously, eliminating a race condition where early telemetry events lacked account metadata

Existing feature improvements

  • Updated plugin marketplace default git timeout from 30s to 120s; added CLAUDE_CODE_PLUGIN_GIT_TIMEOUT_MS to configure it
  • BashTool now skips login shell (-l flag) by default when a shell snapshot is available, improving command execution performance (previously required setting CLAUDE_BASH_NO_LOGIN=true)
  • Tool results larger than 50K characters are now persisted to disk (previously 100K threshold), reducing context window usage and improving conversation longevity
  • The /model picker now shows human-readable labels (e.g., "Sonnet 4.5") instead of raw model IDs for pinned model versions, with an upgrade hint when a newer version is available

Major bug fixes

  • Fixed a security issue where statusLine and fileSuggestion hook commands could execute without workspace trust acceptance in interactive mode
  • Fixed a bug where duplicate control_response messages (e.g. from WebSocket reconnects) could cause API 400 errors by pushing duplicate assistant messages into the conversation
  • Fixed slash command autocomplete crashing when a plugin's SKILL.md description is a YAML array or other non-string type

2.1.52

Major bug fixes

  • VS Code: Fixed extension crash on Windows ("command 'claude-vscode.editor.openLast' not found")

Claude Code changes

Changed documents

Agent teams [Source]

  • Added new section on choosing an appropriate team size: recommends starting with 3–5 teammates, explains that token costs scale linearly per teammate, and that coordination overhead increases with more agents. [line 337] [Source]

MCP [Source]

  • Added GraphOS MCP Tools (Apollo GraphQL) — search Apollo docs, specs, and best practices. [line 672] [Source]
  • Various reordering of remote MCP server entries in the marketplace list

Terminal config [Source]

  • Updated notification setup documentation: Kitty and Ghostty now noted as supporting desktop notifications without additional configuration; simplified iTerm 2 setup steps; clarified that other terminals (including macOS Terminal) do not support native notifications and should use notification hooks instead. [line 32] [Source]

API changes

New Documents

Prompting tools [Source]

New guide to Claude Console prompting tools. Documents the prompt generator (for creating first-draft prompts), prompt templates and variables (using {{double brackets}} for dynamic content), and the prompt improver (which enhances prompts through 4 automated steps: example identification, initial draft, chain-of-thought refinement, and example enhancement).

Upload File (C# beta) [Source]

New C# SDK reference for the Beta.Files.Upload method (POST /v1/files). Documents the FileUploadParams input, all supported beta headers, and the FileMetadata return type (ID, created timestamp, filename, MIME type, size, and downloadable flag).

Changed documents

Customer support chat [Source]

  • Internal links updated from individual prompt engineering topic pages to the consolidated claude-prompting-best-practices.md reference

Remote MCP servers [Source]

  • Added Pendo — connect to Pendo for product and user insights (user-specific URL). [line 425] [Source]
  • Added Open Targets — drug target discovery and prioritisation platform (https://mcp.platform.opentargets.org/mcp). [line 436] [Source]
  • Added Honeycomb and Consensus to the API-level remote MCP list (previously only in the Claude Code MCP page)
  • Various reordering of MCP server entries throughout the list

Bash tool [Source]

  • Added Terminal-Bench 2.0 benchmark results note highlighting performance gains with a persistent bash session. [line 4] [Source]
  • Added new Git-based checkpointing section with a 4-step pattern for using git as a structured recovery mechanism in long-running agent workflows (baseline commit, per-feature commits, state reconstruction, revert-on-failure). [line 239] [Source]

Computer use tool [Source]

  • Added WebArena benchmark results note (state-of-the-art among single-agent systems for autonomous web navigation). [line 4] [Source]
  • Added guidance for multi-session agents: run end-to-end verification at the start of each session (not only after implementation), with a link to the "Effective harnesses for long-running agents" engineering article. [line 278] [Source]

Implement tool use [Source]

  • Added three new tool design best practices: consolidate related operations into fewer tools (use an action parameter instead of separate tools), use meaningful namespacing in tool names (prefix by service), and design tool responses to return only high-signal information (semantic identifiers, minimal fields). [line 50] [Source]
  • Added reference to the "Writing tools for agents" engineering article. [line 58] [Source]

Memory tool [Source]

  • Added explanation of the just-in-time context retrieval principle: store what agents learn in memory and pull it back on demand rather than loading everything upfront. [line 6] [Source]
  • Added new Multi-session software development pattern section: describes a 3-phase workflow (initializer session, subsequent sessions, end-of-session update) using memory files as a structured recovery mechanism, with the principle of completing one feature at a time and verifying end-to-end before marking complete. [line 451] [Source]

Tool use overview [Source]

  • Added benchmark results illustrating the impact of tool access (LAB-Bench FigQA and SWE-bench). [line 4] [Source]
  • Added link to "Advanced tool use" article as a next step for workflows with many tools. [line 249] [Source]

Programmatic tool calling [Source]

  • Added BrowseComp and DeepSearchQA benchmark results showing programmatic tool calling as the key factor unlocking agent performance on multi-step research tasks. [line 4] [Source]
  • Added concrete efficiency example: checking budget compliance for 20 employees goes from 20 model round-trips to a single script, reducing context from hundreds of KB to a few lines. [line 7] [Source]

Tool search tool [Source]

  • Improved context bloat explanation with concrete numbers: a typical multi-server setup (GitHub, Slack, Sentry, Grafana, Splunk) can consume ~55K tokens in definitions; tool search typically reduces this by over 85%. [line 7] [Source]
  • Added best practice recommendation for consistent namespacing in tool names (prefix by service, e.g. github_, slack_). [line 495] [Source]

C# beta SDK reference (beta.md and related files) [Source]

  • Added new code_execution_20260120 beta enum value across code execution tool types
  • Added BetaCodeExecutionTool20260120 class with DeferLoading and Strict fields
  • Internal class renamed from UnionMember2 to All for BetaAllThinkingTurns

C# beta Files (beta/files.md) [Source]

  • Added the Upload File endpoint (Beta.Files.Upload) to the Files API index

Compaction [Source]

  • Expanded description to explain that compaction isn't only about staying under token limits — long contexts cause models to lose focus, and compaction actively keeps the active context performant. Added link to "Effective context engineering" article. [line 7] [Source]

Context editing [Source]

  • Enhanced explanation: context editing is about actively curating what Claude sees, not just cost optimization. Context is a finite resource with diminishing returns, and irrelevant content degrades model focus. Added link to "Effective context engineering" article. [line 8] [Source]

Context windows [Source]

  • Introduced "context rot" concept: accuracy and recall degrade as token count grows, making context curation as important as available context size. [line 11] [Source]
  • Added Claude's state-of-the-art results on MRCR and GraphWalks long-context retrieval benchmarks. [line 14] [Source]
  • Added guidance for multi-session agents: design state artifacts for fast context recovery at session start, with a reference to the memory tool's multi-session pattern. [line 155] [Source]

IP addresses [Source]

  • The four previously "pending phase-out" IP addresses (34.162.46.92, 34.162.102.82, 34.162.136.91, 34.162.142.92) are now confirmed phased out and grouped with the existing phased-out address (34.162.183.95). Users should remove all five from firewall allowlists. [line 22] [Source]

Claude prompting best practices [Source]

  • Major restructure: the document is now the single comprehensive prompt engineering reference, consolidating content previously spread across multiple topic pages
  • Added "golden rule" analogy for instruction clarity: show your prompt to a colleague with minimal context — if they'd be confused, Claude will be too. [line 35] [Source]
  • Added detailed guidance on using examples effectively (few-shot prompting): relevance, diversity, use of <example> tags, and 3–5 examples as a target. [line 31] [Source]
  • Added dedicated sections on XML tag structuring, giving Claude a role, long-context prompting (put longform data at top, use XML for documents, ground responses in quotes), and model self-knowledge prompts. [line 56] [Source]
  • Added new "Output and formatting" section covering format control strategies (tell what to do vs. not to do, XML format indicators, matching prompt style to desired output style). [line 112] [Source]

Prompt engineering overview [Source]

  • Removed the step-by-step ordered list of prompting techniques and replaced with a single pointer to claude-prompting-best-practices.md as the living reference. Added link to the new prompting tools page. [line 24] [Source]