Claude Code 2.1.219

• [new] Initial support for the claude-opus-5 model ID (via CLAUDE_5_OPUS), so Claude Code can target Opus 5 as it becomes available.
• [changed] Memory writes are stricter about what counts as worth saving: durable, applicable lessons (corrections, preferences, non-obvious procedures) rather than trivia, CLAUDE.md/code facts, or transient sandbox/CI quirks. Corrections and preferences are meant to be saved in the same turn they appear, before the reply finishes.
• [changed] Self-correction is toned down in user-facing text. Only mistakes that change the user’s code, conclusions, or decisions get an explicit correction; accurate earlier statements are left alone, and minor slips are fixed quietly without apologies or long post-mortems.
• [changed] Stronger defaults on task scope: deliver what was asked at the intended breadth, finish the full task before reporting done, and keep written deliverables (especially Markdown) substantive without filler sections or boilerplate.
• [changed] Subagent usage guidance: avoid spinning multiple subagents for one small job; use parallelism for genuinely independent, sizeable tracks, and don’t redo a subagent’s work after it reports back.
• [changed] Server-side fallbacks support updated for the Claude API: the array form still uses the earlier beta header; the "default" scalar form pairs with server-side-fallback-2026-07-01. Fallback entries remain limited at launch (substitute model support centered on claude-opus-4-8), with per-hop max_tokens and further overrides rolling out.
• [new] Support for the mid-conversation-tool-changes-2026-07-01 API beta, for tool definitions that can change over the course of a conversation.
• [changed] Focus mode behaviour clarified: when it is on, only the final text of each response is shown—progress chatter between tool calls is skipped so everything the user needs lands in that final message.
• [changed] Fast-mode migration guidance updated away from retired *-fast model strings toward current Opus fast mode (Opus 4.8 / the active fast-capable default), avoiding silent fallback or hard API errors from old IDs.
• [changed] Adding a working-directory root is constrained to strict subdirectories of cwd (or dirs passed at launch via --add-dir / SDK additionalDirectories); duplicates are rejected without re-running registration hooks.
• [changed] SDK compatibility surface extended to include 0.3.218.

Models & prompts

• claude-opus-5 is recognized as a model identifier; built-in guidance and compaction notes also track the current Opus lineup (including Opus 4.8 and related tiers).
• Built-in agent behaviour around memory, corrections, subagents, readability vs. brevity, and “disabled thinking” routes is tightened so tool calls stay real tool calls and user-visible text stays selective but complete sentences—not compressed jargon.
• Plan mode, permission denials, git stash safety across worktrees, and computer-use MCP instructions remain in place with wording cleanups only where noted above.

Under the hood

• New CLAUDE_CODE_DISPATCH_V2S flag for an alternate dispatch path.
• Additional internal feature gates around memory directories (pin/write and pinned injection), org/team memory mounting and recovery, and related experiments—not standalone user-facing commands yet.
• Beta headers added for mid-conversation tool changes and the newer server-side fallback shape described above.

Upgrade notes

• If you call the Messages API with server-side fallbacks, match the beta header to the parameter shape: array form → server-side-fallback-2026-06-01; "default" scalar form → server-side-fallback-2026-07-01. Mixing them returns 400. Still not available on Batches, Bedrock, Vertex, or Foundry (use client-side/SDK fallback there).
• Replace any remaining claude-opus-4-6-fast / claude-opus-4-7-fast (or similar) model strings with supported Opus fast-mode usage so you do not get silent downgrades or hard errors.