Jonathon's AI Wiki

What's New in Claude Code — Week 29 (July 13–19, 2026)

8 min read

Source: raw/anthropic-watch-claude-code-tag-v2-1-208.md + raw/anthropic-watch-claude-code-tag-v2-1-209.md (Anthropic-watch release stubs, reconciled against the official CHANGELOG.md for v2.1.208/209); ai-research/claude-code-docs-whats-new-w29-2026-07-17.md (official W29 digest page, fetched 2026-07-17); ai-research/claude-code-docs-changelog-2026-07-17.md (full changelog snapshot for v2.1.210–v2.1.212, fetched 2026-07-17). Continues Week 28; the reference tables in CLI Reference carry the new flags/env/settings.

The week of July 13–17 ships six releases (v2.1.207–v2.1.212) and two headline features. Artifacts call your MCP connectors — published artifacts can now call MCP connectors live on every page load, turning a one-session snapshot into a persistent live dashboard visible to anyone with the link. Screen reader mode (v2.1.208) ships plain-text rendering for VoiceOver, NVDA, and similar users. The release cluster also delivers /fork→background session + /subtask, claude auto-mode reset, session-wide WebSearch/subagent caps, MCP auto-background for long-running calls, and a large reliability and performance sweep across the full week.

Artifacts call your MCP connectors (headline)

Published artifacts can now call MCP connectors each time someone views them, so a dashboard shows live data and can take actions on demand rather than serving a snapshot from the session that built it. Each call runs through the viewing account’s own connections — not the creator’s — so each viewer sees data scoped to their own access. Viewers approve access before the page’s first connector call.

Companion updates shipping this week: public sharing links for artifacts, editor roles for shared editing on Team and Enterprise plans, and artifact creation from Claude Tag sessions. See Artifacts for the full feature treatment; the MCP-connector capability is the change that makes every page load pull fresh live data.

Example use:

“Build a dashboard artifact of open pull requests that pulls the live list through my GitHub connector when the page loads.”

Docs: code.claude.com/docs/en/artifacts#pull-live-data-with-mcp-connectors

v2.1.212 (July 17)

  • /fork → new background session. /fork now copies your conversation into a new background session with its own row in claude agents, so you can keep working in the original while the fork runs separately. The old in-session forked-subagent behavior (delegate while staying in the same session) is now /subtask. See Scheduled Tasks and Agent Teams for the background-session surface this builds on.
  • claude auto-mode reset [--yes] — new subcommand restores the default auto-mode configuration. --yes skips the confirmation prompt. See Auto Mode.
  • Session-wide WebSearch cap — CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION (default 200) stops runaway loops consuming unlimited search calls.
  • Session-wide subagent cap — CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION (default 200) stops runaway agent spawns. Both caps are tunable via env var.
  • MCP auto-background — MCP tool calls that run longer than two minutes now move to the background automatically so the session stays usable. Tune or disable the threshold with CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS.
  • /resume picker — /resume now opens a session picker when there is no obvious recent session to resume.
  • Fixed plan mode running file-modifying Bash commands without a permission prompt.
  • Fixed SIGTERM not propagating to the full Bash process tree when Claude terminates a command.
  • Prompt caching improvement for commands routed behind Claude Code gateways.
  • The Task tool’s mode parameter is now deprecated; scheduling has moved to the ScheduleWakeup tool.

v2.1.211 (July 15)

  • --forward-subagent-text + CLAUDE_CODE_FORWARD_SUBAGENT_TEXT — include subagent text and thinking blocks in stream-json output, so downstream systems consuming the stream see subagent reasoning rather than only parent-turn output. Directly relevant to automation pipelines built on claude -p --output-format stream-json.
  • Always-allow rules save at repo root. “Always allow” permission rules now save at the repository root (.claude/settings.json at the git root), so approvals granted inside a git worktree persist across all worktrees sharing that repo and across sessions.
  • continue: false halt no longer dropped. A PreToolUse hook returning continue: false now correctly halts the agentic loop — previously the halt was silently ignored when the output arrived mid-stream between two chunks. See Hooks.
  • Safety: auto mode no longer overrides a PreToolUse hook’s ask decision for unsandboxed Bash. A hook returning ask now floors the decision at a permission prompt instead of being silently overridden by the auto-mode classifier. See Auto Mode.
  • Security: permission previews relayed to chat channels now neutralize bidirectional-override, zero-width, and look-alike quote characters before display — prevents visual spoofing of permission approval messages.
  • Fixed parallel Claude Code sessions all logging out simultaneously after wake-from-sleep.
  • Fixed nested .claude/rules/*.md files loading when project setting sources are excluded.
  • Integer env vars (e.g. CLAUDE_CODE_MAX_OUTPUT_TOKENS) now accept scientific notation (1e6 = 1000000, not 1).
  • Fixed /loop sessions hidden from /resume session picker history.
  • Fixed /clear not resetting the session cost counter.

v2.1.210 (July 14)

  • Live elapsed-time counter on the collapsed tool-summary line — long-running tool calls now visibly tick instead of looking stuck.
  • Startup warning for invalid permission rule targets — Write(path), NotebookEdit(path), Glob(path) now warn on startup; use Edit(path) or Read(path) instead. Previously failed silently with no diagnostic.
  • Worktree-isolation fix — isolation: 'worktree' subagents could run git-mutating commands against the main repo checkout instead of their own isolated worktree. Now correctly contained.
  • ultracode keyword fix — no longer fires on non-human-originated input (webhook payloads, relayed PR comments).
  • Fixed a hook callback timeout being misreported to the model as a user rejection, causing unattended sessions to stop and wait.
  • Fixed Claude assuming a cd took effect after its command was moved to the background.
  • Fixed paste markers leaking into external editors (stray È/É characters around pasted text).
  • Fixed claude attach failing with “job not found” or “agent is still starting” during session transitions.

v2.1.209 (hotfix)

Fixed /model and other dialogs being blocked in claude agents background sessions — reverts an overly broad guard introduced by v2.1.208’s background-session work. (v2.1.208 and v2.1.209 are a linked pair: v2.1.208 hardened the background daemon; v2.1.209 undoes the guard from that work that broke dialogs.)

v2.1.208 — accessibility + background agents + performance

New surface (also in CLI Reference):

  • Screen reader mode — opt-in plain-text rendering for screen-reader users. Enable via claude --ax-screen-reader, CLAUDE_AX_SCREEN_READER=1, or "axScreenReader": true in settings. Claude Code’s first real accessibility surface.
  • vimInsertModeRemaps setting — map two-key insert-mode sequences (e.g. jj → Escape) in vim mode.
  • CLAUDE_CODE_PROCESS_WRAPPER — enterprise/managed-launcher control: agent view and the background service route every Claude Code self-spawn through a required corporate wrapper executable.
  • Mouse-click support for multi-select menus and “Other” input rows in fullscreen mode.

Background-agent reliability:

  • Replies typed to a background agent are now saved and delivered on session restart when delivery fails (were silently lost).
  • Fixed permanent attach failure (“Couldn’t start the background daemon”) after an update replaced the binary a running claude agents process launched from.
  • An older daemon no longer silently restarts newer-version workers onto the older binary.
  • Completed background agents stay listed in /tasks until cleanup instead of vanishing on finish; attaching to a stopped agent shows its transcript immediately instead of a blank “Session is starting.”

Correctness fixes (selected):

  • Fixed the context-window / auto-compact indicator briefly resetting to 200k after CLI auto-update.
  • Fixed supervised + background sessions crashing on an HTTP/2 GOAWAY with requests in flight.
  • Fixed truncated stream-json/JSON output and a missing result message when piping large claude -p responses.
  • Fixed CLAUDE_CODE_MAX_OUTPUT_TOKENS (and similar) silently using the mantissa of scientific-notation values (1e6 became 1).
  • Fixed Read reporting empty files as “shorter than offset,” Grep silently returning “No files found” on invalid regex and under-reporting paginated counts, and Glob crashing on null bytes in the pattern/path.
  • Catastrophic removals hidden inside $(…) / backticks / <(…) now prompt in --dangerously-skip-permissions and auto mode, matching the plain form.
  • /release-notes “Show all” no longer injects the entire changelog into every subsequent request’s context.

Performance / memory:

  • Multiple long-session memory-leak fixes: MCP stdio stderr accumulating up to 64 MB/server, LSP documents now LRU-capped at 50, async hook output freed after backgrounding, and unbounded growth from large tool-result payloads in headless/SDK sessions.
  • Up to 7× faster tool rounds in print/SDK sessions with many MCP tools (tool-pool assembly cached).
  • Up to 79× smaller session transcripts in edit-heavy sessions (pruned superseded file-history backups).
  • File-edit read cache bounded to 16 MB; large markdown tables capped at 200 rendered rows with a ”… N more rows” notice.

v2.1.207 (carryover from Week 28, July 11)

  • Auto mode now available without the CLAUDE_CODE_ENABLE_AUTO_MODE opt-in on Bedrock, Vertex AI, and Foundry (disable via disableAutoMode).
  • Security: fixed remote managed settings from a non-interactive run (claude -p, SDK) being permanently recorded as consented without the consent dialog ever showing.
  • Changed Bedrock, Vertex, and Claude Platform on AWS to default to Opus 4.8.
  • Plugin shell-injection fix: ${user_config.*} in shell-form commands is now rejected (use exec-form args or $CLAUDE_PLUGIN_OPTION_<KEY>); plugin option values no longer read from project-level .claude/settings.json.
  • Fixed spurious prompt-injection warnings on benign system-generated conversation updates; several Bedrock AWS-SSO credential fixes. (Already reflected in CLI Reference.)
  • Claude Code CLI Reference — new flags/env/settings: --forward-subagent-text, CLAUDE_CODE_FORWARD_SUBAGENT_TEXT, CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION, CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION, CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS, claude auto-mode reset [--yes], /fork→background-session, /subtask.
  • Artifacts — MCP-connector live-data feature (the W29 headline); also adds public sharing links, editor roles (Team/Enterprise), Claude Tag artifact creation.
  • Auto Mode — claude auto-mode reset [--yes] command (v2.1.212); auto mode no longer overrides PreToolUse hook ask decisions (v2.1.211).
  • Claude Code Hooks — continue: false halt fix (v2.1.211); hook timeout misreported as user rejection fix (v2.1.210); auto mode PreToolUse-ask override fixed (v2.1.211).
  • MCP — Model Context Protocol — MCP auto-background (>2 min threshold, CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS, v2.1.212).
  • Claude Code Scheduled Tasks — /fork now creates a background session; /subtask is the new in-session delegate (v2.1.212).
  • Agent Teams — /fork→background session is a new first-class agent-view primitive; session caps (v2.1.212).
  • What’s New — Week 28 — prior digest: in-app browser, /checkup, v2.1.202–206.

Graph View

Backlinks