Claude Code 2.1.304 has been released. 76 CLI changes Highlights: • tool.check decider:\"person\" makes a mod require human approval; unapproved asks are refused with mod reason • Added /auto-mode-setup: guided scan of project and history so Auto Mode learns environment for smarter actions • Added /claude-api managed-agents-secure: read-only check flags Managed Agents deviations vs Anthropic guide Full details available in thread ↓

Claude Code 2.1.303 has been released. 1 CLI change Highlights: • CLI bug fixes reduce crashes and command failures, improving workflow reliability and predictability Complete details in thread ↓

Claude Code 2.1.302 has been released. 63 CLI changes Highlights: • Read tool refuses image sources over 128 MB to avoid loading large files into memory and crashing • Prompt suggestions omit credit-billed models after a credit-billed reply, avoiding unexpected credit use • Added 'Idle compaction' on/off in /config; post-compaction notices now show how to disable it Complete details in thread ↓

Claude Code 2.1.301 system prompt updates Notable changes: 1) Claude is now explicitly scoped to authorized security testing/defensive work/CTFs/education. It must refuse destructive techniques (incl. DoS, mass targeting, supply chain compromise, malicious detection evasion) and require clear authorization context for dual-use tooling. Links: 1) github.com/marckrenn/claude-…

Claude Code 2.1.301 is now available. 51 CLI changes, 1 system prompt change Highlights: • System prompt adds security-testing scope and must refuse out-of-scope probes, blocking unauthorized testing • Unattended retry mode now aborts requests when the next usage-limit reset is far in the future Complete details in thread ↓

Claude Code 2.1.300 system prompt updates Notable changes: 1) Claude no longer receives the explicit "AskUserQuestion" directive. In its place, the prompt now injects a "claude-mermaid-runtime-begin" marker right before the Harness block, shifting behavior away from forced question-asking and toward a runtime-wrapped/flagged instruction section (likely tied to Mermaid handling). Links: 1) github.com/marckrenn/claude-…

Claude Code 2.1.300 has been released. 88 CLI changes, 1 system prompt change Highlights: • Added virtualText: prompt-box text excluded from draft/submitted prompt, so it stays out of model input • $.ui.view({agentId}) shows agent transcript in main view; $.ui.view({}) returns to prior view for quick review • prompt.change hook fires on non-typed prompt edits (history/clear/undo/plugin fill), lets mods react Complete details in thread ↓

Claude Code 2.1.299 has been released. 190 CLI changes, 1 system prompt change Highlights: • Added --disallow-tools to claude plugin eval to block specified built-in tools during runs and avoid tool bias • Model accepts security-testing/dual-use prompts and uses AskUserQuestion to clarify, so they are not refused • Idle compaction compacts chats over 200K tokens after ≈54m idle, preventing full reprocessing on resume Full details are in thread below

Claude Code 2.1.296 has been released. 79 CLI changes Highlights: • Read tool adds allow_large option to let it ingest a large text file in one call when context allows • Managed-settings PreToolUse/prompt hooks now refuse denied tool calls and end the turn, stopping actions • Fix: redaction strips data after empty keys from shared transcrs & debug logs, incl shell JSON, stopping leaks Full details available in thread ↓

Claude Code 2.1.295 has been released. 143 CLI changes Highlights: • Command and HTTP hooks failing, timing out, or exiting abnormally now block the action, preventing execution • Gateway upstreams can include an optional models list that limits which models are sent, including on failover • Added timeouts.upstream_ttfb_ms for gateway upstreams to cap stream startup time; stalled streams time out Complete details in thread ↓

Claude Code 2.1.294 has been released. 2 CLI changes Highlights: • Instruction-form prompt and agent hooks now correctly block targeted commands, restoring intended enforcement • Improved judging of Stop and SubagentStop prompt hooks written as instructions, reducing premature agent stops Full details are in thread ↓

Claude Code 2.1.293 is now available. 56 CLI changes Highlights: • Haiku default changed to Claude Haiku 5.5 (1M context), allowing larger prompts and reducing per‑token cost • Fixed action-order in context compaction so done work isn't retracted or redone, yielding stable results • Backgrounding with ← preserves queued messages; if a queued message can't move, ← stays put and reports why Full details available in thread ↓

Claude Code 2.1.292 has been released. 92 CLI changes Highlights: • claude plugin install: added --marketplace — auto-adds marketplace if missing and installs plugin • Agent tool adds effort flag: spawned sub-agents run at the specified effort, enabling predictable cost control • macOS/Windows notebook/PDF reads fixed to stop returning files outside approved selection, preventing access Complete details in thread ↓

Claude Code 2.1.291 is now available. 2 CLI changes Highlights: • Fixed regression where cloud sessions could drop answers to permission prompts, preventing lost responses • Fixed regression in 2.1.288 where quitting could drop a session's last messages, preventing data loss Full details in thread below ↓

Claude Code 2.1.290 has been released. 190 CLI changes Highlights: • Sign-in approval page adds a Deny button to end pending sign-ins so waiting terminals stop in seconds • Replies halted by content filters are retried once before an error, reducing abruptly ended turns • claude attach and claude logs accept partial session names instead of full IDs, easing session access Complete details in thread ↓

Claude Code 2.1.289 is now available. 27 CLI changes Highlights: • Agents: teammates can spawn shared agents via agent.spawn; agent IDs unified and idle/waiting states clarified • Fixed user plugins rewriting org-managed MCP server sign-in descriptions, preserving org sign-ins • Read deny rules now apply to @-mentioned files, preventing @-references from bypassing file read restrictions Full details are in thread ↓

Claude Code 2.1.288 has been released. 89 CLI changes Highlights: • Added a tool to run bash commands and return output, enabling direct shell tasks from prompts • Added command to launch an agent for complex multi-step tasks, enabling delegated workflows and orchestration • Pressing Up on an empty prompt restores drafts cleared with Ctrl+C, recovering pasted text and images Complete details in thread ↓

Claude Code 2.1.287 has been released. 106 CLI changes Highlights: • Claude Mods allow plugins to modify deeper assistant behavior, enabling richer extensions and custom workflows • Fixed dangerous rm losing its always-ask prompt on output redirection, preventing accidental deletions Complete details in thread

Claude Code 2.1.286 has been released. 88 CLI changes Highlights: • Permission prompts now show counts like '2 of 5' when requests stack, so users see how many remain • Sessions retry the previous model once when the Anthropic API refuses a model, preventing every-turn failures • Log redaction fixed to mask secrets even when key names include invisible characters, preventing secret leaks Full details are in thread ↓

Claude Code 2.1.285 is now available. 136 CLI changes Highlights: • claude --desktop opens desktop app in the current directory or resumes a session for quick CLI→desktop flow • Added allowedProviders setting to restrict which API providers a machine may use, reducing external calls • Added CLAUDE_CODE_DISABLE_WEB_FETCH to disable WebFetch and stop automatic web fetches during tool runs Complete details in thread ↓