Claude Code v2.1.261: New Skill Doctor and Output Limits
Claude Code v2.1.261 (from Anthropic) introduces several practical additions for developers who rely on AI coding agents. The standout is /skill-doctor, a new diagnostic command that shows which loaded skills are never used and how much context they consume, making it easier to prune them and reclaim tokens.
New Capabilities
- Output limit settings:
bashOutputMaxCharsandtaskOutputMaxCharslet you raise how much command and background-task output Claude receives inline before saving to a file — up to 128K characters. - Large subagent prompts: The new
--append-subagent-system-prompt-fileflag reads the subagent system prompt from a file, accommodating prompts too large for the command line. - Policy diagnostics:
/statusandclaude doctornow include an "Organization policy" line that explains why your org's policy couldn't be loaded (e.g., a proxy not passing the endpoint).
Key Fixes
- Typed or pasted characters occasionally landing out of order or being dropped during fast input — fixed.
/add-dir <subdirectory>no longer prints a false "couldn't be resolved" error on/netautomounts.- Bedrock setup wizard: no more hanging when AWS never responds; now times out with a clear error. Also works behind TLS-inspecting proxies.
- Cloud sessions: plugin synced from claude.ai no longer discarded when managed settings force-enable it; avoids fallback to a marketplace clone.
- Remote Control: fixed stale permission mode, stuck spinners after stopping a turn, and inbound event stream failures behind TLS-inspecting proxies on Windows.
gcpAuthRefreshno longer opens a browser at startup if the credential is still valid but the check is slow.- Fixed high CPU usage when a background agent couldn't be resumed due to a tight retry loop.
- Feature flags gated to newer versions no longer apply to older versions running on the same machine.
/usageand the VS Code usage panel now correctly show model-specific weekly limits even when the endpoint is rate-limited.claude -p --resume <file>with a malformed session ID now resumes under a fresh ID.- Progress indicators in iTerm2, Ghostty, and ConEmu no longer show as finished while background workflows run.
Who It's For
Developers using Claude Code extensively, especially those with complex setups involving cloud sessions, remote control, or corporate proxies.
📖 Read the full source: GitHub Claude-Code
👀 See Also
Altman and Amodei Walk Back AI Job Apocalypse Predictions Ahead of IPOs
OpenAI's Sam Altman and Anthropic's Dario Amodei now admit they were wrong about AI eliminating white-collar jobs, as both companies eye $1 trillion IPOs. Goldman Sachs CEO David Solomon says he was right all along.
Claude Set a Real Alarm on Android via Intent System — No Hacking, But Transparency Issues Remain
Claude AI used Android's normal intent system to create a native alarm in the Samsung Clock app. The user initially feared a breach, but it's standard app-to-app communication. The lack of upfront disclosure about device-level actions raises transparency concerns.
Diagnosing Operational Drift and Task Amnesia in OpenClaw with Gemini 2.5 Flash on Proxmox
OpenClaw users report issues with persistent workflows on a Proxmox VM, citing operational drift and task amnesia. Despite stable performance in one-off tasks, the Gemini 2.5 Flash model struggles with automation and memory in this setup.
OpenAI to deploy AI models on U.S. Department of War classified network
OpenAI has reached a deal to deploy its AI models on the U.S. Department of War's classified network, with implementation scheduled for 2026. The Reuters article generated 15 points and 6 comments on Hacker News.