Claude Code v2.1.238: keybindingFlavor, headersHelper, self-hosted runner fixes

✍️ OpenClawRadar📅 Published: August 21, 2026🔗 Source
Ad

Claude Code v2.1.238 lands with a handful of new settings and a long list of bug fixes. The headline additions: a keybindingFlavor setting for Bash-style Ctrl+W, a headersHelper for plugin marketplaces that require HTTP headers, and two new flags for self-hosted runners.

New settings and commands

If you've ever hit Ctrl+W in the Claude Code prompt expecting to delete a word but got something else, set keybindingFlavor to "readline". The default remains "classic", but readline mode makes Ctrl+W delete back to the previous whitespace, just like in Bash.

Plugin marketplaces now support a headersHelper on a URL marketplace or catalog entry. This command mints HTTP headers—like a short-lived token—for fetching catalog data and same-origin archives. Note that the helper runs only on install or update, after showing the command. claude plugin install/update will prompt [y/N], or you can pass -y to skip.

For self-hosted runners, two new flags:

  • claude self-hosted-runner --defer-shutdown-max-min : on SIGTERM, keeps serving attached sessions, parks remaining work after the given minutes, then exits.
  • claude self-hosted-runner --proxy-authorization-command / --proxy-authorization-file: for egress proxies that need a freshly issued Proxy-Authorization header on every connection.

Fixes that matter

The release fixes unbounded memory growth in long interactive sessions — subagent tool results are now released once they leave the recent display window. No more silent memory leaks.

Output styles for custom, project, and plugin configurations no longer drift back to the default voice mid-session. And CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=true now works even when your account is near (but not over) its usage limit.

Worktree-isolation Bash refusals won't falsely tell you to remove a redirect that isn't there. Self-hosted runners are less likely to be dropped by the server after a slow poll request, avoiding handoffs of healthy sessions.

MCP elicitation dialogs now show URLs longer than 4,096 characters, and permission prompts keep the "don't ask again" option even when the project path is too wide. Leftover /tmp/claude-*-cwd files are cleaned up when a Bash command is killed, times out, or is interrupted.

Held Backspace works on terminals that send Ctrl+H for Backspace during keystroke bursts (slow SSH/mosh). Text wrapping in permission prompt diffs no longer clips wide characters like emoji or tabs. Killing a suspended (Ctrl+Z) session no longer leaves bracketed-paste mode stuck with a hidden cursor.

stdio MCP servers won't receive a server/discover request before initialize, preventing forced lazy server startups. Proxy connection refusals now name the proxy instead of a generic network error. The /model and /effort cache-miss warning no longer appears when the prompt cache has already expired.

Ad

Remote Control and cross-session fixes

Several Remote Control issues are addressed:

  • Per-task Stop from the tasks panel now works on CLI-hosted sessions.
  • Sessions started by claude remote-control no longer inherit session-scoped environment variables from the launching shell.
  • Crash recovery: a crashed session can now be reused on next message, no restart needed.
  • Messages sent from web/Desktop while Claude is mid-turn no longer disappear.
  • Model picks made on phone/web now update the terminal display.
  • "login expired" disconnects during network hiccups are now retried, keeping the connection.
  • Sign-out ends with a clear message instead of a failed reconnect report.
  • ListAgents / SendMessage now work in server mode and Desktop/IDE hosts, reaching Remote Control peers.
  • Idle workers pre-warmed for background sessions are hidden until a task claims them.

Cross-session messaging: if a session refuses inbound messages (e.g., crossSessionInbound: "refuse"), the sender gets a "refused" status. If an inbox drops messages due to rate limiting or full queue, you're notified now.

Finally, bare claude starts faster on macOS, and Bash handling saw improvements (including output handling for killed commands).

📖 Read the full source: GitHub Claude-Code

Ad

👀 See Also

Krasis: Hybrid CPU/GPU Runtime for Large MoE Models Achieves 3,324 tok/s Prefill on RTX 5080
Tools

Krasis: Hybrid CPU/GPU Runtime for Large MoE Models Achieves 3,324 tok/s Prefill on RTX 5080

Krasis is a hybrid CPU/GPU runtime that runs large MoE models by handling prefill on GPU and decode on CPU, achieving 3,324 tokens/second prefill on an RTX 5080 with Qwen3-Coder-Next 80B Q4. It requires ~2.5x model size in system RAM but enables running models too large for VRAM.

OpenClawRadar
TextForge: A Claude Code-built email approval tool for LLM workflows
Tools

TextForge: A Claude Code-built email approval tool for LLM workflows

A developer built TextForge using Claude Code to automate email workflows with mandatory approval gates, preventing LLMs from sending emails without explicit permission. The tool integrates with Pipedrive CRM and required Google CASA2 security audit compliance.

OpenClawRadar
OpenClaw Multi-Agent Workflow Issues: Stalling, Context Loss, and Token Inefficiency
Tools

OpenClaw Multi-Agent Workflow Issues: Stalling, Context Loss, and Token Inefficiency

A developer reports OpenClaw multi-agent workflows frequently stall with agents hanging, experience context leakage despite custom documentation, and consume excessive tokens with no output. The setup used Gemini 3 Pro/Codex models with a COO orchestrator and specialized task agents.

OpenClawRadar
Crime Team: Multi-Agent Orchestrator for OpenClaw — Parallel Code Review with Coder Agent
Tools

Crime Team: Multi-Agent Orchestrator for OpenClaw — Parallel Code Review with Coder Agent

Crime Team v0.1 runs multiple specialist OpenClaw agents in parallel for code review, then integrates findings. Includes per-agent models, a coder agent that applies changes, and a re-audit loop. CLI + GUI.

OpenClawRadar