Claude Code v2.1.311 Fixes Startup Freezes, Long-Line Edits, and Context Overflow Halts
Claude Code v2.1.311 landed with two new self-hosted-runner flags and a stack of bug fixes that hit real developer pain points — startup freezes, context overflow, and edits on large files.
New Flags for self-hosted-runner
The release adds --hold-idle-file and --max-hold-min to self-hosted-runner. These keep an idle session alive while your host reports processes it left running — useful when a spawned process outlives the agent turn.
Context Overflow Now Compacts Instead of Stopping
Standard-context sessions used to halt with Usage credits required for 1M context once a request crossed 200K tokens. In v2.1.311, Claude Code compacts the conversation and continues. No more dead session from a long read or edit.
Startup and File-Edit Freezes
- Long freezes at startup and after reading/editing files when a CLAUDE.md file holds a very long list item.
- Write and Edit stalling up to 5 seconds after completing, and reporting 0 lines changed, when editing thousands of lines.
- Interrupts not taking effect until timeout while an MCP server hadn't answered a resource read or a skill was still downloading.
Subagent and MCP Fixes
- Subagents can now use MCP tools from servers that finish connecting after the subagent starts.
- Claude no longer sees stale MEMORY.md and CLAUDE.md content right after automatic compaction when those files were edited earlier in the session.
- SubagentStop hooks now fire once on finish instead of roughly twice a minute per running subagent.
- A subagent's isolated worktree is no longer removed while a background subagent it started is still working in it.
Session and CLI Fixes
claude -pand SDK sessions no longer save transcript messages out of order or keep retracted messages.claude mcp add,claude mcp remove, and/mcpnow name the project entry in~/.claude.jsonrather than the current folder when run below a repo root.claude mcp listandclaude mcp getno longer duplicate error codes (ENOENT: ENOENT:).claude respawn <id>no longer starts the background service before reporting that no session matches the id.- The "background service did not respond" error now suggests
claude daemon stop --anyinstead of the disabledclaude daemon restart.
Cloud and Remote Control Fixes
- Remote Control sessions no longer hang on "Connecting" for ~2 minutes when a proxy blocks a startup request.
- Messages Claude sends to a cloud or Remote Control session no longer appear as your own in the mobile apps.
- Claude is now told the cause when an attachment fails to save (e.g., disk full) or when attachments past the first 20 are dropped.
- WebFetch and WebSearch in Cowork cloud sessions now fail after 300 seconds on a stalled response instead of hanging.
- Cowork cloud sessions now ignore the
agentsetting so connectors and tools survive restarts.
UI and Tool Details
- The
@file picker now respects.ignore/.rgignorenegation rules that re-include git-ignored files. - The Bash tool description no longer claims the working directory always persists between calls.
- The environment block no longer names a shell that commands don't run in (e.g., fish when Bash uses zsh).
- Agent team sessions show the task list again on
--resume,--continue,/resume, or--session-id, including after/clearor a crash. - Pasting text that mentions an image path (e.g.,
! open /tmp/shot.png) no longer attaches the image and drops the path from the prompt.
If you use Claude Code daily, this release is worth pulling — especially if you've hit the 200K-token stop or the Write/Edit stall on large files.
📖 Read the full source: GitHub Claude-Code
👀 See Also

NYT Magazine covers real-world OpenClaw use in small businesses — gift article shared from Reddit
A New York Times Magazine article features OpenClaw users interviewed about their business use cases, originally posted on r/openclaw. Gift link included.

OpenClaw Docker users: 2026.3.13 update missing Docker tags
OpenClaw version 2026.3.13 has been released, but Docker users should avoid updating as the Docker image lacks both 'latest' and '2026.3.13' tags. Users running from npm or git are unaffected.

Claude vs GPT-4o: Same Double Pendulum Prompt, Different Coordinate Conventions
Claude and GPT-4o produce visually different double pendulum simulations because they interpret theta from opposite verticals — top vs bottom — while using the same renderer. The math is correct in both cases, but the mismatch reveals a subtle ambiguity in prompt interpretation.

Research on Professional Social Networks for AI Agents
Analysis of intent, behavior, and platform trends for professional AI agent social networks, focusing on Moltbook, Agent.ai, and Clawsphere, with examination of Meta's acquisition impact.