Claude Code v2.1.314: blockWritesOutsideWorkingDirectories and 20+ bug fixes
Claude Code v2.1.314 ships a new filesystem guard plus a long list of stability fixes across Bedrock, Vertex, MCP, VS Code, and the CLI. The headline feature is a permission setting that stops Write, Edit, and NotebookEdit from touching paths outside the working directories.
New permission: blockWritesOutsideWorkingDirectories
Add this to your settings to make the file-editing tools refuse any path outside the working directories — in every permission mode, not just the default:
{
"permissions": {
"blockWritesOutsideWorkingDirectories": true
}
}
If you run agents in CI or on shared machines, this closes the gap where a tool could write to /etc, home dotfiles, or a sibling repo depending on the active mode.
Gateway and admin controls
- Claude apps gateway upstreams now accept an optional
groupslist. Only developers in the listed IdP groups are routed to that upstream — and that holds during failover too. - Bedrock upstreams gain
guardrail_trace, which logs which Amazon Bedrock guardrail filter fired on a request without logging the matched text. - Admin TLS trust settings in managed settings now actually apply in Claude Desktop sessions.
forceRemoteSettingsRefreshset only in remote managed settings no longer stops applying once the stored copy expires — previously a session could start without policy if the server was unreachable.
CLI correctness fixes
claude -pused to exit 0 with no output when startup failed. It now prints the error and exits 1, so scripts and CI stop treating a broken start as success.- Sessions behind an org-mandated effort cap no longer fail turn after turn with "Invalid effort level." Claude Code lowers the effort, retries, and tells you which level it used.
claude ultrareviewandclaude remote-controlno longer ask the workspace trust question whenCIis set; they refuse an untrusted folder immediately.claude mcp logoutno longer reports success when the sign-in wasn't removed (for example, a locked macOS keychain over SSH).- Remote Control model switches that take over 8 seconds now return the final result instead of a "still pending" error, and worker registration retries with growing waits when the session service is slow.
Bedrock, Foundry, and TLS
- Bedrock sessions with many images no longer fail every turn and
/compactno longer returns "400 Input is too long" — older images are dropped before compacting. - Microsoft Foundry deployments missing the default Sonnet, Opus, or Haiku model now fall back to an older model in the same family instead of failing.
UNABLE_TO_GET_ISSUER_CERTon TLS connections is fixed when the system cert store holds a renewed root cross-signed with its predecessor.- Gateways that reject tool search's
tool_referenceblocks trigger one retry without tool search, and the session remembers the fallback.
UI and background session fixes
- VS Code and Claude Desktop's Code tab now do idle compaction: a 200K+ token conversation left idle is summarized, and the summary is used on your next message if the prompt cache has expired.
- PreCompact hooks have their output truncated at 10,000 characters, so a verbose hook can't make compaction fail or loop.
- Background diff refreshes no longer briefly take the repo index lock, which was breaking concurrent
gitcommands with "Unable to create '.git/index.lock': File exists." - Markdown tables stop jumping between grid and stacked layouts on half-typed streaming rows; long tables no longer fall back to raw pipes.
- "Continue automatically at usage limit" now stops at three continuations when a task restarts itself between waits.
- Ultracode stays off after being disabled, instead of being re-enabled at the next background or cloud session start.
Smaller fixes
- The
/resumepicker no longer shows CLAUDE.md or CLAUDE.local.md text as a session title when the session only ran slash commands. /btwfrom a subagent's transcript now answers from that transcript, not the main conversation.- Pasting an image over SSH to Linux now prints guidance to copy the file and paste its path instead of doing nothing silently.
- Model switching works when a broken user-supplied
--plugin-diror--plugin-urlplugin exists alongside managed plugin settings. - The Bedrock/Vertex model upgrade offer stops reappearing at every start when managed settings pin the model.
Anyone running Claude Code in CI, across multiple cloud providers, or behind a gateway should pull this one — the claude -p exit-code fix alone will catch failures that were being silently swallowed.
📖 Read the full source: GitHub Claude-Code
👀 See Also
OpenClaw Release Delayed Again: Upgrade Test Issues from v2026.7.1
OpenClaw's planned August 18 release is delayed as testing finds issues with fresh installs and upgrades from v2026.7.1 and earlier. No new date set.
DoltLite Beta: SQLite fork with Git-style version control, built with 2k agent PRs
DoltLite 0.50.0 is a SQLite fork with Git-style branching, merging, and sync, built via 2,000 agent PRs. Beta brings storage format stability and SQL compatibility.

Constraint Decay: Why LLM Agents Fail at Structured Backend Code
New research introduces 'constraint decay': as structural requirements accumulate, LLM agent performance drops drastically — capable agents lose 30 points in assertion pass rates, weaker ones approach zero. Actionable insights for anyone using AI coding agents.

AI Agents Hiring Other AI Agents: From Solo Workers to Networked Economies
A Reddit post argues that AI agents will evolve from isolated tools into networked workers that delegate tasks, specialize, build reputation, and exchange value — shifting the hard problem from intelligence to coordination.