Claude Code v2.1.316 adds subagent effort control, pooled budgets, and fixes for orphaned shells
Claude Code v2.1.316 is out. It adds a handful of environment knobs for controlling subagent behavior and policy merging, extends the Claude apps gateway with shared budgets and Bedrock guardrail tagging, and fixes a long list of shell, hook, and model-fallback bugs.
New environment variables
CLAUDE_CODE_DISABLE_ULTRATHINK— turns theultrathinkkeyword off.CLAUDE_CODE_SUBAGENT_EFFORT— sets the effort level subagents run at when nothing else names one, instead of inheriting the session's effort.CLAUDE_CODE_DISABLE_LOOPED_CROSS_CERT_DROP— keeps system cross-certificates that certify each other in a loop, for servers whose private CA needs one.
Gateway changes
Pooled group budgets are now available on the Claude apps gateway: one monthly budget that a group's members share, on top of each member's own limit. Set it with pooled_budget on the spend limits admin API.
managed.policy_mode: layered makes every policy that matches a developer apply, each merged over the ones listed below it — instead of only the first matching policy.
For Bedrock upstreams, guardrail.input_tagging wraps prompts in the input tags an Amazon Bedrock guardrail's prompt-attacks filter needs. all_messages_and_system also tags the system prompt.
Deprecation: Opus 4.8 fast mode
Fast mode for Opus 4.8 is deprecated as of October 30, 2026, and will be removed on November 30, 2026. After removal, fast mode requests on Opus 4.8 return an error and do not fall back to standard Opus 4.8. To keep the speedup, migrate to Opus 5.5.
Notable fixes
- Shell commands are no longer left running on macOS/Linux when Claude Code exits on its own (e.g. credentials refused, cloud session archived).
- Bash aliases with line breaks (except Git Bash) and functions whose names start with a dash are no longer garbled or dropped in the Bash tool's shell.
- Bash scripts started from the Bash tool no longer stop after the first command when
SHELLOPTSis exported. - PowerShell hooks now block on exit code 2 when the command begins with a type cast,
param,using, or an attribute. - Network requests no longer hang when a compressed response is cut off or corrupted.
sonnet,opus, andhaikuon Vertex AI and Amazon Bedrock now switch to an older model mid-session instead of erroring when the newest model isn't available to the account or region.- The gateway now checks every model a request names against
availableModels,deniedModels, andmodels; a blocked model returns a 400 naming the field. claude auth loginruns where managed settings require a gateway and shows the gateway sign-in in a terminal.--restrictedsessions no longer upload files from outside the working directories when a Read allow rule covered them.- Prompts containing invisible characters send cleaned on the first Enter; commands still wait for review.
- Sub-agents that inherited a parent sub-agent's model no longer jump to the main model when a background job or message wakes them.
- A managed
autoMode.allowlist without"$defaults"keeps the built-in exceptions from host--settings,/permissions, or/auto-mode-setup. - Cloud sessions and cloud scheduled tasks are marked failed when auto mode stops a turn because the server returned no safety verdict.
claude -ptext output now says why auto mode stopped a turn when a background agent is still running.CLAUDE_CODE_AUTO_MODE_SERVER=1is honored in host-app sessions (e.g. Claude Desktop) behind a proxy or LLM gateway with experimental betas off.- Plugin MCP
headersHelperscripts no longer get a corruptedCLAUDE_PLUGIN_ROOTand server URL when a credential-named env var is set to a switch value like1. - The SDK
register_repo_rootcontrol request loads theCLAUDE.mdof a directory outside the working directory whenCLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1is set. - The Artifact tool no longer shows
?forself-level artifact DB access rules, and quotes them back when refusing a republish that would drop them. /model <id>warns when the model isn't recognized by this version and auto-compact will assume a 200K context window.- Grep and Glob return instead of hanging when the search process is killed at timeout without exiting.
If you run subagents with a different effort level than your main session, or you administer a shared gateway, this release is worth pulling.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Claude.ai Experiencing Elevated Errors and Login Issues
Claude.ai is reporting elevated errors affecting the platform, including login issues specifically for Claude Code. The incident was officially posted on March 11, 2026 at 17:19:35 UTC.
Governments Are Betting Big on AI — The Economist Warns of Risks
The Economist argues that governments are making a dangerous bet on the AI boom, risking economic and security pitfalls. Key concerns: over-reliance on tech giants, hasty regulation, and potential job displacement.

AI Interview Platforms Tested: CodeSignal, Humanly, Eightfold in Job Screening
The Verge tested three AI interview platforms including CodeSignal, Humanly, and Eightfold for job screening. The AI avatars conduct one-on-one video interviews, analyze responses, and claim to reduce bias, though bias-free systems remain impossible due to training data limitations.

AI Scans 400K Reddit Posts, Uncovers Hidden Ozempic Side Effects Like Menstrual Changes
University of Pennsylvania researchers used LLMs to analyze 400,000 Reddit posts on GLP-1 drugs, uncovering underreported symptoms like menstrual irregularities and chills.