Capability change
- Before: Failed SDK hook callbacks could let actions run, and auto mode could fall back to its built-in classifier when server checks were unavailable.
- Enabling change: Version 2.1.317 adds onFailure: "block" to Agent SDK hook callbacks; 2.1.319 changes auto mode to ask or deny calls the server cannot check when telemetry or feature flags are off.
- After: Builders can make hook failures block execution, and affected auto-mode sessions require approval or deny tool calls that cannot receive server checks.
Human consequence: People can require failed or unavailable checks to stop unattended execution instead of allowing work to continue through a fallback.
Bottleneck: Hook blocking must be configured, unavailable checks can require human approval, and CLAUDE_CODE_AUTO_MODE_SERVER=0 retains the built-in classifier.
Buildability: B4. Verification: source-inspected. Related shift: verification-ownership.
These releases make failure handling an execution policy: SDK callback failures can block actions, and affected auto-mode configurations no longer silently substitute a local classifier for unavailable server checks.
Builders can use this for workflows that must stop when a required check fails. Watch configuration and approval overhead: hook blocking is explicit, and the documented environment setting can preserve classifier fallback.
Published automatically by the New Runtime weekly knowledge desk from the week's evidence.
Supporting evidence
- claude-code 2.1.317 Claude Code 2.1.317 (external, 2026-11-03)
- claude-code 2.1.319 Claude Code 2.1.319 (external, 2026-11-06)
Provenance
Auto-created by the New Runtime weekly knowledge desk on 2026-11-17 from evidence reviewed in the week of 2026-11-02. It is published through the live overlay and joins the hand-edited catalog at the next release.