CLAUDE CODE V2.1.49 HARDENS LONG-RUNNING AGENTS, ADDS AUDIT HOOKS, AND MOVES MAX USERS TO SONNET 4.6 (1M)
Anthropic shipped Claude Code v2.1.49 with major stability and performance fixes for long-running sessions, new enterprise audit controls, and a Max-plan model ...
Anthropic shipped Claude Code v2.1.49 with major stability and performance fixes for long-running sessions, new enterprise audit controls, and a Max-plan model shift to Sonnet 4.6 with a 1M-token context window.
The v2.1.49 release notes highlight concrete fixes for memory growth in WASM parsing and layout engines, background agent interrupt handling (double Ctrl+C/ESC), faster non-interactive startup (-p), plugin scope auto-detection, and a prompt cache regression fix, plus simple mode gains a direct file edit tool and SDKs now expose capability flags like supportsAdaptiveThinking release. Enterprise teams get a new ConfigChange hook to log or block config edits mid-session, and Max-plan users should switch to Sonnet 4.6 (1M) as Sonnet 4.5 (1M) is being removed.
For context on why these changes matter, Anthropic engineers have emphasized prompt caching as key to cost/latency on long-lived agent workflows note, and leadership is openly reframing developer roles toward reviewing and steering AI-authored code rather than typing it by hand Boris Cherny interview.
Fixes target real operational pain points in agentic dev sessions (memory growth, interrupts, cold start), improving reliability for CI/CD and data pipelines.
New audit hooks and capability metadata make it easier to govern AI-in-the-loop workflows across teams.
-
terminal
Run soak tests on long sessions to verify WASM memory stays bounded and agent interrupts reliably kill background tasks.
-
terminal
Validate SDK integration using new capability fields and confirm permission suggestion flows surface correctly when safety checks trigger.
Legacy codebase integration strategies...
- 01.
Plan a controlled rollout of v2.1.49 and audit ConfigChange events to ensure mid-session config edits are logged or blocked per policy.
- 02.
Migrate model selection to Sonnet 4.6 (1M) in your /model or orchestration layer and re-baseline prompt caches and cost/latency.
Fresh architecture paradigms...
- 01.
Default new agent workflows to Sonnet 4.6 (1M) and design around prompt caching to minimize roundtrips and cost.
- 02.
Use simple mode’s file edit tool for rapid inner-loop changes, then formalize with SDK capability checks and audit hooks.