Claude Code v2.1.128: The Update Where Claude Speaks First#

2 new env vars, a quieter system prompt for subagents, and Sonnet 4.6 — Published on May 4, 2026

Part of the Claude Code Version Tracker series. | Official Env Vars | Official Changelog

For the first time in Claude Code's history, Claude will prompt you instead of waiting. v2.1.128 ships two new environment variables: one that makes Claude speak up when a session goes idle, and one that trims the cost of running background subagents. The model underneath also moved up a generation. Small diff, meaningful shift in how the agent behaves.

New Environment Variables#

CLAUDE_CODE_PROACTIVE — Claude checks in when you've been away#

You break for lunch. You come back 90 minutes later to a session that's been idle with a stale, token-heavy context. Before this release, that was your problem to sort — manually decide whether to /clear, keep going, or lose your place. Now Claude offers a clean restart before you have to ask: a "Start a new task?" prompt, a suggestion for what to do next, and a note on the tokens you'd save by clearing.

This is the first active output from idle detection. The watching infrastructure — CLAUDE_CODE_IDLE_THRESHOLD_MINUTES, CLAUDE_CODE_IDLE_TOKEN_THRESHOLD — was already there in v2.1.126. CLAUDE_CODE_PROACTIVE is what turns observation into action. Claude doesn't interrupt while you're working; it prompts on the return from idle, when the session has been quiet and you're coming back to it.

For everyday users: you'll see this as a gentle nudge when you return to a long-running session. The prompt gives you a direct choice — start fresh or keep going — instead of leaving you to guess whether the context is still useful.

For agentic harnesses: CLAUDE_CODE_PROACTIVE is an explicit on/off switch in the environment. Pair it with CLAUDE_CODE_IDLE_THRESHOLD_MINUTES to control when idle triggers. In CI or unattended batch pipelines, turn it off — no human is there to act on the prompt. In supervised overnight runs where an operator might check in, leave it on: it's a clean signal that the session is alive, oriented, and ready to hand off.

CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT — lower overhead for focused subagents#

The full Claude Code system prompt is large: tool descriptions, permission rules, skill registrations, behavioral guidelines, your CLAUDE.md. For an interactive session where you might reach for any capability, all of it earns its weight. For a background subagent spawned to do one specific job — run a test suite, apply a patch, summarize a file — most of it is overhead you're paying for on every turn.

CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT strips that prompt down for targeted spawns. It lives in the subprocess initialization path, not the interactive session layer — it's a flag for programmatic agent spawning, not something you'd set for your own day-to-day sessions.

For everyday users: if you've used /do, /maistro, or any multi-step skill that spawns background agents, each of those agents carries the full system prompt on every turn — your CLAUDE.md, all skill registrations, every tool description. That's the weight being cut here. The practical tell: a subagent running with this flag won't know about your custom skills or CLAUDE.md conventions, so it's a tradeoff. Narrow, well-scoped tasks get cheaper; tasks that need your project context should stay on the full prompt.

For agentic harnesses: this is the token-cost knob that's been missing from the fleet infrastructure built across v2.1.114–v2.1.126. The system prompt is your largest fixed per-turn cost across parallel subagents. Use CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT on leaf-node workers with a narrow, well-defined job; leave it off on the coordinator that needs full tool and skill awareness. The result: you can run deeper agent trees at lower cost without compromising the agents that actually need the full context.

Model Bump: Sonnet 4.6 Takes Over for Remote Sessions#

Sonnet 4.6 takes over as the embedded default for Claude Code Remote, tracked by undated alias so remote sessions roll forward automatically as Anthropic ships new checkpoints. For anyone building on Claude Code Remote, the model underneath has quietly moved up a generation without requiring a config change on your end.

The Pattern#

Both variables extend the same infrastructure story — and it's worth naming explicitly. The idle detection stack existed before this release; proactive mode is the first user-visible behavior built on top of it. The fleet infrastructure for spawning many agents has been accumulating since v2.1.114; the simple system prompt is the efficiency layer that makes running that fleet at scale economically sensible.

Claude Code is becoming a platform you configure for a specific operating mode, not just a tool you run. These two knobs — when Claude speaks, and how much it costs per subagent turn — are exactly the kind of controls you need when sessions get longer and agent trees get deeper.

Sources#

Extracted from claude v2.1.128 binary at ~/.local/bin/claude. This analysis is conducted for independent security research and interoperability purposes under fair use principles. All trademarks belong to their respective owners. The information presented here documents publicly observable behavior of installed software and is not intended to circumvent any technological protection measures, infringe on intellectual property rights, or encourage unauthorized use. Use these findings at your own discretion.

Related: Context Window Management Guide | Claude Code Productivity Tips | The Agentic Engineering Playbook

Let's Connect

BlueskyLinkedInThreads
← Back to all posts