Gemini CLI v0.42 / v0.43 preview: what shipped in the dual release

"Gemini CLI dropped Stable and Preview the same day. What's the difference? Which one is safe for production?" On 2026-05-12, Google's Gemini CLI released v0.42.0 (Stable) and v0.43.0-preview.0 (Preview) simultaneously. This article unpacks both channels and their impact on AI sales automation, using official GitHub Releases as primary source.

Stable is a production-bound cumulative-improvement release (Voice Mode upgrades, Model Routing improvements, Sandbox Security hardening). Preview is a structural update with Subagent Protocol and Surgical Editsas flagship additions. For products like Sales Claw that use Gemini CLI as a backend, the two-track release gives operators a clear answer to "when do I ship what to production."

Primary sources: gemini-cli GitHub Releases (v0.42.0 / v0.43.0-preview.0), Gemini CLI site, Gemini API Release Notes. Also read the Claude Code 2.1.140 article for comparison.

1. Gemini CLI v0.42.0 + v0.43.0-preview.0 — same-day release overview

Gemini CLI ships on Stable / Preview / Nightly channels. Stable = production. Preview = next-gen features for validation. Nightly = post-merge builds for contributors.

With Stable and Preview shipping the same day, operators can now "receive the features they were just testing in Preview as Stable." That said, Subagent Protocol and Surgical Edits are Preview-only — not yet in Stable.

2. v0.42.0 Stable — primary changes

Voice Mode improvements

v0.42.0 makes Voice Mode UI and backend integration noticeably better. From the release notes:

• Updated mic icon and placeholder text
• Added waveform animation during voice input (visual feedback)
• Integrated privacy / compliance warning UI for the Gemini Live backend
• Improved voice-transcript insertion at cursor position

For sales automation, Voice Mode helps with secondary scenarios (final copy review, on-the-go target-list review). But contact-form copy and other accuracy-critical text should stay text-input (transcription error risk).

Model Routing and API timeouts

Model Routing now distinguishes between auto and explicit fallback chains. Default API timeout is shortened to 60 seconds, with retries enabled for undici timeouts.

• Auto-retry on ERR_STREAM_PREMATURE_CLOSE
• Message queuing during compression
• InvalidStream events handled without throwing (robustness)

For Sales Claw running Gemini CLI through a managed PTYfor long batches, these fixes compound. The "1–2 in 100 submissions experiencing mid-stream cuts" observed before is expected to drop.

Sandbox Security hardening

• Prevent auto-update from switching to less-stable channels
• Stricter sandbox-state resolution at auto-update checks
• Tightened Inquiry constraints to prevent non-approved changes
• Blocked shell-invoked exit_plan_mode

Blocking shell-invoked exit_plan_modeaddresses a real risk in enterprise: shell-driven command injection rewriting the agent's mode. Quiet but important.

3. v0.43.0-preview.0 flagship — Subagent Protocol

LocalSubagentProtocol — same-process parallelism

LocalSubagentProtocol lets the Gemini CLI Main Agent spawn multiple Workers within the same process. No IPC overhead; context shared via shared memory.

Sales Claw Phase A spawns 3–5 lightweight haiku-class subagents in parallel per target. See Workflow overview. With LocalSubagentProtocol standardizing this interface for the Gemini backend, custom parallel-management code shrinks.

RemoteSubagentProtocol — horizontal scale

RemoteSubagentProtocol spawns workers on different machines / containers, communicating over ACP (Agent Coding Protocol). Scales beyond a single machine's CPU / memory.

• Spawn many workers in Kubernetes pods; MAIN AGENT distributes
• Pick workers in geographically close regions for latency
• Reassign failed workers to other nodes automatically

Suitable for "analyze 10,000 companies in 4 hours." Sales Claw currently only supports Local. Adopting the Remote protocol later removes single-machine resource limits (a future consideration).

ACP infinite-thought loop prevention

v0.43.0-preview.0 fixes the ACP-mode infinite-thinking loop. Specifically, disabling nextSpeakerCheckbreaks the loop where subagents keep deciding "is it my turn next?"

Also: streaming-ID collisions on parallel tool calls and a context-manager chat-corruption bug are fixed. Robustness improvements that matter when processing 100+ companies at once.

If you're using Gemini CLI as a Sales Claw backend, start with current Stable.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub

4. Surgical Edits and Shell Command Safety

A quiet but important Preview feature: Surgical Edits. A system-prompt / eval-set update that steers the model toward "edit just the right lines, don't rewrite the whole file."

For AI sales automation, you rarely use Gemini CLI for direct coding. But if Gemini modifies Sales Claw itself or workflow scripts, Surgical Edits matters. "Tune the prompt for one target list" stays contained to the relevant one prompt-function block instead of a full-file rewrite.

5. AI sales automation usage — Sales Claw preferences.aiProvider

Sales Claw is OSS that lowers misfire and policy-violation risk through policy control, pre-send checks, do-not-contact detection, CAPTCHA-on-detect stop, rate limits, audit logging, and automatic stop conditions. The AI CLI backend is selected in data/settings.json's preferences.aiProvider — one of claude / codex / gemini.

Provider switch procedure

# 1. Install Gemini CLI v0.42.0
npm install -g @google/gemini-cli@latest

# 2. Confirm version
gemini --version
# → 0.42.0 (or newer Stable)

# 3. Update Sales Claw preferences
#    data/settings.json::preferences.aiProvider = "gemini"

# 4. Restart Sales Claw dashboard
#    Click [Start Gemini] to relaunch the managed PTY

Setup details: AI CLI setup guide. Gemini auth is OAuth-based; headless Linux hangs are fixed in v0.42.0 (safe in remote-dev environments).

Impact on Phase A parallel analysis

Sales Claw Phase A runs parallel site analysis per target with a lightweight model. The Gemini backend currently parallelizes via the standard Agent tool. Once LocalSubagentProtocol lands in Stable, robustness improves.

6. Comparison with Claude Code 2.1.140 / Codex CLI 0.130

From Sales Claw's perspective:

• Low-cost large batches: Gemini Flash-Lite
• Autonomous loops + dashboard: Claude Code 2.1.140 + Agent View (see 2.1.140 article)
• Headless / remote control: Codex CLI 0.130 + codex remote-control
• Voice + text hybrid: Gemini CLI v0.42.0

7. Risks of autonomous operation and safe-by-design

Even with newer Gemini CLI versions, running AI sales automation unattended for long periods still requires multiple guards. Sales Claw defaults meet these, but running Preview in production departs from policy.

The risk of running Preview in production

Legal / compliance

• Japan's Specified Electronic Mail Act: auto-append 4 sender-info fields via preferences.complianceFooter: true
• Terms compliance: pages marked "no sales," "recruiting only," "IR only" auto-skip
• No CAPTCHA bypass: stop into awaiting_approval, log on detect
• Rate limit: suppress consecutive sends to the same domain
• Opt-out: auto-insert "reply to opt out"
• Voice Mode privacy: Gemini Live processes voice on Google's side. Don't enter confidential info verbally (v0.42.0 added UI warnings)

Automatic stop conditions to prevent runaway

8. Bottom line — run Stable and Preview as a two-track

Gemini CLI's same-day release of v0.42.0 + v0.43.0-preview.0 is a milestone declaration: Google is shipping "production and next-gen on separate channels."The practical setup: keep production on Stable, validate Preview's Subagent Protocol on a separate environment.

Next action: npm install -g @google/gemini-cli@latest to pick up v0.42.0, set Sales Claw preferences.aiProvider to gemini, and validate Phase A on a 5-company sample. Details in AI CLI setup and Workflow overview. Sales Claw itself is free at the download page.

Read this. Then try 5 companies.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub