OpenClaw v2026.3.8 shipped on March 8th, and for a point release, it packs a surprising amount of infrastructure-level work. Here's what matters.

The Headlines

Local state backups. openclaw backup create and openclaw backup verify are now first-class CLI commands. You can archive your config, workspace, and state with --only-config or --no-include-workspace flags, and the system validates manifests and payloads on verify. Destructive flows now prompt you to back up first. This was long overdue — losing your agent config to a bad update was a rite of passage nobody enjoyed.

Talk Mode silence timeout. New top-level talk.silenceTimeoutMs config lets you control how long Talk waits through silence before auto-sending the transcript. Each platform keeps its default pause window when unset, so existing behavior is preserved. If you've been frustrated by Talk Mode cutting you off mid-thought — or waiting forever — this is your fix.

Brave Search LLM Context. With tools.web.search.brave.mode: "llm-context", web_search now calls Brave's dedicated LLM Context endpoint, returning extracted grounding snippets with source metadata instead of raw search results. Useful if you want your agent to ground answers in web content without scraping.

ACP Provenance. openclaw acp --provenance off|meta|meta+receipt adds optional origin metadata and visible receipt injection for ACP sessions. Agents can now retain and report where their ACP context came from, complete with session trace IDs. Enterprise traceability, basically.

TUI workspace detection. Launch the TUI inside a configured agent workspace and it now infers the active agent automatically, while still respecting explicit agent: session targets.

Notable Fixes

The one that will relieve the most people: Telegram DM dedup. Previously, the same DM could trigger duplicate replies when both agent:main:main and agent:main:telegram:direct:<id> resolved for one agent. Fixed.

Cron/Telegram announce delivery now routes through real outbound adapters, so plain Telegram targets no longer silently report delivered: true when no message actually reached the user. If you had cron jobs that seemed to work but never actually sent anything — that's why.

Matrix DM routing gets safer fallback detection for broken m.direct homeservers. Feishu onboarding no longer re-prompts to download the plugin right after you just installed it. macOS Tailscale gateway discovery stays alive when other gateways are present.

Other fixes span browser CDP normalization (Browserless containers, WSL2 setups), config/runtime snapshot integrity after writes, and a TUI light-theme detection via COLORFGBG with WCAG AA-compliant palettes.

Breaking-ish

Mac assets on this release reuse the beta artifact line from v2026.3.8-beta.1. Not a breaking change, but worth noting if you're checksumming downloads.

The openai-codex agent normalizes GPT-5.4 transport back to openai-codex-responses on chatgpt.com/backend-api when config drifts, and now uses the 1,050,000-token context window instead of inheriting stale legacy limits.

Bottom Line

v2026.3.8 is a maintenance-plus release. The backup commands alone make the update worthwhile — everything else is polish on an increasingly solid foundation. The Telegram DM fix will save you the most confusion. Update and move on.

Full changelog: GitHub Release