OpenClaw v2026.10.7 adds Gemini 4 Argon and replay-safe session recovery

3 weeks ago

OpenClaw 2026.10.7 adds Gemini 4 Argon, keeps unfinished tool work when a session is interrupted, and makes updates, backups and multi-file patches recover safely. Release: https://github.com/openclaw/openclaw/releases/tag/v2026.10.7 Changelog: https://github.com/openclaw/openclaw/blob/v2026.10.7/CHANGELOG/2026.10.7.md Compare: https://github.com/openclaw/openclaw/compare/v2026.10.6...v2026.10.7

Ask about this presentation

Chapters

  1. 0:00Gemini 4 Argon
  2. 0:15Argon tool calls
  3. 0:31Argon limits
  4. 0:43GPT-5.6, OpenRouter, plugins
  5. 0:5802: Sessions that survive interruptions
  6. 1:01Unfinished tool work
  7. 1:15After a stop
  8. 1:31Claude CLI cache
  9. 1:44Failed patches
  10. 2:0003: Upgrades and backups
  11. 2:01Update recovery
  12. 2:17Backups
  13. 2:33Gateway CPU
  14. 2:4504: Around the Gateway
  15. 2:47Linux, macOS Talk, passkeys
  16. 3:02Computer control, GitHub, cron
  17. 3:16Android and the changelog
Show transcript
  1. OpenClaw 2026.10.7 adds Google's Gemini 4 Argon to the bundled model catalog. Pick it by name or with the Gemini 4 alias, and newer Gemini 4 models resolve from the same template.

    Slide text
    01 02 03 04 Gemini 4 Argon
    openclaw/openclaw · CHANGELOG/2026.10.7.md
    OpenClaw

    OpenClaw 2026.10.7 adds Gemini 4 Argon

    Bundled model catalog
    gemini-4-argon
    Newer Gemini 4 models resolve too
  2. Pointed at Argon by hand, older builds took the Gemini 2.5 path, and Google's API rejected replayed tool calls with an HTTP 400. Argon now gets its own thinking levels: low, medium and high.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Gemini 4 Argon
    openclaw/openclaw · compare v2026.10.6...v2026.10.7 · #171852

    Argon tool calls stop failing with an HTTP 400

    Before, OpenClaw sent Argon down its Gemini 2.5 path.

    Thinking level Low Medium High
  3. Gemini 4 relies on implicit caching only, so explicit cached contents are off for it. A config asking for minimal thinking, or none, gets low, because Argon rejects minimal.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Gemini 4 Argon
    openclaw/openclaw · compare v2026.10.6...v2026.10.7 · #171852

    Argon caches implicitly, and Low is its lowest thinking level

    Caching implicit only, no explicit cached contents
    Minimal or off sent to Argon as Low
  4. Tool turns on GPT-5.6 no longer fail with an HTTP 400 error over reasoning effort. OpenRouter onboarding keeps a valid setup instead of rejecting it, and external TypeScript plugins start faster.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Gemini 4 Argon
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    GPT-5.6 tool turns go through

    GPT-5.6 tool turns no reasoning effort 400
    OpenRouter onboarding keeps a valid setup
    External TypeScript plugins start faster
  5. Sessions that survive an interruption.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Sessions
    openclaw/openclaw · CHANGELOG/2026.10.7.md
    02
    Sessions
  6. When a session is interrupted, OpenClaw keeps the outcomes of tool calls still in flight, and notes where each recovery came from. The same recovery covers a model returning an empty reply at its output limit.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Sessions
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    Interrupted sessions keep their unfinished tool results

    Each recovery notes where it came from.

    Tool work also recovers after an empty reply at the output limit.

  7. After you stop a conversation, OpenClaw clears leftover recovery state before your next message. Rejected Claude CLI drafts stay out of final replies, and Codex previews stop holding up replies after a context refresh.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Sessions
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    Your next message after a stop starts clean

    Stopped conversations clear leftover recovery state first
    Rejected Claude CLI drafts stay out of final replies
    Codex previews stop blocking replies after a context refresh
  8. On Claude CLI, a finished child task used to swap out the running session's cached prefix. Now the prefix stays put, and the session keeps its context when caller-only tools come and go.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Sessions
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    Claude CLI keeps its cache prefix when a child task finishes

    Context stays when caller-only tools come and go.

  9. A multi-file patch with one bad hunk could rewrite or delete the earlier files before it failed. Now apply_patch checks the whole patch before changing anything. A disk error partway through the writes can still leave earlier edits.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Sessions
    openclaw/openclaw · compare v2026.10.6...v2026.10.7 · #160409

    A multi-file patch that fails leaves your files alone

    OpenClaw checks the whole patch before it changes any file.

    A failure during the writes can still leave earlier edits.

  10. Upgrades and backups.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Upgrades and backups
    openclaw/openclaw · CHANGELOG/2026.10.7.md
    03
    Upgrades and backups
  11. An interrupted package update now resumes safely. A verified activation stays successful even if later maintenance fails. On Windows, 8.3 short names and deep state paths no longer abort onboarding.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Upgrades and backups
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    An update that stops halfway resumes safely

    Interrupted package updates resume where they stopped
    Verified activations stay successful through later errors
    Windows 8.3 and deep paths no longer abort onboarding
  12. A backup of an older exec approval policy used to restore as deny-all, wiping your allowlist. Now the rules come back. Interrupted R2 restores resume, and Memory Wiki re-imports keep your handwritten notes.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Upgrades and backups
    openclaw/openclaw · compare v2026.10.6...v2026.10.7 · #164297

    Restoring a backup brings back your exec approval rules

    Older supported policies used to restore as deny-all.

    R2 verification and restores resume after an interruption
    Memory Wiki re-imports keep the notes you wrote
  13. If the Gateway's temporary log folder was deleted or went stale, the Gateway could get stuck at full CPU. It happened when the date rolled over, and this release fixes it.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Upgrades and backups
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    A stale log folder can’t pin the Gateway at full CPU

    The hang hit at date rollover.

  14. Around the Gateway.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Around the Gateway
    openclaw/openclaw · CHANGELOG/2026.10.7.md
    04
    Around the Gateway
  15. Saved Gateway dashboards on Linux now keep their pairing, instead of asking for approval after every restart. Talk on macOS keeps replies from slow agent runs, and passkey prompts stop blocking headless browser clicks.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Around the Gateway
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    Linux Gateway dashboards stay paired after a restart

    Saved Linux dashboards keep pairing across restarts
    macOS Talk keeps replies from slow runs
    Headless browser passkey prompts stop blocking clicks
  16. Native computer control now recovers when its session expires. A GitHub operator attached partway through a task can still publish the pull request, and cron announcements leave your CLI chats intact.

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Around the Gateway
    openclaw/openclaw · CHANGELOG/2026.10.7.md

    Computer control comes back after a session expires

    Native computer control recovers after session expiry
    Late-attached GitHub operators can publish pull requests
    Cron announcements keep your CLI chats
  17. This release ships without a new Android app build, so Android stays on its August version. The rest of the fixes are in the release changelog. So which model does your OpenClaw run on today?

    Slide text
    OpenClaw 2026.10.7
    01 02 03 04 Around the Gateway
    openclaw/openclaw · GitHub release v2026.10.7 · Release verification

    The Android app stays on its August version

    Everything else: CHANGELOG/2026.10.7.md

    Which model does your OpenClaw run on?

What happened in AI today, in your inbox

One short email each evening. Read it and you’re caught up.

Built with ❤️ in New York