OpenClaw Recounts 2026.10.7 at 61 Pull Requests as Main Lets Owners Bind Older Automations to Their Conversations
The project's corrected 2026.10.7 release notes count 61 pull requests, not the 70 in the GitHub notes we reported Saturday, and spell out that managed Claude CLI turns stop loading CLAUDE.md and user hooks. On main, owners can bind older automations to the conversations that created them, Knowledge memory ships bundled but off, and Android gains Gateway folders. Plus: check your container bind and read the test-audit skill.
The Official 2026.10.7 Notes Count 61 Pull Requests, Not 70
The official @openclaw announcement of v2026.10.7 went out at 10:29 PM Pacific on Sunday with a scale of "61 PRs · 54 contributors". We reported 70 pull requests on Saturday, so we rechecked the primary records.
Both numbers come from the project. The GitHub release notes for openclaw 2026.10.7, published at 2:53 AM Pacific on Saturday, October 24, still say the audited record covers "49 in-range PRs + 21 retained seed-only PRs = 70 unique PRs". The OpenClaw v2026.10.7 release notes on the docs site now say "61 pull requests, 26 direct commits, and 54 contributors". The docs page arrived through pull request #183785 manually correcting the 2026.10.7 release page, merged at 6:13 PM Pacific on Sunday. Its author says the draft "counted analysis groups as PRs"; the corrected count expands four backport bundles into their original pull requests without counting the wrappers twice. It also removes a Codex-preview claim and two other descriptions that were not new since 10.6, and does not authorize editing the live GitHub release text.
Correction to our October 24 post. We wrote that 2026.10.7 counted 70 unique pull requests and that Codex previews stop blocking replies after a context refresh. Both came from the GitHub notes. The corrected release page counts 61 underlying pull requests, 26 direct commits and 54 contributors, drops the Codex-preview fix, and does not list the Gemini 4 Argon support we also repeated. Our release date stands: the build was published Saturday, October 24, Pacific time; the announcement came Sunday.
The rewritten page also states side effects the GitHub version did not. Managed Claude CLI turns now use OpenClaw's own tool surface, and after upgrading they "stop loading Claude user settings, CLAUDE.md and rules, auto-memory, user hooks and plugins, Chrome, and Claude-only MCP servers". Existing conversations start one fresh native Claude session on their first turn after the upgrade, while MCP servers configured in OpenClaw stay available through the Gateway. It also confirms that package updates now stop before staging a release that needs a newer Node than the Gateway service uses, leaving the current Gateway running.
The Claude CLI line is the one to check before or right after upgrading. If your agents quietly relied on a CLAUDE.md file, a personal Claude hook or a Claude-only MCP server, that behavior is gone after 2026.10.7. Move those instructions and tools into OpenClaw configuration, where the Gateway can see and govern them, and expect one cold session per conversation.
Owners Can Now Bind Older Automations to Their Conversations
On Sunday we noted that Doctor's new automation review had no fix for unbound jobs short of recreating them. That gap closed the same day. Pull request #182561 binding older automations to the conversation that created them, by jalehman and merged at 10:18 AM Pacific on Sunday, adds a bind action to the automations tool. Older automations are linked to their conversation by session key alone, so each run uses the run session's board, memory and workspace instead of the conversation's own.
From the conversation that owns such a job, an owner can now ask the agent to bind it. The job then runs as a linked child of that conversation, and nothing changes until someone binds it. Bind refuses jobs already bound or tied to another conversation, turns without owner authority and archived parents. The OpenClaw automations CLI reference documents the same rules; openclaw automations bind run from a shell is refused because it has no conversation. This is on main, not in a release.
Binding moves an automation from a shared run context into one conversation's memory and workspace, so it is a scope change worth reviewing, not just cleanup. When this ships, start from the Doctor list of unbound jobs, bind each from the chat that genuinely owns the work, and delete jobs whose owning conversation nobody can identify.
Knowledge Memory Moves Into OpenClaw, Off by Default
The biggest memory change on main is packaging. Pull request #180139 bundling the Knowledge memory plugin, merged at 10:30 AM Pacific on Sunday, moves Knowledge, which the author calls "the evidence-backed memory plugin", into the OpenClaw repository. It ships off by default. The OpenClaw Knowledge plugin guide says setting plugins.slots.memory to "knowledge" turns it on and replaces memory-core as the active memory provider, since only one plugin owns that slot at a time. A new knowledge_attach_file tool attaches a workspace file by path instead of copying its bytes through the model's context.
Existing standalone users need care. An installed copy takes precedence over the bundled one, and the guide warns that openclaw plugins uninstall removes plugins.entries.knowledge.config, including its stores, so copy that block somewhere safe before switching. The pull request says existing stores open without a migration step. Separately, pull request #183475 reporting all-zero query embeddings in memory-core, merged at 5:42 PM Pacific on Sunday, replaces a silent failure: when an embedding provider returned an all-zero vector, vector recall was skipped and reduced recall looked like a quiet corpus. Searches now warn and name the provider.
Android Picks Up Gateway Folders and Effort Controls on Main
Android fixes are waiting too, though the GitHub notes say the 2026.10.7 Android APK was skipped because the app's version file still read 2026.8.2. Three pull requests by contributor Daniil Karpenko merged overnight Saturday. Pull request #164857 showing Gateway session groups in the Android sidebar displays Gateway folders, including empty ones, alongside channel groups; folder editing is deferred. Pull request #163516 honoring write-authorized effort controls on Android lets connections with write access adjust Thinking and Fast Mode without administrator access, matching what the Gateway already allowed. Pull request #174152 keeping photo and voice notes in one Android message stops combined photo-and-voice messages from losing attachments after transcription.
Security Practice: Check Your Bind Before You Check Anything Else
Container installs listen on every interface unless you say otherwise
A weekend X post said Kaspersky found nearly 1,000 OpenClaw instances open to strangers. The figure is old: the Kaspersky analysis of OpenClaw's exposed installations, published in February, describes a late-January Shodan scan that found nearly a thousand installations running without authentication. The OpenClaw network exposure guide says gateway.bind defaults to "loopback" on regular hosts, but "auto" is the default in a detected container and resolves to 0.0.0.0. The "lan", "tailnet" and "custom" modes should only be used with Gateway auth and a real firewall.
- If you run OpenClaw in Docker, set
gateway.bindexplicitly, confirm auth is configured, and runopenclaw security audit. The OpenA2A October report on exposed AI agent services counted 205,638 reachable OpenClaw gateways in its October 19 sweep. Reachable is not the same as unauthenticated, but the count shows how many gateways sit on public addresses. - On main, pull request #179899 letting operators keep Gateway environment variables out of ACP agents adds opt-in
agentProcessEnvAllowandagentProcessEnvBlocksettings, because agents the acpx plugin starts inherited the whole Gateway environment, includingOPENCLAW_GATEWAY_TOKEN. The OpenClaw ACP agents setup guide shows a block list for that token and warns that blocking a provider key an agent needs stops it from starting.
Skill Card: test-audit, the Rulebook Behind the Test Purge
Why OpenClaw keeps merging "remove low-value tests" batches
Of the 819 pull requests merged between Sunday and Monday mornings Pacific, 77 were "remove low-value tests" batches. The rules behind them live in the test-audit agent skill in the OpenClaw repository. Before a test is added, the skill asks what observable behavior it protects, what credible regression would make it fail, and why existing tests miss it. It rejects self-comparisons, source greps, duplicate contracts and production hooks that exist only for tests.
API, protocol, config, migration, storage, security and release contracts are kept even when slow, and a failing baseline test may be a real bug. For subsystem-wide campaigns, its test-audit campaign guide has reviewers make one deliberate mutation of the production code for each restored contract and confirm the keeper test goes red before restoring the source byte for byte. In the Telegram campaign, that review found nine real gaps.
The mutation step is the part worth copying into your own skills and agent repositories. Deleting tests is easy for an agent and dangerous at scale. Requiring a deliberate break of the real code, and a red test, before a deletion counts as safe turns cleanup into evidence you can review.
Upgrading agents that carry real work?
SEN-X helps teams audit Claude CLI and automation changes before upgrades, scope memory plugins, and lock down container binds and agent environments.
Contact SEN-X →