Originally published: 2026-10-02 14:08 UTC / 16:08 Berlin / 10:08 EDT Last verified: 2026-10-02 14:08 UTC (GitHub releases.atom fetched 2026-10-02T14:08:00Z; v2026.8.35 entry updated 2026-10-02T14:06:34Z) No corrections at this time.
OpenClaw shipped v2026.8.35 on October 2, 2026, the same day as v2026.8.34, as a follow-up release on the extended-stable channel. The release ships one new model integration, one durable-worker fix that is production-paging material, and 49 PRs of targeted improvements across the update path, agent delivery, channels, runtime reliability, secrets, and WebChat. v2026.8.35 explicitly preserves the v2026.8.34 publication contract — operators on extended-stable do not need to re-validate what v2026.8.34 already delivered.
This is a documentation-comparison report and a direct follow-up to today's pillar (openclaw-v2026-8-34-extended-stable-august-rollup-oct-2026). It is the same-day news-impact for the deltas: GPT-6.1 Sol support, the durable-worker retry fix, plugin-settings preservation across updates, and the channel/IMAP/IMAP backlog changes that are not in v2026.8.34. No live gateway was upgraded during this run.
Every quote below is verbatim from the v2026.8.35 release page. The release is documented as "the audited record covers the complete v2026.8.34..741d94f history: 49 in-range PRs + 0 retained seed-only PRs = 49 unique PRs."
1. GPT-6.1 Sol support. From the release notes (verbatim): "GPT-6.1 Sol support: add the current Sol model across OpenAI routing, discovery, reasoning, harness, and Reef guard-model boundaries. (#161400, #162955)." This is the headline new capability. The integration covers the five surfaces an OpenAI model lands on inside OpenClaw: routing (where the request goes), discovery (how the model is enumerated to the user), reasoning (the harness-level model handling), the harness itself (the agent loop), and the Reef guard-model boundary (the safety/quality classifier that screens model output). If you run OpenClaw with OpenAI as your provider, GPT-6.1 Sol becomes selectable on v2026.8.35 without additional configuration. v2026.8.34's release notes mentioned "features like new model support" in aggregate but did not name the model; v2026.8.35 names it.
2. Durable worker retries stop freezing hosts. From the runtime-reliability fixes (verbatim): "stop durable worker retries from freezing hosts (#160812)." Pre-fix: a durable worker that retried a long-running operation could hold the host's resources open until the retry budget was exhausted. On a small host, a single stuck retry could pin a CPU, a database connection, or a socket and degrade every other agent on the same machine. Post-fix: the retry path yields the host. This is the production-paging fix of the release. If you have ever had a host go non-responsive during a high-volume cron window or a long-running batch agent, this is the documented path that closes the incident class.
3. Plugin settings survive updates and migrations. From the update-safety fixes (verbatim): "preserve incompatible-plugin settings through recovery (#160344)" and "keep retained plugin records during migration (#161485)." Pre-fix: a plugin whose settings schema did not match the new gateway version could have its settings silently dropped during recovery or migration. Post-fix: the settings are preserved through the upgrade, and the inventory record is retained. This is the update-hygiene fix for teams that ship a managed OpenClaw fleet with a curated plugin set.
4. Failed requests can no longer replace earlier questions with a steer. From the agent-delivery fixes (verbatim): "keep failed requests from replacing earlier questions with a steer (#162439)." Pre-fix: a turn that failed could overwrite the prior user question with a steering message in the next turn, hiding the original ask. Post-fix: a failed turn is recorded as a failure; the prior question stays intact. This is a correctness fix with real implications for multi-turn agents and for compliance audit trails that need to reconstruct what the user actually asked.
5. Cron output recovery is uncapped and complete. From the agent-delivery fixes (verbatim): "recover uncapped cron replies (#160520)" and "preserve complete CLI subagent answers (#162843)." Pre-fix: a cron reply longer than the buffer cap was being truncated, and a CLI subagent's full answer could be cut off in the parent. Post-fix: cron output is recovered at full size, and the CLI subagent answer is delivered whole. For teams running scheduled agents that report on long output (nightly research, log analysis, batch summaries), this is the fix for the silent-truncation class.
6. Session, MCP, and tools hardening. Multiple smaller fixes from the v2026.8.34..741d94f window: "reject malformed session targets (#161533)" — defends against session ID forgery; "preserve spawned working directories (#162308)" — agents launched into a working directory no longer forget it; "project deeply nested MCP results safely (#160825)" — fixes a path-traversal-shape bug in deeply nested MCP result directories; "keep remote MCP bundles working in native sessions (#162376)" — restores remote MCP bundle support in native sessions that broke in a prior release.
7. Channels: Gmail, IMAP, Matrix, Telegram. From the channels fixes (verbatim): "recover Gmail transient binds (#161503)" — fixes a Gmail watcher that bound transiently and lost state on reconnect; "bound IMAP backlog processing (#160413)" — caps how much of an IMAP backlog is processed in one go so a long-stale inbox does not stall the watcher; "restore Matrix direct mappings (#161727)" — fixes Matrix direct-message routing that was lost in a prior version; "restore Telegram progress when hooks suppress previews (#161546)" — Telegram progress messages now render even when hook-based preview suppression is in effect.
8. Secrets, sandboxing, runtime, and WebChat. The security-and-ownership-hardening bucket: "retain secret entry kinds during value rotation (#160926)" — a secret-store kind (API key vs OAuth vs service-account) was being lost across rotations, breaking re-issuance; "restore admitted secret-egress execution (#160760)" — recovers a path where the secret-egress was being denied on a previously admitted flow; "repair read-only copied skills without escaping their confined roots (#162304, #162295)" — closes a sandbox escape path in the read-only skill copy path. The runtime-reliability bucket adds "prevent redaction stalls (#161089)", "cancel stale thinking-catalog waits (#161319)", "pass the responding agent to TTS model selection (#161454)", and "install the required Visual C++ runtime after a managed llama.cpp launch failure (#163093)" — the last item is the one most likely to break a Windows-onboarding path. The performance bucket adds "reduce Codex fleet heap use (#162912)" and "bound model-catalog waits (#161132)". The WebChat bucket adds "keep saved replies visible when stale history responses race with live messages (#162763)".
9. Update and pnpm safety. "avoid pnpm terminal failures (#161920)" — fixes a path where a pnpm update could leave the gateway in a broken state on Windows. "carry the existing Gateway lock through container/direct startup (#160193)" — prevents two gateways from racing on startup in containerized deployments.
For OpenAI providers on OpenClaw extended-stable: GPT-6.1 Sol is now a one-upgrade change. v2026.8.34's pillar mentioned "new model support" in aggregate; v2026.8.35 names the model. If you run on the extended-stable track, GPT-6.1 Sol becomes available without leaving your stability channel. The integration is across the full surface — routing, discovery, reasoning, harness, and the Reef guard boundary — so the model is treated as a first-class citizen on the gateway, not a routing add-on.
For operators who have ever had a host freeze during a cron window or a long-running agent. The durable-worker-retry fix is the change that ends the "host is unresponsive, what changed" ticket. The release notes do not describe the exact mechanism (whether the fix is a backoff with yield, a context-switch point, or a worker pool change), but the user-facing behavior is documented: retries no longer freeze hosts. This is the production-paging fix of the release.
For platform teams running curated plugin fleets. Plugin-settings preservation across updates and plugin-record retention during migration are the two fixes that eliminate the most common "I lost my plugin config after upgrade" support path. If you have ever had a user complain that their plugin settings vanished on an OpenClaw upgrade, this release addresses the underlying update-safety issue, not the symptom.
For compliance and audit. The "keep failed requests from replacing earlier questions with a steer" fix matters for any compliance audit trail that needs to reconstruct the user-facing question history. Pre-fix, a turn that failed could overwrite the prior question in the next turn's prompt. Post-fix, the prior question is preserved and the failure is recorded as a failure. For regulated workloads (HIPAA, SOC 2, FINRA), this is the kind of audit-trail correction that auditors care about.
For multi-channel deployers (Gmail, IMAP, Matrix, Telegram). Four separate channel fixes in one release is unusual. If any of your monitored channels has been flaky — Gmail reconnects losing state, IMAP backlogs stalling the watcher, Matrix direct messages misrouted, Telegram progress messages hidden by hook-based preview suppression — v2026.8.35 addresses all four.
For Windows operators. The "install the required Visual C++ runtime after a managed llama.cpp launch failure" fix is the change that prevents a managed llama.cpp launch from failing on a fresh Windows host that does not have the Visual C++ runtime installed. The fix installs the runtime on the failure path, so a managed llama.cpp launch no longer needs a pre-flight check.
This is a documentation-comparison report. Verification steps taken:
openclaw-v2026-8-34-extended-stable-august-rollup-oct-2026): the v2026.8.34 release notes mention "new model support" in aggregate but do not name the model. v2026.8.35 names GPT-6.1 Sol. The 113 audit-selected fix units in v2026.8.34 are distinct from the 49 in-range PRs in v2026.8.35; the v2026.8.35 release is built on the v2026.8.34..741d94f window.No live gateway upgrade was performed during this run. The "new model support" claim in the v2026.8.34 release notes is now substantiated by the named GPT-6.1 Sol in v2026.8.35.
Upgrade cost: Zero for the open-source gateway. For operators using the extended-stable channel, the upgrade is a version bump. No pricing change documented.
Upgrade risk: v2026.8.35 explicitly preserves the v2026.8.34 publication contract ("align OpenClaw and every publishable official plugin to 2026.8.35 while preserving the 2026.8.34 publication contract"). The 49-PR scope is materially smaller than v2026.8.34's 113-fix rollup and is targeted at specific failure modes rather than a broad rescan. Operators on extended-stable who have not yet upgraded to v2026.8.34 should land v2026.8.34 first, then v2026.8.35; operators already on v2026.8.34 can land v2026.8.35 directly.
Risk — durable-worker change. The fix that "stops durable worker retries from freezing hosts" changes the retry behavior for durable workers. If you have a custom durable worker that depends on the pre-fix behavior (e.g., holding a resource through the retry window), review the change in PR #160812 before upgrading. For the standard agent fleet, the change is a strict improvement.
Risk — secret-store kinds during rotation. The "retain secret entry kinds during value rotation" fix is a security-and-ownership correction. Operators who have manually worked around a secret-store-kind loss during rotation can clean up the workaround after upgrading.
Limitations of this report:
v2026.8.35 is a same-day, follow-up, targeted release on the extended-stable channel. It is not a second pillar — it is the natural next chapter of today's v2026.8.34 pillar, and the four changes worth acting on are: GPT-6.1 Sol support (new model on extended-stable without leaving the channel), the durable-worker fix (no more host freezes during retries), the plugin-settings preservation across updates (no more "I lost my plugin config" tickets), and the "keep failed requests from replacing earlier questions with a steer" correctness fix (audit-trail integrity for regulated workloads).
For operators on extended-stable who already moved to v2026.8.34, v2026.8.35 is a direct upgrade. For operators still on v2026.8.33 or earlier, the upgrade path is v2026.8.34 → v2026.8.35, with v2026.8.34's 113-fix rollup landing first. For operators on the bleeding-edge stable channel (v2026.9.7), the GPT-6.1 Sol support and the durable-worker fix are not automatically back-ported; monitor the bleeding-edge channel for an equivalent release.
Today, for operators on OpenClaw extended-stable:
1. Upgrade to v2026.8.35: npm install -g openclaw@2026.8.35 (or download the appropriate installer from the GitHub releases page). The release preserves the v2026.8.34 publication contract, so the upgrade is additive. 2. After upgrade, verify that GPT-6.1 Sol is now listed in your model picker. If you use OpenAI as your provider, the model should be selectable on the routing surface immediately. 3. Verify the durable-worker fix on a representative cron or batch agent. A retry that previously pinned a CPU or stalled the watcher should now yield the host. 4. If you have a curated plugin fleet, verify the plugin-settings preservation across the upgrade — settings that were in place before the upgrade should still be in place after, and any plugin record that was retained during a prior migration should still be retained.
Today, for compliance-sensitive deployers:
5. Verify the "keep failed requests from replacing earlier questions with a steer" fix on a multi-turn agent. A turn that fails should be recorded as a failure; the prior user question should remain in the prompt history.
Today, for multi-channel deployers:
6. Verify the Gmail, IMAP, Matrix, and Telegram fixes on the channels you monitor. Reconnect a Gmail watcher that was losing transient binds; queue a long-stale IMAP mailbox and verify the backlog is processed incrementally; send a Matrix direct message and verify routing; trigger a Telegram progress message and verify it renders under hook-based preview suppression.
Skip if not in scope:
openclaw-v2026-8-34-extended-stable-august-rollup-oct-2026 — published 2026-10-02 12:22 UTC