OpenClaw v2026.9.8 and Hermes pinned inputs add model support and reproducibility

What shipped

From 28 September to 5 October 2026 open agent projects released incremental updates focused on reliability and model compatibility. OpenClaw delivered multiple extended-stable builds with security and workflow changes. Hermes Agent focused on content-addressed input archives to support consistent agent behavior.

Hermes Agent

Nous Research issued three sequential releases that archive reviewed project-management inputs as immutable sha256 objects. The approach prevents accidental edits during agent development cycles. Scripts manage the archive process to keep assets stable for repeated runs.

• •Hermes Agent inputs-d release Nous Research published content-addressed copies of reviewed inputs using sha256 naming. The release keeps assets fixed so agent scripts cannot alter them later. Archive scripts handle storage to maintain workflow consistency.
• •Hermes Agent inputs-e release Nous Research published content-addressed copies of reviewed inputs using sha256 naming. The release keeps assets fixed so agent scripts cannot alter them later. Archive scripts handle storage to maintain workflow consistency.
• •Hermes Agent inputs-f release Nous Research published content-addressed copies of reviewed inputs using sha256 naming. The release keeps assets fixed so agent scripts cannot alter them later. Archive scripts handle storage to maintain workflow consistency.

OpenClaw

OpenClaw issued over thirty updates across extended-stable lines that backport 113 fixes and add support for models including GPT-6.1 Sol and Claude Opus 5.5. Releases emphasize credential preservation during upgrades and tighter sandbox controls. Multiple builds target gateway reliability and plugin state recovery for production agent deployments.

• •OpenClaw backport 113 fixes The update rolls up corrections for upgrades, authentication, and Doctor recovery. It preserves credentials and agent state during changes while hardening filesystem boundaries. Users gain more reliable long-running agent sessions.
• •OpenClaw GPT-6.1 Sol support Routing now includes the GPT-6.1 Sol model with gateway reliability patches. Plugin settings and locks stay intact during updates. Channel integrations for Gmail, Matrix, and Telegram become more stable.
• •OpenClaw v2026.9.8 release The version improves update safety and Doctor recovery by backing up state before migration. Plugin settings remain intact across restarts. Builders get fewer broken agent configurations after upgrades.
• •OpenClaw v2026.8.35 release This gateway-only build adds GPT-6.1 Sol support plus security and performance patches. It serves as the current extended-stable equivalent to long-term support. Teams running production agents receive a tested baseline.
• •OpenClaw boundary security tightening Controls now cover remote filesystem changes, plugin scanning, and browser authentication. Sandbox rules limit private skill ingress. Agent operators reduce exposure to unauthorized actions.
• •OpenClaw upgrade and Doctor hardening Workflows now retain credentials, transcripts, and plugin inventory across restarts and repairs. Agent state survives migration steps. Daily operators experience fewer lost sessions after maintenance.
• •OpenClaw 2026.8.34 extended-stable The build backports 113 fixes across sessions, channels, and model runtimes. It targets correctness in sandboxing and auth flows. Production teams obtain a consolidated patch set without full version jumps.
• •OpenClaw v2026.8.34 release This gateway-only extended-stable version includes critical security and reliability updates. It matches the end-of-August codebase plus fixes. Deployments gain a stable reference point for agent fleets.
• •OpenClaw macOS launch crash fix A rebuilt notarized DMG replaces the prior 2026.9.6 desktop build. The change resolves ticket-reported startup failures. Desktop users regain reliable local agent access.
• •OpenClaw OpenAI OAuth discovery check Model discovery must succeed before GPT-6 Astra appears via OAuth. The gate prevents misconfigured endpoints. Administrators avoid broken model selections during setup.
• •OpenClaw Anthropic Fable 5.1 continuity Reasoning state persists across dynamic model switches. Multi-step tasks keep thought progression intact. Complex agent chains avoid partial result loss.
• •OpenClaw Discord voice session handling Speaker ownership and playback state survive lifecycle transitions. Cancelled realtime consults record correctly while the session stays active. Voice agents maintain continuity in group calls.
• •OpenClaw metrics and asset hardening Operator scope checks apply to Prometheus scrapes and Discord uploads follow media policies. Nodemailer dependency updates close DoS vectors. Monitoring setups become more secure.
• •OpenClaw GPT-5.6 family preservation Catalog routing, image, vision, and thinking behaviors stay supported for Sol, Terra, and Luna variants. Bedrock images and Ultra reasoning continue across boundaries. Existing GPT-5.6 workflows require no migration.
• •OpenClaw flagship model additions Extended-stable now lists Meta Muse Spark 1.3, Anthropic Fable 5.1, GPT-6 Astra, and GPT Image 2.5. Long-term support deployments gain current models. Teams standardize on one release line.
• •OpenClaw Slack Huddles and Gemini TTS Users join authenticated Slack huddles and run Gemini 3.8 Flash text-to-speech. Dual-voice dialogues become available. Voice-enabled agents expand to Slack workspaces.
• •OpenClaw Claude Sonnet and Opus 5.5 Support arrives for Claude Sonnet 5.5 while default Anthropic flows use Claude Opus 5.5. Bare opus references update automatically. Reasoning agents access newer model variants.
• •OpenClaw video provider integrations Kie AI, Z.AI, and Novita join as video generation options with bundled artwork. Agents can call these providers through standard plugins. Content teams add video output without custom code.
• •OpenClaw decision_evaluate tool The new tool scores Boolean, Choice, and Score prompts using the agent's configured decision model. Fallback logic handles rejected inputs. Decision agents gain built-in evaluation steps.
• •OpenClaw Telnyx provider support Telnyx becomes an official external model provider with standard onboarding. Agents route requests through the new endpoint. Users diversify model sources without extra configuration.
• •OpenClaw desktop audio streaming Managed panels now stream per-application audio on Linux with authorization controls. Isolated playback supports focused agent sessions. Desktop operators monitor audio output cleanly.
• •OpenClaw isolated worktree sessions Chat sessions start inside managed worktrees from web, iOS, or Android. Parallel repository sessions no longer collide. Multi-device agent work stays separated.
• •OpenClaw Apple chat app updates macOS and iOS displays now show message times, models, and agent identities plus file and audio attachments. macOS gains a command palette. Native clients improve daily agent interaction.
• •OpenClaw Sign in with ChatGPT Beta OAuth option joins Codex login and API keys with restricted credentials for unsupported media. Capability notes clarify limits. OpenAI users gain simpler authenticated starts.
• •OpenClaw OpenAI Agents API plugin Agents run on hosted or self-hosted OpenAI environments with streamed replies, steering, and tool history. Persona context and token usage tracking arrive bundled. Production agents connect directly to the API.
• •OpenClaw v2026.9.7 release Updates now back up every state and agent database before migration. Recovery paths improve after partial failures. Operators reduce data loss during version bumps.
• •OpenClaw July LTS gateway hardening Command parsing, browser origin checks, and diagnostics tighten across plugins. Message loss drops on Discord, Matrix, Telegram, and Slack. Gateway stability rises for connected agents.
• •OpenClaw Doctor registry repair Default plugins including Browser, Canvas, and Talk voice stay available after registry fixes. Partial prior builds recover cleanly. Agent environments retain core capabilities post-repair.
• •OpenClaw v2026.9.6 macOS rebuild The release replaces a launch-crashing binary with a notarized build containing 178 direct commits. Over 2,600 merged PRs cover gateway and runtime. Desktop users receive a stable installer.
• •OpenClaw security rollup and Prometheus auth Repository advisories close, metric scraping gains authorization, and IMAP dependency updates patch DoS issues. Monitoring and asset policies strengthen. Secure deployments become simpler to maintain.
• •OpenClaw Meta Muse Spark and Fable support Extended-stable adds Meta Muse Spark 1.3, Anthropic Fable 5.1, GPT-6 Astra, and GPT Image 2.5 with full GPT-5.6 routing. Migration paths stay intact. Flagship models reach stable lines.
• •OpenClaw v2026.8.33 release The gateway-only build adds current flagship models plus security and performance updates. It serves as the extended-stable reference from late August. Teams lock to a tested model set.

What this means for you

For Vibe Builders: You can now pin reviewed inputs in Hermes Agent to keep agent runs reproducible without manual file management. OpenClaw releases add model options like GPT-6.1 Sol and Claude Opus 5.5 plus safer update paths that preserve plugin settings. Test the decision_evaluate tool or video providers on a small workflow this week to see which features reduce manual steps in your agent setups.

For Non-techies: OpenClaw updates make agent updates safer by backing up state and keeping credentials intact, so daily business automations break less often. New model support and Slack or Discord voice features let you expand simple chat agents into voice or team channels. Start with the latest extended-stable build and enable one new provider such as Telnyx to test reliability on routine tasks.

For Developers: OpenClaw backports 113 fixes across sandboxing, Doctor recovery, and model routing while adding Agents API plugin support for streamed replies and tool history. Hermes input pinning gives a content-addressed baseline you can reference in CI. Benchmark the new extended-stable line against your current stack and watch gateway credential handling before rolling into production pipelines.

What to watch next

Track whether OpenClaw extends the Agents API plugin to additional providers beyond OpenAI. Watch for Hermes Agent follow-up releases that link pinned inputs to live agent evaluations. Note any new Doctor recovery patterns that surface in the next gateway build.