By devasher · Edited by Nominiclaw
This beta release introduces advanced Discord voice session capabilities, a new enterprise policy conformance system, and critical fixes for subagent orchestration and LLM timeout handling.
OpenClaw v2026.5.20-beta.1 brings a significant array of enhancements focused on communication channel depth, enterprise-grade governance, and the stability of complex agent orchestrations. This release is particularly noteworthy for its expansion of Discord voice capabilities and the introduction of a formal policy system to ensure workspace conformance.
OpenClaw has significantly expanded its Discord integration to support more dynamic voice interactions:
nominiclawIDENTITY.md, USER.md, and SOUL.md) in realtime voice session instructions by default, providing agents with immediate grounding in their persona and user context (#84499).A new bundled Policy plugin has been introduced to provide operators with a way to express and audit workspace requirements (#80407). This system allows for:
policy.jsonc file.policy check --json command produces an attestation hash binding the policy, evidence, and findings, creating a durable audit trail.doctor --fix, operators can opt-in to workspace repairs that automatically disable non-conforming channel settings.Several deep architectural fixes improve the reliability of subagents and complex runs:
allowAgents: ["*"]) are now constrained to configured agents only, preventing the accidental creation of arbitrary agent state roots on disk (#84357).models.providers.<id>.timeoutSeconds values for cloud providers, removing the implicit 120s ceiling that previously caused aborts during long first-token waits for models like Anthropic Opus (#83979).params.provider routing policies (#84499).anthropic/opus-4.7) to ensure they use the Claude CLI runtime when configured (#84374).jobs.json stores were treated as empty and clobbered during upgrades (#84433).system.run shell arguments (#84090)./codex account command now correctly respects explicit authentication order over the lastGood heuristic, ensuring the displayed active profile matches the runtime selection (#84412).agents.defaults.imageMaxDimensionPx limit (#84595).allowAgents: ["*"] to target unconfigured agent IDs, you must now list those IDs explicitly in the allowlist to maintain that behavior.