What’s New

Release notes for the documentation and platform-facing behavior described on this site.

v0.5.1 — February 19, 2026

Stability and compatibility

• Improved SQLite compatibility defaults in generated configs for runtimes where node:sqlite is not available.
• Documented fail-open-vs-fail-closed control with store.required across approval/rate/repetition stores.
• Preserved backward compatibility for existing v0.4.x configs and module ordering.

Docs updates

• Added module docs navigation and index entries for new v0.5.x modules.
• Updated OpenClaw quick start notes for mixed v0.4.x + v0.5.x fleets.

v0.5.0 — February 19, 2026

New optional hardening modules

• Added self_defense (immutable control-plane protection + tamper detection).
• Added tripwire_guard (deterministic honeytoken tripwires for file/env probing).
• Added repetition_guard (identical-call loop detection).
• Added verdict_provider (external deterministic provider integration).

Network and sandbox controls

• Added egress_guard.bindingMode to support tool-specific network binding (intersect with tool_policy egress rules).
• Added exec_sandbox.childPolicy.network control for child-process network restrictions.

Approval and bridge behavior

• Added HTTP pending/polling flow support for bridge architectures.
• Clarified single-bot Telegram topology constraints for polling consumers.

v0.4.0 — February 18, 2026

Module deep dives (new)

• Added dedicated pages for all 11 modules:
  • kill_switch, skill_scanner, tool_policy
  • fs_guard, command_guard, exec_sandbox
  • egress_guard, output_dlp, rate_budget
  • approval_gate, audit
• Converted /docs/modules/ into a module index with direct deep-dive links
• Added practical defaults and config snippets for each module page

OpenClaw compatibility notes (new)

• Added validated OpenClaw + RADIUS compatibility snapshot in docs (real hook runs)
• Added production-safe baseline radius.yaml for OpenClaw with egress_guard and moduleConfig.audit
• Documented runtime caveat for agentradius@0.4.0: prefer moduleConfig.audit in hook flows

v0.3.0 — February 18, 2026

New docs sections

• Added Modules with per-module purpose, config, and posture notes
• Added Features as a capability map
• Added this What’s New page for versioned updates

Website updates

• Hero and interaction polish aligned to the current “Calm Engineering” design language
• Removed noisy motion patterns (ticker/marquee style blocks)
• Improved tether interaction with cleaner label placement under the control
• New SVG brand assets: circular favicon and circle mark near wordmark

UX and accessibility

• Added keyboard-visible focus states for interactive elements
• Improved readability and spacing consistency across docs and landing blocks

v0.2.0 — February 2026

Security architecture baseline

• 11-module deterministic pipeline documented and reflected in examples
• Profile model stabilized: local, standard, unbounded
• OpenClaw quick start and generic quick start published

Operator tooling

• doctor workflow for config/runtime checks
• pentest workflow for attack-scenario smoke tests
• audit and approval sections expanded in configuration reference

v0.1.0 — January 2026

Initial public docs release:

• Overview, quick start, and configuration foundation
• Deterministic-enforcement positioning (no LLM in policy decisions)

Upgrade checklist (v0.4.x -> v0.5.x)

1. Keep your current modules array as-is first; upgrade runtime without enabling new modules yet.
2. Re-run npx agentradius doctor and npx agentradius pentest.
3. If you use SQLite-backed state in hook subprocess mode, set explicit store.required per environment policy.
4. Enable new modules one-by-one (self_defense, tripwire_guard, repetition_guard) and validate audit output.
5. If using tool-specific egress policy, switch egress_guard.bindingMode to intersect and define tool-level egress bindings in tool_policy.

Upgrade checklist (v0.3.x -> v0.4.0)

1. Read Modules and verify your enabled module order.
2. Re-run npx agentradius doctor on your existing radius.yaml.
3. Re-run npx agentradius pentest and compare results with previous baseline.
4. If you use custom policies, confirm fail-closed behavior still matches your intent.