ur-agent 1.80.9 → 1.81.0

package/CHANGELOG.md

@@ -1,5 +1,37 @@
 # Changelog
 
+## 1.81.0
+
+- Added `/mode redteam`, a warning-gated, session-only UR policy mode for
+  complete authorized security research. The mode changes UR's dynamic system
+  and file-analysis guidance without claiming to override the selected
+  provider/model's independent policies.
+- Kept operational controls intact: security scope approval is now bound to the
+  current session, active Bash and PowerShell security tools are blocked outside
+  approved hosts/ports/tools/intensity, and security state moved to the
+  gitignored `.ur/security/` with legacy `.309/security/` migration.
+- Added the first-party `reverse-skills` plugin, an UR-native adaptation of the
+  MIT-licensed `zhaoxuya520/reverse-skill` research methodology. Plugin commands
+  and skills declare `requiredMode: "redteam"` and are disabled outside it;
+  cross-client bootstrap code and separately licensed components are excluded.
+- Documented activation, warnings, provider boundaries, manifest gating, scope
+  behavior, provenance, and the full research capability map, with regression
+  tests for policy switching, classification, scope enforcement, storage, and
+  manifest validation.
+
+## 1.80.10
+
+- Fixed the underlying cause of task-free research still returning
+  `TaskListRequired` in the published CLI: the public bundle compiled the
+  shipped `Explore` and `Plan` definitions out of the active agent registry.
+  Research classification was correct, but there was no protected worker to
+  receive the capability-reduced call.
+- Made `Explore` and `Plan` deterministic core agents for public, local, API,
+  and subscription-CLI use, aligned the development and release feature sets,
+  and added tests against the real built-in registry plus the bundle script.
+  Anthropic-internal experimentation remains isolated to internal builds.
+  Write-capable delegation, task enforcement, and Approve All are unchanged.
+
 ## 1.80.9
 
 - Fixed the remaining provider-independent `TaskListRequired` failure for

package/README.md

@@ -114,6 +114,14 @@ available, and the approval decision in task evidence. Security-research prompts
 need scoped targets and authorization confirmation; local/lab/test targets are
 preferred unless the user confirms authorized external scope.
 
+Authorized researchers can explicitly activate `/mode redteam`. The first use
+shows a mandatory risk warning; `/mode redteam --accept-risk` acknowledges it
+and enables UR's unrestricted security-research prompt policy for the current
+session only. UR's scope, permissions, sandbox, approvals, and audit controls
+remain enforced, and the selected model/provider keeps its own policies. The
+bundled UR-native `reverse-skills` plugin is runtime-gated to this mode. See the
+[Redteam Mode Guide](docs/REDTEAM.md).
+
 Outside-workspace reads are allowed when requested or clearly required, and the
 outside path is recorded in the task board/evidence. Modifying or deleting
 outside-workspace files requires approval before execution.
@@ -560,6 +568,7 @@ ur plugin search --capability skills --json
 ur plugin show github@ur-plugins-official
 ur plugin doctor
 ur plugin install engineering-discipline@ur-plugins-official
+ur plugin install reverse-skills@ur-plugins-official
 ur plugin install hello@ur-plugins-official
 ur plugin update <plugin>
 ur plugin disable <plugin>