Arc 1.8.3

• Skill exposure and task outcome are now recorded separately and folded into ranking behind trusted verdicts and relevance. A project knowledge report shows which skills are load-bearing, inert, failure-correlated, or still under-reviewed.
• An operator can send a note to a running agent at the next safe model boundary. Peers can also send notes, but they are authenticated, clearly attributed, delivered after operator notes, and presented only as claims to verify.
• Administrative tools can be preauthorized one by one for roles that genuinely need them. Unknown names and ordinary non-administrative tools are rejected rather than silently broadening access.
• Model certification now exercises the same nested tool envelope used by room work. Arc recovers object-shaped JSON strings, double-encoded payloads, and misplaced parameters without inventing or renaming semantics.

• Polling a live background job no longer trips the no-progress guard. Long test suites can be waited on normally, bounded by the job timeout and a generous safety backstop.
• When waiting stalls, the stop message now names the right primitive: arc_wake_when for room activity or shell_output for a background job, including when the inner call is routed through arc_call_tool.
• An empty, unrecoverable tool envelope now reports the envelope error directly instead of surfacing a misleading missing parameter from the inner tool.
• Host discovery reports every interpreter it can find, including versions, and linked git worktrees carry an explicit warning when an installed package could shadow the checkout under test.

Windows 10/11 (64-bit), or macOS 11+ on Apple Silicon.