DynosAI 0.18.0 — Secure Autonomous Runtime

0.18.0 sits on the 0.17.0 eval-intelligence baseline and places named execution policy outside the model without claiming a container sandbox.

What this release is

Longer local work needs a host-owned security profile. Strict, Balanced and Autonomous select network, dependency and secret materialization. They never skip human gates. Local network policy remains decision-only.

Shipped

• Execution profiles: Strict (offline), Balanced (default-deny), Autonomous (allowlist — not unrestricted).
• Runtime-only secret vault under .dynosai/runtime/secret-vault.json. Values never enter model context, evidence, logs or Studio.
• Policy evidence on Studio code/merge review, project overview, dynosai_stats and diagnostic bundles.
• Host-owned Studio Project settings control. The coding agent cannot change the profile through MCP.

Deliberately not shipped

• Docker, VM or remote ExecutionRuntime implementations.
• OS-level network interception for child-process sockets.
• Secret materialization into model prompts.
• Skipping specification, plan, code or merge gates because a profile is named Autonomous.

Authority unchanged

Schema remains v6. Git is source truth. .dynosai/knowledge.db is workflow truth. The vault is a runtime file, not a new authority table. Cursor ACP and Codex app-server transports are unchanged. 0.16 leases, the 0.15 harness and 0.17 eval intelligence are preserved.

Validation

Release still requires repository checks, Studio asset sync, the stable pytest suite, and website check/typecheck/lint/build when web files change.