Agentic workstation and RegesMode

Planned. The native release does not execute agent tasks, shell commands, browser actions or repository patches. The Project Workbench, Orchestrator and RegesMode atlas views are interactive local simulations, not connected execution systems.

Intended workbench

The target is a local coding workstation with a real folder tree, multi-file editor, inspectable changes, terminal, owned browser preview and model-neutral agent runs. A model should propose structured operations through a broker rather than execute arbitrary strings from its output.

Patch application needs path validation, revision checks, a reviewable diff, atomic application where possible and rollback/recovery. A file must not become a parent directory, traversal must be rejected, and a stale revision must not silently overwrite newer work.

RegesMode policy

RegesMode is the planned policy layer for effective permissions. It should combine the workspace’s rules, selected role, tool scope and explicit user grants. Separate reads, local writes, command execution, network access, credential use and external publication. Model output, repository text and retrieved documents cannot grant themselves a higher role.

An actual sandbox must state its isolation guarantees. Running a command in the same privileged host process with a decorative badge is insufficient. Tool children need environment allowlists, deadlines, output bounds, cancellation and tracked lifecycle.

Orchestration

Agent roles are workflow responsibilities, not internal MoE experts. A planner, implementer and reviewer may use the same or different serving models. Budgets, serial/parallel execution, pause/stop/reset, handoff state and durable audit events should be explicit.

Every tool result needs provenance and a request/run identity. A run is complete only when the required verification and recovery conditions are met, not when a model writes a confident summary. Preserve original failures rather than repeatedly masking them with patches.

References and gates

LM Studio Bionic is an independent reference for agent-oriented workflows; it is not a RegesARC component. The REA repository is another research reference and must be assessed at a pinned revision before reuse.

Native release gates include actual scoped tools, permission enforcement, structured calls, durable run state, cancellation, stale-patch tests, owned browser/process cleanup and end-to-end task verification. Until those pass, the native sidebar correctly labels the agent platform as planned.