One workspace.
Every model.
Real control.

The operator may choose the exact OpenRouter model used by a pane and can see what was requested and what provider actually served the request.

The application can delegate model selection through Auto/Smart Select behavior or apply routing profiles for cost, reasoning, coding, research, and provider preference.

Every pane preserves separate conversation state, model choice, usage, attachments, authority, errors, and history so one branch never silently alters another.

Google Drive, VPL authority, projects, attachments, and reusable context can be applied intentionally and remain traceable to the request that used them.

V4 introduces a shared mode/scope registry so existing modes remain isolated and unknown scopes fail explicitly instead of silently resolving to BUILD.

The preferred Windows path uses getUserMedia + MediaRecorder + server-side transcription, while Stop ends capture without submitting and dictated text remains in the compose field for manual review.

Useful per-request activity detail moves into Session History so history becomes the primary drill-down surface rather than maintaining a separate Activity path.

V4 extends Library Context toward multiple selected sources with preserved provenance, bounded context ceilings and backward-compatible one-item behavior.

Each bounded implementation wave can verify the exact source produced by the previous banked wave before edits begin, preventing silent drift across agents or sessions.

The V4 design includes a Windows troubleshooting-log collector pattern that can pull scrubbed diagnostic evidence into the local Google Drive-synced project evidence plane without embedding Drive credentials in the app.

Build is not a blank prompt box. GPT is calibrated as the Build Director / Governor, receives the MySmartRouter operating context, and is expected to guide the next action rather than make the user manage the process.

A context-hygiene control shows when the Governor context is fresh, aging or stale. Green is fresh, amber begins around 30 minutes, and red at about 45 minutes signals elevated drift risk. Refresh re-anchors GPT to the latest governed Build state.

Handoff controls identify the recommended next specialist/tool. The recommended handoff is visually ON so the user can see what should happen next instead of interpreting a row of equal choices.

Ralph is a bounded execution loop, not another AI model. GPT determines whether Ralph is appropriate and prescribes task scope, acceptance criteria, loop class, iteration ceiling, time/cost ceiling and tests.

Returned work from another tool or agent can be brought back into the active Build session, reviewed, imported, saved and used to drive the next governed step.

Independent QA can be sent through the configured relay to an AI reviewer. The Build workflow keeps QA separate from the model or tool that performed the primary work.

The user enters the work and MySmartRouter uses the configured OpenRouter automatic-routing path so the user does not have to manually select a model for every request.

A separate creation mode uses an available zero-token-price OpenRouter route when available, giving the user a practical place for drafts, exploration, brainstorming and other lightweight work without consuming paid model credits.

Users can dictate long prompts, see live transcription, stop and submit, and save the result as JSON. Voice makes complex direction faster when typing would slow the work down.

Build, Auto Router and Free Model each maintain their own Session History. V2.0 keeps the last 25 sessions per section and supports Open and Delete without mixing history across modes.

Build and Auto Router expose current-session cost and per-session cost in history. Each mode also shows a rolling Total Cost · Past 30 Days based on the current date rather than a calendar-month reset.

Build, Auto Router and Free Model keep independent server-side OpenRouter and MySQL configuration so keys, database access and section history are not forced into one shared configuration.

V2.0 adds plain-language help for Build, Ralph, handoffs, Drift Risk, relay/QA, Auto Router, Free Model, voice, sessions, OpenRouter setup, MySQL setup, hosting setup, security and troubleshooting.

The Build header keeps a visible summary of the current session, working owner/model and latest activity so the user can return after time away and quickly understand where the build stands.

The control surface is designed to remain usable in Windows Chrome and on narrow mobile screens without turning the interface into an unreadable desktop-only layout.

GPT serves as the sole product/build governor and adjudicator for the governed Build path. It independently reviews evidence, authorizes the next bounded wave, directs rework when required, and keeps implementation sequencing out of the operator's hands unless a true product decision is needed.

The operator remains product owner rather than technical courier, debugger or project manager. The workflow is designed so the user does not have to shuttle files between models, interpret stack traces, reconstruct recorded context, compare technical artifacts manually or decide sequencing that the control plane can determine.

Execution/preparation engineering and implementation agents work only within the current authorized packet. Allowed files, forbidden scope, acceptance evidence and stop conditions are explicit, preventing an AI coding agent from silently broadening a task or turning a bounded repair into an uncontrolled rewrite.

The normal operator cadence is PROCEED → READY → REVIEW → PROCEED. PROCEED executes only the next authorized wave; READY ends execution; REVIEW triggers an independent evidence inspection; STATUS reports position without advancing work; ERROR records a controlled failure; STOP halts at a safe persisted state.

Google Drive functions as the shared evidence plane while chat remains the control surface. Build Control records the current phase, banked state, authorization, actors, permitted scope, residuals and next gate. Once work is PASS / BANK, it becomes protected authority and cannot be silently reopened by the execution agent.

Each review resolves to one of four explicit dispositions: PASS / BANK, PASS WITH RESIDUAL, REWORK REQUIRED or BLOCKED. The supervisory reviewer then records what passed, issues, current position, implementation readiness, remaining work and the exact next command.

Later implementation waves can be gated against the exact source state produced by earlier work. Continuity hashes, allowed-file lists and source checks make it possible to detect drift before another agent edits the codebase and to stop rather than continue from an unverified baseline.

If a model session loses conversational context, the project is rehydrated from Drive authority: Build Control, latest banked review, Current Review evidence and the relevant source package. The control model deliberately prefers persistent authority over reconstructing the build from model memory.

Existing Ralph loops and handoffs remain subordinate to the Governor. Ralph stays bounded by approved task scope, acceptance criteria, iteration and cost/time ceilings, tests and explicit START / PAUSE / STOP / REVIEW RESULTS controls; handoffs remain recommended actions rather than unmanaged model-to-model autonomy.

AI-assisted development can move quickly but can also drift when several models, tools and sessions participate. The V4 control plane makes authority, sequencing, review ownership and evidence persistent, so speed does not require the user to become the human message bus between AI systems.

This is governed human-in-the-loop execution, not unrestricted autonomous software development. External tools and implementation agents remain bounded by explicit authorization, evidence requirements, review gates and stop conditions, with the operator retaining product authority.

Each pane owns its own conversation ID, title, messages, model, system instructions, streaming state, abort controller, draft, errors, route metadata, and accumulated cost.

Responses stream independently. Stopping or regenerating one pane affects only that pane and leaves neighboring work untouched.

Conversation history is stored locally in IndexedDB, enriched with projects, tags, archive/pin state, usage totals, and Drive mapping where available.

Supported local text/Markdown files and Drive references can become explicit pane context. The application records which attachments were associated with the request.

Conversations can be exported as JSON or Markdown, while schema-aware import preserves older V1/V2 records without destructive identifier collisions.

The product deliberately avoids becoming a generic sidebar-plus-single-thread clone. Large side-by-side AI workstreams remain a core interaction principle.

The command is treated as an application action. MySmartRouter loads the VPL master, uses it to route to the needed subordinate authority, and injects only relevant governed context.

Conversations can be mapped to Drive files and reopened into normal panes. Remote modification metadata is checked so a newer Drive copy is not silently overwritten.

When local and Drive copies diverge, the product favors explicit user decisions such as opening the remote version, saving a new copy, or deliberately overwriting.

The user enters one source prompt, selects two or more models, and launches them concurrently. MySmartRouter creates one normal independent pane per selected model and records a shared launch ID for provenance and combined analytics.

After the common starting question, each branch becomes fully independent. The user can challenge one model, redirect another, or stop one branch without affecting siblings.

Each branch keeps its normal usage and cost ledger while the parent launch record aggregates total request count, tokens, and combined spend across the fan-out.

Copy Full Chat, Copy From Here, and Copy Context Only create clean portable context. Paste Into Chat makes the destination independent while preserving optional source provenance separately.

Shared launch IDs and copy provenance never create live model-to-model synchronization. A destination chat cannot silently alter the source or sibling branches.

V2 deliberately stops before autonomous agent collaboration, generalized workflow orchestration, arbitrary shell control, or unattended infrastructure changes.

V2A establishes the architectural boundary for future operational tools such as Drive, Hostinger, Git/GitHub, and VPL Build Control. Only explicitly authorized capabilities are active.

OpenRouter and Drive connection state are presented separately. Credentials are not embedded into the standalone HTML and are kept apart from normal local settings.

Governed work records what authority and context were used for the request, making later review possible without relying on model memory.

The product distinguishes requested model from actual model/provider where route metadata is available, preserving operator visibility into how a result was produced.

Build and Auto Router show the accumulated cost of the active session when authoritative provider cost metadata is available.

Historical Build and Auto Router sessions carry their saved cost in the session metadata/pill so users can see which workstreams consumed spend.

Each mode shows Total Cost · Past 30 Days using a true rolling date window. Free Model will ordinarily remain at zero when the provider reports zero-token-price usage.

Input, output, total tokens, cost, reasoning tokens, cached tokens, model and provider are captured when returned.

Each pane accumulates request count, total tokens, and cost so the user can see the operational footprint of a workstream.

V2A can roll usage by model, provider, conversation, project, context profile, routing profile, and time period while monthly budgets provide a spend guardrail.

Established the standalone HTML, OpenRouter authentication, dynamic model catalog, independent pane state, streaming, local persistence, export/import, usage metrics, budgets, and two-pane mobile/desktop identity.

Added governed BUILD execution, Auto Router and Free Model creation paths, Drive/context concepts, provenance, handoffs, Ralph, independent QA, session history, cost visibility, voice and operator-minimizing controls.

Consolidated the working implementation into the authoritative stable source baseline with QA/finalization evidence, protected behavior, per-mode isolation and continuity artifacts that V4 must preserve.

Adds the locked five-mode structure, scope-safety registry, desktop voice repair, Save JSON/session hardening, activity/history consolidation, multi-select Library Context, diagnostic collection, stronger build governance and staged Multi / Image-Video expansion.