# Project Roadmap

This roadmap outlines completed development milestones, ongoing stabilization efforts, and scheduled future capabilities for Antigravity Swiss Knife.

---

## 1. Completed Milestones (v0.1.0 to v0.2.0)

- **Milestone 1–5 (Core Daemon & Secret Rotation)**: Pure-Go background daemon, Linux Secret Service integration (`libsecret`), atomic credential swapping, SQLite WAL concurrency protection, and active conversation layout preservation.
- **Milestone 6–9 (Layout & MD3 Design System)**: Window geometry enforcement (1152x648 px, strict 16:9 aspect ratio, 4px layout token grid), Google Material Design 3 dark theme, and dual concentric circular quota gauges.
- **Milestone 10–13 (Cross-Platform Packaging & VM Validation)**: Automated release packaging pipeline producing native `.deb`, NSIS `.exe`, and macOS `.dmg` bundles; automated validation inside Windows 11 and macOS Ventura VMs using Quickemu; Git worktree development scripts.
- **Milestone 14–19 (Host Injection & Workspace Overhaul)**: Chrome DevTools Protocol runtime script injection (`persistent_script.js`), 6-probe custom model security auditing, full-lifecycle GitHub Workspace Kanban board, and active conversation continuation latch (`pkg/revival`).

---

## 2. Near-Term Objectives (v0.2.x Cycle)

- **Headless Daemon Autonomy**: Enhancing standalone CLI daemon mode (`swiss daemon --web`) to support full quota rotation and notification hooks without launching the Electron desktop GUI.
- **Wayland / PipeWire Coordinate Precision**: Refining display frame acquisition on Linux Wayland compositors to support high-refresh-rate display captures for Computer Use agents.
- **Worktree Session Decoupling**: Enabling independent conversation state tracking when switching between multiple concurrent Git worktrees on the same repository.

---

## 3. Planned Future Feature: Gemini Bridge (Utilities First Tab)

**Gemini Bridge** will be added as the first tab in the **Utilities** section (`Utilities` $\rightarrow$ `Gemini Bridge`, ahead of Storage Manager, ACP Agent Mesh, Chat & Project Importer, and Computer Use). It establishes a two-way communication and control link between **Gemini App / Web (Gemini Spark)** and all local **Antigravity surfaces**:
- **Antigravity 2.0 Desktop App** (via Chrome DevTools Protocol injection and local `~/.gemini/antigravity/brain/` state inspection)
- **Antigravity CLI (`agy`)**
- **Antigravity VS Code Extension**

### Core Capabilities
1. **Two-Way Agent Orchestration**:
   - Exposes a Streamable HTTP Model Context Protocol (`/mcp`) gateway and authenticated tunnel/relay managed by the Swiss Knife daemon.
   - Allows Gemini's cloud agent (from mobile or web) to consistently dispatch tasks to local Antigravity agents, monitor live step progress, answer approval prompts, and read back generated artifacts.
2. **Local Computer & Workspace Information Retrieval**:
   - Enables Gemini App/Web to query local project state, conversation transcripts, quota status, and local browser/GUI tools (`obscura`, `open-computer-use`, `Agent-S`) on demand.
3. **Account-Switch Resilience**:
   - Decouples the Gemini App/Web connection identity from local Antigravity account rotations so active remote sessions survive laptop account switches and IDE relaunches without interruption.

---

## 4. Planned Future Feature: Fine-Grained Token Breakdown (Skills vs Harness)

> [!NOTE]
> This capability is a planned future feature currently in architectural design. It is not yet implemented in the current production release.

Autonomous coding agents generate significant context token overhead simply by maintaining environment definitions, tool schemas, and conversation histories. Developers currently lack visibility into how many tokens drive actual problem-solving versus the structural framework running the agent.

The planned **Skills vs Harness Token Breakdown** will measure and categorize token consumption across two distinct layers:

### Categorization Model
1. **Harness Overhead Tokens**:
   - System Prompts & Guardrails: Core operational instructions, persona definitions, and project rules.
   - MCP Tool & Skill Schemas: Function signatures, parameter type descriptions, and environment capability manifests sent on every turn.
   - Conversation History Framing: Prior chat turns, message wrappers, and repeated context snapshots.
2. **Skill Execution Tokens**:
   - Reasoning & Planning: Internal chain-of-thought and decision tokens generated by the model.
   - Tool Invocation Payloads: Exact arguments emitted to execute code edits, run terminal commands, query databases, or automate browsers.
   - Execution Results: File contents, diff outputs, test failure traces, and compiler feedback returned by tool invocations.
   - Subagent Delegations: Context handed off to downstream child agents.

### Planned Integration Surfaces
- **Token Monitor Subsystem**:
  - A dedicated "Context Allocation" chart in the Token Monitor displaying the ratio of Harness vs Skill tokens over time.
  - Spend analysis calculating dollar expenditure driven by tool schema overhead versus productive code generation.
- **Main Chat View Overlay**:
  - A per-turn token badge positioned in each assistant message header.
  - Hovering or clicking the badge reveals an interactive breakdown showing the exact percentage of tokens in that turn dedicated to harness framing versus skill execution.

---

## 5. Long-Term Vision (v0.3.0 and Beyond)

- **Federated Local Agent Clustering**: Distributing subagent tasks across multiple workstations on a local network via Agent Client Protocol mesh networking.
- **Offline Conversation Vaulting**: Cryptographically secured, zero-knowledge local backups of entire conversation trees with full-text search indexing.