AI coding agents inside IDEs make broad, multi-file edits from short, vague prompts — and can silently break relationships between files the agent isn't tracking (e.g., resizing a frontend button accidentally desyncs the backend route it calls).
Watchtower passively builds and maintains a JSON map of how your codebase's files and functions connect (within a language and across languages) alongside a running change log. It uses this shared foundation to power four core capabilities.
- Passive Connection-Break Detection (
Alt+A) — Flags when a change breaks a tracked cross-file relationship. - Change Logging Over Time — Records what changed, where, and whether it was reverted or repeated ("thrash").
- Prompt Refinement (
Alt+P) — Rewrites a vague prompt into an explicit one, grounded in the map + log. Output via a pseudo-terminal and auto-copied to your clipboard to paste into Antigravity. - Context-Aware Chat (
Alt+C) — Answers questions about the codebase grounded in real project structure and history, not generic explanation.
(Placeholder for Screen Recording/Demo)
A small circular HUD shows passive-watch status at a glance (Off / Scanning / Watching / Flagged).
Note: Mistral (cloud API) is used only as a reasoning layer on top of this data — refining prompts, answering questions, explaining flags. It never edits the relationship map or the codebase; map-building stays mechanical and deterministic.
watchtower/
├── package.json VS Code extension manifest (commands, keybindings, settings)
├── README.md This file
├── assets/ Static assets including logos and demo GIFs
├── src/
│ ├── extension.js Entry point — wires Alt+A / Alt+P / Alt+C + HUD together
│ ├── core/ Shared file names, HUD states, storage and watcher
│ ├── parsers/ Plugin architecture + JS/Python parsers
│ ├── matcher/ Language-agnostic route/string cross-language bridge
│ ├── map/ Full scan + incremental update
│ ├── log/ Change recording, revert detection, thrash detection
│ ├── detection/ Connection break detection logic
│ ├── hud/ Circular status webview (Off/Scanning/Watching/Flagged)
│ ├── prompt/ Context-scoping for Alt+P
│ ├── chat/ Context-retrieval for Alt+C
│ └── mistral/ Cloud API wrapper (advisory-only)
├── test/ Zero-dependency assertion test suite (43 checks)
├── test-fixtures/ Small JS+Python fixture project used by tests
└── archive/ Contains older files like the phase0-spike
For more detailed diagrams on how Watchtower works under the hood, see ARCHITECTURE.md.
Watchtower has zero npm dependencies, so there's no npm install step.
- Copy (or symlink) the
watchtower/folder into your extensions directory:- macOS/Linux:
~/.vscode/extensions/watchtower-dev.watchtower-0.1.0 - Antigravity should follow the same VS Code extensions convention.
- macOS/Linux:
- Reload the IDE window.
- Run Watchtower: Set Mistral API Key from the command palette and paste in your Mistral cloud API key.
- Open a project folder and press Alt+A. The HUD should appear and enter "Scanning," then settle into "Watching" once the first scan completes.
| Key | Action |
|---|---|
Alt+A |
First press: one-time scan + build relationship map. Later presses: toggle passive watching on/off. |
Alt+P |
Select a rough prompt or code, then press to get a refined prompt (copied to clipboard). Requires Alt+A to have run at least once. |
Alt+C |
Open the context-aware chat prompt box. |
watchtower.mistralModel— Mistral model name (defaultmistral-large-latest).watchtower.flagThreshold—strict/balanced(default) /lenient— how aggressively connection breaks are flagged.watchtower.maxLogEntriesInPromptContext— caps how much change-log history is pulled into Alt+P / Alt+C context to avoid token-wasteful over-stuffing.
No install step needed — the extension has zero runtime dependencies outside the Node/VS Code standard library.
node test/runTests.jsRuns 43 checks including JS/Python parser testing, cross-language route matcher, HUD rendering, prompt context scoping, and connection-break detection.