Anchr

With ⚓, Drift is Okay.

Anchr is a VSCode extension and repo-level protocol framework for guarding AI coding agents against drift, hallucinated verification, scope creep, and false completion.

What It Does

Anchr installs a .anchr/ protocol folder into a workspace. AI agents read the manifesto, work one atomic step at a time, and write .anchr/out/signal.json after each step. The VSCode daemon watches that signal file and runs deterministic checks before the agent continues.

On success, the status bar stays active. On violation, Anchr enters HARD STOP, writes a lock file, shows a VSCode notification, and prevents further signal processing until a human clears the lock.

The current hardening layer adds locked-goal evidence, bounded decomposition, rule/domain-confidence evidence, two-run test attestation, high-impact action approval, raw signal rate/blocklist protection, procedure memory, and graph regression checks. It also records session-isolated quality/relevance/style telemetry, enforces deterministic locked-goal relevance at the signal boundary, requires ABA re-anchor evidence after style drift, evaluates six drift metrics, and ships a production-code red-team latency suite. These are local feature vectors, not provider embeddings. Provider ensembles, provider embeddings, and multi-agent routing remain runtime-bound unless Anchr owns that execution boundary.

Requirements

Anchr's protocol checks run through the anchr_tools.py CLI, so Python 3.10+ must be on your PATH (python3/python on macOS/Linux, python/py on Windows). The extension shows a warning on activation if no Python launcher is found. A Git repository is also required — the daemon refuses to run outside one.

How to Run

1. Install from the Marketplace. In VSCode open the Extensions view, search Anchr (publisher Vivartan Enterprises), and Install — or get it from the VS Code Marketplace / Open VSX. Reload the window after installing so the new version is active.
2. Initialize the workspace. Open your project (it must be a Git repo with Python 3.10+ on PATH — see Requirements), then run Anchr: Initialize in Workspace from the Command Palette or the Anchr panel. This scaffolds .anchr/ (manifesto, config, tools) and wires your agent's rule files (CLAUDE.md / AGENTS.md / Copilot / Cline / Roo / Continue / Aider).
3. Start the daemon. Run Anchr: Start Daemon and confirm the status bar shows $(eye) Anchr (active). If it shows Anchr Off, agents would run unguarded — start it before working.
4. Point your agent at the protocol. Tell your coding agent: "Execute .anchr/start.md." It will read the manifesto, work one atomic step at a time, and write .anchr/out/signal.json after each step; the daemon validates every step and HARD STOPs on a real violation.
5. (Optional) Enrich the graph. Use the panel's Enrich graph (local AI) to add a free semantic layer (local Ollama, or docstrings) — never required, never sends code anywhere.

When Anchr updates, reload the VSCode window (Command Palette → "Developer: Reload Window") so the new version takes effect.

Main Commands

• Anchr: Initialize in Workspace
• Anchr: Start Daemon
• Anchr: Stop Daemon
• Anchr: Show Status
• Anchr: Open Panel
• Anchr: Clear Hard Stop Lock
• Anchr: Reset Session
• Anchr: Remove from Workspace
• Anchr: Sync - Re-verify Audit State

Runtime Files

.anchr/
  manifesto.md
  config.yml
  anchr_tools.py
  out/
    signal.json
    audit.rpt
    plan.rpt
    session.log
    sync.log
    web_sources.jsonl
    graph_manifest.json
    LOCK

Source of Truth

Configuration is intentionally kept in .anchr/config.yml. VSCode settings are not used for daemon behavior in v1, so the workspace protocol remains auditable and portable across agents.

Privacy & Telemetry

Anchr collects no telemetry. The daemon, the core anchr_tools.py checks, and the panel never contact any server — in standard mode the guard runs entirely locally. The only network calls Anchr makes are ones you or your agent explicitly trigger: the WEB_VERIFY tool fetches the source URLs you ask it to verify; optional paid graph semantic enrichment sends function snippets to the Anthropic API only when you opt in with --yes (local Ollama enrichment stays on your machine); and enterprise sign-in, when enabled, uses Google OAuth and the Anchr auth service at https://auth.vivartanenterprises.dev for entitlement and billing. No code, signal, audit finding, or workspace content is sent anywhere otherwise.

Build

npm install
npm run check
npm run compile
npx @vscode/vsce package

Publisher

Publisher ID: VIVARTAN

Author: Vivartan Enterprises