git:20260820.2ef9899 to git:20260824.6d915b2

8 added, 7 removed. Audit A to A.

# AIDE Sovereign Workbench

> Offline-first, open-source development workbench: real code editor, local AI models, Git review, debugging, and optional online providers — without sending source code to the cloud.

## Key facts

- Two-process architecture: browser frontend (Monaco editor) + Node.js daemon on loopback (HTTP/WebSocket, ports 4173/4777).

+ - Cockpit orchestrator UI: Describe -> Plan -> Approve -> Build -> Verify; approval-gated agent loop with diff previews, live git-diff workbench, verification rail.

+ - Model Hub in-app: Hugging Face GGUF search, quantization listing, resumable downloads, one-click engine registration; local import by path.

+ - Harness scaffold v2.1: sized operating layer per model (micro for small contexts, full Code+Lens for strong budgets); byte-deterministic; 20-task effectiveness battery published in docs/evidence/.

+ - Backend auto-select: probes llama-server builds via --list-devices; Vulkan/CUDA/CPU chosen by operator profile > benchmark verdict > hardware heuristic.

- Local AI chat via llama.cpp OpenAI-compatible loopback; GGUF weights downloaded separately and SHA-256 verified; never bundled.

- - Model handoff: every chat-capable model (local GGUF + cloud providers) is a route with its own context window; health-gated per-conversation model picker, history fitted to the target window, explicit fallback + answer labeling.

- - LSP client for TypeScript (diagnostics, completion, hover, go-to-definition); DAP client for Python/debugpy (foundation).

- - Git integration via CLI: status, diff, stage, commit, branches, history.

- - Online providers opt-in: OpenAI, Anthropic, Google Gemini, Mistral, Groq, OpenRouter; credentials encrypted at rest with Windows DPAPI in the daemon; host allowlist with explicit approval for custom hosts.

- - ChatGPT/Claude conversations.json export import into local chat history (additive, 10 MB cap).

- - Veritas evidence gates: AI-generated changes must pass compile, tests, Git whitespace, manifest validation, path boundaries, secret scanning before release; models never approve or apply their own changes.

- - Verified continuously: 188 architecture tests, 17 end-to-end browser tests, 10/10 doctor preflight.

+ - LSP client for TypeScript; DAP client for Python/debugpy. Git via CLI. Tasks service with problem matchers.

+ - Online providers opt-in (OpenAI, Anthropic, Gemini, Mistral, Groq, OpenRouter); DPAPI-encrypted credentials; host allowlist.

+ - Veritas evidence gates: models never approve or apply their own changes.

+ - Verified continuously: 188 architecture tests, 17 e2e browser tests, 10/10 doctor preflight, plus unit batteries (facade 12, scaffold 5, agent loop 14+5).

## Links

- Repository: https://github.com/AnonymousNomad/aide-sovereign-workbench

- License: Apache-2.0 (LICENSE)

- Security: SECURITY.md

- Contributing: CONTRIBUTING.md

- Roadmap: docs/RELEASE_ROADMAP.md

- Research log: docs/RESEARCH_LOG.md

- Model packs: models/PACKS.md

- Benchmarks: benchmarks/

- Docs: docs/