AetherStack for VS Code

One chat surface for local models, authenticated AI CLIs, and multi-model services.

AetherStack brings Codex, Antigravity, Grok, Cursor, Claude, Ollama, and configured API providers into VS Code. It keeps model routing, service state, memory, and local runtime controls in one place, then shows which CLI or model is actually working instead of hiding execution behind a generic spinner.

What is new in 0.3.327

• Durable graph artifacts use consistent save/reload validation. Invalid lineage or corrupt metadata stops consumption explicitly, without rewriting your saved files. Valid artifacts and repeated identical saves remain compatible.

Recent improvements

• Editable node profiles reach named-service contributors and final synthesis without silent shortening. Graph and pipeline instructions use the same explicit validation, and an invalid import cannot replace your saved valid pipeline. Accepted profile text retains its original whitespace and content.

• The CLI bridge preserves complete system instructions. Antigravity carries long reviewer profiles over stdin; native-argument and Cursor system-context limits now reject oversized input explicitly instead of silently shortening the instructions. Missing native-role metadata uses the provider's declared system flag.

• Long Antigravity requests use stdin, avoiding Windows command-line size failures without trimming the requested context. Only a successful terminal result counts as its answer; progress and partial failures remain distinct.

• Malformed checkpoint history and incorrectly typed iteration identities fail before saving or claiming resume state. Valid empty histories remain compatible.

• Durable graph checkpoints reject invalid lease settings and clock readings before they can corrupt resume state or bypass expiry. The saved state stays recoverable after a rejected operation.

• Parallel evidence workers retain distinct claim IDs through review and synthesis, including long IDs and collisions with another worker's IDs.

• Oversized evidence merges return a visible stage failure with diagnostics; claims are never silently dropped.

• Auto now runs the ordered multi-model synthesis service it advertises. Single-model priority/failover remains available as an explicitly labeled Direct mode instead of being presented as synthesized Auto.

• Provider, embedding, capability-matrix, discovery, and Redis traffic now authorizes every resolved and connected address. Connections are pinned to the checked address, closing DNS-rebinding gaps after policy approval.

• HTTP redirects are bounded and reauthorized. Cross-origin redirects lose bearer/cookie/Host secrets, external-to-local redirects and HTTPS downgrades are rejected, and unsafe request bodies are never silently replayed.

• Reused Redis pool sockets are rechecked immediately before every command, so an old connection cannot bypass a new request's classification policy.

• Host CLI timeout and cancellation remain bounded even if an operating-system process-tree kill produces no child close event.

• Release builds are byte-reproducible across Windows, macOS, and Linux, and Marketplace publication is checked against the exact tested VSIX before the matching GitHub Release becomes public.

• Answering retains the truthful Codex-style one-line process display with the current CLI, stage, elapsed time, attempt position, and real failover.

How answering works

Open AetherStack: Show Chat View or AetherStack: Open Chat in Editor and type normally.

• Auto classifies the task, assigns distinct lead/worker/review roles using your configured model priority, and returns one synthesized answer. The live process line always names the model and stage currently running.
• Direct explicitly sends the task to one suitable model from that same priority order and fails over only after a real failure or incomplete answer. It is visibly labeled Direct and never reported as synthesized.
• Named services such as Research, Planning, Coding, Testing, and Technical Writing run explicit lead/worker/review stages and return one synthesized answer. A single fallback response is not presented as successful synthesis.
• LLM sends the task to one model you choose explicitly.

Use /direct for the compatibility path, or /research, /plan, /code, /test, /bugfix, and /preset <name> to select a specialist workflow. Use the graph editor to build your own validated sequence.

Existing host CLI logins count. If Codex, Antigravity, Grok, Cursor, or Claude is already authenticated, AetherStack can use it through a loopback bridge protected by a random bearer token. Provider API keys and local Ollama models can participate in the same ordered model catalog.

Getting started

Docker Desktop or Docker Engine is the only required system prerequisite. Keep its daemon running, install the extension, then run AetherStack: Start All Services or press Start all services in the Control Center.

code --install-extension AetherStack.aetherstack

The extension installs its matching runtime from the bundled archive, verifies the SHA-256 checksum, and promotes only a complete versioned runtime. It then starts the local Hub, LiteLLM gateway, and Open WebUI. Docker itself is not silently installed because accepting its vendor license and administrator changes belongs to the user.

Core commands

The full command list remains available in the Command Palette.

Privacy and local inference

Private mode projects the request to RESTRICTED: cloud models are withheld, and model, embedding, storage, and tool effects must stay local. Private project and session data uses isolated namespaces and is not merged into the common memory pool on release. These are GDPR-ready privacy-by-design controls, not a claim that every deployment is legally compliant or that the local store is an encrypted vault.

For local inference, AetherStack prefers an existing host Ollama before its container fallback:

Use a model large enough for the task. qwen2.5-coder:7b is the default local-default backend and a practical baseline for coding work; smaller smoke models remain available but can visibly reduce answer quality.

ollama pull qwen2.5-coder:7b

Settings

Provider keys are read from the selected runtime's .env; the gateway key is stored in SecretStorage. AetherStack does not scan unrelated project .env files. Open WebUI is published only on 127.0.0.1:3000, and its backend is not exposed directly on a host port.

License

Source-available under the PolyForm Noncommercial License 1.0.0 included with the extension. Commercial use, including resale, is not permitted. Redistributions must retain the license and the required original-author notice.