Summary
Add a first-class admin dashboard for llama-swap that gives local operators a lightweight control center for model/process state, routing visibility, logs, metrics, recent requests, and common model operations.
This should borrow the useful parts of oMLX’s dashboard pattern: fast local visibility, model management, stats, and setup guidance, without turning llama-swap into a full model studio.
Scope
Implement a dashboard landing view under the existing UI that shows:
- server health, uptime, version/build info where available
- running and configured model counts
- currently running models and lifecycle state
- recent model load/unload/swap/failure events
- recent proxied requests with model, endpoint family, status, and duration where available
- links into logs, metrics, request inspection, playground, configuration, and integrations
Add model operations from the dashboard:
- load configured model
- unload one running model
- unload all non-protected running models
- show pending/success/error state for operations
Add read-only configuration visibility:
- configured models
- aliases
- TTL/process settings
- upstream endpoint or assigned port where safe
- redaction for API keys, auth headers, secret-like environment variables, and secret-like command arguments
Add lightweight integration setup guidance when llama-swap has enough configuration to generate it.
Non-Goals
- Do not add model downloading or Hugging Face model management.
- Do not add arbitrary command editing from the browser.
- Do not attempt full oMLX-style runtime tuning such as sampling parameters, chat template kwargs, model type overrides, KV cache controls, or inference-engine memory controls.
- Do not add backend-specific metrics unless they are exposed through a stable upstream API.
Acceptance Criteria
- The dashboard renders useful partial data even if one backing API fails.
- Running models are visually distinct from stopped models.
- Recent failures are visible without opening logs.
- Request overview never shows full request or response bodies.
- Sensitive config values are redacted.
- Dashboard controls are keyboard reachable.
- UI remains fully local/offline-capable.
- The dashboard feels lightweight and operational, not like a full model studio.
PRD.md
SPEC.md
Summary
Add a first-class admin dashboard for llama-swap that gives local operators a lightweight control center for model/process state, routing visibility, logs, metrics, recent requests, and common model operations.
This should borrow the useful parts of oMLX’s dashboard pattern: fast local visibility, model management, stats, and setup guidance, without turning llama-swap into a full model studio.
Scope
Implement a dashboard landing view under the existing UI that shows:
Add model operations from the dashboard:
Add read-only configuration visibility:
Add lightweight integration setup guidance when llama-swap has enough configuration to generate it.
Non-Goals
Acceptance Criteria
PRD.md
SPEC.md