AIWorkHubOne controlled engineering loop
AIWorkHub connects repository context, a manager plan, dependency-aware task cards, model routing, isolated execution, evidence collection and independent review. The parts are not separate dashboards: each transition belongs to one explicit task state machine.
Task truth before model prose
Creating a card does not mean a model ran. A task becomes processing only after an exact claim and successful launch. Workers stop at review; the verified manager inspects the diff, tests, logs, artifacts and tool receipts before accepting or returning residual work.
Use the model routes you already have
Editor routes use models visible through VS Code. CLI routes reuse their own authenticated sessions. Codex, Claude, Copilot, DeepSeek and GLM are the workforce—not products AIWorkHub replaces. Runtime preflight separates model visibility, route readiness and optional fallbacks so an unavailable redundant route does not become a false repository blocker.
Optimize the model portfolio
AIWorkHub can keep premium frontier models for architecture, hard judgment and independent review while assigning bounded throughput to lower-cost capable routes. Source Graph, bounded reads and focused replacement output reduce avoidable tokens inside each route; routing reduces the share of remaining work paid at premium-model rates. Deterministic local tools handle hash checks, edit application and validation when model judgment is unnecessary.
Economy is measured at accepted quality: a cheaper failed attempt is not a saving. Current evidence shows a large observed price spread and substantial retry spend, while a quality-adjusted fleet-wide ROI remains an explicit open benchmark.
Repository-native authority
Every repository owns its .aiworkhub/ state. Tasks, callbacks, indexes, context stores and audit evidence do not move into a central AIWorkHub cloud service.