Repository navigation
Home
A model-agnostic, multi-agent framework built on LangGraph that turns a prompt into product-ready software.
Today SCIO takes one software task, from a Markdown spec or a GitHub Issue, to a merged, tested and documented pull request, and does it reliably with real models. Product Mode (milestones M6 to M11) is extending that into building whole products from a prompt, in any stack, one approved milestone at a time.
Status (2026-10-12): v0.4.0 released. Product Mode M6 (Foundations) merged, M7 (Greenfield bootstrap,
scio new) in review. Task mode is proven live: ten regression tasks at pass^5 = 1.00 offline, nightly live suites green, first live debugger and advisor runs recorded.
| You want to | Read |
|---|---|
| Install and run a task or bootstrap a product | Getting Started |
| Understand the design: layers, tiers, the graph, escalation | Architecture |
| See the Product Mode roadmap and what each milestone delivers | Product Mode |
| Add a stack (Node, Go, ...) | Toolchain Profiles |
| See what passes, what it costs, how the evals work | Evals and Live Results |
| Know why a decision was made | Decision Records |
| Contribute, or run a session the way the project expects | Contributing and Protocols |
This wiki is the map; the territory is docs-as-code in the repository and is checked by CI (links, model-ID leaks, diagrams versus the compiled graph, changelog shape). When the wiki and a repository document disagree, the repository wins.
- README: landing page, usage, acceptance status
- docs/ARCHITECTURE.md: the design, state schema, agent contracts, generated graph diagrams
- docs/PRODUCT_MODE.md: the Product Mode roadmap
- docs/adr/: architecture decision records 0001 to 0010
- docs/LESSONS_LEARNED.md: one retrospective per milestone
- docs/ISSUE_LEDGER.md: every problem hit, with root cause, fix and sweep
- docs/EVAL_HISTORY.md: nightly live results
- Project board and milestones
A cheap router classifies each task into an effort tier (T0, T1 or T2). Opus plans, advises and checks that the objective was met before finishing. Sonnet does the work: coding, testing, debugging, documenting. Every loop is bounded (iterations, stalls, tool calls, tokens, dollars), every side effect is idempotent under crash and resume, tests and lint are ground truth that no model verdict can override, and escalation to the advisor is capped and logged with its trigger, cost and outcome. One writer at a time; parallel work is read-only.