code-engine

AI-native Software Intelligence and Generation Platform — an AI software engineering organization that researches, designs, generates, verifies, and continues to understand the software it builds. Third engine on the research-engine substrate. A MoreSalamander StudioLabs production.

User Idea → Product Intelligence → Parallel Software Research
→ Software Context Graph → Software Knowledge Graph
→ Architecture Intelligence → AI Engineering Team
→ Generated Software → Continuous Verification
→ Impact Analysis → Evolution → Reusable Knowledge

The generated repository carries its own self-model: .code_engine/graph.json maps every file to the component, feature, and requirement it exists for — written during generation, not inferred after. Impact analysis and evolution checks read the sidecar, so the software stays understood long after it is generated.

Verification is execution

Every build runs ruff and the generated project's own pytest suite as real subprocesses, plus three deterministic lint agents (Security, Performance, Code Review). Failures map through the graph — file → component → feature → requirement — and land as problem nodes. The repair loop is bounded and gated: patches must parse, stay inside the project, and never touch tests. A red suite is reported red.

Semantic change intelligence

`impact`	before touching a component: transitive dependents, affected APIs, tests to re-run, requirements at risk, suspect docs, infrastructure touchpoints — pure graph traversal
`evolve`	reality vs model: ast-observed imports vs declared dependencies (drift), OSV.dev advisories (live), PyPI staleness (live), documentation drift — findings persist into the sidecar
`patterns`	cross-project memory: engineering patterns ranked by observed verification success and repair cost

Honest scope

One stack, deep and real: FastAPI + SQLAlchemy/SQLite + pytest + static JS — generated projects run, and their tests execute in the loop. Custom endpoints beyond CRUD are visible implemented: false stubs recorded as known problems. Semantic modeling covers generated projects by construction plus Python-ast drift analysis; arbitrary polyglot repos are out. Parallel.ai ships as a keyed adapter, fail-closed without a key. Learning uses our own verification outcomes — no invented telemetry.

Run it

.venv/bin/python -m code_engine build "Build an AI research platform"
.venv/bin/python -m code_engine impact <project_dir> <component>
.venv/bin/python -m code_engine evolve <project_dir>

github.com/MoreSalamander/code-engine · 25 offline tests (incl. executing a generated project's own suite) · substrate: research-engine · part of MoreSalamander