Architecture
Design goal
Deconnected correlates evidence that normally lives in separate tools. It does not try to replace browser discovery, static-flow analyzers, SQL parsers or code agents. It normalizes their output into a graph and applies refactoring-specific policy.
Layers
Discovery
Sources include:
- repository files and imports;
- LibCST facts for Python;
- sqlglot table references;
- Docker Compose and Dockerfiles;
- SQLAlchemy database reflection;
- TestQL runtime topology;
- optional Semcod analyzers.
Evidence graph
Typical node kinds:
file, module, route, service, config, table, page, URL, asset, test
Typical edge kinds:
imports, calls_or_defines_api, uses_table, depends_on,
foreign_key, observes_route, loads, defines_service
Every edge can carry evidence, confidence, runtime and test-only flags.
Correlation
Runtime URLs are normalized and matched to static routes. This enables paths such as:
page → browser request → route → source file → table
A complete path increases confidence. A name match by itself does not.
Classification
Tables are classified from:
- entrypoint reachability;
- static incoming references;
- runtime references;
- test-only references.
Refactoring planner
The rule-based planner chooses the highest-ranked extraction candidate and emits a constrained RefactorPlan. The plan includes:
- affected symbols;
- allowed paths;
- ordered steps;
- validations;
- assumptions;
- blockers;
- rollback instructions.
Agent providers
LiteLLMProvider refines structured plans;ClaudeCodeProvider can plan or edit a repository through claude -p;- providers return
AgentRunResult rather than directly controlling acceptance.
Policy and verification
After an agent run:
- snapshots are compared;
- changed files are checked against plan scope;
- forbidden operations are searched in the patch;
- validations run only if policy passes;
- all artifacts remain auditable.
Current limitations
- JavaScript/TypeScript analysis is primarily source-pattern based rather than TypeScript Compiler API based;
- graph diff before/after is not yet a separate acceptance gate;
- runtime SQL traces are available as a library helper but are not automatically merged from arbitrary external test runs;
allowed_paths enforcement is post-execution, not a filesystem sandbox;- model cost usage is not yet collected from provider responses;
- destructive database operations are blocked by policy but no approval workflow is implemented.