How it works
How repository evidence becomes engineering workflow, dependency, business-domain, and architecture context
Unoplat Code Confluence turns repository evidence into focused, reusable context for developers and coding agents. Framework-aware discovery identifies relevant files, constructs, and source spans first; generation then works from that evidence instead of beginning with an unrestricted repository-wide search.
Engineering Workflow
Verified repository commands and configuration for everyday development.
Dependency Guide
A concise explanation of the packages the codebase depends on.
Business Domain
A source-backed summary of the domain represented by detected data models.
Repository Architecture
A validated, high-level diagram grounded in detected application boundaries.
Engineering Workflow
The workflow guide inspects package-manager scripts and relevant configuration files to identify the canonical install, build, development, test, lint, and type-check commands. It records the working directory and configuration source when available and explicitly marks stages for which no command was detected.
The result is written to the managed Engineering Workflow section of each codebase's AGENTS.md. If that section already agrees with current repository evidence, it is preserved rather than rewritten.
Dependency Guide
Detected dependencies are normalized into a current codebase inventory. Existing entries are reused when a dependency has not changed; only new dependencies require additional research. Public packages are described from official or vendor-authored documentation, while dependencies that cannot be verified confidently are conservatively identified as internal.
The validated result is written to dependencies_overview.md, with AGENTS.md linking to it as the detailed dependency reference.
Business Domain
Business-domain generation starts from detected data-model identifiers and their exact source spans. It reads each identified span and nearby context, then synthesizes a concise description of the dominant domain and the main concepts represented by those models.
The summary is written to the managed Business Domain section of AGENTS.md. A deterministic companion file, business_domain_references.md, preserves the supporting model identifiers and line-level source links so readers can trace the summary back to code.
Repository Architecture
Architecture generation builds on fresh interface evidence produced for the current repository run. Framework and library operations are first classified by the published DIDS capability enum, then normalized into architectural roles. Each codebase's app_interfaces.md records the resulting inbound, outbound, bidirectional, and internal constructs as codebase-relative anchors in this form:
path: L<line>: <match_text>These anchors provide a focused starting point for modeling application boundaries. The architecture process reads only the listed fresh app_interfaces.md artifacts, then inspects related source sections and relevant deployment evidence when needed. It runs when fresh evidence includes an external boundary, avoiding an architecture claim when the required foundation is absent.
The result is a repository-root architecture.md containing one focused Mermaid architecture-beta diagram. Before completion, the diagram is rendered and its SVG is inspected; the current artifact is accepted only after validation succeeds.
Precision and token efficiency by design
Source spans narrow business-domain reading to detected models. Fresh interface anchors narrow architecture analysis to relevant boundaries and code. Dependency deltas reuse unchanged entries, while no-change checks preserve accurate workflow content. Deterministic rendering and final validation keep the published artifacts stable and verifiable.
Together, these mechanisms reduce unnecessary model input and repeated repository exploration. Actual token usage varies with repository structure, detected evidence, and the configured model.
Alpha
Unoplat Code Confluence is in active development. Generated context should be reviewed alongside the linked source evidence, especially for frameworks or application patterns that are not yet covered.