Evidence sources
Git repositories, GitHub collaboration data, backlog systems, CI and test results, incidents, agent sessions, ADRs, and documentation.
How it works
Versioned evidence first. Intelligence and coordination on top.
The architecture preserves the current local archive workflow while adding semantic identity, historical relationships, and live coordination in later phases.
Ingestion and normalization
The current archive CLI plus future importers, language indexers, backlog connectors, and transcript parsers preserve source pointers and completeness.
Temporal intelligence plane
Symbol index, lineage, evidence graph, risk and staleness analysis, context building, and coordination are introduced progressively.
Delivery
CLI, MCP, HTTP, IDE, live web, and portable Astro archive experiences consume shared evidence-backed view models.
Identity through time · In development
A function is not permanently identified by a path and line range.
KodeAtlas is designed to preserve a stable logical symbol identity while recording each version of that symbol at a specific commit. The resolver can then describe whether the symbol is unchanged, modified, renamed, moved, signature-changed, split, merged, deleted, ambiguous, or unresolved.
Line ranges remain valuable historical evidence and display locations. They are not the permanent identity.
Domain model
Model software work as versioned entities and cited relationships.
Repository
Source identity, Git history, configuration, and external-system mappings.
Symbol
A stable logical identity for a function, method, class, type, route, test, configuration key, or other code element.
Symbol version
Where and how a symbol existed at a particular commit, including path, range, signature, fingerprints, and relationships.
Code anchor
A durable reference from a task, decision, incident, document, or claim to a symbol or structural code region.
Evidence
The Git object, PR, issue, review, test, incident, transcript, document, or human assertion supporting a fact or inference.
Context packet
A bounded answer joining current structure, history, risk, planned work, and active work, with citations and uncertainty.
Structured retrieval before fuzzy retrieval
Compiler, AST, Git, and graph evidence define identity.
Embeddings can help find a related review or transcript. They should not decide whether two methods are the same symbol, whether a signature changed, or whether a task targets a deleted function.
KodeAtlas uses structured evidence for identity and dependency questions, then adds semantic search where narrative retrieval is useful.
Local and team modes
One local command first. Shared infrastructure only when value requires it.
The first intelligence workflows are intended to use an embedded relational store and local object storage. Team mode can later use PostgreSQL, an S3-compatible artifact store, access control, and a coordination service without forcing the local product to depend on cloud infrastructure.
Take the next step
Build memory on evidence that survives the model session.
The architecture is deliberately incremental: the archive remains useful independently while each intelligence layer proves its value.