Technical reference / Setsumei

Trace a conclusion back to the analyzed source.

Setsumei builds a source-derived component graph, measures structural importance and reports ranked candidates across security, performance, reliability and structural health. Inspect the source, detector evidence and relationships behind a result, then reuse the analysis in briefs, documentation and agent queries.

ILLUSTRATIVE DEPENDENCY VIEW / RELATIONSHIP STATE RESOLVED · PARTIAL · UNKNOWN

ENTRY_POINT + EVENT_HANDLERResolved through compiler and language-service evidence.

↓ RESOLVED RELATIONSHIPS

TARGETThe component under architectural review.

↓ METHOD + CONFIDENCE VISIBLE

BOUNDARY_APartial confidence is visible to the reviewer.

UNRESOLVED EDGEUnknown relationships are retained and visible.

FOUNDATIONShared foundation: code other components depend on.

Resolved Partial confidence Unknown retained

01 / Pipeline

Analyze once, then review, query or investigate.

SOURCE + CONFIGURATION

Choose the repository state, included projects and analysis configuration. Results carry run information; this is not a promise of immutable historical graph storage.

DETERMINISTIC ANALYSIS

Language-aware resolvers reconstruct components and cross-file relationships. Core analysis and static detection do not require an LLM or model API key; optional AI-assisted steps, such as investigation, call a language model. Reproducibility depends on matching source, configuration, toolchain and analysis version.

GRAPH + DETECTORS

Measure dependencies, reachability, subsystem structure, bottlenecks and risk propagation. Flag pattern-based and structural candidates, including cycles and excessive coupling.

PERSISTED ANALYSIS

Reports, the generated HTML GUI brief and MCP queries can use the analysis without an AI investigation step.

OPTIONAL INVESTIGATION

A configured model examines pre-identified candidates with relevant source and graph context. Its verdicts, reasoning and evidence are available when investigation has run.

02 / Relationship graph

Language-aware resolution with uncertainty attached.

Setsumei uses compiler and language-service toolchains appropriate to each analyzed language, with lower-confidence fallback strategies where needed. The graph connects code components across files. Relationships retain their resolution method and confidence; unresolved or external targets remain visible instead of appearing equally certain.

SOURCE ANALYSISVISIBLE UNKNOWN RELATIONSHIPSSNAPSHOT-BASED RESULTS

Six languages today; depth varies, so confirm fit.

Setsumei currently analyzes JavaScript, TypeScript, C#, Java, Go and Python. Depth varies by language and framework, so fit is confirmed per repository. Language and build versions, included projects, generated-source handling and dynamic behavior affect coverage. Confirm version-specific prerequisites for your evaluation.

Check repository fit Review local deployment

03 / Detection and priority

Find candidates and prioritize them in structural context.

Core detectors cover security, performance, reliability and structural health. Representative candidates include unsafe query construction, hardcoded secrets, queries inside loops, blocking operations, swallowed exceptions, circular dependencies and excessively coupled components. Available patterns and resolution vary by language and framework; these are review candidates, not a guarantee of a defect.

Per-repository baselines help identify unusual complexity and coupling. Structural importance and propagation put a candidate in context and help reviewers distinguish an isolated issue from one on a consequential path. Dead-code analysis identifies unreachable components with confidence levels and public-API awareness.

04 / Architecture and importance

Find the subsystems, shared foundations and important paths.

COMMUNITIES

Graph-based community detection groups components by their relationships, rather than folder layout alone. Use cohesion and related measures to inspect those boundaries. Optional model-assisted naming is separate from structural grouping.

IMPORTANCE

Consequence measures how much a component sets in motion. Foundation measures how much other code rests on it. Use these two axes to choose where to start reading and where to focus investigation.

PATHS

Entry points, structural bottlenecks and forward/backward dependency slices (what a component depends on and what depends on it) help orient engineers and examine possible change impact. A structural bottleneck is not proof of a runtime performance bottleneck.

05 / Optional investigation

Investigate candidates with supporting context.

INPUT

A pre-identified candidate, relevant source, graph neighborhood and detector evidence from static analysis.

PROCESS

Assess whether the candidate is a meaningful security, performance, reliability or structural issue, a false positive, or still unresolved.

RECORD

Retain the recorded verdict, reasoning and evidence for review. Inspect which analysis and investigation run the information describes.

REVIEW

Engineers decide what to change and how to verify it. Investigation supplies evidence and recommendations; it does not implement an approval gate.

06 / Outputs and reuse

Use the analysis in the form the task needs.

Local GUI brief

Open the scan’s generated HTML to review the queue, finding detail, structural views and supporting context in a browser.

Reports and source evidence

Inspect relationships, candidates, available investigation outcomes, source references and analysis coverage.

Architecture documentation

Generate a structural architecture document with subsystem detail and a Mermaid dependency diagram. Structural-only generation is available; optional model-written descriptions are checked against structural data.

Coding-agent queries

Let a compatible agent query stored relationships, findings, subsystem context and source through read-only local MCP.

07 / Runtime and storage

Local execution and customer-controlled results.

ANALYZE

Your engineers run source analysis through the CLI in the local or on-premises environment they control. Use the package’s version-specific prerequisites.

STORE

The analysis persists in a customer-managed PostgreSQL store. Your organization manages the data, access and retention; when your team runs Setsumei, Syntaxa does not receive or access it.

VIEW

The scan generates HTML which opens the local GUI scan brief. Opening that HTML is separate from connecting an MCP client to the analysis store.

MCP

Current MCP runs locally, per developer, with one repository per server instance and the analysis on the same machine and PostgreSQL store. The client’s model and independent tools have their own permissions and data handling.

Optional investigation and model-written descriptions use the provider configured for that workflow. Review that provider and any connected agent’s data handling separately from local analysis. Review security and data handling.

08 / Refresh and freshness

Refresh the model after relevant changes.

Analysis is run on demand; it is not a continuously updating graph. Incremental analysis reparses changed files and affected dependents. MCP freshness checks can identify analyzed files that have changed. Check the run and freshness information before reusing an answer.

Dependency completeness after incremental analysis is best effort: deep call-only chains can differ from a clean rebuild. Run periodic full analysis when dependency and architecture completeness matters.

09 / Limits and provenance

Inspect coverage, query limits and provenance.

Dynamic behavior, generated code, external dependencies, exclusions and incomplete resolution can limit conclusions. Resolution coverage is not an accuracy or safety score. Static evidence does not establish runtime behavior, business intent or release safety.

Source changed? Check freshness and refresh the analysis after relevant edits.
Analysis in progress? Some queries return results with warnings while stricter queries may be unavailable. When completeness matters, wait for completion and re-query.
Before acting: inspect query bounds or truncation, unresolved relationships and run provenance. Graph tables update in place, without immutable snapshot isolation; investigation findings may come from a different run than the structural analysis.

Discuss an evaluation Explore MCP workflows and limits