Sruja Architecture

Sruja Architecture Skill

Deterministic, evidence-first workflow for generating and maintaining repo.sruja (or architecture.sruja). The skill guides the AI to write valid architecture from real code evidence; no DSL expertise required.

Quick Start

1. Evidence — Prefer .sruja/author_evidence.json when present and recent; else run sruja sync -r . (do not ask the user to run it). Use .sruja/context.json only for the small fields needed for synthesis (communities, suggested_questions, architecture_digest) and avoid loading .sruja/graph.json unless the user explicitly asks for deep dependency debugging.
2. Synthesis (Mandatory) — Propose a domain architecture map from evidence. Communities are import-graph hints, not containers; merge/split/rename into domain boundaries.
3. Questions — Ask 2–5 targeted questions only when evidence is ambiguous (boundaries, externals, datastores, deployment, runtime/data flows).
4. Output (Never write truth by default) — Write either (a) .sruja/proposals/<id>.json (preferred, incremental) or (b) repo.sruja.working (greenfield / big bang). Do not write repo.sruja until the user explicitly says to promote.
5. Promote + Validate — For proposals: sruja propose approve <id> then sruja lint repo.sruja. For working DSL: user promotes by renaming/copying to repo.sruja then sruja lint.
6. Refine — Run sruja drift -r . and iterate; keep unknowns as open questions instead of guessing.
7. Impact (Optional) — Before refactoring code, run sruja impact <target> -r . --depth 3.

Workflow checklist: [ ] Evidence gathered → [ ] Synthesis done → [ ] Questions (if needed) → [ ] Proposal/working DSL written → [ ] Promotion path used → [ ] sruja lint passed → [ ] Open questions listed

Core Principles

• Evidence-first — Gather evidence before modeling.
• No guessing — Surface open questions instead of fabricating answers.
• Synthesis is required — Import clusters and manifests are facts; architecture boundaries and names are hypotheses backed by evidence.
• Minimal DSL — Generate only what evidence supports; prefer small, reviewable proposals.
• Validation — Always lint and fix errors before considering complete.

Evidence and Discovery

Discovery is backed by a static analysis graph (Tree-sitter): modules, imports, and dependencies from supported languages (TypeScript, Python, Go, Rust, Java, C#, Ruby, etc.). Use this graph as the single source of truth; prefer it over guessing.

Progressive loading (large or multi-repo):

If evidence is missing or stale, run sruja sync -r . yourself. After generating, suggest Sruja: Refresh repo context or sruja sync -r . for next time.

Evidence provides: repository structure, detected technologies, module boundaries, entry points, external dependencies, scan scope.

Workflow (Detail)

1. Collect Evidence

Prefer .sruja/author_evidence.json when present and recent. If missing or stale, run:

sruja sync -r .

Optional: sruja author evidence -r . (writes .sruja/author_evidence.json)

2. Ask Targeted Questions

Only when evidence is ambiguous:

• Main system boundaries? External services? Datastores? Deployment? Main data flows?

Do not ask about information already clear from evidence.

3. Generate Minimal DSL

Default output is a proposal or working DSL, not repo.sruja. C4 context and container levels first; component level only when evidence justifies it.

import { * } from 'sruja.ai/stdlib'

Person = person "Person"
ExternalAPI = system "External API" { description "Third-party or backend service" }

System = system "System" {
  Container = container "Container" {
    technology "Technology"
    description "Description"
  }
}

Person -> System.Container "Protocol"
System.Container -> ExternalAPI "HTTPS"

Component knowledge (opt-in): Ask the user if they want per-component markdown under .sruja/knowledge/. If yes: one file per component (e.g. .sruja/knowledge/PaymentService.md), add doc ".sruja/knowledge/<Id>.md" on the element, use references/KNOWLEDGE_TEMPLATE.md. Scope: containers and key systems; quality over coverage.

Doc references — markdown only: In .sruja files, doc must point to markdown (e.g. .sruja/knowledge/Component.md, docs/README.md, or a generated knowledge base). Do not set doc to source code paths (e.g. .rs, .ts, crates/...). The markdown itself may link to or reference code; the separation is: Sruja DSL elements reference markdown via doc, and that markdown can refer to files.

Sources (specs and infra artifacts): Use source <kind> "<path>" inside an element body to bind code-adjacent artifacts (OpenAPI specs, Kubernetes manifests, Terraform modules, docs) to the architecture element they describe. This keeps architecture grounded in real, reviewable files without requiring the user to write DSL.

Payments = container "Payments API" {
  technology "Go"
  description "Handles payment processing"

  source openapi "./specs/payments.yaml"
  source kubernetes "./k8s/payments/"
  source terraform "./infra/payments/"
  source docs "./docs/payments.md"
}

Supported kinds include: openapi, asyncapi, kubernetes (k8s), dockerfile (docker), terraform (tf), docs (doc), readme, proto (protobuf), config, graphql (gql), helm, and custom.

4. Validate and Repair

sruja lint repo.sruja

Fix all errors. Common: E201 (invalid kind), E204 (circular deps), E205 (orphans), E206 (invalid refs).

5. Refine (Optional)

sruja drift -r .

Use drift results to refine repo.sruja. Use -a path to specify a different architecture file.

Operating Modes

1. Local authoring — Gather .sruja/author_evidence.json (or sruja sync), synthesize a proposal or repo.sruja.working, promote to repo.sruja only when the user asks; run sruja lint after promotion.
2. System context — Prefer sruja_get_author_evidence (MCP) or author evidence file; use .sruja/context.json for drift/truth only—not full graph dumps.
3. Drift refinement — After code/intent changes, use sruja drift -r . (and optionally sruja intent propose) to turn deltas into DSL updates or open questions; do not invent without evidence.
4. Multi-repo — When system.index.json exists (from sruja compose), load only the impacted slice; use canonical IDs repo_id::local_id; check conflicts and ownership. See docs/FEDERATION.md.

When to Apply

• Discovering architecture from a new codebase
• Generating initial repo.sruja from requirements
• Refactoring or updating existing architecture
• Validating architecture against code
• Maintaining architecture documentation

What NOT to Do

• Do not guess missing information — list open questions.
• Do not add framework/domain/security narratives without evidence.
• Do not create components for completeness only.
• Do not add relationships unsupported by evidence.
• Do not skip linting.
• Do not set doc on elements to source code paths (e.g. crates/foo/src/lib.rs, extension/src/bar.ts). Use markdown only; the markdown can reference code.

Open Questions

Surface missing or unclear information explicitly in comments:

// OPEN QUESTIONS:
// - How is authentication implemented?
// - What message queue is used for async operations?
// - External API integrations not detected?

Evidence References

When synthesizing non-trivial boundaries, actors, externals, or runtime edges, include evidence references and unknowns:

• Proposal path: populate evidence_refs[] and open_questions[] on the Proposal JSON.
• Working DSL path: include // OPEN QUESTIONS: blocks when inventing actors/externals/flows, and list evidence refs in descriptions (do not rely on .sruja/graph.json as a first-turn input).

Multi-repo (Publish and Compose)

Setup: See FEDERATION_SETUP_GUIDE.md for step-by-step instructions.

Workflow:

1. Per-repo: Generate repo.sruja in each repository (same as single-repo workflow)
2. Publish: In each repo: sruja publish -r . -o repo.bundle.json
3. Collect: Copy bundles to shared location (rename to avoid collisions: api.repo.bundle.json)
4. Compose: sruja compose -i <dir> -o system.index.json builds unified graph with canonical IDs
5. Resolve conflicts: Check conflicts array for duplicate kind+label across repos

When working across repos:

• Use canonical IDs: repo_id::local_id (e.g., api-service::PaymentAPI)
• Load only impacted slice from system.index.json, not full graph
• Check ownership before modifying cross-repo relationships

Prompt for multi-repo setup:

Use sruja-architecture. Help me set up federation across multiple repos.
Guide me through: 1) generating repo.sruja in each repo, 2) publishing bundles,
3) composing system.index.json. Reference docs/FEDERATION_SETUP_GUIDE.md for steps.

See docs/FEDERATION.md for artifact schemas and retrieval behavior.

Progressive Discovery — What to Load When

Do not load references/AGENTS.md, references/REFERENCE.md, or entire rules/ at once. Load only what the task needs:

Auto-Capturing Knowledge from Conversations

When users share conventions, patterns, or workflows in conversation, automatically record them using the existing sruja_record_learning MCP tool.

Detect these patterns:

• "Always do X", "Never do Y", "We use Z for..."
• "When you see X, do Y", "For X, use Y"
• "The convention is...", "The pattern is..."
• "We follow X approach", "Our team does Y"

Capture using MCP tool sruja_record_learning:

{
  "context": "user convention: [brief description]",
  "hypothesis": "the convention/pattern shared",
  "outcome": "success",
  "guardrail_advice": "how to apply this"
}

Example:

{
  "context": "user convention: error handling pattern",
  "hypothesis": "Always use Result<T, E> with thiserror for error handling",
  "outcome": "success",
  "guardrail_advice": "Never use unwrap() in production code"
}

How it works:

1. User shares knowledge in conversation ("We always use X pattern")
2. You detect the knowledge-sharing pattern
3. You call sruja_record_learning MCP tool
4. Knowledge is stored in .sruja/agent_memory.json
5. Next sessions retrieve it via sruja_get_agent_learnings or sruja focus

This creates an auto-learning loop without explicit user action, using existing tools.

Related References

• Discovery and refinement: references/REFERENCE.md
• Modeling rules: rules/ (per-task; see table above)
• Prompts: references/PROMPTS.md; references/AGENTS.md only for full design/refactor
• Multi-repo setup: docs/FEDERATION_SETUP_GUIDE.md (step-by-step)
• Multi-repo reference: docs/FEDERATION.md (schemas, commands)

Prerequisites and Installation

CLI: Workflow uses sruja sync, sruja discover, sruja lint, sruja drift, and (optionally) sruja impact. Install a CLI that includes sync and discover:

curl -fsSL https://sruja.ai/install.sh | bash

Or from repo: cargo install --path crates/sruja-cli. Verify with sruja --help (must list sync and discover). If the CLI lacks these, the skill falls back to repo structure and codebase only; upgrade via the install script or repo build.

Extension (optional): Sruja extension — syntax, diagnostics, Sruja: Run validation, diagram preview, export to Markdown, Sruja: Refresh repo context. When installed, run Refresh once (or after big changes); the skill prefers .sruja/context.json.

Install skill:

npx skills add https://github.com/sruja-ai/sruja --skill sruja-architecture

Retrieval Order (Architecture-Aware Codegen/Review)

1. Local repo truth — repo.sruja (or architecture.sruja) after promotion
2. Fresh synthesis evidence — .sruja/author_evidence.json or MCP sruja_get_author_evidence; if missing/stale (git_commit / updated_at), run sruja sync -r .. Do not load .sruja/graph.json unless debugging dependencies.
3. Drift/truth — .sruja/context.json (violations, truth_status) or sruja status -r . --format json
4. Slice from system index — If system.index.json exists, load only impacted slice; use canonical IDs repo_id::local_id
5. Intent and contract refs — ADRs, intent files from repo or bundle

Prefer canonical IDs; do not invent when context is missing — ask or mark unknown.

Quick Start Prompt

Single repo:

Use sruja-architecture skill. Prefer `.sruja/author_evidence.json` (or run `sruja sync -r .`). Synthesize domain boundaries from communities (import hints, not containers). Write `.sruja/proposals/<id>.json` with evidence_refs and open_questions, or `repo.sruja.working` for greenfield—do not write repo.sruja until I say promote. After promotion: `sruja propose approve <id>` or copy working → repo.sruja, then `sruja lint` until clean.

Multi-repo:

Use sruja-architecture skill. Help me set up federation across multiple repos.
1) First, help me generate repo.sruja in each repository.
2) Then guide me through publishing bundles and composing system.index.json.
Reference docs/FEDERATION_SETUP_GUIDE.md for detailed steps.

Optional Cursor rule: "For architecture authoring, use .sruja/author_evidence.json or MCP sruja_get_author_evidence; run sruja sync -r . if stale. Avoid .sruja/graph.json unless debugging deps. Synthesize proposals with evidence_refs; promote to repo.sruja only on request."