backendpro 0.6.1

Backend Pro Max — BM25-searchable backend & distributed-systems engineering intelligence as an AI skill / CLI.

🚀 Backend Pro Max

A staff-engineer-in-a-box for your AI coding assistant

Curated, BM25-searchable backend & distributed-systems intelligence across 34 domains and 12 language stacks — every row sourced, auditable, and freshness-tracked. Drop it into Claude Code, Cursor, Windsurf, GitHub Copilot, Gemini, Continue, or any AI assistant.

Quick Start · Domains · Stacks · Install as a Skill · Examples · Contributing

---

✨ What is this?

Backend Pro Max grounds your AI coding assistant in opinionated, source-citable, senior-engineer-grade knowledge for backend & distributed-systems work — and forces it to search before answering.

LLMs know surface-level facts about backend tech, but they:

• 🎯 Recommend the trendy pattern instead of the right one for your team / scale.
• ⏱️ Forget timeouts, retries, idempotency, backpressure, and graceful shutdown.
• 🧩 Don't know your stack's idioms — Spring lazy-init pitfalls, FastAPI sync-in-async, Express vs Fastify, sqlx compile-time queries, EF Core change tracking, …
• 🔀 Mix up consistency models, replication modes, and partition strategies.
• 🛡️ Skip the boring-but-critical stuff: SLOs, error budgets, runbooks, PII in logs.

This skill fixes that with a structured, searchable knowledge base that the model is instructed to consult — so its advice cites a row, not a vibe.

---

🎁 What you get

📚 34 domain knowledge bases	Languages · Patterns · Databases · Messaging · Cache · Cloud · IaC · Containers · Observability · API · Auth · Security · CI/CD · Testing · Architecture · Scaling · Consistency · Performance · Reliability · Data · Anti-patterns · Cost · Migration · Incident · Capacity · Compliance · Multi-tenant · Release · ML Platform · Edge · Mobile Backend · API Contract · Interview · Latency Numbers
🛠️ 12 stack guidelines	Go · Java/Spring · Python/FastAPI · Node/Express · Rust/Axum · C#/ASP.NET · Kotlin/Spring · Scala/Akka · Elixir/Phoenix · Ruby/Rails · PHP/Laravel · C++
🔎 Pure-Python BM25 + synonyms	No installs, no models, no network — partial failure → finds Saga / Circuit Breaker automatically
⚖️ compare mode	backendpro compare "Kafka" "RabbitMQ" --domain messaging → side-by-side markdown table for ADRs
💬 Interactive REPL	backendpro --interactive for design sessions: /d, /s, /all, /cmp, /stale
📅 Freshness tracking	Last Updated column + --max-age-months filter + --stale audit mode
🎯 Confidence scores	Every result carries a BM25 score + high/medium/low label so the agent knows when to trust
⚡ mtime-cached index	Sub-millisecond repeat queries — agent loops stay snappy
🧠 Intent classifier	Auto-detects query intent (comparison, troubleshoot, migration, incident, definition, best-practice, checklist) and applies a structured template to the output
🔀 Hybrid retrieval	pip install backendpro[semantic] adds embedding-based search via sentence-transformers + RRF fusion with BM25 — graceful fallback to pure BM25 when not installed
🏆 Cross-encoder re-ranking	pip install backendpro[rerank] adds cross-encoder re-ranking for precision-critical queries — optional, graceful fallback
⚠️ Anti-patterns domain	15 common distributed-systems anti-patterns (Distributed Monolith, God Service, Dual Writes, …) with symptoms, root causes, and fixes
🤖 Drop-in skill files	SKILL.md for Claude Code · skill-content.md for Cursor / Windsurf / Copilot / Gemini / Continue
📐 Do / Don't + Code examples	Each row contains good vs bad code, severity, and a docs URL
🧠 Auto domain detection	Skip --domain and the engine picks the right CSV from your query
⚙️ JSON output mode	First-class integration with tool-calling agents and MCP servers
🔌 MCP server	pip install backendpro[mcp] → 8 tools on stdio, works with Claude Desktop, Cline, Cursor, Zed
📎 Citation tokens	Every result carries [BPM:domain.slug] — greppable provenance for PR review
📡 JSONL streaming	--jsonl for agent loops that consume results incrementally
✅ CI-enforced	backendpro-validate schema-checks every CSV; 390 pytest cases run on Py 3.9 / 3.11 / 3.12
🔗 100% source citations	Every row carries Source URL, Source Type, Last Updated — official docs, RFCs, papers, OWASP. --strict mode fails on gaps
⚠️ Conflict detector	backendpro conflicts surfaces 12 architectural tensions (retry vs latency, cache vs consistency, etc.) with citation tokens
🕐 Auto-freshness audit	Weekly GitHub Action flags stale rows (>18mo) and broken URLs. --check-urls for local runs
📜 Provenance tracking	Git-blame-based Added By + Version columns. --show-provenance flag in search output
🔍 backendpro lint	Scans source files for 18 backend anti-patterns (missing timeouts, sync-in-async, secrets in .env, SQL injection, etc.). Human, JSON, and SARIF output. Pre-commit hook + GitHub Action included
📤 backendpro export	Export the entire KB to Obsidian (wikilinks + MOC), Notion CSV, or Org-mode. --domain filter
🧠 backendpro learn	Spaced-repetition flashcards (SM-2 algorithm) from the KB. --domain, --daily N, --stats, --reset
💻 VS Code extension	Search, Explain Selection (right-click), CodeLens stack guidelines. CLI or MCP mode. extensions/vscode/
🌐 Web playground	backendpro.cc — search-as-you-type, domain filter, permalinks, compare view. Zero backend

---

⚡ Quick start

Option A — install once, type backendpro

# Install from PyPI — pure stdlib, zero dependencies
pip install backendpro

backendpro --list
backendpro "kafka exactly once delivery"
backendpro "circuit breaker" --domain pattern
backendpro "virtual threads" --stack java-spring
backendpro "idempotency" --all

# v0.2 power tools
backendpro compare "Kafka" "RabbitMQ" --domain messaging   # ADR-style table
backendpro --interactive                                    # REPL for design sessions
backendpro --stale --domain pattern --max-age-months 18     # freshness audit
backendpro "redis cluster" --json                           # MCP / agent-friendly
backendpro-validate                                         # schema-check every CSV

💡 pipx install backendpro works too if you prefer an isolated venv. You can also install from source: pip install git+https://github.com/shashankswe2020-ux/backend-pro-max-skill

Option B — run the script directly (no install)

# 0. No install needed — pure Python 3.8+ stdlib
python3 src/backend-pro-max/scripts/search.py --list

# 1. Auto-detect the domain from the query
python3 src/backend-pro-max/scripts/search.py "kafka exactly once delivery"

# 2. Constrain to a specific domain
python3 src/backend-pro-max/scripts/search.py "circuit breaker" --domain pattern

# 3. Stack-specific guidance
python3 src/backend-pro-max/scripts/search.py "virtual threads" --stack java-spring

# 4. Cross-domain search
python3 src/backend-pro-max/scripts/search.py "idempotency" --all

# 5. JSON output (great for agents / MCP)
python3 src/backend-pro-max/scripts/search.py "redis cluster" --json

💡 Tip: the search engine ranks results with BM25 over the search columns of each CSV, with light keyword-based domain auto-detection when --domain is omitted.

---

🧭 Example queries

A taste of what to ask — these all return ranked, citable rows:

Domain	Try this
🧬 pattern	"saga vs 2pc for distributed transactions"
🗄️ database	"postgres index on jsonb" · "dynamodb single table design"
📨 messaging	"kafka exactly once" · "sqs vs sns vs eventbridge"
⚡ cache	"thundering herd" · "negative caching with redis"
☁️ cloud	"aws gcp azure equivalent of pubsub"
🛰️ observability	"slo error budget alerting" · "otel trace context propagation"
🔐 security	"ssrf prevention" · "sigstore supply chain"
🧪 testing	"contract testing pact" · "testcontainers postgres"
🏗️ architecture	"modular monolith vs microservices"
📈 scaling	"hedged requests" · "backpressure"
⚖️ consistency	"linearizability vs sequential" · "PACELC"
🛡️ reliability	"graceful shutdown" · "circuit breaker timeouts"

---

🧠 Decision Intelligence (v0.3)

Three new commands — decide, adr, design — turn the search engine into a decision advisor. Instead of returning raw rows, they extract constraints from your query, score candidates, and produce structured output an LLM can cite directly.

# Quick decision with trade-off analysis
backendpro decide "Kafka vs Pulsar vs Pub/Sub for high-throughput event streaming on GCP"

# Architecture Decision Record (Markdown)
backendpro adr "Redis vs Memcached for session cache on AWS"

# Capacity-aware design document
backendpro design "Postgres vs DynamoDB for 50M DAU e-commerce with strong consistency"

Why does this matter?

Dimension	Plain LLM	Backend Pro Max + LLM
Candidates	Mentions 2–3 obvious options	Returns 5 ranked candidates from the knowledge base, including less-obvious picks
Constraints	Ignores or hallucinates requirements	Extracts facets (throughput:high, cloud:gcp, consistency:strong) and scores every candidate against them
Trade-offs	Vague "it depends"	Structured pro/con per candidate with severity ratings
Citations	None	Every recommendation cites domain:key rows the reviewer can verify
Reproducibility	Different answer each time	Deterministic BM25 ranking — same query, same results

5 hard demos

These are real outputs from the decide command — the kind of nuanced comparisons that catch LLMs off-guard:

1. Kafka vs Pulsar vs Pub/Sub — decide "Kafka vs Pulsar vs Pub/Sub for high-throughput event streaming on GCP with ordering guarantees"

graph LR
    Q["🔍 Query: Kafka vs Pulsar vs Pub/Sub<br/>high-throughput · GCP · ordering"] --> E["⚙️ Constraint Extraction"]
    E --> C1["throughput: high"]
    E --> C2["cloud: gcp"]
    E --> C3["ordering: true"]
    C1 & C2 & C3 --> S["📊 Score & Rank"]
    S --> R1["🥇 Apache Pulsar<br/>2/2 constraints ✅"]
    S --> R2["🥈 Apache Kafka<br/>2/2 constraints ✅"]
    S --> R3["🥉 Google Pub/Sub<br/>2/2 constraints ✅"]
    S --> R4["4. Redpanda"]
    S --> R5["5. NATS JetStream"]

    style R1 fill:#22d67a,color:#000,stroke:#16a34a
    style R2 fill:#5b9cf6,color:#000,stroke:#3b82f6
    style R3 fill:#f5c842,color:#000,stroke:#d97706

• Recommendation: Apache Pulsar (highest combined score)
• 5 candidates returned across messaging domain
• Constraints extracted: throughput:high, cloud:gcp, ordering:true
• Top 3 all score 2/2 on constraint match — decision comes down to BM25 relevance to the specific query
• 🔑 Plain LLM would say "use Pub/Sub because you're on GCP" — misses that Pulsar and Kafka score equally well on constraints and may be better fits for ordering guarantees

2. Redis vs Memcached — decide "Redis vs Memcached for low-latency session cache on AWS"

graph LR
    Q["🔍 Query: Redis vs Memcached<br/>low-latency · session cache · AWS"] --> E["⚙️ Constraint Extraction"]
    E --> C1["cloud: aws"]
    E --> C2["throughput: high"]
    E --> C3["latency: low-ms"]
    C1 & C2 & C3 --> S["📊 Score & Rank"]
    S --> R1["🥇 Redis<br/>3/3 constraints ✅"]
    S --> R2["🥈 Redis Cluster<br/>3/3 constraints ✅"]
    S --> R3["🥉 Memcached<br/>3/3 constraints ✅"]
    S --> R4["4. In-process LRU"]
    S --> R5["5. L1+L2 Hybrid"]

    style R1 fill:#22d67a,color:#000,stroke:#16a34a
    style R2 fill:#5b9cf6,color:#000,stroke:#3b82f6
    style R3 fill:#f5c842,color:#000,stroke:#d97706

• Recommendation: Redis (highest score)
• 5 candidates across cache + database domains
• Constraints extracted: cloud:aws, throughput:high, latency:low-ms
• Redis scores 3/3 on all constraint columns
• 🔑 Plain LLM gives a generic feature comparison. Backend Pro Max surfaces that Redis wins on every extracted constraint axis, and also pulls in database-domain rows for Redis Cluster that an LLM wouldn't think to include.

3. Monolith vs Microservices — decide "monolith vs microservices for a 10-person startup"

graph LR
    Q["🔍 Query: monolith vs microservices<br/>10-person startup"] --> E["⚙️ Constraint Extraction"]
    E --> C0["No constraint columns<br/>on architecture domain"]
    C0 --> S["📊 Pure BM25 Ranking"]
    S --> R1["🥇 Monolith<br/>Best relevance match"]
    S --> R2["🥈 Modular Monolith<br/>Strong #2"]
    S --> R3["🥉 Microservices"]
    S --> R4["4. Serverless/FaaS"]
    S --> R5["5. Hexagonal Architecture"]

    style R1 fill:#22d67a,color:#000,stroke:#16a34a
    style R2 fill:#5b9cf6,color:#000,stroke:#3b82f6
    style R3 fill:#f5c842,color:#000,stroke:#d97706

• Recommendation: Monolith (highest BM25 score)
• 5 candidates across architecture + patterns domains
• No constraint columns on architecture domain — pure relevance ranking
• 🔑 Plain LLM hedges with "it depends on your team." Backend Pro Max commits to monolith as #1, returns modular monolith at #2, and includes the specific trade-off rows (coupling, deployment, team autonomy) that justify the ranking.

4. Postgres vs DynamoDB — decide "Postgres vs DynamoDB for 50M DAU e-commerce with strong consistency on multi-cloud"

graph LR
    Q["🔍 Query: Postgres vs DynamoDB<br/>50M DAU · strong consistency · multi-cloud"] --> E["⚙️ Constraint Extraction"]
    E --> C1["throughput: high"]
    E --> C2["cloud: multi-cloud"]
    E --> C3["consistency: strong"]
    C1 & C2 & C3 --> S["📊 Score & Rank"]
    S --> R1["🥇 Spanner / TiDB<br/>3/3 constraints ✅"]
    S --> R2["🥈 CockroachDB<br/>3/3 constraints ✅"]
    S --> R3["🥉 Postgres<br/>2/3 constraints ⚠️"]
    S --> R4["4. DynamoDB<br/>1/3 constraints"]
    S --> R5["5. Cassandra"]

    style R1 fill:#22d67a,color:#000,stroke:#16a34a
    style R2 fill:#5b9cf6,color:#000,stroke:#3b82f6
    style R3 fill:#f5c842,color:#000,stroke:#d97706
    style R4 fill:#f56060,color:#000,stroke:#dc2626

• Recommendation: Spanner / TiDB (surprise — neither of the two named!)
• 5 candidates across database + consistency + scaling domains
• Constraints extracted: throughput:high, cloud:multi-cloud, consistency:strong
• Spanner scores 3/3; Postgres lands at #3 with 2/3 (not cloud-native multi-cloud)
• 🔑 This is the killer demo. A plain LLM would compare only Postgres and DynamoDB. Backend Pro Max widens the search and surfaces Spanner/TiDB as a better fit for the actual constraints — exactly what a staff engineer would do in a design review.

5. REST vs GraphQL — decide "REST vs GraphQL for a public API with low latency"

graph LR
    Q["🔍 Query: REST vs GraphQL<br/>public API · low latency"] --> E["⚙️ Constraint Extraction"]
    E --> C1["latency: low-ms"]
    C1 --> S["📊 Score & Rank"]
    S --> R1["🥇 REST<br/>api domain"]
    S --> R2["🥈 GraphQL<br/>api domain"]
    S --> R3["🥉 Tail Latency<br/>⚡ performance domain"]
    S --> R4["4. gRPC"]
    S --> R5["5. Cold Starts<br/>⚡ performance domain"]

    S -.->|"cross-domain"| PD["⚡ performance"]
    S -.->|"primary"| AD["🔌 api"]

    style R1 fill:#22d67a,color:#000,stroke:#16a34a
    style R2 fill:#5b9cf6,color:#000,stroke:#3b82f6
    style R3 fill:#f5c842,color:#000,stroke:#d97706
    style PD fill:#fb8c3e,color:#000,stroke:#d97706
    style AD fill:#a78bfa,color:#000,stroke:#7c3aed

• Recommendation: Based on API-domain BM25 ranking
• 5 candidates across api + performance domains
• Constraints extracted: latency:low-ms
• Pulls in tail-latency and performance rows that a pure API comparison would miss
• 🔑 Plain LLM gives a feature matrix. Backend Pro Max crosses into the performance domain and flags tail-latency concerns for GraphQL that most comparisons overlook.

Note: Constraint scoring works best on domains with constraint columns (databases, messaging, cache). Other domains use pure BM25 relevance. More constraint columns are planned for future tiers.

---

🧬 How is this different from RAG + LLM?

At first glance Backend Pro Max looks like a RAG system — it retrieves knowledge and feeds it to a model. But the architecture is fundamentally different, and the difference matters.

The standard RAG pipeline

User query ──▶ Embed ──▶ Vector DB ──▶ Top-K chunks ──▶ LLM ──▶ Answer

RAG embeds your docs into a vector store, retrieves the nearest chunks, and stuffs them into the prompt. The LLM then generates an answer from those chunks. This works, but has well-known failure modes:

Problem	What happens
Chunk boundary corruption	A CSV row about "Saga pattern" gets split mid-sentence; the LLM sees half a trade-off
Semantic drift	"connection pool exhausted" embeds near "swimming pool" — wrong chunks retrieved
No structure in, no structure out	Chunks are flat text; the LLM has to infer which part is "when to use" vs "when not to"
Hallucination laundering	The LLM cites a retrieved chunk but subtly changes the recommendation
Non-determinism	Same query, different embedding model version → different chunks → different answer
Cold start cost	Embedding 21 CSVs requires a model download, GPU/CPU time, and a vector DB
Dependency weight	sentence-transformers + faiss + chromadb = 500 MB+ before you write a line of code

The Backend Pro Max approach

User query ──▶ BM25 (stdlib) ──▶ Structured rows ──▶ Agent context
                                  (not chunks)        (not generation)

Design choice	Why
BM25 over structured CSV rows	Each row is a complete, self-contained knowledge unit with named columns (Name, When to Use, When NOT to Use, Trade-offs, Severity). No chunk splitting, no boundary corruption.
Deterministic ranking	Same query → same BM25 scores → same results. Every time. Reviewable, testable, reproducible.
Zero dependencies by default	Pure Python stdlib. No embedding model, no vector DB, no GPU. pip install backendpro and you're done in 2 seconds.
Intent classification, not generation	The classifier routes your query to a structured template (troubleshoot → Symptom/Root Cause/Fix). The model fills in from retrieved rows — it doesn't hallucinate a structure.
Constraint columns, not vibes	Throughput Tier: very-high, Latency Tier: sub-ms, Cloud Native: aws,gcp — structured metadata that the decide command scores programmatically. No embedding similarity guesswork.
The model is the consumer, not the generator	Retrieved rows go into the agent's context as facts to cite. The model's job is to explain and compare them — not to generate knowledge from noisy chunks.
Optional semantic upgrade	pip install backendpro[semantic] adds embedding-based hybrid search for conceptual queries. But it's an addition to BM25, not a replacement — and BM25 alone covers 95% of queries.

Side by side

	RAG + LLM	Backend Pro Max
Retrieval	Embedding similarity (approximate)	BM25 keyword match (exact) + optional hybrid
Unit of retrieval	Text chunk (arbitrary boundary)	Structured row (complete knowledge unit)
Ranking	Cosine similarity (non-deterministic across model versions)	BM25 score (deterministic, testable)
Output structure	LLM-generated (variable)	Template-driven by intent (consistent)
Constraint matching	None (hope the LLM reads the chunk correctly)	Programmatic column-level scoring
Cold start	Download model + build index (minutes)	Zero (pure stdlib, sub-ms first query)
Dependencies	500 MB+ (transformers, vector DB, numpy)	0 (stdlib only)
Reproducibility	Low (embedding drift, LLM temperature)	High (same query = same rows = same scores)
Auditability	"The LLM said so"	"Row pattern:Saga scored 8.42 on BM25"

What actually breaks vs what this fixes

Problem in RAG + LLM	What Backend Pro Max does
Different answers every run	Same input → same decisions
Loses architectural context	Persists system decisions
Retrieves noisy chunks	Retrieves complete decision units
LLM hallucinates tradeoffs	Explicit scoring + constraints
No explainability	Every decision is traceable

When you'd still want RAG

• You have unstructured documents (PDFs, wikis, Confluence) that can't be easily tabulated into CSV rows with named columns.
• Your knowledge base is huge (100K+ documents) and keyword matching produces too many false negatives for conceptual queries.
• You need cross-document reasoning where the answer spans multiple sources.

Backend Pro Max is designed for curated, structured, expert-reviewed knowledge — the kind a staff engineer would put in a decision matrix, not in a wiki page. For that use case, structured retrieval beats RAG on every dimension that matters: precision, reproducibility, auditability, and speed.

---

📚 Domains

Domain	What's in it
🧠 language	Go, Java, Kotlin, Python, Rust, Node.js/TS, C#, Scala, Elixir, Ruby, PHP, C++
🧩 pattern	Saga, CQRS, Event Sourcing, Outbox, CDC, Circuit Breaker, Bulkhead, Retry, Idempotency, Leader Election, Sidecar, Strangler Fig, ACL, BFF, API Gateway, Rate Limiting, Sharding, Read Replica, Materialized View, Process Manager, Outbox+Inbox, Fan-out / Scatter-Gather
🗄️ database	Postgres, MySQL/Vitess, CockroachDB, Spanner/TiDB, MongoDB, Cassandra/Scylla, DynamoDB, Redis, Memcached, Elastic/OpenSearch, ClickHouse, DuckDB, Snowflake/BigQuery/Redshift, Neo4j/Memgraph, Influx/Timescale/VictoriaMetrics, vector DBs, S3/GCS/Blob, etcd/ZK/Consul, SQLite
📨 messaging	Kafka, Redpanda, Pulsar, RabbitMQ, NATS/JetStream, MQTT, SQS, SNS/EventBridge, Kinesis, Pub/Sub, Service Bus / Event Grid / Event Hubs, ZeroMQ
⚡ cache	In-process LRU, Redis (single + cluster), Memcached, CDN, HTTP cache, read/write/write-back, materialized views, negative caching, Bloom filters, L1+L2 hybrid
☁️ cloud	AWS / GCP / Azure / Cloudflare service mapping & equivalents
🏗️ iac	Terraform/OpenTofu, Pulumi, AWS CDK, CloudFormation, Bicep, Ansible, Crossplane, Helm, Kustomize, Packer
📦 container	Docker/OCI, Podman, containerd, Kubernetes, EKS/GKE/AKS, Helm, Kustomize, ArgoCD/Flux, Istio/Linkerd/Cilium, Envoy, Karpenter, Nomad, Compose, Testcontainers
📊 observability	Prometheus, Mimir/Cortex/Thanos/VM, Grafana, Loki, ELK/OpenSearch, Tempo/Jaeger/Zipkin, OpenTelemetry, Pyroscope/Parca, Datadog, New Relic / Honeycomb / Dynatrace, Sentry, Fluent Bit / Vector, PagerDuty / Opsgenie, SLO frameworks
🔌 api	REST, GraphQL, gRPC, gRPC-Web/Connect, WebSocket, SSE, HTTP/2, HTTP/3, Webhooks, WebSub/ActivityPub, JSON-RPC, SOAP
🔑 auth	OAuth 2.0 + PKCE, OIDC, JWT, SAML, mTLS, API keys, HMAC signing, sessions, passkeys/WebAuthn, magic links, RBAC/ABAC/ReBAC, SCIM, workload identity (IRSA / WIF)
🛡️ security	OWASP Top 10, CSRF, XSS, SSRF, deserialisation, secrets, supply chain (SLSA, Sigstore), zero-trust, TLS hardening, PII/logging, rate limiting, CORS, SBOM, SAST, DAST/fuzz
🔁 cicd	GitHub Actions, GitLab CI, Jenkins, CircleCI, Buildkite, Drone, Tekton, Argo Workflows, ArgoCD, Flux, Spinnaker, Argo Rollouts, Renovate/Dependabot, SonarQube, GHAS
🧪 testing	Unit, component/slice, integration (Testcontainers), contract (Pact), E2E, property-based, fuzz, snapshot, mutation, load, stress/soak, chaos, smoke / synthetic monitoring
🏛️ architecture	Monolith, modular monolith, microservices, serverless/FaaS, event-driven, hexagonal/ports-and-adapters, clean/onion, DDD, CQRS+ES, service mesh, BFF, lambda/kappa, actor model, cell-based
📈 scaling	Vertical, horizontal, autoscaling (HPA/KEDA/Karpenter), sharding, read replicas, multi-tier caching, connection pooling, backpressure, bulkhead, hedged requests, load balancing, CDN, geo-distribution, async/queue load levelling, indexing, materialized views, partitioning
⚖️ consistency	Linearizability, sequential, causal, read-your-writes, eventual, SEC/CRDTs, CAP, PACELC, Raft, Paxos, 2PC, snapshot isolation/SSI, quorum, Lamport/vector/HLC clocks
🚀 performance	N+1, missing indexes, plan regressions, pool exhaustion, GC pauses, hot keys, tail latency, thundering herd, async-blocking, cold starts, leaks, hot-path allocations, JSON serialisation, chatty interfaces, TLS overhead
🛟 reliability	SLO/SLI/error budget, timeouts, retries+backoff, circuit breaker, bulkhead, idempotency, graceful shutdown, liveness/readiness, capacity & headroom, RPO/RTO, multi-AZ/region, backups + PITR, chaos engineering, runbooks, blue/green & canary, feature flags, per-tenant quotas, postmortems
🧮 data	Spark, Flink, Kafka Streams/ksqlDB, Airbyte/Fivetran/Stitch/Meltano, dbt, Airflow, Dagster, Prefect, Iceberg/Delta/Hudi, ClickHouse/Druid/Pinot, Spark Streaming + Delta, Debezium, Kafka Connect, LakeFS/Nessie, vector DBs, feature stores
⚠️ antipattern	Distributed Monolith, Shared Database Integration, God Service, Sync-over-Async, Dual Writes, Chatty Microservices, Unbounded Retry, Missing Idempotency Key, Premature Microservices, Log-and-Throw, Generic Error Swallowing, N+1 Query, Secrets in Env Vars, Time-Based Cache Invalidation Only, Polling Instead of Events

---

🛠️ Stacks

Each stack file contains tight, opinionated, "what would a staff engineer say in code review" guidelines — categorised by Concurrency, HTTP, Errors, Persistence, Tooling, Observability, Performance, Testing, Build, … — with ✅ Do / ❌ Don't plus good vs bad code examples.

Stack	Highlights
🐹 go	context.Context, errgroup, http.Client reuse, pgx/sqlc, table-driven tests
☕ java-spring	Virtual threads (Loom), constructor DI, OSIV off, Flyway, Testcontainers, native image
🐍 python-fastapi	async-all-the-way, Pydantic v2, httpx, uv, ruff/mypy, structlog, Testcontainers
🟢 nodejs-express	Fastify > Express, zod at boundaries, Undici pool, pino, OTel, Vitest
🦀 rust-axum	Tokio + Axum + Tower, sqlx compile-time queries, thiserror/anyhow, tracing, tokio-console
🟪 csharp-aspnet	Minimal APIs, async-all-the-way, HttpClientFactory, Polly v8, EF Core AsNoTracking, Native AOT
🟧 kotlin-spring	Coroutines + structured concurrency, Spring Boot Kotlin DSL, Exposed/jOOQ, kotest
🔺 scala-akka	Pekko (Akka fork), Typed actors, Pekko Streams, Cats Effect / ZIO
💧 elixir-phoenix	OTP supervision, GenServer, Task.async_stream, Phoenix LiveView, Broadway, libcluster
💎 ruby-rails	Modular Rails (Packwerk), Sidekiq, Puma tuning, Bullet, Rails 7+ defaults, Solid Queue/Cache
🐘 php-laravel	Octane (Swoole/RoadRunner/FrankenPHP), OPcache+JIT, Horizon, eager loading, PHPStan
➕ cpp	C++20+, RAII, jthread/stop_token, coroutines, sanitizers, CMake presets, Conan/vcpkg, GoogleTest, clang-tidy

---

🤖 How AI agents use this

A typical interaction inside Claude Code, Cursor, Copilot, etc.:

👤  "Add retries to our outbound HTTP client without melting the dependency."

🤖  → search.py "retry backoff jitter circuit breaker" --domain reliability
    → search.py "http client retries" --stack <your stack>
    → answers with: exponential backoff + jitter, max attempts, idempotency
      key requirement, circuit breaker around it, budgeted timeout, plus
      a code snippet using the right library for your stack — and cites
      the row(s) it pulled from.

The skill files (SKILL.md / skill-content.md) instruct the agent to:

1. Search first — never guess when a row exists.
2. Cite the row (domain + key) so reviewers can verify.
3. Prefer stack guidelines for code-shaped answers.
4. Combine multiple domains for cross-cutting concerns (e.g. a "saga" answer pulls from pattern + messaging + consistency + reliability).

---

📁 Repository structure

See CLAUDE.md for the full layout. TL;DR:

src/backend-pro-max/
├── data/                       # 34 domain CSVs + stacks/ (12 stack CSVs)
│   ├── languages.csv  patterns.csv  databases.csv  messaging.csv …
│   └── stacks/
│       └── go.csv  java-spring.csv  python-fastapi.csv  …
├── scripts/
│   ├── core.py                 # BM25 engine + domain auto-detection
│   ├── search.py               # CLI entry point
│   ├── lint.py                 # Anti-pattern linter (18 rules)
│   ├── export.py               # Obsidian / Notion / Org-mode export
│   ├── learn.py                # SM-2 spaced-repetition flashcards
│   ├── calc.py                 # Back-of-envelope calculators
│   ├── conflicts.py            # Architectural tension detector
│   ├── freshness.py            # Stale-row & URL-health checker
│   └── provenance.py           # Git-blame provenance populator
└── templates/base/
    ├── skill-content.md        # Drop-in rules for any AI assistant
    └── quick-reference.md      # Cheatsheet

extensions/
├── vscode/                     # VS Code extension (TypeScript)
│   ├── src/extension.ts        # Commands, CodeLens, MCP client
│   └── package.json            # Marketplace metadata + settings
└── jetbrains/                  # JetBrains plugin stub (Java)
    └── src/main/resources/META-INF/plugin.xml

lint-rules.yml                  # 18 lint rules with BPM citations
.pre-commit-hooks.yaml          # Pre-commit framework integration
.github/actions/lint/action.yml # GitHub Action for SARIF upload
.claude/skills/backend-pro-max/ # Claude Code skill (SKILL.md)
docs/                           # ARCHITECTURE.md & USAGE.md

Visual architecture

flowchart TD
    U["👤 User Query\n"Design a URL shortener with caching""]:::user
    S["📜 SKILL.md / skill-content.md\nInstructs model to search before answering"]:::skill
    C["🔎 backendpro CLI\nBM25 search engine · pure Python stdlib"]:::cli
    D["📚 21 Domain CSVs\napi · cache · database\nscaling · reliability …"]:::data
    K["🛠️ 12 Stack CSVs\ngo · java-spring\npython-fastapi …"]:::data
    A["🌐 Auto-detect / --all\nCross-domain search"]:::data
    R["📋 Ranked Results\nCited rows · do/don't · code\nseverity · docs URL"]:::result
    G["✅ Grounded, Citable Answer"]:::answer

    U --> S --> C
    C --> D & K & A
    D & K & A --> R --> G

    classDef user     fill:#6366f1,color:#fff,stroke:#4f46e5,stroke-width:2px
    classDef skill    fill:#8b5cf6,color:#fff,stroke:#7c3aed,stroke-width:2px
    classDef cli      fill:#0ea5e9,color:#fff,stroke:#0284c7,stroke-width:2px
    classDef data     fill:#f59e0b,color:#fff,stroke:#d97706,stroke-width:2px
    classDef result   fill:#10b981,color:#fff,stroke:#059669,stroke-width:2px
    classDef answer   fill:#22c55e,color:#fff,stroke:#16a34a,stroke-width:2px

---

🔌 Installation as an AI skill

🟣 Claude Code

Symlink (or copy) the src/backend-pro-max directory into your repo at .claude/skills/backend-pro-max/ — the SKILL.md already lives there. The agent will discover it automatically.

mkdir -p .claude/skills
ln -s "$(pwd)/src/backend-pro-max" .claude/skills/backend-pro-max

🟦 Cursor / Windsurf / Continue / GitHub Copilot / Gemini

Copy src/backend-pro-max/templates/base/skill-content.md into your editor's rules file:

Tool	Rules file
Cursor	.cursor/rules/backend.mdc
Windsurf	.windsurfrules
Continue	AGENTS.md
GitHub Copilot	.github/copilot-instructions.md
Gemini Code Assist	GEMINI.md

Make sure the assistant can run python3 src/backend-pro-max/scripts/search.py … in your repo.

⚙️ Anywhere else (CLI / scripts / MCP)

The CLI is pure Python 3 standard library. Either install it:

pip install git+https://github.com/shashankswe2020-ux/backend-pro-max-skill
backendpro --list
backendpro "redis cluster" --json

…or just clone and run the script directly:

python3 src/backend-pro-max/scripts/search.py --list
python3 src/backend-pro-max/scripts/search.py "redis cluster" --json

The --json output makes it trivial to wire into an MCP tool, a custom agent loop, or any CI step.

---

✅ Prerequisites

• Python 3.8+ — that's it. No pip install, no virtualenv, no models.
• Works on Linux, macOS, Windows (WSL/native), and inside containers.

---

🧪 Smoke test

# Installed CLI
backendpro --list
backendpro "circuit breaker"
backendpro "virtual threads" --stack java-spring
backendpro "idempotency" --all
backendpro compare "Postgres" "DynamoDB" --domain database
backendpro decide "Kafka vs Pulsar" --constraints throughput=high
backendpro adr "Redis vs Memcached for session cache on AWS"
backendpro design "Postgres vs DynamoDB for 50M DAU e-commerce"
backendpro-validate                       # ✅ All CSVs valid (21 domains + 12 stacks).

# v0.6 DX tools
backendpro lint src/                                        # scan for anti-patterns
backendpro lint src/ --format sarif                         # SARIF for GitHub Code Scanning
backendpro export --format obsidian --out /tmp/bpm-vault    # export to Obsidian vault
backendpro export --format notion --out /tmp/notion         # export to Notion CSV
backendpro learn --domain consistency --daily 5             # spaced-repetition flashcards
backendpro learn --stats                                    # learning progress

# Or, without installing
python3 src/backend-pro-max/scripts/search.py --list
python3 src/backend-pro-max/scripts/search.py "circuit breaker"

Run the test suite (contributors)

pip install -e ".[dev]"
pytest                                    # 390 tests
ruff check src tests                      # lint
backendpro-validate                       # schema validation

---

� MCP / Agent Integration (v0.3.1)

Backend Pro Max ships as a Model Context Protocol (MCP) server — any MCP-aware IDE or agent framework can consume all 8 tools natively.

Install

pip install backendpro[mcp]    # adds the mcp SDK as an optional extra

Core pip install backendpro still works with zero dependencies — MCP is fully optional.

Run the MCP server

backendpro-mcp                 # starts on stdio transport

Claude Desktop config

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "backendpro": {
      "command": "backendpro-mcp"
    }
  }
}

Available MCP tools

Tool	Description
backendpro_search	BM25 search in a domain (query, domain?, max_results?, min_score?)
backendpro_search_all	Cross-domain search (query, max_results?)
backendpro_search_stack	Stack-specific guidelines (query, stack, max_results?)
backendpro_compare	Side-by-side comparison (names[], domain?)
backendpro_decide	Opinionated recommendation (requirement)
backendpro_adr	Architecture Decision Record (title, context_domains[])
backendpro_design	System design scaffold (description)
backendpro_find_stale	Freshness audit (domain, months)

Every tool returns structured JSON with [BPM:…] citation tokens — stable, greppable keys that let reviewers verify grounding in PRs:

grep -r '\[BPM:' .   # find all citations in generated code / docs

JSONL streaming

For agent loops that consume results incrementally:

backendpro "kafka" --domain messaging --jsonl     # one JSON object per line
backendpro "idempotency" --all --jsonl             # cross-domain, streamed
backendpro compare "Kafka" "RabbitMQ" --jsonl      # per-field lines

Function-calling manifest (tools.json)

A pre-generated tools.json at repo root provides dual-format schemas (OpenAI functions + Anthropic tool_use) for LangGraph, CrewAI, AutoGen, and other agent frameworks:

import json
tools = json.load(open("tools.json"))["openai"]    # or ["anthropic"]

MCP Inspector validation

npx @modelcontextprotocol/inspector backendpro-mcp
# All 8 tools tested ✅ — see docs/mcp-inspector-report.md

---

🛡️ DX & Distribution (v0.6)

Five features that turn the knowledge base into an active developer tool:

Linter (backendpro lint)

Scan source files for backend anti-patterns — 18 rules across Go, Python, Java, TypeScript, and .env files. Each finding links to a BPM citation.

backendpro lint src/                             # human-readable output
backendpro lint src/ --format json               # machine-readable
backendpro lint src/ --format sarif              # GitHub Code Scanning
backendpro lint src/ --severity error            # errors only

Pre-commit: Add to .pre-commit-config.yaml:

repos:
  - repo: https://github.com/shashankswe2020-ux/backend-pro-max-skill
    rev: v0.6.0
    hooks:
      - id: backendpro-lint

GitHub Action: Use .github/actions/lint/action.yml to upload SARIF inline annotations.

Export (backendpro export)

Export the KB to your preferred knowledge management tool:

backendpro export --format obsidian --out ./vault     # Obsidian vault with wikilinks + _Index.md
backendpro export --format notion --out ./notion       # Notion-importable CSVs
backendpro export --format org --out ./org             # Org-mode files
backendpro export --format obsidian --domain cache --out ./vault  # single domain

Learn Mode (backendpro learn)

Spaced-repetition flashcards (SM-2 algorithm) generated from the KB:

backendpro learn --domain consistency --daily 5   # 5 flashcards from consistency domain
backendpro learn --stats                          # learning progress
backendpro learn --reset                          # start over

VS Code Extension

Install from extensions/vscode/:

cd extensions/vscode && npm install && npm run compile
npx @vscode/vsce package    # produces .vsix
code --install-extension backendpro-*.vsix

• Cmd+Shift+P → "Backend Pro Max: Search" — query the KB from VS Code
• Right-click → "Explain with Backend Pro Max" — explain selected text
• CodeLens — stack guidelines shown at top of Go, Python, Java, TS, Rust files
• MCP mode — set backendpro.useMcp: true to use backendpro-mcp instead of CLI

Web Playground

Live at backendpro.cc — search-as-you-type, domain filter, shareable permalinks, compare view. Zero backend.

---

🧱 Extending

Adding a new row, a new domain, or a new stack takes ~2 minutes.

Want to add…	Steps
📝 A new row	Append to the relevant data/<domain>.csv (keep column order; quote any field with commas). Set Last Updated to today (YYYY-MM-DD) where the column exists.
🆕 A new domain	Add data/<domain>.csv, register in CSV_CONFIG + _DOMAIN_KEYWORDS in core.py.
🧱 A new stack	Add data/stacks/<stack>.csv, register in STACK_CONFIG in core.py.
🔁 A search synonym	Add a token → expansions entry in _SYNONYMS in core.py (keep it conservative).

Always run before pushing:

pytest                  # ranking-quality + unit tests
backendpro-validate     # CSV schema check
ruff check src tests    # lint

Full details in CONTRIBUTING.md, CLAUDE.md, and docs/ARCHITECTURE.md.

---

🤝 Contributing

PRs welcome — especially for:

• 🆕 New stacks: Swift on the server, Erlang/OTP, Zig, Crystal, Gleam, Deno, Bun, …
• 🌐 New domains: FinOps, ML platform, edge / WASM, blockchain infra, mobile-backend, …
• 🧠 More rows in existing CSVs (with Do, Don't, code, severity, and a docs URL).
• 🐛 Corrections — if a recommendation is dated or wrong, open a PR with the source.

Please follow the Git workflow in CLAUDE.md:

1. Branch from main (feat/... or fix/...).
2. Commit with a clear message.
3. Open a PR — never push directly to main.

---

❓ FAQ

Does this need an internet connection?

No. The CLI is offline-first, pure Python stdlib. The only network calls are whatever the AI assistant itself makes.

Why CSV instead of YAML / JSON / SQLite?

CSV diffs cleanly in PRs, opens in any editor / spreadsheet, and is trivial to parse with stdlib. Search is BM25 over the configured columns.

How is this different from a generic "rules" file?

A flat rules file forces the model to keep everything in context. This skill makes the model search a structured KB on demand — so it scales to hundreds of rows across 20+ domains without bloating the prompt.

Can I use it from an MCP server / tool-calling agent?

Yes — pip install backendpro[mcp] and run backendpro-mcp. It exposes 8 tools via stdio transport, compatible with Claude Desktop, Cline, Cursor, Zed, and any MCP-aware client. See the MCP section for details. For non-MCP agents, use --json or --jsonl, or import tools.json for OpenAI / Anthropic function-calling schemas.

---

📜 License

MIT © 2025 contributors

_{Built for the engineers who actually ship distributed systems. If this saves you one outage, ⭐ the repo.}