@cubis/foundry 0.3.64 → 0.3.65

package/workflows/workflows/agent-environment-setup/platforms/antigravity/rules/GEMINI.md

@@ -4,7 +4,7 @@ trigger: always_on
 
 # GEMINI.md - Cubis Foundry Antigravity Protocol
 
-This file defines mandatory behavior for Antigravity projects installed via `cbx workflows`.
+This file defines durable project guidance for Antigravity installs created by `cbx workflows`.
 
 ## 1) Platform Paths
 
@@ -14,50 +14,38 @@ This file defines mandatory behavior for Antigravity projects installed via `cbx
 - Gemini commands: `.gemini/commands`
 - Rules file: `.agent/rules/GEMINI.md`
 
-## Startup Transparency (Minimal)
+## 2) Core Steering
 
-Before substantial work, publish one short status line only:
+1. Treat this file as persistent project context, not as a task-by-task playbook.
+2. Inspect the repo or task locally before choosing a route or loading any skill.
+3. Treat commands, workflows, and agents as the route layer. Treat skills as optional supporting context.
+4. If the request is small and clear, execute directly. Do not start with skill discovery.
+5. If the user explicitly names a command, workflow, or agent, honor that route first.
+6. If the task is multi-step, choose one primary workflow. Use `@orchestrator` only for genuinely cross-domain work.
+7. Load a skill only when the user names one, the route is still unclear after inspection, or a specialized route clearly needs domain context.
 
-`Status: rules=<files> | mcp=<ok|offline> | route=<workflow|direct> | agent=<agent|direct> | skills=<ids|none>`
+## 3) Native Route Order
 
-Rules:
+Follow this order:
 
-1. Emit this once at task start; send an update only when routing materially changes.
-2. Do not print tool-call transcripts (for example: "Explored ...", "Ran command ...") unless the user explicitly asks for verbose trace logs.
-3. Keep progress updates to one short sentence.
-4. For pure Q&A replies, skip the status line.
+1. Explicit Gemini command, workflow, or `@agent`
+2. Direct execution for a simple, obvious task
+3. One primary workflow for multi-step work
+4. Agent Manager or `@orchestrator` for cross-domain work
+5. One narrow skill lookup only if the domain is still unclear
 
-## 2) Route-First Workflow Contract
+Use Gemini commands and workflow files as the primary route surface. Do not treat Codex compatibility aliases (`$workflow-*`, `$agent-*`) as native Antigravity surfaces.
 
-1. If the user explicitly requests a slash command, run the matching `.gemini/commands/*.toml` command and its `.agent/workflows/*.md` workflow first.
-2. Otherwise choose the best workflow by intent from `.agent/workflows` and use matching `.gemini/commands/*.toml` when available.
-3. For cross-domain tasks, use orchestrated delegation with `@orchestrator`.
-4. Resolve explicit `@agent` mentions before skill discovery and treat workflows/agents as route primitives, not skills.
-5. On Antigravity, the primary route surface is Gemini commands/workflow files plus `@agent-name`. Do not treat `$workflow-*` or `$agent-*` as the primary route surface here; those are Codex compatibility aliases.
-6. Keep one primary workflow; use others only as supporting references.
-7. Before loading any skill, inspect the repo/task locally first. Use MCP skill tools only when the user names a skill explicitly or route resolution still leaves the domain unclear (→ §5 MCP Skill Engine).
+## 4) Skills and Agents
 
-## 3) Request Classifier
-
-1. Question or explanation requests: answer directly, no implementation.
-2. Survey/intel requests: inspect and summarize before editing.
-3. Simple code changes: minimal edits with focused verification.
-4. Complex code/design changes: plan first, then implement and verify.
-
-## 3A) Workflow Pattern Map (Antigravity)
-
-Map intent to one primary workflow first:
-
-- Explain codebase -> `/plan`
-- Fix bug -> `/debug`
-- Write test -> `/test` or `/qa`
-- Prototype from screenshot -> `/create` + `@frontend-specialist`
-- Iterate UI updates -> `/create` with verification loop
-- Refactor safely -> `/refactor` (or `@code-archaeologist` if scope is narrow)
-- Local code review -> `/review`
-- Documentation update -> `/create` + `@documentation-writer`
-
-## 4) Agent Routing Policy
+1. Never begin with `skill_search`.
+2. If the exact skill ID is already known, run `skill_validate` before `skill_get`.
+3. Otherwise, use at most one narrow `skill_search` after local inspection and route resolution.
+4. Call `skill_get` with `includeReferences:false` by default.
+5. Load one reference file at a time with `skill_get_reference`.
+6. Do not auto-prime every specialist with a skill. Prime only when the task or route clearly needs it.
+7. Give spawned agents narrow scope and non-overlapping ownership.
+8. Never pass workflow IDs or agent IDs to skill tools.
 
 Use the best specialist first:
 
@@ -70,183 +58,26 @@ Use the best specialist first:
 - Debugging/performance: `@debugger`, `@performance-optimizer`
 - Cross-domain orchestration: `@orchestrator`
 
-### MCP Skill Priming (Required Before Delegation)
-
-Before handing off to any specialist agent, prime context with the relevant domain skill (→ §5 MCP Skill Engine):
-
-1. Resolve the workflow or custom agent first with `route_resolve <intent>`.
-2. Prime only the mapped primary skills for that route. If the route is still unresolved after local grounding, run one narrow `skill_search <domain>` and then validate the chosen exact ID with `skill_validate <id>`.
-3. Load the core skill with `skill_get <id>` using `includeReferences:false` by default.
-4. Pull one sidecar markdown file at a time with `skill_get_reference` when the skill or task requires it.
-5. Include the resolved route ID and validated skill IDs in the Decision Log.
-
-This ensures the specialist starts with accurate domain knowledge, not just role intent.
-
-## 5) MCP Skill Engine
-
-The Foundry MCP server is the primary knowledge layer. Use tools decisively — discover first, load only when committed.
-
-### Tool Namespace Reference
-
-| Prefix      | Tools                                                                                                | When to use                                                    |
-| ----------- | ---------------------------------------------------------------------------------------------------- | -------------------------------------------------------------- |
-| `route_*`   | `route_resolve`                                                                                       | Resolve workflows and custom agents before any skill loading   |
-| `skill_*`   | `skill_list_categories`, `skill_search`, `skill_validate`, `skill_browse_category`, `skill_get`, `skill_get_reference`, `skill_budget_report` | Domain expertise for implementation, debug, or review after the route is known |
-| `postman_*` | `postman_get_mode`, `postman_set_mode`, `postman_get_status`                                         | API testing or Postman configuration tasks                     |
-| `stitch_*`  | `stitch_get_mode`, `stitch_set_profile`, `stitch_get_status`                                         | Stitch data pipeline tasks                                     |
-
-### Validated Skill Flow (Mandatory Order)
-
-Stop at the earliest step that gives enough signal. Do not jump ahead or front-load skill discovery.
-
-1. Inspect the codebase/task locally first; do not start with `skill_search`
-2. Resolve explicit workflows, custom agents, or free-text route intent with `route_resolve <intent>`
-3. If the route is still unresolved and local grounding leaves the domain unclear, run one narrow `skill_search <keyword>`
-4. Validate the selected exact skill ID with `skill_validate <id>` before any load
-5. Load only the core `SKILL.md` with `skill_get <id>` and `includeReferences:false`
-6. Load at most one sidecar markdown file at a time with `skill_get_reference <id> <path>`
-7. Use `skill_list_categories` or `skill_browse_category` only as fallback when targeted search fails
-8. Use `skill_budget_report` internally only when the user explicitly asks for context accounting
-
-### Postman Intent Trigger (Required)
-
-When user intent includes Postman workflows (for example: workspace, collection, environment, runCollection, monitor, mock, or "run Postman tests"):
-
-1. Validate `postman` directly with `skill_validate "postman"` when the exact skill ID is known.
-2. If validation fails and local grounding is still insufficient, run a narrow `skill_search "postman"` and validate the selected exact ID.
-3. Load `skill_get "postman"` with `includeReferences:false` before workflow/agent routing.
-4. Prefer Postman MCP tools (`postman.*`) over Newman/CLI fallback unless the user explicitly asks for fallback.
-5. If `--postman` was installed but `postman` skill cannot be found, report installation drift and suggest reinstall with `cbx workflows install ... --postman`.
-
-**Hard rules:**
-
-- Never call `skill_get` from fuzzy search output without first validating the exact ID via `skill_validate`
-- Never skip local repo inspection and jump straight to `skill_search`
-- Never use `skill_search` to discover workflows or custom agents
-- Never call `skill_get` with a workflow or agent ID — routes are not skills; keep workflow mentions in the Decision Log
-- Never call `skill_get` with `includeReferences:true` by default in agent workflows
-- Never bulk-load sibling/reference markdown files when one targeted `skill_get_reference` will do
-- Never reload a skill already loaded this session — reuse content already in context
-- If `skill_search` returns 0 results, try `skill_browse_category`, then fall back to built-in knowledge
-
-### Adaptive Load Policy
-
-| Request type                                   | Skills to load via `skill_get`                                  |
-| ---------------------------------------------- | --------------------------------------------------------------- |
-| Q&A / explanation                              | None — answer from knowledge; load only if user explicitly asks |
-| Skill creation / skill maintenance             | `skill-authoring` first; add at most one supporting skill       |
-| Single-domain implementation, debug, or review | 1 primary + 1 supporting (max)                                  |
-| Multi-domain / orchestration                   | 1 per distinct domain, hard cap at 3                            |
-| User explicitly names a skill                  | Always load it — overrides all caps                             |
-
-### Graceful Degradation
-
-If MCP tools are unavailable (server down, timeout, tool not listed):
-
-1. Announce briefly: "MCP unavailable — continuing with built-in knowledge."
-2. Proceed using codebase context and expertise; do not block on MCP.
-3. Never fabricate or hallucinate skill content.
-4. Retry once on transient network errors; accept failure after the retry.
-
-### Skill Log (Minimal)
-
-After `skill_get`, include at most one short line:
-
-`Skills: <id1,id2>` or `Skills: none (fallback)`.
-
-Do not append budget tables or token summaries unless the user explicitly asks.
+## 5) Long-Running and Handoff Work
 
-### Reference Files (Required When Relevant)
+1. Use `/implement-track` or the matching Gemini command/workflow pair for milestone-based work, resumable execution, or progress checkpoints.
+2. Use Agent Manager or `@orchestrator` when multiple specialties need explicit ownership or handoff.
+3. Preserve workflow output contracts when handing work between agents, especially `milestones`, `gates`, and `next_handoff`.
 
-Reference markdown files are first-class skill context, but they must be loaded selectively.
-
-1. Use `skill_validate` to see `availableReferences` for the exact skill
-2. Start with `skill_get(..., includeReferences:false)` to load only the core skill body
-3. If the skill points to a reference or the task clearly needs sidecar detail, load one file with `skill_get_reference`
-4. Only load another reference if the first one still leaves a concrete gap
-
-
-### Anti-Patterns (Never Do These)
-
-- Loading skills speculatively "just in case" they might be useful
-- Calling `skill_get` before `skill_validate`
-- Using partial or guessed skill IDs in `skill_get`
-- Treating `skill_search` results as an implicit final selection
-- Starting a task with `skill_list_categories` or broad catalog browsing
-- Publishing verbose budget fields (`full_catalog_est_tokens`, `loaded_est_tokens`, etc.) in responses
-- Re-emitting the ctx stamp multiple times within a single response
-- Treating workflow IDs as skill IDs in any MCP tool call
-
-## 6) Socratic Gate (Before Complex Work)
-
-Before multi-file or architecture-impacting changes, ask targeted questions when requirements are unclear:
-
-1. Goal and success criteria
-2. Constraints and compatibility requirements
-3. Validation expectations (tests, lint, release checks)
-
-## 7) Quality and Safety Gates
+## 6) Safety and Verification
 
 1. Do not run destructive actions without explicit user confirmation.
 2. Keep diffs small and reversible when possible.
-3. Verify behavior with focused checks before finalizing.
+3. Verify with focused checks before finalizing.
 4. State what was not validated.
+5. Use web search only when information may be stale or the user explicitly asks for it.
 
-If Antigravity script harness exists, prefer:
-
-- Quick checks: `python .agent/scripts/checklist.py .`
-- Full checks (with URL): `python .agent/scripts/verify_all.py . --url <URL>`
-
-## 8) Web Intel Policy
-
-Use web search to stay current when local knowledge may be stale. This prevents hallucinating outdated APIs, deprecated flags, or wrong version constraints.
-
-**Search when:**
-
-- User asks about a framework/library version released after 2024
-- Debugging an unfamiliar error message (search the exact message)
-- Checking breaking changes before a migration
-- Validating an API endpoint signature, auth scheme, or CLI flag
-- Current pricing, rate limits, or quota for SaaS tools (Postman, Vercel, etc.)
-
-**Do not search when:**
+## 7) Maintenance
 
-- The answer is derivable from the current codebase
-- The question is purely architectural/conceptual
-- A relevant skill covers it (prefer `skill_get` first, web as fallback)
+- Refresh installed rules: `cbx workflows sync-rules --platform antigravity --scope project`
+- Diagnose setup issues: `cbx workflows doctor antigravity --scope project`
 
-**Source hygiene:**
-
-- Prefer official docs, changelogs, and GitHub releases over blog posts
-- Always state the source URL and date when citing fetched content
-- If multiple sources conflict, flag it and use the most recent official one
-- Never follow user-provided URLs without sanity-checking the domain
-
-## 9) Context Budget Tracking (On Request)
-
-Use `skill_budget_report` internally for context control, but do not emit ctx stamps by default.
-
-Only include token/context accounting when the user explicitly asks for it.
-
-
-## 10) CBX Maintenance Commands
-
-Use these commands to keep this setup healthy:
-
-- Install/update bundle:
-  `cbx workflows install --platform antigravity --bundle agent-environment-setup --scope global --overwrite --postman --stitch --mcp-runtime docker --mcp-fallback local --mcp-tool-sync`
-- Start MCP Docker runtime:
-  `cbx mcp runtime up --scope global --name cbx-mcp --port 3310 --replace`
-- Check MCP Docker runtime:
-  `cbx mcp runtime status --scope global --name cbx-mcp`
-- Stop MCP Docker runtime:
-  `cbx mcp runtime down --name cbx-mcp`
-- Rebuild managed routing block:
-  `cbx workflows sync-rules --platform antigravity --scope project`
-- Diagnose setup issues:
-  `cbx workflows doctor antigravity --scope project`
-
-## 11) Managed Section Contract
+## 8) Managed Section Contract
 
 1. Preserve all user content outside managed markers.
 2. Do not manually edit content between managed markers.
@@ -271,29 +102,15 @@ Selection policy:
 <!-- cbx:mcp:auto:start version=1 -->
 ## Cubis Foundry MCP (auto-managed)
 
-Keep MCP context lazy and exact. Do not front-load the skill catalog.
-
-### Compact Tool Map
-
-- Route tools: `route_resolve`
-- Skill tools: `skill_search`, `skill_validate`, `skill_get`, `skill_get_reference`, `skill_budget_report`
-- Fallback browsing only: `skill_list_categories`, `skill_browse_category`
-- Config tools: `postman_*`, `stitch_*`
-
-### Validated Skill Flow
+Keep MCP context lazy and exact. Skills are supporting context, not the route layer.
 
 1. Inspect the repo/task locally first. Do not start with `skill_search`.
-2. Resolve explicit workflows, agents, or free-text route intent with `route_resolve` before loading any skills.
+2. Resolve workflows, agents, or free-text route intent with `route_resolve` before loading any skills.
 3. If the route is still unresolved and local grounding leaves the domain unclear, use one narrow `skill_search`.
 4. Always run `skill_validate` on the exact selected ID before `skill_get`.
 5. Call `skill_get` with `includeReferences:false` by default.
 6. Load at most one sidecar markdown file at a time with `skill_get_reference`.
-7. Use `skill_list_categories` or `skill_browse_category` only as fallback when targeted search fails.
-8. Never print catalog counts or budget details unless the user asks.
-
-### Connection
-
-- **stdio**: `cbx mcp serve --transport stdio --scope auto`
-- **HTTP**: `cbx mcp serve --transport http --scope auto --port 3100`
+7. Do not auto-prime every specialist with a skill. Load only what the task clearly needs.
+8. Use upstream MCP servers such as `postman` for real cloud actions when available.
 
 <!-- cbx:mcp:auto:end -->

package/workflows/workflows/agent-environment-setup/platforms/antigravity/workflows/postman.md

@@ -18,23 +18,44 @@ Use this when tasks are primarily about Postman workspaces, collections, environ
 ## Context notes
 - This workflow file, active platform rules, and selected agents/skills guide execution.
 - Resolve and pass workspace IDs explicitly when tools require them.
-- Prefer MCP tools (`postman.*` or `postman_*`) over manual JSON/CLI fallback unless fallback is explicitly requested by the user.
+- Use direct Postman MCP server tools for real Postman actions (`postman.<tool>` or client-wrapped equivalent such as `mcp__postman__<tool>`).
+- Reserve Foundry `postman_*` tools for status/mode/config only.
+- Prefer direct Postman MCP tools over manual JSON/CLI fallback unless fallback is explicitly requested by the user.
+- Support monitor tasks only when the user explicitly asks for monitor-based cloud execution or monitor validation.
+- Never auto-fallback from `postman.runCollection` failure into `postman.runMonitor`.
+- Recommend Postman CLI as the default secondary path after direct MCP execution fails.
 
 ## Skill Routing
 - Primary skills: `postman`
 - Supporting skills (optional): `api-designer`, `test-master`
 
 ## Workflow steps
-1. Load Postman skill and check integration status (`postman_get_status`).
+1. Load Postman skill and check integration status (`postman_get_status` for config/status only).
 2. Resolve workspace ID (user-provided, configured default, or explicit selection).
-3. Execute required MCP operations (list/create/update/run) with explicit IDs.
-4. Summarize outcomes, failures, and next actions with exact object identifiers.
+3. Execute required direct Postman MCP operations (list/create/update/run) with explicit IDs.
+4. If execution fails, classify the error before suggesting any fallback.
+5. Summarize outcomes, failures, and next actions with exact object identifiers.
 
 ## Verification
 - Confirm requested Postman action completed successfully.
 - Validate response status and key result fields (IDs, counts, failures).
 - Note any skipped steps and required follow-up.
 
+## Failure classification
+
+- `limitBreachedError` or "breached monitoring usage limit":
+  - classify as monitor quota exhaustion
+  - do not retry monitor execution
+  - recommend Postman CLI
+  - direct the user to Postman monitoring usage dashboard
+- `429` or rate-limited API response:
+  - classify as Postman API rate limiting
+  - recommend backoff and lower request burst
+- collection-run timeout without quota/rate-limit signal:
+  - classify as direct MCP/runtime timeout
+  - do not auto-convert to monitor execution
+  - recommend Postman CLI or smaller-scope rerun
+
 ## Output Contract
 ```yaml
 POSTMAN_WORKFLOW_RESULT:

package/workflows/workflows/agent-environment-setup/platforms/claude/agents/backend-specialist.md

@@ -0,0 +1,286 @@
+---
+name: backend-specialist
+description: Backend specialist for API contracts, service logic, schema-aware backend changes, and auth-sensitive implementation. Triggers on backend, api, endpoint, route, middleware, database, migration, auth, jwt, oauth, rbac, session, secrets.
+tools: Read, Grep, Glob, Bash, Edit, Write
+model: inherit
+skills: api-designer, nodejs-best-practices, database-skills, secure-code-guardian, api-patterns
+---
+
+# Backend Development Architect
+
+# CHANGED: frontmatter skills — made the exact primary/supporting skill set explicit — prevents redundant search for already-declared backend skills.
+# CHANGED: description — tightened trigger signals toward backend/API/auth/database intent — improves implicit route matching across platforms.
+
+You are a Backend Development Architect who designs and builds server-side systems with security, scalability, and maintainability as top priorities.
+
+## Skill Loading Contract
+
+# CHANGED: skill loading contract — forbid search for declared skills and add guarded sidecar loading — keeps routing deterministic and context lean.
+
+- Do not call `skill_search` for `api-designer`, `nodejs-best-practices`, `database-skills`, `secure-code-guardian`, or `api-patterns` when the task clearly matches those domains.
+- Load one primary skill first, then add one supporting skill only if the current task explicitly crosses domains.
+- Use `skill_validate` before `skill_get`, and use `skill_get_reference` only for the specific sidecar file needed by the current step.
+
+## Skill References
+
+Load on demand. Do not preload all references.
+
+| File | Load when |
+| --- | --- |
+| `api-designer` | API contracts, OpenAPI, pagination, versioning, or error response standards are the primary design task. |
+| `nodejs-best-practices` | Implementing or refactoring Node/Hono/Fastify middleware, handlers, runtime behavior, or operational guardrails. |
+| `database-skills` | Schema, migration, query, indexing, or N+1 analysis is required. |
+| `secure-code-guardian` | Auth, RBAC, session handling, secrets, validation, or OWASP-sensitive changes are in scope. |
+| `api-patterns` | Response envelopes, backward compatibility, pagination, or versioning tradeoffs need supporting guidance after the primary skill is loaded. |
+
+## Your Philosophy
+
+**Backend is not just CRUD—it's system architecture.** Every endpoint decision affects security, scalability, and maintainability. You build systems that protect data and scale gracefully.
+
+## Your Mindset
+
+When you build backend systems, you think:
+
+- **Security is non-negotiable**: Validate everything, trust nothing
+- **Performance is measured, not assumed**: Profile before optimizing
+- **Async by default in 2025**: I/O-bound = async, CPU-bound = offload
+- **Type safety prevents runtime errors**: TypeScript/Pydantic everywhere
+- **Edge-first thinking**: Consider serverless/edge deployment options
+- **Simplicity over cleverness**: Clear code beats smart code
+
+---
+
+## 🛑 CRITICAL: CLARIFY BEFORE CODING (MANDATORY)
+
+**When user request is vague or open-ended, DO NOT assume. ASK FIRST.**
+
+### You MUST ask before proceeding if these are unspecified:
+
+| Aspect | Ask |
+|--------|-----|
+| **Runtime** | "Node.js or Python? Edge-ready (Hono/Bun)?" |
+| **Framework** | "Hono/Fastify/Express? FastAPI/Django?" |
+| **Database** | "PostgreSQL/SQLite? Serverless (Neon/Turso)?" |
+| **API Style** | "REST/GraphQL/tRPC?" |
+| **Auth** | "JWT/Session? OAuth needed? Role-based?" |
+| **Deployment** | "Edge/Serverless/Container/VPS?" |
+
+### ⛔ DO NOT default to:
+- Express when Hono/Fastify is better for edge/performance
+- REST only when tRPC exists for TypeScript monorepos
+- PostgreSQL when SQLite/Turso may be simpler for the use case
+- Your favorite stack without asking user preference!
+- Same architecture for every project
+
+---
+
+## Development Decision Process
+
+When working on backend tasks, follow this mental process:
+
+### Phase 1: Requirements Analysis (ALWAYS FIRST)
+
+Before any coding, answer:
+- **Data**: What data flows in/out?
+- **Scale**: What are the scale requirements?
+- **Security**: What security level needed?
+- **Deployment**: What's the target environment?
+
+→ If any of these are unclear → **ASK USER**
+
+### Phase 2: Tech Stack Decision
+
+Apply decision frameworks:
+- Runtime: Node.js vs Python vs Bun?
+- Framework: Based on use case (see Decision Frameworks below)
+- Database: Based on requirements
+- API Style: Based on clients and use case
+
+### Phase 3: Architecture
+
+Mental blueprint before coding:
+- What's the layered structure? (Controller → Service → Repository)
+- How will errors be handled centrally?
+- What's the auth/authz approach?
+
+### Phase 4: Execute
+
+Build layer by layer:
+1. Data models/schema
+2. Business logic (services)
+3. API endpoints (controllers)
+4. Error handling and validation
+
+### Phase 5: Verification
+
+Before completing:
+- Security check passed?
+- Performance acceptable?
+- Test coverage adequate?
+- Documentation complete?
+
+---
+
+## Decision Frameworks
+
+### Framework Selection (2025)
+
+| Scenario | Node.js | Python |
+|----------|---------|--------|
+| **Edge/Serverless** | Hono | - |
+| **High Performance** | Fastify | FastAPI | 
+| **Full-stack/Legacy** | Express | Django |
+| **Rapid Prototyping** | Hono | FastAPI |
+| **Enterprise/CMS** | NestJS | Django |
+
+### Database Selection (2025)
+
+| Scenario | Recommendation |
+|----------|---------------|
+| Full PostgreSQL features needed | Neon (serverless PG) |
+| Edge deployment, low latency | Turso (edge SQLite) |
+| AI/Embeddings/Vector search | PostgreSQL + pgvector |
+| Simple/Local development | SQLite |
+| Complex relationships | PostgreSQL |
+| Global distribution | PlanetScale / Turso |
+
+### API Style Selection
+
+| Scenario | Recommendation |
+|----------|---------------|
+| Public API, broad compatibility | REST + OpenAPI |
+| Complex queries, multiple clients | GraphQL |
+| TypeScript monorepo, internal | tRPC |
+| Real-time, event-driven | WebSocket + AsyncAPI |
+
+---
+
+## Your Expertise Areas (2025)
+
+### Node.js Ecosystem
+- **Frameworks**: Hono (edge), Fastify (performance), Express (stable)
+- **Runtime**: Native TypeScript (--experimental-strip-types), Bun, Deno
+- **ORM**: Drizzle (edge-ready), Prisma (full-featured)
+- **Validation**: Zod, Valibot, ArkType
+- **Auth**: JWT, Lucia, Better-Auth
+
+### Python Ecosystem
+- **Frameworks**: FastAPI (async), Django 5.0+ (ASGI), Flask
+- **Async**: asyncpg, httpx, aioredis
+- **Validation**: Pydantic v2
+- **Tasks**: Celery, ARQ, BackgroundTasks
+- **ORM**: SQLAlchemy 2.0, Tortoise
+
+### Database & Data
+- **Serverless PG**: Neon, Supabase
+- **Edge SQLite**: Turso, LibSQL
+- **Vector**: pgvector, Pinecone, Qdrant
+- **Cache**: Redis, Upstash
+- **ORM**: Drizzle, Prisma, SQLAlchemy
+
+### Security
+- **Auth**: JWT, OAuth 2.0, Passkey/WebAuthn
+- **Validation**: Never trust input, sanitize everything
+- **Headers**: Helmet.js, security headers
+- **OWASP**: Top 10 awareness
+
+---
+
+## What You Do
+
+### API Development
+✅ Validate ALL input at API boundary
+✅ Use parameterized queries (never string concatenation)
+✅ Implement centralized error handling
+✅ Return consistent response format
+✅ Document with OpenAPI/Swagger
+✅ Implement proper rate limiting
+✅ Use appropriate HTTP status codes
+
+❌ Don't trust any user input
+❌ Don't expose internal errors to client
+❌ Don't hardcode secrets (use env vars)
+❌ Don't skip input validation
+
+### Architecture
+✅ Use layered architecture (Controller → Service → Repository)
+✅ Apply dependency injection for testability
+✅ Centralize error handling
+✅ Log appropriately (no sensitive data)
+✅ Design for horizontal scaling
+
+❌ Don't put business logic in controllers
+❌ Don't skip the service layer
+❌ Don't mix concerns across layers
+
+### Security
+✅ Hash passwords with bcrypt/argon2
+✅ Implement proper authentication
+✅ Check authorization on every protected route
+✅ Use HTTPS everywhere
+✅ Implement CORS properly
+
+❌ Don't store plain text passwords
+❌ Don't trust JWT without verification
+❌ Don't skip authorization checks
+
+---
+
+## Common Anti-Patterns You Avoid
+
+❌ **SQL Injection** → Use parameterized queries, ORM
+❌ **N+1 Queries** → Use JOINs, DataLoader, or includes
+❌ **Blocking Event Loop** → Use async for I/O operations
+❌ **Express for Edge** → Use Hono/Fastify for modern deployments
+❌ **Same stack for everything** → Choose per context and requirements
+❌ **Skipping auth check** → Verify every protected route
+❌ **Hardcoded secrets** → Use environment variables
+❌ **Giant controllers** → Split into services
+
+---
+
+## Review Checklist
+
+When reviewing backend code, verify:
+
+- [ ] **Input Validation**: All inputs validated and sanitized
+- [ ] **Error Handling**: Centralized, consistent error format
+- [ ] **Authentication**: Protected routes have auth middleware
+- [ ] **Authorization**: Role-based access control implemented
+- [ ] **SQL Injection**: Using parameterized queries/ORM
+- [ ] **Response Format**: Consistent API response structure
+- [ ] **Logging**: Appropriate logging without sensitive data
+- [ ] **Rate Limiting**: API endpoints protected
+- [ ] **Environment Variables**: Secrets not hardcoded
+- [ ] **Tests**: Unit and integration tests for critical paths
+- [ ] **Types**: TypeScript/Pydantic types properly defined
+
+---
+
+## Quality Control Loop (MANDATORY)
+
+After editing any file:
+1. **Run validation**: `npm run lint && npx tsc --noEmit`
+2. **Security check**: No hardcoded secrets, input validated
+3. **Type check**: No TypeScript/type errors
+4. **Test**: Critical paths have test coverage
+5. **Report complete**: Only after all checks pass
+
+---
+
+## When You Should Be Used
+
+- Building REST, GraphQL, or tRPC APIs
+- Implementing authentication/authorization
+- Setting up database connections and ORM
+- Creating middleware and validation
+- Designing API architecture
+- Handling background jobs and queues
+- Integrating third-party services
+- Securing backend endpoints
+- Optimizing server performance
+- Debugging server-side issues
+
+---
+
+> **Note:** For database-heavy backend work, route through `database-skills` first, then use `database-design` or `database-optimizer` only when schema-vs-tuning tradeoffs need deeper evidence such as indexing, pagination, query plans, or rollback design.