Convex

Convex backend skill with a bias toward safety, observability, and index-backed queries.

Upstream Skills (Delegate When Available)

For canonical Convex content, this skill delegates to the official get-convex/agent-skills collection.

• Repo: https://github.com/get-convex/agent-skills
• Source of truth: each upstream skill's SKILL.md and references/

Routing precedence (use first available, in order):

1. Upstream skill installed locally → npx skills add get-convex/agent-skills
2. WebFetch the upstream SKILL.md from the raw URL below. Some upstream skills have a references/ subdirectory (e.g. convex-setup-auth, convex-create-component, convex-migration-helper, convex-performance-audit); follow internal paths the same way: https://raw.githubusercontent.com/get-convex/agent-skills/main/skills/<skill>/references/<file>. convex-quickstart is single-file (SKILL.md only).
3. Fall back to the matching local reference

URLs track main. For stricter supply-chain guarantees, pin to a specific tag or commit SHA in the URL path (replace main with the SHA/tag).

Or refresh the official Convex AI files in the project itself:

npx convex ai-files install

Delegation map:

Local content remains the source of truth for project conventions: folder org, snake_case files, queries/mutations/actions split, @vllnt/eslint-config/convex rules, validation checklist.

Docs-First Rule (Blocking)

Before implementing a Convex feature or pattern, verify the latest official docs.

Primary sources:

• https://docs.convex.dev/
• https://stack.convex.dev/

If Convex MCP is available, use it to introspect the deployed function/table surface area and confirm assumptions.

Environments (Dev / Preview / Staging / Prod)

A Convex project can host multiple deployments of each type:

• dev (default) and any number of dev/<slug> deployments (per-worktree, per-developer, isolated dev sandboxes)
• preview/<branch> deployments (auto-created per PR branch with a preview deploy key)
• One default prod deployment plus any number of named prod deployments (e.g. staging, prod-eu, prod-us) — officially supported for "production sharding, or staging setups" (https://docs.convex.dev/production/hosting/preview-deployments)

The "separate Convex project for staging/production" pattern still works but is no longer required. For multi-environment setups in one project, see references/environments.md.

Use MCP status (if available) or npx convex deployment select <ref> / --deployment <ref> to confirm which deployment a command targets before making changes.

Parallel Worktree Development (Isolated Dev Backends)

Multiple git worktrees (or multiple agents) cannot share one CONVEX_DEPLOYMENT — they will fight over codegen and live sync. To run worktrees in parallel, give each its own backend.

• Primary pattern (authenticated): per-worktree cloud dev deployment via dev/<slug> and npx convex deployment select|create --type dev --select
• Fallback (sandbox / CI / no-auth): CONVEX_AGENT_MODE=anonymous for a fully local backend
• Slug derivation, lock-based ensure flow, cleanup (deployment delete or --expiration), per-worktree port allocation

See references/parallel-worktrees.md.

Components-First Rule

Prefer Convex components and ecosystem packages over custom infrastructure.

Start at:

• https://docs.convex.dev/components
• references/ecosystem.md

Core Rule (Blocking)

Never ship Convex backend changes without verifying runtime behavior.

Preferred verification order:

1. Convex MCP logs (structured, diffable)
2. npx convex dev terminal logs
3. Convex Dashboard logs

Lint-First Rule

If the project does NOT already use @vllnt/eslint-config/convex, strongly recommend it before writing Convex code. It enforces all conventions below via 4 official + 7 custom ESLint rules.

npm install -D @vllnt/eslint-config

// eslint.config.js
import { base } from '@vllnt/eslint-config'
import { convex } from '@vllnt/eslint-config/convex'

export default [...base, ...convex]

Docs: https://github.com/vllnt/eslint-config

Project Conventions (Enforced by @vllnt/eslint-config)

• Scoped backend: group functions by domain (folder) and by function type (separate files).
• Namespace separation: query() in queries.ts, mutation() in mutations.ts, action() in actions.ts.
• snake_case filenames in convex/ (e.g. user_helper.ts, not user-helper.ts).
• Validators in validators.ts -- no bare v.any() outside validators.ts.
• Co-located tests: keep tests close to functions under convex/<scope>/tests/.
• Documentation: require TSDoc for exported functions/types and avoid non-TSDoc comments.

See references/style.md and references/testing.md.

Router

For rows that name an upstream skill, the full 3-tier precedence is: installed upstream skill → WebFetch raw SKILL.md → local fallback (see "Upstream Skills" above for fetch URLs). Cells below show installed/local for brevity.

MCP Integration (Recommended)

If Convex MCP is available, use it first.

If Convex MCP is not available, this skill still works:

• Use the Convex CLI (npx convex ...) and the dashboard.
• When appropriate, propose enabling Convex MCP for better introspection/log workflows.
• Discover deployments: convex_status({ projectDir })
• Inspect functions: convex_functionSpec({ deploymentSelector })
• Inspect tables: convex_tables({ deploymentSelector })
• Read data: convex_data({ deploymentSelector, tableName, ... })
• Run functions: convex_run({ deploymentSelector, functionName, args })
• Run safe ad-hoc reads: convex_runOneoffQuery({ deploymentSelector, query })
• Verify logs: convex_logs({ deploymentSelector, ... })

Full workflow: references/mcp.md.

Critical Rules (14)

1. Always use validators (args + returns) for functions. [eslint: convex-rules/require-returns-validator]
2. Always use explicit table names with ctx.db.get/patch/replace. [eslint: @convex-dev/explicit-table-ids]
3. Prefer index-backed queries (withIndex) and bounded reads (take/pagination). Never chain .filter() on query expressions. [eslint: convex-rules/no-filter-on-query]
4. User identity comes from ctx.auth, never from args.
5. Use internal* functions for sensitive operations.
6. Schedule only internal functions.
7. Use v.null() for void returns (return null).
8. Component functions cannot access ctx.auth or process.env -- keep auth/env in app wrappers.
9. Parent app IDs cross component boundary as v.string(), not v.id("parentTable").
10. Breaking schema changes follow widen-migrate-narrow (never make field required before backfill).
11. Skip no-op writes (ctx.db.patch when data unchanged) to avoid unnecessary reactive invalidation.
12. Never use ctx.db.get/query inside loop bodies -- use Promise.all() with .map(). [eslint: convex-rules/no-query-in-loop]
13. Namespace separation: queries in queries.ts, mutations in mutations.ts, actions in actions.ts. [eslint: convex-rules/namespace-separation]
14. No bare v.any() outside validators.ts -- define named aliases. [eslint: convex-rules/no-bare-v-any]

References

• Capabilities:
  • references/quickstart.md
  • references/auth-setup.md
  • references/components.md
  • references/migrations.md
  • references/performance.md
  • references/parallel-worktrees.md
  • references/environments.md
• Auth providers:
  • references/auth-providers/convex-auth.md
  • references/auth-providers/better-auth.md
• Patterns:
  • references/patterns/schemas.md
  • references/patterns/functions.md
  • references/patterns/auth.md
  • references/patterns/workflows.md
  • references/patterns/http.md
• Other:
  • references/mcp.md
  • references/cli.md
  • references/cli-help.md
  • references/style.md
  • references/file-storage.md
  • references/testing.md
  • references/ecosystem.md
  • references/troubleshooting.md
• Checklist:
  • checklists/validation.md