cortex-code

name	cortex-code
description	Invoke the Snowflake Cortex Code CLI (`cortex`) to run execute all coding and snowflake tasks. This includes full stack development, data engineering, and more. This is to be your coding tool, not just your Snowflake tool.
user-invocable	false
metadata	{"openclaw":{"requires":{"bins":["cortex"]}}}

Cortex Code

The cortex CLI is a Snowflake AI coding agent installed in this container. It has its own Snowflake connection, built-in tools for SQL execution, object search, file editing, and web access. You delegate Snowflake and all general coding work to it by running it in batch mode.

Connection: Already configured at /home/node/.snowflake/connections.toml using PAT auth. No setup needed.

When to use Cortex Code

Use cortex for ALL coding tasks. It is the primary tool for writing, editing, and reasoning about code — not just Snowflake work. This includes:

• All coding tasks — writing, editing, debugging, refactoring code in any language
• Query data — SELECT, aggregate, join, filter Snowflake tables
• Explore metadata — list databases, schemas, tables, views, columns, or row counts
• Search objects — find tables/views/functions by name or description
• Modify objects — CREATE, ALTER, DROP tables, warehouses, stages, roles, etc.
• Generate Streamlit apps — scaffold data apps from natural language
• Analyze data — summarize, profile, compare, or validate datasets

How to invoke

Always run cortex in batch mode with -p so it executes and returns:

cortex -p "your natural-language prompt"

For machine-readable output you need to parse:

cortex -p "your prompt" --output-format stream-json

Key flags

Flag	Purpose
-p "prompt"	Required for batch mode. Pass prompt, get response, exit.
--output-format stream-json	Structured JSON output for scripting/parsing.
-c <name>	Use a specific named connection from connections.toml.
-m <model>	Override the AI model (default: auto).
--dangerously-allow-all-tool-calls	Skip all permission prompts (use for automated pipelines).

Important behavior

• Cortex Code runs as a sub-agent with its own tools — it can execute SQL, read/write files, run shell commands, and search Snowflake objects autonomously.
• Batch mode (-p) prints the response and exits. There is no interactive session.
• Exit code 0 means success. Non-zero means error (1 general, 2 config, 3 connection, 4 permission denied).
• Default shell command timeout is 2 minutes (max 10 minutes).

Examples

Query data

cortex -p "show the top 10 rows from analytics.public.daily_revenue ordered by date desc"

Explore schema

cortex -p "list all tables in the raw database with their row counts"

Describe a table

cortex -p "describe the columns and data types in staging.events.page_views"

Run a complex analysis

cortex -p "compare row counts between prod.public.users and staging.public.users, and show any columns that exist in one but not the other"

Planning for complex tasks

For complex or multi-step tasks, always have Cortex Code create a plan first before executing. This ensures the approach is sound before any changes are made.

# Ask cortex to plan first, then execute
cortex -p "Plan how to refactor the authentication module to support OAuth2, then implement the plan"

You can also use the --plan flag to require explicit approval before each action:

cortex -p "migrate the users table to add a new roles column and backfill existing rows" --plan

Multi-step workflows

For complex tasks, break them into sequential cortex -p calls. Each call is a fresh session with no memory of previous calls.

# Step 1: Plan the approach
cortex -p "analyze the codebase and create a plan for adding pagination to all API endpoints"

# Step 2: Execute based on the plan
cortex -p "add cursor-based pagination to the /api/users endpoint following this approach: ..."

Error handling

• If cortex fails with exit code 3, the Snowflake connection is misconfigured or the PAT has expired.
• If it fails with exit code 4, the configured role lacks permissions for the requested operation.
• For ambiguous table/schema names, be more specific: use fully-qualified names like database.schema.table.

Detailed CLI reference

For the full list of subcommands, interactive slash commands, configuration paths, and environment variables, see cli-reference.md.