Reference

Everything RadarModel 0.2 does, in one page: radarmodel-ai first, then radarmodel-schema. The README on GitHub says the same in the repository.

Commands

radarmodel-ai init  [-config FILE]           write a starter config
radarmodel-ai check [-config FILE] [flags]   scan once, compare, report
radarmodel-ai watch [-config FILE] [flags]   scan every interval until stopped
radarmodel-ai show  [-config FILE] SOURCE    list the stored models of a source
radarmodel-ai version

The config is radarmodel.json in the current folder, or the file in $RADARMODEL_CONFIG, or -config FILE.

Flags for check and watch:

Exit codes

Breaking wins: a run with a breaking change and a failed provider exits 1, and the report shows both.

Configuration

{
  "interval": "1h",
  "timeout": "30s",
  "min_fraction": 0.5,
  "confirm_removals": true,
  "state_dir": ".radarmodel",
  "providers": [
    {"type": "openai"},
    {"type": "anthropic"},
    {"type": "google"},
    {"type": "openrouter"},
    {"type": "openai", "name": "proxy", "base_url": "https://llm-proxy.example.com", "api_key_env": "PROXY_KEY"}
  ],
  "watch": ["openai:gpt-4o", "anthropic:claude-sonnet-4-5", "google:gemini-2.5-flash", "openrouter:anthropic/*", "proxy:gpt-4o"],
  "alerts": {
    "min_severity": "warning",
    "webhook": {"url_env": "RADAR_WEBHOOK_URL"},
    "slack": {"url_env": "SLACK_WEBHOOK_URL"},
    "email": {"host": "smtp.example.com", "port": 587, "from": "[email protected]",
              "to": ["[email protected]"], "username_env": "SMTP_USER", "password_env": "SMTP_PASS"}
  }
}

Unknown fields are errors, and so are watch entries naming a source that isn’t configured, source names that differ only in case or punctuation, and secret variables that are named but not set.

What counts as what

Only watched models can be breaking or warnings. Every other change is info.

When a watched model disappears and a model with the same display name appears in the same scan, the report adds “possibly renamed to …”.

What each provider shows

What a failed scan is

A failed scan reports an error for that provider, changes nothing in its stored snapshot, and lets the other providers scan as usual. These are all failed scans:

• an HTTP error. 408, 429 and 5xx get three attempts in all, honouring Retry-After up to 30 seconds
• a body that isn’t valid JSON, is cut off, or has no model list
• an empty list
• a page of a paged list failing, or a paged answer where one page was expected
• fewer models than min_fraction of the last list
• a redirect to another host, refused so keys can’t follow it
• a timeout

Files

state_dir holds latest/SOURCE.json, the snapshot the next scan compares with, and the last 50 snapshots in history/SOURCE/. Each file is written whole or not at all. A lock file keeps two runs from using the folder at once; one older than 30 minutes is taken to be left from a crash.

RadarModel Schema

radarmodel-schema watches Postgres, MySQL, MariaDB and SQLite schemas. It shares the exit codes, the -min, -json, -no-alerts and -dry-run flags, the alerts block and the state folder with radarmodel-ai.

Schema commands

radarmodel-schema init   [-config FILE]                  write a starter config
radarmodel-schema check  [-config FILE] [flags] [DB...]  scan, compare, report
radarmodel-schema watch  [-config FILE] [flags]          scan every interval until stopped
radarmodel-schema accept [-config FILE] DB...            take the current schema as the baseline
radarmodel-schema diff   [-config FILE] DB1 DB2          compare two databases now
radarmodel-schema show   [-config FILE] DB               list the stored objects of a database
radarmodel-schema version

The config is radarmodel-schema.json in the current folder, or the file in $RADARMODEL_SCHEMA_CONFIG, or -config FILE. diff exits 0 when the two databases match and 1 when they don’t. When each side uses a single schema under a different name, such as app_staging and app, the names are lined up first.

Schema configuration

{
  "interval": "1h",
  "timeout": "2m",
  "min_fraction": 0.5,
  "ignore": ["*.tmp_*"],
  "databases": [
    {"name": "prod", "driver": "postgres", "dsn_env": "PROD_DATABASE_URL", "schemas": ["public", "billing"]},
    {"name": "app", "driver": "mysql", "dsn_env": "APP_MYSQL_DSN", "baseline": "pinned"},
    {"name": "local", "driver": "sqlite", "path": "app.db", "ignore": ["main.sessions"]}
  ],
  "alerts": {"min_severity": "warning", "slack": {"url_env": "SLACK_WEBHOOK_URL"}}
}

Adding an ignore rule never shows up as objects being dropped: the rules apply to the stored snapshot too.

What counts as what in a schema

What each database shows

SQLite foreign keys have no names, so they’re named after their columns and the table they reference, such as main.orders.fk_customer_id_customers. MySQL integer display widths (int(11)) are left out, so an upgrade past 8.0.19, which stopped printing them, isn’t a type change. When the server’s version changes, the report says so and notes it on the other changes of that scan.

When a schema scan fails

A failed scan reports an error for that database, changes nothing in its stored snapshot, and lets the other databases scan as usual. These are all failed scans:

• a lost connection, an unreachable server, a wrong password, a connection string that’s not set or can’t be parsed
• a scan longer than timeout
• a SQLite file that’s missing, isn’t a database or is a folder
• a schema that kept changing while it was read. Every scan reads twice and only accepts two identical reads, with three tries
• fewer tables than min_fraction of the last scan
• on MySQL, a table, column or trigger that vanished from information_schema but answers “access denied” instead of “doesn’t exist”

On Postgres, the catalog lists every object whatever the user’s privileges, so a revoked GRANT changes nothing. Each read sets its own search_path, IntervalStyle and DateStyle, so a role or connection string with other settings doesn’t change how definitions read.

Database privileges

On Postgres any role that can connect is enough. On MySQL and MariaDB:

GRANT SELECT, SHOW VIEW, TRIGGER ON app.* TO 'radar'@'%';

information_schema only lists what the user can access. With SELECT, a dropped table and a hidden one answer differently, so the first scan checks that every table can be read and fails if one can’t. TRIGGER is what lets MySQL list the triggers. If a privilege was taken away on purpose and the objects really are gone for good, radarmodel-schema accept DB takes the schema as it is now.