Security

View Source

Jido validates data and owns execution processes. The host application still owns authorization, effects, secrets, and resource policy.

Keep Source And Stored Data Separate

The module DSL is trusted compile-time Elixir source. Do not evaluate user or generated Elixir source to create a Flow.

Use Jido.Flow.Codec for database, API, UI, or AI-authored Flow data. The Codec accepts decoded map data and a host-owned trusted Registry.

Stored data cannot create atoms or derive module names. It can select only the Actions, Flows, schemas, and atoms that the Registry contains.

Validate Before Effects

Use Action schemas to reject malformed input before run/2. A schema is not authorization. Check tenant access, object ownership, and current permissions at the application boundary or inside the Action before an effect.

Make file, network, database, and process effects clear in Action names, schemas, and tests. Give an Action only the capabilities that it needs.

Treat context as sensitive caller data. Do not copy secrets into Flow results, error details, component metadata, or telemetry.

Apply Storage Limits

The Codec decoder rejects:

  • invalid UTF-8;
  • depth greater than 100;
  • one collection wider than 10,000 entries; and
  • more than 100,000 data nodes in one document.

These checks occur after the caller decodes JSON. Apply HTTP byte limits, parser limits, and request timeouts before Codec.

Registry size is limited to 10,000 entries. Registry lookup is inert. It does not load or execute an Action.

Apply Runtime Limits

Use a finite complete-call timeout to bound execution while its controller is alive. Synchronous supervisor startup and telemetry handlers can delay timeout and cancellation responses. This is not a hard response-time limit. See Process Ownership for these limits. Use max_concurrency to bound one concurrent Flow execution. Also validate collection sizes in application input. Runtime Map does not use the Codec collection limit.

Each Iterate has a bound from 1 through 10,000. Select a smaller application limit when body work is expensive.

The owner of an async handle can cancel active work with Jido.Exec.cancel/1. Cancellation is in-memory and cannot undo completed side effects. This package does not provide automatic retry, per-node timeout, durable cancellation, durable checkpoints, or exactly-once effects.

Design Effects For Repetition

A caller, process restart, or higher-level runtime can repeat work. Use idempotency keys, conditional writes, deduplication, or transactions when an effect must not occur twice.

Step-wise stale-revision checks stop one old Execution value from dispatching again. They do not create a durable exactly-once guarantee.

Protect Errors And Telemetry

Errors can contain module names, node names, details, and in-memory stacktraces. Telemetry contains execution, Flow, node, target, item, and iteration identifiers.

Redact errors before external logging. Do not place credentials, tokens, personal data, or full context maps in error messages or telemetry metadata. Treat telemetry handlers as a data-access boundary.