Roadmap
What is designed but not built, what is built but not verified everywhere, and what is not planned.
shipped
What the repository build does today is in the changelog; check npm view sealmark version for what is published. This page lists what is not built, what is built but not verified everywhere, and what is not planned. No dates are promised.
Designed, not built#
| Item | What it would do | Status |
|---|---|---|
| DNS rebinding probes | Test whether a URL tool re-resolves a name to a loopback address. Needs a DNS zone that answers differently per query, which the harness does not own. | planned |
| DuckDB fixtures and grading | The sqlite pack's ideas on a DuckDB file. {duckdb_path} is a placeholder only; the engine is an optional native dependency that is not installed or tested. | planned |
| MySQL | A database option that takes a DSN the way --pg does. There is no embedded MySQL for Node, so it would need a service. | planned |
| A badge service | A hosted badge for a published version of a server, with rules for who may publish and when a result is held. | planned |
| Marketplace listing of the action | Needs the action in its own public repository. | planned |
| Prompt-argument probing | Call prompts/get with malformed arguments. Today prompts are listed and scanned as text only. It needs a prompts/get path in the client and a decision on a prompts surface. | planned |
Built, not verified everywhere#
- Windows. The cleanup paths that kill a whole process tree, the
cmd.exequoting in the launcher and the--fixture fsjunctions were written for Windows and have not been run on it. Runnpm teston a Windows machine before relying on them. - The GitHub Action upload step has not run on a hosted runner: it works only once the repository is public. See Use the action.
Not planned#
- Prompt injection carried in tool output. Detecting it needs a model in the loop, and Sealmark has no model in the loop by design.
- Authentication flows that need real credentials. They are out of scope.
- Supply chain and install-time behavior.
The full list of what the tests do not cover is in the known blind spots of the methodology.