# Compatibility

## Declared peer range

```text
@deepseek-ai/cordis        >=4.0.1-rc.1 <5
@deepseek-ai/dsh-tools     >=0.1.0-rc.1 <0.2.0
@deepseek-ai/schemastery   >=3.18.0 <4
Node                       ^22.19.0 || >=24
```

A declared range is not the same as a tested matrix.

## Release matrix

| Package version | DSH line | Node | Unit/runtime adapter | Real profile install | Headless run | Nested dispatch |
|---|---|---|---|---|---|---|
| 0.1.0-next.0 | 0.1.0-rc.6 | 24.11.1 local; 22.19.0/24.x GitHub CI pass | 32 tests pass | tarball and pinned GitHub source add/repeat/dump pass; GitHub uninstall/reinstall pass with pnpm 11.7.0 | read-only dry-run preview pass | real rc.6 registry test passes |

Evidence date: August 14, 2026. The headless run used an isolated DSH home, a synthetic read-only file, one observed dry-run decision, no injected fault, and an unchanged successful tool result. The GitHub-source gate pinned commit `30221b008cd4b7a52291cdffbf8e26ce80180b2c`; pnpm required the exact codeload package key under profile-local `allowBuilds` before running the Git dependency's `prepare` script.

Update this table only with stored CI or release evidence.

## Developer-preview policy

DSH explicitly allows breaking changes during developer preview. This project therefore:

- publishes prereleases until the consumer matrix is green;
- pins exact DSH versions in compatibility CI;
- keeps the runtime adapter thin;
- treats public event/type changes as release-significant;
- records unsupported lines honestly rather than broadening a peer range speculatively.