omdsh-dev/fabric

RFC: dsh-external-fabric — repository purpose, architecture, and decision record

• Status: living document (each section records the decision and its history)
• Scope: this standalone Fabric extension workspace
• Upstream anchors: deepseek-harness snapshots 7b9644f2 (0812) / 9f9e2782a4 (0813), fork tip 65bcaf9902 (feat-fabric)

This document explains why this repository is shaped the way it is. Every non-obvious arrangement below was reached through a concrete failure recorded in the commit history; the sections follow the repository’s evolution rather than its file layout.

1. Purpose: an external Fabric extension, not a fork

deepseek-harness is a private monorepo. The Fabric/Mixin extension layer lives there as three extension packages, but a consumer cannot install them from the registry. This repository externalizes exactly those three packages so they can be installed through the official plugin channel:

dsh plugin --profile <p> add https://github.com/omdsh-dev/fabric/releases/latest/download/pkg.tgz

Boundary (hard rule): the workspace ships exactly three complete packages — cordis-fabric (pure transformation service), cordis-fabric-api (pure compat facade), cordis-fabric-dsh (DSH-facing facades, invariant, profile bootstrap). Anything else — the official @deepseek-ai/dsh-tool-cordis toolset included — is never added as a fourth package; official packages remain upstream dependencies and are not republished here.

2. Host integration: launcher-provided wiring

The three packages install hooks and mount facades through the compiled launcher. src/fabric-dsh.ts compiles to lib/fabric-dsh.js, while src/fabric-dsh-preload.ts compiles to lib/fabric-dsh-preload.js; the launcher injects that compiled preload before the official CLI loads. No host patch checkout is required.

Everything the official channels already cover is deliberately excluded: installing the trio (dsh plugin add), bundle roster rows and dependencies, catalog generation, invariant/gate exemptions for trio-in-workspace, and all documentation (README*, docs/, .agents/). What remains is what no channel can provide: launcher bootstrap (apps/cli/src/profile-boot.ts calls installFabricBootstrap before any target import and checkFabricRequiredPatches after boot), the clientBundle source-transform build seam (packages/client/tsdown.client.ts), catalog entries compiled into the official tool-cordis package, their tests, and the pnpm-policy seams.

2.1 The disabled opt-in rows

The web-app bundle layer inserts cordis-fabric / cordis-fabric-dsh rows as disabled opt-ins: the pure cordis-fabric package is a library with no plugin apply, so an enabled row fails every boot (“invalid plugin”). A profile opts in by enabling the rows; the bundle layer applies on every boot, so pre-existing profiles are covered without edits.

2.2 The TSX dead end (recorded and reverted)

The dsh source launch (node --import tsx/esm apps/cli/src/bin.ts) once appeared to need TSX_TSCONFIG_PATH or a register preload: FiberState (a const enum, only in vendor/cordis/src) failed to resolve. Both workarounds shipped and were then reverted — the real cause was a stale TSX_TSCONFIG_PATH in the shell pointing at an old staging checkout. With a clean environment tsx auto-discovers the entry’s tsconfig (extending the base) and resolves the aliases to src. The official script runs unchanged; no host-specific workaround is required.

3. Install model: self-contained release bundle

The root bundle records the published trio tarballs and ships the prebuilt trio inside bundledDependencies:

https://github.com/omdsh-dev/fabric/releases/latest/download/cordis-fabric.tgz
https://github.com/omdsh-dev/fabric/releases/latest/download/cordis-fabric-api.tgz
https://github.com/omdsh-dev/fabric/releases/latest/download/cordis-fabric-dsh.tgz

This keeps the release installation to one package:

dsh plugin --profile web add https://github.com/omdsh-dev/fabric/releases/latest/download/pkg.tgz

pnpm does not need to resolve nested Git or URL packages. At launch, fabric-dsh asks DSH’s module-fallback healer to map the bundle’s own dependency closure into $DSH_HOME/profiles/node_modules, so the Profile and the bundled preload resolve the same trio copies.

• Host source installs declare the bundle in apps/cli/package.json; run the harness workspace’s pnpm install and pnpm run build, then install the release bundle through the plugin channel (joining cordis-fabric-bundle to dsh.profile.bundles) and enable the cordis-fabric-dsh row. Launches go through the compiled lib/fabric-dsh.js.
• Consumer-side builds use the explicit root build script. The trio and the launcher are built with tsdown before packing; no install-time prepare build is required.

3.1 pnpm 11 supply-chain seams

The self-contained bundle does not require blockExoticSubdeps: false, a Git prepare allowlist, or dangerouslyAllowAllBuilds in the Profile. The workspace still allows the native esbuild build and excludes the fast-moving DSH rc train from minimum-release-age checks:

• allowBuilds: esbuild in this workspace;
• minimumReleaseAgeExclude: ['@deepseek-ai/dsh-*'] — the dsh-* rc train ships inside the 24h window and a name-only entry exempts all versions.

4. Registry dependency policy

The dsh-* host packages publish fast rc trains; this repository tracks them through registry ranges, and each lesson below came from a real breakage.

4.1 The dsh-compact trap

@deepseek-ai/dsh-client-runtime@0.0.1-rc.1 depended on @deepseek-ai/dsh-compact, which was never published (upstream deleted the package after publishing that runtime). The 0.1.0-rc.x series dropped the dependency; verified installable end-to-end.

4.2 The missing rc.5

Upstream code is versioned 0.1.0-rc.5, but the registry jumps rc.3 → rc.6 — rc.5 was never published. Ranges therefore read ^0.1.0-rc.0 (resolving the newest published rc, and rc.0 keeps stable releases in range too). Peers use the same range, which the host workspace’s rc.5 satisfies — host installs reuse workspace packages instead of registry copies.

4.3 Real host types, not a local contract

The trio once declared a host-contracts.ts facade plus a global @deepseek-ai/cordis Events injection. That broke type-checking across host packages and was deleted in favor of importing the real @deepseek-ai/dsh-* types (declared as peers + devDeps) — exactly the upstream shape. ctx.slots typing comes from dsh-client-runtime’s declaration, as upstream.

4.4 Runtime peers of the published libs

With autoInstallPeers: false, the published dsh-* libs’ load-time imports (dsh-scope, dsh-llm, dsh-timeout, dsh-typert-protocol) must be listed as devDependencies explicitly — each was added after a “Cannot find package” at test load.

5. Browser client format: the closure factory

The web shell loads /plugins/<id>/client.js as a classic script and resolves value imports through the loader module table (a synchronous require inside the factory). Plain ESM bundles cannot load there at all. Consequently both trio browser halves ship as closure factories:

window.__ModuleLoader__.load({ id: "cordis-fabric", factory: (require) => { ...; return module.exports; } })

with @deepseek-ai/cordis external (a platform seed) and everything else inlined. cordis-fabric was converted first; cordis-fabric-dsh followed (the same gap, fixed after the ex-setting install exposed the first one). Upstream never notices this — its monorepo builds both through the shared clientBundle() preset.

5.1 ex-setting’s three lessons (same contract, external repo)

The sibling omdsh-dev/ex-setting bundle hit the same contract three times:

1. Its dsh.client manifest must be nested ("dsh": { "client": ... }), not a top-level dshClient field — client-modules scans the nested form;
2. its consumer-side build must use the prepare config, not just the local one, or git installs serve the old artifact;
3. cross-bundle value imports must not rely on a disabled row’s factory — ex-setting inlines/avoids what the module table cannot answer, and installs static styles directly instead of routing them through a Fabric publish the transform could not produce (browser-transform cannot match inside the closure artifact).

6. Test strategy

The upstream suite resolves src through tsconfig paths; this repository only has registry lib artifacts, which drove the evolution below.

• serve.spec uses a test-local node:http adapter for the host webServer service, so exact/prefix routing and real HTTP responses stay covered without a DSH host-webserver test dependency.
• hmr-e2e-runner drives config HMR by toggling the row’s disabled flag in cordis.yml: the vendored fork’s hmr.registerConfig and include internal/update are fork-private and exist in no registry version (verified against latest 1.0.16/1.0.6).
• client specs originally faked CommandUiRuntime/SlotRegistry because the runtime rc.1 tree was uninstallable and the bundles are closure factories. After rc.6 became installable the real reason remained the factory format, so the specs now mount the real services through a test module loader (packages/cordis-fabric-dsh/tests/browser/module-loader.ts): happy-dom provides window; the __ModuleLoader__ sink installs at helper module load; platform seeds (cordis, ui-slots, react) preload as ESM namespaces (the factory require is synchronous and node cannot require ESM); ui-primitives — a render-only heavy package — is stubbed; materialize() executes a factory with the module-table require (recursing into other registered bundles, memoized, stripClientSuffix normalizing pkg/client). Loader baseUrl and fixture URLs are pinned to file paths because happy-dom’s location is http://localhost:3000.

7. Timeline (abridged)

8. Future work

• If the registry ever publishes node-importable builds (plain ESM or the src halves), the test module loader disappears and the specs import packages directly.
• Upstream promoting createSnapshotStore out of dsh-client-runtime shrinks the seed table.
• Upstream publishing hmr.registerConfig / internal/update would let the HMR runner mirror the in-tree config flow again.