Demo: real-world use cases
A runnable script lives at
examples/demo.sh:
bash examples/demo.sh
It needs Bun (the dev runtime) and this repo; the optional eval step also needs an
authenticated claude CLI. Everything else runs offline. Below is the actual output,
annotated.
1. A solo dev ships one capability to every harness
You author a plugin once (weft.yaml + a SKILL.md), with no per-harness JSON, and compile
it to all five harnesses. install only touches the harnesses you actually have.
com.acme/code-helper: valid (1 components)
Built com.acme/code-helper -> out/
claude: 3 files (catalog at out/claude)
codex: 4 files (catalog at out/codex)
cursor: 3 files (catalog at out/cursor)
copilot: 3 files (catalog at out/copilot)
opencode: 2 files (catalog at out/opencode)
skipped codex: harness not detected
skipped cursor: harness not detected
skipped copilot: harness not detected
skipped opencode: harness not detected
Installed com.acme/code-helper@0.1.0 (project)
2 artifacts placed
2. A company curates an internal marketplace
One marketplace.yaml packages many plugins into each harness’s native catalog. This is the
platform-team workflow.
Built marketplace acme-tools (2 plugins) -> market/
claude: 5 files ...
Claude catalog lists every plugin:
"name": "code-tools", "source": "./plugins/code-tools",
"name": "weather-tools", "source": "./plugins/weather-tools",
3. Trust before you ship
The publish gate runs static validation + a security scan + the deterministic eval tier, and computes badges. Signing makes the compiled artifacts tamper-evident.
valid: yes
scan: clean
badges: valid, scanned
Publish gate passed.
Signed 2 artifacts -> weft.sig (key kept in .weft/, public key in weft.pub)
signature: valid
tampered artifacts: 0
signed badge verified.
Now tamper with a placed file and re-verify; it must fail:
verify correctly FAILED (tamper detected)
4. Enterprise managed mode
The same mechanism a solo dev uses, restricted by policy: an org pins installs to its own namespaces, and a foreign namespace is blocked.
install correctly BLOCKED (com.acme not in the com.enterprise allowlist)
5. Evals against the REAL harness
The headline of the eval system: drive the actual headless Claude and assert over what it
did. Here Weft installs the code-review skill into a scratch project, asks Claude to
review a file with a planted bug, parses the live stream-json trace, and checks that
Claude read the file and named the bug. It’s a real pass, not a mock:
Eval: com.acme/sample-plugin:code-review
claude: PASS
- reads-before-reviewing: pass
trace: pass (1/1 samples passed (minPassRate 1))
output: pass (1/1 samples matched)
A harness with no available CLI is reported UNTESTED rather than faked. Copilot,
which has no structured trace, degrades its trace assertions to output.