Skip to main content

Execution Classes

Every skill that carries scripts is assigned an execution class describing what it needs at runtime. The classification is derived by a script, not asserted by hand, and a --check mode reports local source drift. The classifier is a heuristic, not runtime acceptance.

bash skills/process/adversarial-review/scripts/classify-mobile-execution.sh
bash skills/process/adversarial-review/scripts/classify-mobile-execution.sh --check

The classes​

ClassNameNeedsMobile path
—manifest-onlyNothing — no scripts/Instructions; declared tools still required
E0Process-boundbash, python3, a binaryNone on-device; use remote execution
E1Capability-boundFilesystem and/or clockWasm component with granted capabilities
E2PureComputation onlyWasm component, no capabilities
RRemoteA full desktop environmentExplicit authorized remote executor required

Historical classification snapshot​

ClassCount
manifest-only249
E028
E118
E22
R13
Total310

These historical counts do not certify current skills as mobile-ready. A manifest-only skill is instructions a model reads; there is nothing to execute and therefore nothing to port. Instruction-only workflows may still require unavailable tools, data or services. Current counts come from the final generated inventory.

E1 was wrong for every member — a lesson​

E1 was originally defined as "pure text/JSON transformation." In practice it was the residual: whatever E0, E2, and R did not match fell into it.

When a skill was later needed for a Wasm port, an audit found all 18 of 18 E1 members touch the filesystem or the clock. Not one was a pure transformation. The residual had silently absorbed every skill no other rule matched, and then presented itself as a positive finding.

Two things make this worth remembering:

  1. A --check drift test cannot catch it. Drift checks compare the committed file to a freshly generated one. Both come from the same wrong rule, so they agree perfectly and the check passes forever.
  2. The risk was written down and shipped anyway. The script's own header said E1 was "the class most likely to be wrong, which --check cannot detect."

Corrective action: E1 now carries an explicit needs_capabilities field. A skill labelled portable must state its price.

Best practice

Never let an else-branch stand as a verdict. If a class is defined as "everything left over," hand-verify a sample before trusting it — the automated check that would normally catch drift is structurally blind to this failure.

Reading the output​

{
"skill": "some-skill",
"class": "E1",
"needs_capabilities": ["filesystem", "clock"],
"evidence": "scripts/run.sh reads $PWD and calls date"
}

evidence names the file and construct that drove the decision, so a classification can be argued with rather than merely trusted.

What to do per class​

manifest-only — nothing. This is the target state; prefer it when authoring.

E2 — port to a Wasm component against prometheus:component. Two skills qualify today.

E1 — portable, but the host must grant capabilities. Decide whether the capability is essential or incidental; a skill that calls date for a log line can often drop the dependency and become E2.

E0 — no on-device path. Either rewrite as manifest-only, or accept remote execution.

R — requires an explicitly authorized remote executor and destination result evidence. A paired sync transport is not itself an executor. See current portability limits.