Skip to content

skill-construction.md

Latest commit

 

History

History
48 lines (30 loc) · 3.6 KB

skill-construction.md

File metadata and controls

48 lines (30 loc) · 3.6 KB
Raw

Constructing skills from local resources

The public repository includes the construction workflow and common protocol. It does not include the paper's collected corpus or authored skill packages. Supply your own local resources and keep generated corpora outside the checkout.

An embodied skill is a package ⟨m, I, S, R, A⟩: discoverable metadata (m), operational instructions (I), scripts (S), references (R), and assets (A). SKILL.md holds metadata and instructions; optional resources go into scripts/, references/, and assets/.

1. Collect modules

Copy inputs/sources.json to your external workspace and edit workspace, source patterns, extra files, and paper_zip to point to resources you have locally. The supplied paths illustrate the collection recipe; they are not directories distributed in this public repository. Remove inapplicable entries.

python scripts/01_collect.py \
  --config ../esb-runtime/sources.json --output ../esb-runtime/corpus

The collector creates three-field cards (type, description, source), inventory/provenance records, and deduplicated source payloads stored by SHA256. A paper or repository locator records provenance; it does not imply that the upstream full text or code was retrieved. Model descriptors retain external weight/runtime dependencies.

2. Structure components and contracts

python scripts/02_structure.py --data ../esb-runtime/corpus

The common protocol maps control and perception interfaces to scripts; method references and experience material to references; and model descriptors/configuration to assets. Native bindings retain input schemas, dependencies, coordinate conventions, model profiles, and execution budgets. Source fragments requiring integration remain marked as source-only.

3. Prepare the authoring handoff

Adapt inputs/request.json to the operational purpose. It is an example request, not a pre-authored skill.

python scripts/03_assemble.py --data ../esb-runtime/corpus prepare \
  --request ../esb-runtime/request.json

Give prompts/assemble_skill.md and the generated 03_handoff.json to the authoring agent, with access to the corpus. The agent selects compatible modules and writes applicability, invocation, feedback interpretation, verification, and bounded recovery/stopping rules. Its JSON plan must satisfy protocol.json:assembly_plan_schema.

4. Materialize the plan

python scripts/03_assemble.py --data ../esb-runtime/corpus materialize \
  --plan ../esb-runtime/plan.json --repo ../esb-runtime/authored-skills

Materialization writes SKILL.md, copies selected payloads and declared dependencies, relocates links, verifies hashes and local references, and registers the result in an assembly index.json. Existing packages are not overwritten. Component checks do not establish physical success or implement a missing robot adapter.

The assembly index.json is distinct from the runtime skills-index.json. Runtime discovery additionally requires each package's metadata.json and its execution/interaction contract; do not assume an arbitrary materialized source-only package is runnable in the simulator. Consult catalog.py and the selected native provider when integrating a new package.

For resource access, use scripts/access.py --data ../esb-runtime/corpus --request path/to/request.json. read returns a resource and prepare checks arguments and returns a native call descriptor. Programmatic invoke requires a caller-supplied runtime dispatcher.