| name | operate-modly-asset |
| description | Run one reviewed Modly extension for one approved M5 asset specification through modly-cli-mcp under GPT orchestration. Use for the optional local route only after capability research, not direct model calls, selection, refinement, or QA. |
Operate a Modly asset extension
Scope
Delegate one bounded local 2D, 2.5D, or 3D production operation through
modly-cli-mcp. This route is enabled only after the OpenAI flow and neutral
contracts are established.
Inputs
- Explicit
FORGE_ROOT and external WORLD_ROOT.
- Exact world/target/inventory/specification hashes and authorized references.
- Reviewed evidence that one named Modly extension/revision supports the exact
operation, input/output contract, model, weights, licenses, and platform.
- A strict request with orchestrator
gpt, route modly, executor
modly_cli_mcp, and a live-discovered capability_execute, workflow_run,
or process_run operation. Before execution it binds the hash-addressed
discovery snapshot, canonical workflow_run or process_run surface,
CLI-MCP/Modly versions, capability ID, exact extension identity and hashes,
exact model/weights identity, reviewed setup, and bounded arguments.
Outputs
- Hash-addressed authoring candidates returned by the reviewed extension.
- Sanitized successful Modly receipt containing CLI-MCP/Modly versions,
canonical surface, capability/run identity, supported discovery snapshot hash,
extension version/revision/manifest/workflow hashes, model/version/weights
hash, reviewed setup, replayability, outputs, and parent receipt hashes.
Invariants
- GPT remains the orchestrator;
modly-cli-mcp is the only permitted local
model bridge and the named extension is the executor capability.
- Refuse unknown/unavailable capabilities, floating revisions, unverified
weights, unreviewed setup, undeclared downloads, incompatible licenses,
arbitrary code, and direct model invocation.
- Treat the
0.1.1 wrapper baseline narrowly: TripoSG and Hunyuan3D
image-to-mesh, mesh optimizer, and default-path mesh exporter are the only
capability_execute operations currently known as executable. UniRig,
unknown/UI-only capabilities, generic chains, and explicit exporter paths are
unavailable; 2D/2.5D requires a separately discovered reviewed extension.
- Require every receipt identity and discovery hash to match the pre-execution
request exactly; runtime
run_id and supported state do not authorize a
different extension, model, workflow, or setup.
- Keep tokens, local paths outside scope, raw transcripts, configuration, models,
weights, workflows, and extensions out of bundles and games.
- Do not select, refine, process, approve, or package the returned candidate.
Completion
Complete when the exact extension contract was reviewed before execution and the
request/receipt plus all candidate hashes validate. Otherwise fail closed and
record the missing capability evidence.