lifecyclemodel-recursive-orchestrator（生命周期模型递归协调器）

Lifecycle Model Recursive Orchestrator

Use this skill when one request needs more than a single builder call. The orchestrator validates a request, resolves each node to reuse/build/cutoff, invokes the native tiangong lifecyclemodel orchestrate actions, and writes a full local run directory.

What The Implementation Does

• validates assets/request.schema.json
• normalizes root, optional extra nodes, and dependency edges
• chooses per-node resolution in this order:

- reuse existing resulting process - reuse existing process - reuse existing lifecycle model - build with lifecyclemodel-automated-builder - build with process-automated-builder - unresolved / cutoff

• records invocation results for native CLI builder slices
• emits:

- assembly-plan.json - graph-manifest.json - lineage-manifest.json - boundary-report.json - publish-bundle.json via publish

Request Shape

Required top-level fields:

• goal
• root
• orchestration
• publish

Optional top-level fields:

• candidate_sources
• nodes
• edges
• notes

The root block may directly carry:

• candidate lists
• process_builder
• submodel_builder
• projector

Supported process_builder fields are intentionally narrow:

• flow_file
• flow_json
• run_id

Removed legacy fields such as mode=langgraph and python_bin are not part of the supported request surface anymore.

Use extra nodes only for additional dependencies or subsystems beyond the root node.

Downstream Builders

• process_builder reuses the same native slice as tiangong process auto-build
• submodel_builder reuses the same native slice as tiangong lifecyclemodel auto-build
• projector reuses the same native slice as tiangong lifecyclemodel build-resulting-process

The wrapper does not call other skills directly. It delegates to tiangong lifecyclemodel orchestrate, and the CLI orchestrator owns request normalization, node resolution, invocation ordering, and final manifests.

Current limitation:

• candidate discovery must already be resolved into the request or into local artifacts upstream
• the submodel builder path only covers the CLI-backed auto-build slice
• validate-build and publish-build for lifecyclemodels now exist in the CLI, but this orchestrator does not invoke them yet

Typical Workflow

1. Read assets/request.schema.json and prepare a manifest.
2. Run plan and inspect the generated assembly-plan.json.
3. Run execute to invoke native CLI builder slices and persist invocation logs.
4. Run publish only to prepare a local handoff bundle. Remote writes remain outside this skill and should later go through tiangong publish run.

Commands

node scripts/run-lifecyclemodel-recursive-orchestrator.mjs plan \
  --request assets/example-request.json \
  --out-dir /abs/path/run-001 \
  --json

node scripts/run-lifecyclemodel-recursive-orchestrator.mjs execute \
  --request assets/example-request.json \
  --out-dir /abs/path/run-001 \
  --allow-process-build \
  --allow-submodel-build \
  --json

node scripts/run-lifecyclemodel-recursive-orchestrator.mjs publish \
  --run-dir /abs/path/run-001 \
  --publish-lifecyclemodels \
  --publish-resulting-process-relations \
  --json

Load References On Demand

• references/workflow.md: stage-level orchestration flow and policy knobs
• references/data-model.md: graph/lineage fields and version contract
• references/invocation-contracts.md: exact handoff contract to each downstream skill
• references/minimal-relation-fields.md: minimum relation tuple to preserve