machina-sports/skills/manifest-generator

Project Manifest Generator

Generates a draft project.manifest.yml for a Machina template by statically scanning a list of workflows and aggregating the credentials, connectors, datasets and agents they reference.

Built to scale Sprint 1A of the Pipeline Platform Cleanup (the hand-written botandwin manifest at entain-templates) to every template in machina-templates without writing each one by hand.

What it does

1. Deterministic extraction (always) — a pyscript connector walks each workflow_object in your list and emits the union of:

   • TEMP_CONTEXT_VARIABLE_* → credentials
   • task.connector.name for type=connector|prompt tasks → connector deps
   • task.config.action == "search" → dataset reads (external deps)
   • task.config.action == "update|insert|save" → dataset writes (deps_on)
   • type=agent references → agents
   • type=workflow references → workflow calls (for extends-like inheritance)

2. Optional LLM enrichment — pass enrich_with_llm=true and a Gemini/Vertex AI call fills in:

   • source_label ("OpenAI API key (platform.openai.com)" etc)
   • test_workflow ("<connector>-test-credentials")
   • validation rule (http for *_API_KEY, json_object for service-account JSON, non_empty_string for generic strings)
   • dataset description + populated_by guesses

3. Output — writes a <template_name>-manifest-draft document into the document collection. Operator opens it in Studio, eyeballs the suggested values, edits, then copies the manifest field into the real project.manifest.yml for commit.

How to run

Install this skill into your project (via Studio template browser), then:

POST /workflow/executor/generate-project-manifest
Content-Type: application/json

{
  "workflow_names": [
    "machina-assistant-thread-create",
    "machina-assistant-thread-respond",
    "machina-assistant-kb-search"
  ],
  "template_name": "machina-assistant",
  "description": "Foundational AI assistant for Machina platform",
  "enrich_with_llm": true
}

Response carries manifest_draft_doc_name. Fetch with:

POST /document/search
{ "filters": { "name": "machina-assistant-manifest-draft" }, "page_size": 1 }

Limitations (V1)

• Doesn't fetch workflow YAMLs from a Git repo — works against the workflows already imported into your project. To scan a template you haven't installed yet: import it first.
• Heuristic dataset detection (same caveats as core/workflow/dependency_graph.py — $.get('xxx') is mostly inferred from type=document tasks now, not just regex).
• LLM enrichment is best-effort. Always review the draft before commit.

Roadmap

• V2: clone-from-git input so you can manifest a template without installing it.
• V3: bulk mode — scan an entire repo and emit one PR per template.
• V4: hook into core/dataset/controller.py:process_install_file so installing a template auto-generates + offers a draft manifest.

Related

• agent-templates/machina-assistant/project.manifest.yml (hand-written reference)
• docs/project-manifest-guide.md (schema reference)
• machina-client-api /project/health, /bootstrap-check, /workflow/<id>/dependencies