workflow-orchestration

name	workflow-orchestration
description	Model routing and structured-output rules for go-orca workflow personas (Director, Project Manager, Architect).

Workflow Orchestration Skill

Use this skill when configuring or improving go-orca workflow personas — especially Director model routing and Project Manager constitution JSON.

Director — persona model routing

The Director assigns persona_models for every downstream persona. Follow these rules when editing prompts/personas/director.md or tuning routing:

Persona	Model guidance
project_manager	Never assign sub-4B models (e.g. llama3.2:1b, *:1b, *:3b). PM output is strict JSON; small models emit invalid shapes (audience as object, spaced field names). Prefer the workflow default or ≥ 7B models (qwen3.5:9b, gpt-oss:20b, etc.).
architect	Mid/large models (≥ 7B) for design + task graphs.
pod, qa	Must use tools=yes models only.
matriarch	Mid-size, tools=yes when available.
finalizer	Largest available model for synthesis.

Do not document that PM should prefer sub-4B or “small models for planning” — that causes production failures on Ollama homelab catalogs.

Project Manager — constitution JSON contract

PM responses must parse into internal/state.Constitution and Requirements. When editing prompts/personas/project_manager.md:

• constitution.audience — plain string only (e.g. "general readers"). Never {"type":"general"} or other objects.
• Field names — exact snake_case keys: acceptance_criteria, output_medium, out_of_scope. No spaces in keys.
• out_of_scope — JSON array of strings, not a single string (engine normalizes common mistakes).
• acceptance_criteria — array of strings, not objects, unless the engine flex parser applies.
• Keep requirements at the top level (requirements.functional, requirements.non_functional), not nested inside constitution.

The engine runs normalizePMOutput for common drifts; prompts should still target the correct schema to avoid retry loops and 90m handoff waits.

Homelab / Ollama

When the default provider is Ollama, pin provider and model on workflow create (qwen3.5:9b or your cluster default). Director still chooses persona_models, but engine normalizePersonaModels swaps sub-4B models away from PM.