Analysis Playbooks and Composition

Minara retired its standalone Scenario Registry in May 2026. Request routing, analysis guidance, and learned investment rules now live in separate systems. Each can evolve without turning one global scenario catalog into a bottleneck.

The current model

Skills own request routing

A Skill describes what it does, when it should appear, and which tools it may expose. The registry loads direct SKILL.md packages, filters missing environment requirements, and gives the router a compact catalog. The active SkillSession controls what guidance and tools enter the current turn.

This keeps ordinary routing close to the capability itself. Adding a new class of user request usually means improving a Skill's description, references, or tools. It does not require a central scenario definition.

Institution playbooks own analytical substance

Institution playbooks are stable, named bodies of research guidance. A role can receive different playbooks for stocks, digital assets, indices, commodities, or other asset classes. roles.ts maps built-in roles to playbook IDs, and tests verify every reference.

A playbook states what evidence to collect, how to weigh disagreement, and how to proceed when live data is incomplete. Reusable thresholds come from the methodology store through {{methodology:...|default:...}} placeholders. Universal safety and citation rules stay in the shared system prompt.

Roundtable pipelines own collaboration

The Roundtable Builder composes analysis through five phase types:

1. context establishes constraints and prior decisions.
2. parallel lets independent specialists collect evidence without anchoring one another.
3. debate runs multiple rounds of explicit disagreement.
4. synthesis gives one manager responsibility for reconciling the record.
5. final produces the decision and monitoring plan.

Each pipeline stores its phase order, Agents, prompts, skills, loop limits, and budgets. Validation requires unique IDs, at least two Agents in a debate, one manager for synthesis and final phases, and exactly one final phase at the end. A run stores a pipeline snapshot so later edits do not rewrite history.

Users can build and save their own Roundtables without changing the repository. Built-in Agent, phase, and Roundtable templates live in builder-templates.ts and serve as reviewed starting points.

Deep Research keeps its own report contract

Deep Research runs a fixed six-step outer flow for understanding, planning, collecting, analyzing, visualizing, and writing. Topic-specific guidance lives in references/scenarios.md. These references shape source selection and synthesis while the Skill keeps progress reporting and final artifact behavior consistent.

Methodologies remain independent

A methodology is a standalone investment rule with evidence, scope, exclusions, and an auditable lifecycle. Playbooks may use graduated methodologies through a placeholder or methodology_lookup, but they do not own the rule's confidence or outcome history.

This separation prevents a successful rule from being trapped inside one workflow. It also prevents a weak learned rule from entering an Institution prompt before it passes the graduation gate.

Contributor guidance

• Adding a Skill
• Adding an Analysis Playbook
• Adding a Methodology
• Institution Mode