version: "1.0.1"

name: request-routing description: > Route SAS Viya requests for finding resources,reading tables, running queries, scoring MAS/job/jobdef/SCR/cas models, listing resources, and describing metadata. Use when the task involves choosing the correct SAS Viya tool or deciding whether to verify, read, query, score, list, or describe a resource.

Request Routing — Canonical SKILL

Purpose: single source-of-truth for routing SAS Viya actions (read, query, score, describe, list).

Dotted Resource Reference Parsing — a.b Notation

When any resource reference appears in the form a.b, parse b to determine the resource type:

Rule: if b is not one of {mas, job, jobdef, scr, sas, casl}, treat a.b as a table reference where a is the library name and b is the table name.

Examples:

• Public.customers → table, lib=Public, table=customers
• SASHELP.cars → table, lib=SASHELP, table=cars
• churnRisk.mas → MAS model, name=churnRisk
• simplejob.job → Job model, name=simplejob
• myScorer.jobdef → JobDef model, name=myScorer
• loanModel.scr → SCR model, name=loanModel

Quick workflow

• Verify — confirm resources exist (use find-*).
• Execute — run the mapped execution tool (read, query, score, describe, list).
• Format — return results and append a short Strategy Summary.

Important Reminder: If the category is "Find resource" do not use the list- rules below, use the specific find- tool for that resource type. The list-* tools are for discovery when the user does not have a specific resource in mind.

Classification

Verification rules

• Always verify resources using the appropriate find-* tool before executing actions

Execution rules

• Execute only after verification; choose the execution tool per the Execute mapping section.
• Scoring flows:
• Inline scenario: verify model → call scoring tool.
• Table rows: verify model + table → read rows → map columns to model inputs → score → merge predictions with rows.
• Read/query flows:
• Use sas-score-read-table for all row reads, including those with a WHERE filter. A filter condition is not a reason to use sas-query.
• Use sas-score-sas-query only when the request requires SQL aggregation (COUNT, SUM, AVG, MIN, MAX), GROUP BY, JOIN across tables, or computed columns.
• When mapping between table columns and model inputs is ambiguous, ask the user for explicit mapping.

Defaults & exceptions

• Default model type: MAS unless the user specifies otherwise.
• Skip find-* verification for SCR models; SCR endpoints may be scored directly.
• If server determination is ambiguous, prompt the user for clarification.
• Pagination: always pass start=1 and limit=10 when calling any tool that accepts these parameters, unless the user specifies different values. Never omit them and rely on tool-level defaults.

Execute mapping (concise)

• Read rows (including WHERE filter): sas-score-read-table (lib, table, server, where)
• SQL aggregation/join only: sas-score-sas-query (lib.table, query, sql)
• MAS scoring: sas-score-mas-score (mas, scenario)
• Job scoring: sas-score-job-score (name, scenario)
• JobDef scoring: sas-score-jobdef-score (name, scenario)
• SCR scoring: sas-score-scr-score (name, scenario)
• Program scoring: sas-score-program-score (src, scenario, folder, output, limit)
• CAS Program scoring: sas-score-cas-program-score (src, scenario, folder, output, limit)
• Describe: sas-score-*-describe (mas/job/jobdef/scr/table)

Combined Read + Score (short)

1. Verify: find table (server) and find the model (mas,job,jobdef,scr). If table/model not found, ask user to clarify resource and server.
2. Read: fetch rows (read-table or sas-query).
3. Map: ensure table columns match model inputs; ask for mapping if needed.
4. Score: call the appropriate scoring tool.
5. Merge: attach predictions to rows and return.

Strategy Summary (append to responses)

Strategy Summary:

• Classification: [Find / Read / Score / List / Describe]
• Verification: [Resources verified / skipped]
• Tool(s): [Primary tool(s) invoked]
• Decision: [Server chosen, model type, mapping]
• Next steps: [Follow-ups or clarifications]

Error handling (short)

• Resource not found → ask for exact resource name and server (for tables).
• Column/input mismatch → request mapping from user.
• Empty result → ask to relax filters or confirm criteria.
• Execution error → return tool error verbatim.

Examples (minimal)

• Read: "read customers in Public" → find Public (CAS) → read-table → return rows.
• Score inline: "score a=1,b=2 with job simplejob" → find job → job-score → return merged result.
• Score table: "score Public.customers with model risk" → find table (CAS) & model (MAS) → read rows → score → return merged.

Notes: Keep this SKILL as the canonical, compact router; agent wrappers should be short and reference this document for details and examples.