epam/timebase-agent-plugins/qql-generator

QQL Generator

Mission

Generate QQL that is task-correct, schema-grounded, and safe to execute. Start from user intent, compose only the needed concept modules, and never treat parser success as semantic correctness.

How to Route Requests

1. Identify the request type:

• build a new query,
• repair a broken query,
• create/alter/modify/drop schema,
• explain QQL behavior.

1. Detect data shape needs:

• polymorphic/PackageHeader/array payloads,
• simple fixed-schema tabular projections.

1. Open the smallest set of relevant references:

• start with one main workflow file,
• add concept references only when needed,
• use recipes only as compact starters.

1. Decide whether Python is needed:

• Stay in this skill when the deliverable is QQL, a small MCP execution preview, or a direct answer from QQL plus MCP.
• Do not switch to the Python client skill when the final Python would only connect and execute the same QQL and return the same result.
• Escalate to Python only when the deliverable changes: local file persistence, large result extraction, reusable script, app/service, or downstream Python computation or visualization.

Mandatory Policy

• Ground schema-dependent QQL from MCP schema/message evidence when available.
• If MCP is unavailable or unreachable, include this exact warning in user-facing output:

Warning: TimeBase MCP is not available, results are expected to be significantly worse than MCP-grounded output.

• compile_query is parser-only (syntax/parse diagnostics). Compile success does not prove class/field/type or other semantic correctness.
• execute_query is non-default and potentially mutating. Use it only when explicitly needed/requested by the user.
• When the user asks which QQL functions are available, or a function call fails because the function may be unknown or unsupported on the connected server, inspect server capabilities before proposing a fix. Prefer the TimeBase MCP list_qql_functions tool (v0.2.0+) when available, if the installed MCP server is older and lacks that tool, fall back to QQL introspection via stateless_functions() and/or stateful_functions(). Treat the connected server as the source of truth for availability. See references/mcp-workflow.md and references/concepts/stateless-functions.md.
• Never invent final schema facts (classes, fields, enum values, entry types). Ask for schema or clearly mark assumptions.

Style and Safety Rules

• Quote stream/class/field identifiers with double quotes.
• Use single quotes for string literals.
• Use == for equality and === for strict equality.
• Stateful functions use braces (avg{}(...)); stateless functions do not (abs(...)).
• Keep clause order valid (WITH, SELECT, FROM, ARRAY JOIN, OVER, WHERE, GROUP BY, HAVING, LIMIT, UNION).
• OVER TIME(...) / OVER COUNT(...) must appear before WHERE.
• Do not use SQL-only constructs (ORDER BY, SQL JOIN, SQL window OVER (PARTITION BY ...), CREATE TABLE).
• Do not declare built-in identity fields (timestamp, symbol, type) in DDL.
• Never output schema-dependent queries when schema is unknown.
• Distinguish parser validity from semantic validity in notes.
• Keep destructive DDL and confirm modes explicit and user-intent driven.

Pre-Output Safety Gates

Before finalizing any answer, verify all applicable gates:

1. Relevant references are loaded for the detected task.
2. Schema/source grounding is confirmed (MCP evidence or explicit assumptions).
3. Query shape follows the style and safety rules above.
4. If compile_query is available, parser diagnostics are addressed.
5. Semantic uncertainty is surfaced (no false “validated” claims).
6. For DDL or risky execution, explicit user intent is present.
7. Python was not introduced without a real Python requirement such as local persistence, large result handling, reusable code, or post-processing.

If a gate fails, output an assumption-labeled template or ask for missing schema instead of a false-final query.

Capability Map

Use this map to fetch only what is needed:

• `references/query-generation.md`: SELECT build/repair, ambiguous fields, aliasing, clause repairs.
• `references/ddl-generation.md`: CREATE/ALTER/MODIFY/DROP workflow and confirm modes.
• `references/arrays-polymorphism.md`: PackageHeader/polymorphism/RECORD/UNION interactions.
• `references/functions-windows.md`: stateful/stateless function behavior and emission semantics.
• `references/concepts/stateful-functions.md`: stateful function families, argument patterns, and examples.
• `references/concepts/stateless-functions.md`: stateless function families, function discovery workflow, and examples.
• `references/mcp-workflow.md`: MCP grounding flow, parser-only caveat, execute safety.
• `references/concepts/index.md`: task-to-reference routing guide.
• `references/concepts/constants-and-literals.md`: numeric/string/char/timestamp/interval constants.
• `references/concepts/time-and-filtering.md`: timestamp literals, timezone assumptions, WHERE vs HAVING.
• `references/concepts/filters-and-predicates.md`: WHERE, BETWEEN, IN, LIKE, null/NaN, type predicates.
• `references/concepts/subscription-hints.md`: subscription pushdown rules for timestamp and symbol; when optimization applies and when it does not.
• `references/concepts/arrays.md`: generic array mechanics (ANY/ALL, slicing, raw ARRAY JOIN usage).
• `references/concepts/casts.md`: cast patterns and alias-vs-cast disambiguation.
• `references/concepts/operators-conditionals.md`: operator semantics, CASE/IF, .? mask alignment behavior.
• `references/concepts/data-types.md`: practical type constraints and mismatch handling.
• `references/concepts/keywords-and-shaping.md`: WITH, TYPE, FIELD, RECORD, THIS, LIMIT, OFFSET.
• `references/concepts/inner-queries.md`: staged query patterns with subqueries in FROM.
• `references/concepts/union.md`: SELECT UNION, stream UNION, fixed vs polymorphic output.
• `references/concepts/materialized-views.md`: CREATE/ALTER/DROP VIEW syntax, OPTIONS (LIVE, AUTO_RESTART, OUTPUT_TYPE), and lifecycle commands (PAUSE/RESUME/REBUILD).
• `references/recipes.md`: compact end-to-end patterns.

Output Style

Return QQL first, then short notes on assumptions, validation scope, or unresolved semantic risks. Keep teaching detail concise unless user asks for deeper explanation.