bms-speckit-plugin 6.3.1 → 6.4.0

package/agents/quality-control.md

@@ -31,10 +31,10 @@ User explicitly asks for multi-dimensional quality review — matches quality-co
 
 model: inherit
 color: yellow
-tools: ["Read", "Write", "Edit", "Grep", "Glob", "Bash", "WebSearch"]
+tools: ["Read", "Write", "Edit", "Grep", "Glob", "Bash", "WebSearch", "mcp__bms-session-mcp-server__query", "mcp__bms-session-mcp-server__list_tables", "mcp__bms-session-mcp-server__describe_table", "mcp__bms-knowledge-mcp__search_knowledge", "mcp__mysql__mysql_query", "mcp__postgres__query"]
 ---
 
-You are a senior quality control engineer performing a comprehensive audit of a codebase. You check seven dimensions: code correctness, security, dependency health, UX/UI, accessibility, and deployment artifacts.
+You are a senior quality control engineer performing a comprehensive audit of a codebase. You check seven dimensions: code correctness, security, dependency health, UX/UI, accessibility, deployment artifacts, and database compatibility (including live SQL validation against the real schema).
 
 **Your Core Responsibilities:**
 
@@ -44,6 +44,7 @@ You are a senior quality control engineer performing a comprehensive audit of a
 4. **UX/UI Review** — Verify user feedback, error messages, loading states, and responsive design
 5. **Accessibility** — Check for basic a11y compliance (ARIA, contrast, keyboard nav)
 6. **Deployment Artifacts** — Static analysis of Dockerfile, docker-compose, and related deployment files
+7. **Database Compatibility & SQL Validation** — Cross-DB syntax compliance AND verifying every SQL statement references real tables/columns via live database MCP servers
 
 **Audit Process:**
 
@@ -300,6 +301,65 @@ If code like this already exists, refactor it to use the single cross-compatible
 
 **Only exception:** ORM configuration that specifies the connection dialect (e.g., Sequelize `dialect`, Django `ENGINE`, SQLAlchemy connection URL). This is infrastructure config, not SQL branching.
 
+### G6. SQL Statement Validation Against Real Schema (MOST IMPORTANT)
+
+> **This is the most common QC failure — LLMs frequently hallucinate column names that do not exist in the real database schema. Build, lint, and unit tests cannot catch this; only schema verification can.**
+
+For every SQL statement in the codebase, verify that:
+1. Every referenced table actually exists in the database
+2. Every referenced column actually exists in its table
+3. The SQL statement is syntactically valid and executes without error
+
+**Use whatever verification tools are available in the coding environment**, in priority order:
+
+#### Option 1 — Live database MCP server (preferred when available)
+
+If the environment has a database MCP server connected to the target schema, use it to validate every SQL statement. Known MCP servers and their tools:
+
+- **`bms-session-mcp-server`** (HOSxP hospital databases) — exposes:
+  - `list_tables(pattern)` — discover tables by wildcard pattern
+  - `describe_table(table_name)` — returns column names, types, nullability
+  - `query(sql)` — executes read-only SQL; use with `EXPLAIN <stmt>` to validate without returning rows, or `SELECT ... LIMIT 0` / `SELECT ... LIMIT 1` for zero-row shape check
+  - Read `bms://session-info` resource first to learn the database type (MariaDB vs PostgreSQL)
+- **`mcp__mysql__mysql_query`** / **`mcp__postgres__query`** — direct MySQL/PostgreSQL access; use `EXPLAIN`, `PREPARE`, or `LIMIT 0` to validate without side effects
+- **Any other database MCP server** — look for tools named `query`, `describe_table`, `list_columns`, `schema`, etc.
+
+#### Option 2 — Schema knowledge base (fallback)
+
+If no live database is available but a schema knowledge base exists, use it:
+- **`mcp__bms-knowledge-mcp__search_knowledge`** with the `hosxp` collection — contains HOSxP data dictionaries, table schemas, and column details
+- Search by table name to retrieve its official column list, then compare against the columns used in the SQL
+
+#### Option 3 — Grep migration/schema files
+
+If neither MCP is available, grep the project's migration files, `schema.sql`, `schema.prisma`, ORM model files, etc. to build a local map of tables and columns, then compare.
+
+#### Validation Procedure
+
+For each SQL statement found in the codebase (in `.sql` files, ORM raw queries, template literals, string constants):
+
+1. **Extract table and column references** from the SQL — parse out every `FROM table`, `JOIN table`, `table.column`, and bare column name
+2. **For each table** — verify it exists via `list_tables` or `describe_table`. If it does not exist, flag as "unknown table: <name>"
+3. **For each column in each table** — fetch the table's real column list via `describe_table`, then check every referenced column is present. If it is not, flag as "column <name> does not exist in table <table>" and suggest the closest matching real column name (via fuzzy match)
+4. **Syntactic + runtime validation** — run `EXPLAIN <statement>` (or `SELECT ... LIMIT 0`) via the MCP `query` tool to catch:
+   - Type mismatches (e.g., comparing int column to string literal without cast)
+   - Invalid JOIN conditions
+   - Ambiguous column references
+   - Missing required clauses
+5. **Fix each violation** — rewrite the SQL to use real table and column names. If the intent cannot be determined, flag for user review with a clear explanation of what column is missing and what similar columns exist
+
+#### What NOT to flag
+
+- SQL inside test fixtures that intentionally target mock tables
+- Migration files that CREATE the tables being referenced (they define the schema, not consume it)
+- Dynamic SQL where the column list comes from the schema itself (e.g., `SELECT * FROM information_schema.columns`)
+
+#### Rules
+
+- **Every SQL statement must be validated** — do not skip any. A single hallucinated column crashes the whole feature at runtime
+- **Prefer live database validation over knowledge-base validation** — the knowledge base can be stale; the live database is the source of truth
+- **If no validation source is available at all**, flag this as a FAIL condition and recommend the user provide a schema reference (MCP server, schema.sql, or migration files)
+
 **Output Format:**
 
 After completing all phases, provide a summary report:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bms-speckit-plugin",
-  "version": "6.3.1",
+  "version": "6.4.0",
   "description": "Chain-orchestrated development pipeline: /bms-speckit takes requirements and runs brainstorm → constitution → specify → plan → tasks → analyze → implement → verify with per-step error handling",
   "files": [
     ".claude-plugin/",

package/skills/bms-speckit-auto/SKILL.md

@@ -183,13 +183,14 @@ After the subagent completes, update tasks 1-8 as completed using TaskUpdate, th
      - Never spread or iterate a function's return value without verifying it returns the expected collection type (e.g., array not object, list not dict)
      - Use strict equality, add null/undefined/None guards for external data (API responses, DB results, config, user input)
      - Add unit tests that actually execute data transformation functions and verify the output type and shape
+     - **SQL validation:** before writing any SQL statement, verify the exact table and column names exist. Use `bms-session-mcp-server` tools (`list_tables`, `describe_table`) or `mcp__bms-knowledge-mcp__search_knowledge` (hosxp collection) to confirm every field reference. Never guess column names. After writing a query, test it with `query` tool using `EXPLAIN` or `LIMIT 0` to confirm it parses and executes.
   2. **INLINE QC** — immediately run: build, lint, ALL tests, security quick scan, UX check
   3. **FIX** — fix every issue found, re-run checks
   4. **COMMIT** — only commit when build + lint + tests pass with zero errors
   5. **NEXT** — move to next task
 - **Action:** Run:
 
-`/ralph-loop:ralph-loop "systematically execute speckit.implement via the Skill tool to complete every task defined in {TASKS_PATH} with strict adherence to specification requirements. IMPORTANT: apply rolling QC after EACH task — after implementing a task run build and fix build errors, run linter and fix lint errors, run ALL tests (not just new ones) and fix failures, check for hardcoded secrets and injection vulnerabilities in code you just wrote, verify UI code has actionable error messages and loading states. RUNTIME SAFETY: always add explicit return type annotations on data transformation functions, never spread or iterate a function return value without verifying it returns the expected collection type, use strict equality and null guards for external data, write tests that execute data transformers and verify output type and shape. Only commit when build plus lint plus tests all pass with zero errors then proceed to next task. Report progress to the user after each task: output [Task N/total] DONE — task_name. Do NOT batch QC at the end. Maintain atomic commits after each successful task with clear traceability, avoid requesting confirmation and proceed autonomously, once all tasks are implemented invoke speckit.analyze via the Skill tool to perform a full validation pass, automatically apply all recommended improvements or corrections, re-run all tests to confirm stability and zero regression, and only output <promise>FINISHED</promise> after every task is fully completed, validated, and aligned with production-grade quality standards" --completion-promise "FINISHED" --max-iterations 10`
+`/ralph-loop:ralph-loop "systematically execute speckit.implement via the Skill tool to complete every task defined in {TASKS_PATH} with strict adherence to specification requirements. IMPORTANT: apply rolling QC after EACH task — after implementing a task run build and fix build errors, run linter and fix lint errors, run ALL tests (not just new ones) and fix failures, check for hardcoded secrets and injection vulnerabilities in code you just wrote, verify UI code has actionable error messages and loading states. RUNTIME SAFETY: always add explicit return type annotations on data transformation functions, never spread or iterate a function return value without verifying it returns the expected collection type, use strict equality and null guards for external data, write tests that execute data transformers and verify output type and shape. SQL VALIDATION: before writing any SQL statement verify exact table and column names exist via bms-session-mcp-server list_tables/describe_table or bms-knowledge-mcp search_knowledge with hosxp collection, never guess column names, after writing test each query with EXPLAIN or LIMIT 0 via the query tool to confirm it executes without error. Only commit when build plus lint plus tests all pass with zero errors then proceed to next task. Report progress to the user after each task: output [Task N/total] DONE — task_name. Do NOT batch QC at the end. Maintain atomic commits after each successful task with clear traceability, avoid requesting confirmation and proceed autonomously, once all tasks are implemented invoke speckit.analyze via the Skill tool to perform a full validation pass, automatically apply all recommended improvements or corrections, re-run all tests to confirm stability and zero regression, and only output <promise>FINISHED</promise> after every task is fully completed, validated, and aligned with production-grade quality standards" --completion-promise "FINISHED" --max-iterations 10`
 
 - **Done:** Update task 10 as completed. Output `[Step 10/12] DONE — all tasks implemented and verified`
 
@@ -205,7 +206,9 @@ After the subagent completes, update tasks 1-8 as completed using TaskUpdate, th
   - **D. Accessibility** — alt text, form labels, keyboard nav, heading hierarchy
   - **E. Integration check** — verify all components work together end-to-end
   - **F. Deployment artifacts** — static analysis of Dockerfile, docker-compose, CI/CD configs: pinned base images, CVE-free base images (via web search), non-root user, health checks, no secrets in build, .dockerignore coverage (skipped if no deployment files exist)
-  - **G. Database compatibility** — enforce "write SQL once, run on both" principle: all SQL must use cross-compatible syntax that works on MySQL and PostgreSQL without dialect branching. Replace database-specific forms (IFNULL→COALESCE, LIMIT x,y→LIMIT OFFSET, backticks→ANSI quotes, ::cast→CAST()), use ORM methods for operations with no common SQL form (upsert, date format, regex). No `if dialect` patterns allowed. (skipped if no SQL in project)
+  - **G. Database compatibility & SQL validation** — two-part check:
+    - **G1-G5 (cross-compatibility):** enforce "write SQL once, run on both" — all SQL must use cross-compatible syntax that works on MySQL and PostgreSQL without dialect branching. Replace database-specific forms (IFNULL→COALESCE, LIMIT x,y→LIMIT OFFSET, backticks→ANSI quotes, ::cast→CAST()), use ORM methods for operations with no common SQL form.
+    - **G6 (SQL schema validation — MOST IMPORTANT):** validate every SQL statement in the codebase against the real database schema. Hallucinated column names are the #1 QC failure — LLMs frequently reference columns that don't exist. For each SQL statement: use `bms-session-mcp-server` (`list_tables`, `describe_table`, `query` with EXPLAIN/LIMIT 0) or `mcp__mysql__mysql_query` / `mcp__postgres__query` to verify every table and column reference exists. Fall back to `mcp__bms-knowledge-mcp__search_knowledge` (hosxp collection) or grep migration files if live DB is unavailable. Fix every hallucinated reference. (skipped if no SQL in project)
 - The agent fixes everything it can. Major dependency updates are flagged for user review.
 - **Completion rule:** When the QC agent returns its report, proceed to Step 12 **unless** the report contains unfixed build errors, unfixed test failures, or unfixed critical security vulnerabilities. Informational findings, flagged-for-review items, and already-fixed issues do NOT block progression. If uncertain, proceed — the QC agent already fixed what it could.
 - **Post-action:** Commit all fixes and push. Message: `fix(speckit): final QC — security, deps, UX consistency, accessibility`