diff --git a/docs/docs/usage/tool.md b/docs/docs/usage/tool.md
index 40985774f5..af5c667c68 100644
--- a/docs/docs/usage/tool.md
+++ b/docs/docs/usage/tool.md
@@ -51,7 +51,7 @@ Only the data tools are exposed. Coding tools (`bash`, `read`, `edit`, ...), `db
 
 Runs are read-only by default. Reads such as `sql_execute` on a SELECT and `schema_inspect` are allowed; anything else is refused with `Permission denied: <kind>. Re-run with --allow-write to enable writes.` Tools that change local configuration (`warehouse_add`, `warehouse_remove`, `warehouse_install_driver`) also need `--allow-write`.
 
-`DROP DATABASE`, `DROP SCHEMA` and `TRUNCATE` are refused even with `--allow-write`. The block reads the SQL text, so it is best effort: statements built at run time, for example by a stored procedure, are not seen. For a hard guarantee, connect with a database user that lacks `DROP` and `TRUNCATE`.
+`DROP DATABASE`, `DROP SCHEMA` and `TRUNCATE` are refused even with `--allow-write` when the SQL check recognises them. The check reads the SQL text, so it is best effort: a statement the warehouse builds or decodes at run time (dynamic SQL assembled by concatenation, a stored routine, an unusual literal encoding) can get past it. For a hard guarantee, connect with a database user that has no `DROP` or `TRUNCATE` privilege.
 
 Tool results go to the model of the agent that called the tool. `sql_execute` returns result rows, `warehouse_list` shows one sample row per table of a local database, and `schema_inspect` and `sql_execute` with `tables` return value profiles per column; set `ALTIMATE_DISABLE_SCHEMA_VALUES=1` to leave the value profiles out. Keep passwords and keys out of `warehouse_add` arguments, which end up in the calling agent's transcript: reference credentials from a connections file or `ALTIMATE_CODE_CONN_*` variables instead.
 
diff --git a/packages/opencode/src/altimate/native/connections/dbt-adapter-type.ts b/packages/opencode/src/altimate/native/connections/dbt-adapter-type.ts
index 7c25d82b49..b2a25328a4 100644
--- a/packages/opencode/src/altimate/native/connections/dbt-adapter-type.ts
+++ b/packages/opencode/src/altimate/native/connections/dbt-adapter-type.ts
@@ -25,13 +25,15 @@ export async function dbtProjectAdapterType(): Promise<{ project: boolean; type?
       try {
         if (existsSync(config.path())) saved = JSON.parse(readFileSync(config.path(), "utf8"))
       } catch {}
-      projectRoot = saved === undefined ? config.findProjectRoot() : (saved?.projectRoot ?? null)
+      projectRoot = saved === undefined ? config.findProjectRoot() : typeof saved?.projectRoot === "string" ? saved.projectRoot : null
     } catch {
       projectRoot = null
     }
   }
-  // A root without dbt_project.yml is rejected by the routing, which then uses the registry default.
-  if (!projectRoot || !existsSync(path.join(projectRoot, "dbt_project.yml"))) return { project: false, targetTypes: [] }
+  // A root without dbt_project.yml is rejected by the routing, which then uses the registry default;
+  // an adapter already created keeps running even if the file is removed since.
+  if (!projectRoot || (dbtRoute.root === undefined && !existsSync(path.join(projectRoot, "dbt_project.yml"))))
+    return { project: false, targetTypes: [] }
   const targetTypes = await profileTargetTypes(projectRoot)
   try {
     const { resolveDbtTargetPath } = await import("../../validators/validator-utils")
diff --git a/packages/opencode/src/altimate/tools/data-diff.ts b/packages/opencode/src/altimate/tools/data-diff.ts
index c317286376..2a883ff4ad 100644
--- a/packages/opencode/src/altimate/tools/data-diff.ts
+++ b/packages/opencode/src/altimate/tools/data-diff.ts
@@ -119,7 +119,7 @@ export const DataDiffTool = Tool.define("data_diff", {
         target: args.target,
         key_columns: args.key_columns,
         source_warehouse: args.source_warehouse,
-        target_warehouse: args.target_warehouse,
+        target_warehouse: args.target_warehouse || undefined, // "" is the source warehouse, as the ask resolves it
         extra_columns: args.extra_columns,
         algorithm: args.algorithm,
         where_clause: args.where_clause,
diff --git a/packages/opencode/src/altimate/tools/sql-classify.ts b/packages/opencode/src/altimate/tools/sql-classify.ts
index 45eee7d1ef..0b8ec02c69 100644
--- a/packages/opencode/src/altimate/tools/sql-classify.ts
+++ b/packages/opencode/src/altimate/tools/sql-classify.ts
@@ -171,7 +171,7 @@ function otherIsReadOnly(sql: string, parsed: Array<{ category: string }>, diale
 // Hard-deny patterns — blocked regardless of permissions
 const HARD_DENY_TYPES = new Set(["DROP DATABASE", "DROP SCHEMA", "TRUNCATE", "TRUNCATE TABLE"])
 
-const HARD_DENY_PATTERN = /^\s*(DROP\s+(DATABASE|SCHEMA)\b|TRUNCATE(\s+TABLE)?\b)/i
+const HARD_DENY_PATTERN = /^\s*(DROP\s+((EXTERNAL|PLUGGABLE)\s+)?(DATABASE|SCHEMA|CATALOG|NAMESPACE)\b|TRUNCATE(\s+TABLE)?\b)/i
 
 /**
  * Fail closed when the parser is unavailable or rejects the input. Lexical
@@ -192,16 +192,19 @@ function classifyFallback(sql: string, dialect?: string): { queryType: "read" |
   // comment), so the words alone decide; the TRUNCATE() function does not count.
   const lexical = tokenizeSql(sql, dialect)
   if (!lexical.complete) {
-    if (mentionsHardDeny(sql)) blocked = true
+    // Escapes decoded too: dynamic SQL may run a literal such as '\x44ROP SCHEMA s'.
+    if (mentionsHardDenyAnywhere(sql)) blocked = true
   } else {
     // Read with certainty: the statement words in code (literals, identifiers and comments masked,
     // so DROP/**/SCHEMA reads as DROP SCHEMA) decide too, since T-SQL runs `SELECT 1 DROP DATABASE x`
     // as two statements with no ; and a parser may read a literal differently from the engine.
-    const masked = maskSql(lexical.tokens)
-    if (CODE_HARD_DENY.test(masked)) blocked = true
-    // Dynamic SQL (EXEC('...'), EXECUTE IMMEDIATE, PREPARE ... FROM, DO $$...$$) runs a string, so
-    // the words in its literals count too. A statement built by concatenation is not seen.
-    else if (DYNAMIC_SQL.test(masked) && mentionsHardDeny(sql)) blocked = true
+    const names = namesKept(lexical.tokens)
+    // Snowflake's ->> chains statements (elsewhere a JSON operator), and Snowflake Scripting's ASYNC (,
+    // := ( and DEFAULT ( hold one (elsewhere DEFAULT(col) is a function).
+    const chained =
+      (/^snowflake$/i.test(dialect ?? "") || !isModeledDialect(dialect)) &&
+      /(->>|\bASYNC\s*\(|:=\s*\(|\bDEFAULT\s*\()\s*TRUNCATE\b(?!\s*\()/i.test(names)
+    if (CODE_HARD_DENY.test(names) || handlerTruncate(names) || chained) blocked = true
   }
   return { queryType: "write", blocked }
 }
@@ -217,10 +220,10 @@ function batchedStatement(masked: string, dialect?: string): boolean {
 // or clauses (USE, FETCH, OPEN, CLOSE, ENABLE, DISABLE, RENAME, THROW, ADD, MOVE, SEND, RECEIVE, END)
 // count only in their statement forms.
 const BATCH_STATEMENT = new RegExp(
-  "\\b(INSERT|UPDATE|DELETE|MERGE|DROP|ALTER|CREATE|TRUNCATE|EXEC|EXECUTE|GRANT|REVOKE|DENY|BACKUP|RESTORE|DBCC|KILL|SHUTDOWN|BULK|WAITFOR|DECLARE|SET|INTO|" +
+  "\\bTRUNCATE\\b(?!\\s*\\()|\\b(INSERT|UPDATE|DELETE|MERGE|DROP|ALTER|CREATE|EXEC|EXECUTE|GRANT|REVOKE|DENY|BACKUP|RESTORE|DBCC|KILL|SHUTDOWN|BULK|WAITFOR|DECLARE|SET|INTO|" +
     "OPENROWSET|OPENQUERY|OPENDATASOURCE|OPENXML|CHECKPOINT|RECONFIGURE|RAISERROR|GOTO|BEGIN|COMMIT|ROLLBACK|SAVE|DEALLOCATE|" +
     "READTEXT|UPDATETEXT|WRITETEXT|SETUSER|REVERT|PRINT)\\b" +
-    "|\\b(ENABLE|DISABLE)\\s+TRIGGER\\b|\\bRENAME\\s+(OBJECT|DATABASE)\\b|\\bTHROW\\s*($|;|[\\d@])" +
+    "|\\b(ENABLE|DISABLE)\\s+TRIGGER\\b|\\bRENAME\\s+(OBJECT|DATABASE)\\b|\\bTHROW\\s+[\\d@]" +
     // USE except OPTION (USE HINT/PLAN) and USE INDEX; FETCH except OFFSET paging (FETCH NEXT 10 ROWS ONLY).
     "|\\bUSE\\b(?!\\s+(HINT|PLAN|INDEX|KEY)\\b)|\\bFETCH\\b(?!\\s+(FIRST|NEXT)\\s+([\\w@().+-]+\\s+)?(PERCENT\\s+)?ROWS?\\b)" +
     "|\\b(OPEN|CLOSE)\\s+(GLOBAL\\s+)?[\\w@#]+\\s*(;|\\b(SELECT|WITH)\\b)|\\bADD\\s+(SENSITIVITY|SIGNATURE|COUNTER)\\b" +
@@ -228,15 +231,102 @@ const BATCH_STATEMENT = new RegExp(
   "i",
 )
 
-const CODE_HARD_DENY = /\bDROP\s+(DATABASE|SCHEMA)\b|\bTRUNCATE\b(?!\s*\()/i
-const DYNAMIC_SQL = /\bEXEC(UTE)?\b|(^|;)\s*(PREPARE|DO)\b/i
-// Sequence advances outside PostgreSQL's nextval(): NEXT VALUE FOR s (SQL Server, DB2), s.NEXTVAL
-// (Oracle, Snowflake).
-const SEQUENCE_ADVANCE = /\bNEXT\s+VALUE\s+FOR\b|\.\s*NEXTVAL\b/i
+/** A literal's text as an engine may run it, one string per reading: numeric escapes (\xHH, \uHHHH,
+ * \UHHHHHHHH, octal \ooo, U&'...' \HHHH and \+HHHHHH) decoded, and a single-character escape read as a
+ * space when it stands for whitespace (\n \r \t \f, and \v in BigQuery) and as itself otherwise;
+ * and SQL Server's reading, where a backslash before a line break joins the two lines. */
+function escapeReadings(sql: string): string[] {
+  const char = (hex: string, base = 16) => String.fromCodePoint(Math.min(parseInt(hex, base), 0x10ffff))
+  let numeric = sql
+  if (/U&/i.test(numeric)) numeric = numeric.replace(/\\\+([0-9a-fA-F]{6})|\\([0-9a-fA-F]{4})/g, (_, six, four) => char(six ?? four))
+  const decode = (whitespace: RegExp) =>
+    numeric.replace(
+      /\\(?:x([0-9a-fA-F]{1,2})|u([0-9a-fA-F]{4})|U([0-9a-fA-F]{8})|([0-7]{1,3})|([\s\S]))/g,
+      (_, x, u, U, octal, single) =>
+        single !== undefined ? (whitespace.test(single) ? " " : single) : octal ? char(octal, 8) : char(x ?? u ?? U),
+    )
+  return [sql, sql.replace(/\\\r?\n/g, ""), decode(/^[nrtf]$/), decode(/^[nrtfv]$/)]
+}
+
+// TRUNCATE as a statement: TRUNCATE TABLE, or TRUNCATE starting a statement or a block (BEGIN [NOT
+// ATOMIC], DO, LOOP, REPEAT, a handler's condition). After THEN or ELSE it must be a name and the end
+// of the statement, since CASE ... THEN truncate ... is a column. A column or alias named truncate is not.
+const CODE_HARD_DENY = new RegExp(
+  "\\bDROP\\s+((EXTERNAL|PLUGGABLE)\\s+)?(DATABASE|SCHEMA|CATALOG|NAMESPACE)\\b|\\bTRUNCATE\\s+TABLE\\b" +
+    // DuckDB accepts TRUNCATE after EXPLAIN (options) and PREPARE p AS ([^()] keeps the scan linear).
+    "|(^|;|\\bPREPARE\\s+\\w+\\s+AS\\b|\\bEXPLAIN\\s*\\([^()]*\\))\\s*TRUNCATE\\b(?!\\s*\\()" +
+    // After a block or handler word (or EXPLAIN ANALYZE), a target must follow: these words are also
+    // column names, so `SELECT atomic truncate FROM t` is an alias.
+    "|\\b(BEGIN|NOT\\s+ATOMIC|DO|LOOP|REPEAT|SQLEXCEPTION|SQLWARNING|FOUND|ANALYZE|ANALYSE)\\s+TRUNCATE\\s+" +
+    "(?!(FROM|AS|WHERE|ORDER|GROUP|HAVING|UNION|LIMIT|WINDOW|INTO|JOIN|ON|AND|OR|END)\\b)[\\w\"`$\\u0080-\\uffff]" +
+    // After THEN/ELSE: [IF EXISTS] a (dotted, quoted or IDENTIFIER(...)) name, then the statement's end.
+    "|\\b(THEN|ELSE)\\s+TRUNCATE\\s+(IF\\s+EXISTS\\s+)?(?!END\\b)" +
+    "(IDENTIFIER\\s*\\([^()]*\\)|[^\\s;,().]+(\\s*\\.{1,2}\\s*[^\\s;,().]+)*)(\\s+(NOWAIT|WAIT\\s+\\d+))?\\s*(;|$)" +
+    // Anywhere else, TRUNCATE followed by a target (a routine or task body, BEGIN ATOMIC, ALTER TABLE
+    // ... TRUNCATE PARTITION). A column named truncate is followed by a clause word, an operator, `,`
+    // or `)`, or qualified (t.truncate). A name truncate followed by any other word is refused, e.g. an
+    // alias without AS (`SELECT truncate x`, `FROM truncate t`).
+    // The t.truncate check runs only after TRUNCATE matched, over the whitespace just before it (comments
+    // are masked to spaces), so each run of whitespace is scanned once.
+    "|\\bTRUNCATE(?<!\\.\\s*TRUNCATE)\\s+(?!(FROM|AS|WHERE|ORDER|GROUP|HAVING|UNION|EXCEPT|INTERSECT|MINUS|LIMIT|OFFSET|FETCH|" +
+    "QUALIFY|WINDOW|INTO|JOIN|INNER|LEFT|RIGHT|FULL|CROSS|OUTER|ON|USING|AND|OR|NOT|IS|IN|LIKE|ILIKE|RLIKE|REGEXP|SIMILAR|" +
+    "BETWEEN|AT|COLLATE|ESCAPE|ASC|DESC|NULLS|OVER|FILTER|WITHIN|THEN|ELSE|WHEN|END|DIV|MOD|XOR|OVERLAPS|" +
+    "ISNULL|NOTNULL|GLOB|MATCH|SOUNDS|ROWS|RANGE|GROUPS|WITH|START|CONNECT|SETTINGS|OPTION|FOR|OPERATOR|ASCENDING|DESCENDING)\\b)[\\w\"`$\\u0080-\\uffff]",
+  "i",
+)
 
-/** The word DROP with DATABASE or SCHEMA anywhere after it, or TRUNCATE not directly followed by `(`
- * (the TRUNCATE() function; a comment before the `(` does not count, since where a comment ends is
- * part of what is uncertain). Only words count: quotes and comments are ignored, since how they read
+/** A handler's statement follows its conditions (a code, SQLSTATE '...', a name): a TRUNCATE after
+ * HANDLER FOR within the same statement counts. One pass over the statements. */
+function handlerTruncate(masked: string): boolean {
+  return masked.split(";").some((part) => {
+    const handler = part.search(/\bHANDLER\s+FOR\b/i)
+    return handler >= 0 && /\bTRUNCATE\b(?!\s*\()/i.test(part.slice(handler))
+  })
+}
+
+/** Code with literals and comments masked, and each quoted identifier as a placeholder name, so a
+ * statement keeps its target (TRUNCATE `t`) without the identifier's text reading as code. */
+function namesKept(tokens: ReturnType<typeof tokenizeSql>["tokens"]): string {
+  return tokens
+    .map(({ kind, text }) =>
+      kind === "code" ? text : kind === "identifier" ? ` ${"_".repeat(Math.max(text.length - 2, 1))} ` : text.replace(/[^\r\n]/g, " "),
+    )
+    .join("")
+}
+
+/** Whether a write's quoted text holds a hard-denied statement. A write can run its literals as SQL:
+ * dynamic SQL (EXEC('...'), sp_executesql N'...', EXECUTE IMMEDIATE, PREPARE ... FROM, dblink_exec),
+ * or a function, procedure or trigger body created and called in the same call. Each literal is read
+ * word by word with its escapes decoded, as uncertain SQL is, so nothing inside it (nested quotes,
+ * comments) can separate the words. A statement built by concatenation is not seen, and a custom
+ * UESCAPE character is not decoded, so it blocks. */
+function quotedHardDeny(tokens: ReturnType<typeof tokenizeSql>["tokens"], dialect?: string): boolean {
+  // UESCAPE counts with a U&'...' literal: U& in code right before a literal token.
+  const unicode = tokens.some(({ kind, text }, i) => kind === "code" && /\bU&$/i.test(text) && tokens[i + 1]?.kind === "literal")
+  if (unicode && /\bUESCAPE\b/i.test(maskSql(tokens))) return true
+  // "..." is a string in Spark, Databricks and Hive, under SQL Server's QUOTED_IDENTIFIER OFF and in
+  // MySQL without ANSI_QUOTES; other identifiers never are.
+  const quotedStrings =
+    /^(sqlserver|tsql|mssql|fabric|mysql|mariadb|spark|databricks|hive)$/i.test(dialect ?? "") || !isModeledDialect(dialect)
+  return tokens.some(
+    ({ kind, text }) =>
+      (kind === "literal" || (kind === "identifier" && quotedStrings && text[0] === '"')) &&
+      ((/\bUESCAPE\b/i.test(text) && /\bU&'/i.test(text)) || mentionsHardDenyAnywhere(text)),
+  )
+}
+
+// Sequence advances outside PostgreSQL's nextval(): NEXT VALUE FOR s (SQL Server, DB2), and in Oracle,
+// Snowflake and MariaDB (Oracle mode) s.NEXTVAL, quoted or not (elsewhere s.nextval is a column).
+function advancesSequence(tokens: ReturnType<typeof tokenizeSql>["tokens"], masked: string, dialect?: string): boolean {
+  if (/\bNEXT\s+VALUE\s+FOR\b/i.test(masked)) return true
+  if (!/^(oracle|snowflake|mysql|mariadb)$/i.test(dialect ?? "") && isModeledDialect(dialect)) return false
+  // Comments masked to spaces and quoted names unwrapped, so s./* c */"NEXTVAL" reads as s.NEXTVAL.
+  return /\.\s*NEXTVAL\b/i.test(maskSql(tokens, true))
+}
+
+/** The word DROP with DATABASE, SCHEMA, CATALOG or NAMESPACE anywhere after it, or TRUNCATE not
+ * directly followed by `(` (the TRUNCATE() function; a comment before the `(` does not count, since
+ * where a comment ends is part of what is uncertain). Only words count: quotes and comments are ignored, since how they read
  * is what is uncertain, so no comment or quote placement can separate the words. One forward pass. */
 function mentionsHardDeny(sql: string): boolean {
   let drop = false
@@ -250,7 +340,7 @@ function mentionsHardDeny(sql: string): boolean {
     while (j < sql.length && /\w/.test(sql[j])) j++
     const word = sql.slice(i, j).toUpperCase()
     if (word === "DROP") drop = true
-    else if (drop && (word === "DATABASE" || word === "SCHEMA")) return true
+    else if (drop && (word === "DATABASE" || word === "SCHEMA" || word === "CATALOG" || word === "NAMESPACE")) return true
     else if (word === "TRUNCATE" && !/^[ \t\r\n]*\(/.test(sql.slice(j, j + 64))) return true
     i = j
   }
@@ -349,7 +439,21 @@ export function classifyAndCheck(sql: string, dialect?: string): {
   blocked: boolean
   parseError?: string
 } {
-  return classifySql(sql, dialect, 0)
+  const result = classifySql(sql, dialect, 0)
+  if (result.queryType === "write" && !result.blocked) {
+    // SQL read with certainty: a write may still run the SQL in its literals. Otherwise the word scan
+    // of the whole text already covered them.
+    const lexical = tokenizeSql(sql, dialect)
+    if (lexical.complete && quotedHardDeny(lexical.tokens, dialect)) return { ...result, blocked: true }
+  }
+  return result
+}
+
+/** Whether DROP DATABASE/SCHEMA/CATALOG/NAMESPACE or TRUNCATE appears anywhere in the text, read word by word
+ * with quotes and comments ignored and escapes decoded (for text another layer renders, such as dbt's
+ * Jinja). */
+export function mentionsHardDenyAnywhere(sql: string): boolean {
+  return escapeReadings(sql).some(mentionsHardDeny)
 }
 
 function classifySql(sql: string, dialect: string | undefined, depth: number): {
@@ -385,7 +489,14 @@ function classifySql(sql: string, dialect: string | undefined, depth: number): {
         !functions.write &&
         !blocked &&
         !batchedStatement(masked, dialect) &&
-        !SEQUENCE_ADVANCE.test(masked) &&
+        // Snowflake's ->> runs the statements it chains, DML included.
+        !(/^snowflake$/i.test(dialect ?? "") && masked.includes("->>")) &&
+        // Spark/Hive SELECT TRANSFORM(...) USING 'cmd' runs a command on the cluster; any USING not
+        // followed by "(" counts (comments and the script masked), while JOIN ... USING (col) is a join.
+        !((/^(spark|databricks|hive)$/i.test(dialect ?? "") || !isModeledDialect(dialect)) &&
+          /\bTRANSFORM\s*\(/i.test(masked) &&
+          /\bUSING\b(?!\s*\()/i.test(masked)) &&
+        !advancesSequence(lexical.tokens, masked, dialect) &&
         otherIsReadOnly(sql, result.statements, dialect)
           ? "read"
           : "write",
diff --git a/packages/opencode/src/altimate/tools/sql-execute.ts b/packages/opencode/src/altimate/tools/sql-execute.ts
index aa0a3b3bda..78a5a11d41 100644
--- a/packages/opencode/src/altimate/tools/sql-execute.ts
+++ b/packages/opencode/src/altimate/tools/sql-execute.ts
@@ -9,7 +9,13 @@ import { Tool } from "../../tool/tool"
 import { Dispatcher } from "../native"
 import type { SqlExecuteResult } from "../native/types"
 // altimate_change start - SQL write access control + fingerprinting
-import { classifyAndCheck, classifyMongoCommand, computeSqlFingerprint, isScratchSafeWrite } from "./sql-classify"
+import {
+  classifyAndCheck,
+  classifyMongoCommand,
+  computeSqlFingerprint,
+  isScratchSafeWrite,
+  mentionsHardDenyAnywhere,
+} from "./sql-classify"
 import { statementsOf } from "./sql-tokenize"
 
 const MONGO_WAREHOUSE_TYPES = new Set(["mongodb", "mongo"])
@@ -231,7 +237,12 @@ export const SqlExecuteTool = Tool.define(
               )
             }
             const classification = mongo ?? classifyEvery(sql, dialects)
-            const { queryType, blocked } = classification
+            // dbt renders Jinja before it runs SQL, wherever the Jinja sits (a comment or literal too),
+            // so {% do run_query('...') %} runs while the query compiles and {# ... #} removes text.
+            // A Mongo command routed to dbt is rendered too.
+            const jinja = target?.dbt === true && /\{[{%#]/.test(sql)
+            const queryType = jinja ? "write" : classification.queryType
+            const blocked = classification.blocked || (jinja && mentionsHardDenyAnywhere(sql))
             const query = mongo?.normalized ?? sql
             if (blocked) {
               throw new Error(
@@ -671,14 +682,18 @@ function resolveStepRefs(
  * type and every target of the project's profile; when the profile's targets cannot be read, or one
  * is an adapter with no supported dialect, every supported dialect and the generic lexing count.
  * Read without starting dbt or loading the handlers. The registry warehouse is named for the ask. */
-async function defaultTarget(forceReadOnly: boolean): Promise<{ dialects: (string | undefined)[]; warehouse?: string }> {
+async function defaultTarget(
+  forceReadOnly: boolean,
+): Promise<{ dialects: (string | undefined)[]; warehouse?: string; dbt?: boolean }> {
   const fallback = Registry.list().warehouses[0] // before any await, as the call resolves it now
   const dialectOf = (type?: string) => (type ? (Registry.canonicalType(type) ?? undefined) : undefined)
   const dialects: (string | undefined)[] = fallback ? [dialectOf(fallback.type)] : []
+  let viaDbt = false
   if (!forceReadOnly) {
     const { dbtProjectAdapterType } = await import("../native/connections/dbt-adapter-type")
     const dbt = await dbtProjectAdapterType()
     if (dbt.project) {
+      viaDbt = true
       const types = [dbt.type, ...dbt.targetTypes].filter((type): type is string => Boolean(type))
       const known = types.map(dialectOf)
       dialects.push(...known.filter(Boolean))
@@ -686,7 +701,11 @@ async function defaultTarget(forceReadOnly: boolean): Promise<{ dialects: (strin
       if (!dbt.targetTypes.length || known.includes(undefined)) dialects.push(...EVERY_DIALECT, undefined)
     }
   }
-  return { dialects: dialects.length ? [...new Set(dialects)] : [undefined], ...(fallback && { warehouse: fallback.name }) }
+  return {
+    dialects: dialects.length ? [...new Set(dialects)] : [undefined],
+    ...(fallback && { warehouse: fallback.name }),
+    ...(viaDbt && { dbt: true }),
+  }
 }
 
 const EVERY_DIALECT = ["postgres", "redshift", "duckdb", "snowflake", "bigquery", "mysql", "sqlserver", "databricks", "clickhouse", "oracle", "trino", "sqlite"]
diff --git a/packages/opencode/src/altimate/tools/sql-explain.ts b/packages/opencode/src/altimate/tools/sql-explain.ts
index 77fdbbf221..1ee673d029 100644
--- a/packages/opencode/src/altimate/tools/sql-explain.ts
+++ b/packages/opencode/src/altimate/tools/sql-explain.ts
@@ -8,6 +8,7 @@ import * as Precedence from "../workspace/precedence"
 import { validateWarehouseName } from "./input-validation"
 import * as Registry from "../native/connections/registry"
 import { classifyAndCheck, isSingleQuery } from "./sql-classify"
+import { statementsOf, tokenizeSql } from "./sql-tokenize"
 
 /**
  * Detect SQL input that cannot be meaningfully EXPLAIN'd.
@@ -90,14 +91,26 @@ export const SqlExplainTool = Tool.define("sql_explain", {
         output: `Invalid input: ${sqlError}`,
       }
     }
-    // EXPLAIN ANALYZE executes the statement. Only a single read-only query may be analyzed,
-    // so a plan request can never perform a write on engines without a read-only transaction.
-    if (args.analyze) {
-      const dialect = Registry.canonicalType(args.warehouse ? Registry.getConfig(args.warehouse)?.type : undefined) ?? undefined
+    // EXPLAIN is prefixed to the text, and some drivers run every statement in it (Redshift, DuckDB),
+    // so only one statement is explained: a second one would run, not be planned. EXPLAIN ANALYZE
+    // executes the statement, so only a single read-only query may be analyzed.
+    {
+      const warehouse = args.warehouse || Registry.list().warehouses[0]?.name
+      const dialect = Registry.canonicalType(warehouse ? Registry.getConfig(warehouse)?.type : undefined) ?? undefined
       const { queryType, blocked } = classifyAndCheck(args.sql, dialect)
-      if (blocked || queryType !== "read" || !isSingleQuery(args.sql, dialect)) {
-        const error =
-          "EXPLAIN ANALYZE executes the statement, so only a single read-only query can be analyzed. Use analyze: false for a plan of this statement."
+      const statements = tokenizeSql(args.sql, dialect).complete ? statementsOf(args.sql, dialect) : []
+      // The tool adds EXPLAIN itself: SQL starting with an EXPLAIN option (ANALYZE, VERBOSE, an option
+      // list such as ("analyze")) would turn the plan into EXPLAIN ANALYZE, which runs the statement. A
+      // leading "(" is a query only when a query keyword follows it.
+      const option = /^(EXPLAIN|ANALY[SZ]E|VERBOSE)\b|^\((?![(\s]*(SELECT|WITH|VALUES|TABLE|FROM)\b)/i.test(statements[0] ?? "")
+      const error = blocked
+        ? "DROP DATABASE, DROP SCHEMA and TRUNCATE are blocked for safety."
+        : statements.length !== 1 || option
+          ? "Only a single statement can be explained, without EXPLAIN or its options. Run each statement separately."
+          : args.analyze && (queryType !== "read" || !isSingleQuery(args.sql, dialect))
+            ? "EXPLAIN ANALYZE executes the statement, so only a single read-only query can be analyzed. Use analyze: false for a plan of this statement."
+            : undefined
+      if (error) {
         return {
           title: "Explain: INVALID INPUT",
           metadata: { success: false, analyzed: false, warehouse_type: "unknown", error, error_class: "input_validation" },
diff --git a/packages/opencode/src/altimate/tools/sql-function-policy.ts b/packages/opencode/src/altimate/tools/sql-function-policy.ts
index 6341156e29..675bfe026c 100644
--- a/packages/opencode/src/altimate/tools/sql-function-policy.ts
+++ b/packages/opencode/src/altimate/tools/sql-function-policy.ts
@@ -20,7 +20,7 @@ const PURE_PG_FUNCTION = new RegExp(`^pg_(?:${PURE_PG_FUNCTIONS})$`, "i")
 const SIDE_EFFECT_PG_FUNCTION =
   /^pg_(?:write_file|append_file|file_write|file_rename|file_unlink|try_advisory\w*|advisory_unlock\w*|advisory_lock\w*|advisory_xact\w*|terminate_backend|cancel_backend|reload_conf|rotate_logfile|promote|switch_wal|logical_emit_message|create_\w*_replication_slot|copy_\w*_replication_slot|sync_replication_slots|drop_replication_slot|replication_slot_advance|replication_origin_\w+|notify|stat\w*reset\w*|\w*_reset\w*|stat_clear\w*|stat_force\w*|wal_replay_\w+|logical_slot_\w+|replication_slot_\w+|xlog_replay_\w+|log_backend_memory_contexts|create_restore_point|backup_\w+|start_backup|stop_backup)$/i
 const SIDE_EFFECT_FUNCTION =
-  /^(?:nextval|setval|set_config|load_extension|writefile|dblink\w*|lo_\w*|lowrite|loread|checkpoint|force_checkpoint|get_lock|release_lock|release_all_locks|is_free_lock|is_used_lock|master_pos_wait|source_pos_wait|sleep|benchmark|postgres_execute|mysql_execute|sqlite_execute|SYSTEM\$\w+)$/i
+  /^(?:nextval|setval|set_config|load_extension|writefile|dblink\w*|lo_\w*|lowrite|loread|checkpoint|force_checkpoint|get_lock|release_lock|release_all_locks|is_free_lock|is_used_lock|master_pos_wait|source_pos_wait|sleep|benchmark|postgres_execute|mysql_execute|sqlite_execute|bigquery_execute|ducklake_\w+|jdbc|reflect|java_method|getnextval|http_request|remote_query|postgres_attach|sqlite_attach|SYSTEM\$\w+)$/i
 
 /** Accepts one decoded function identifier, without schema names or SQL text. */
 export function hasSideEffectFunction(name: string, rejectUnknownPg = true): boolean {
diff --git a/packages/opencode/src/altimate/tools/sql-tokenize.ts b/packages/opencode/src/altimate/tools/sql-tokenize.ts
index ad4598b6ad..c980b3734e 100644
--- a/packages/opencode/src/altimate/tools/sql-tokenize.ts
+++ b/packages/opencode/src/altimate/tools/sql-tokenize.ts
@@ -18,7 +18,7 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
   const tokens: SqlToken[] = []
   const mysqlComments = !dialect || /^(mysql|mariadb)$/i.test(dialect)
   const hashComments = /^(mysql|mariadb|bigquery)$/i.test(dialect ?? "")
-  const clickhouseComments = /^clickhouse$/i.test(dialect ?? "")
+  const clickhouse = /^clickhouse$/i.test(dialect ?? "")
   // Snowflake also starts a line comment with //.
   const slashComments = /^snowflake$/i.test(dialect ?? "")
   // E'...' escape strings exist only in PostgreSQL and the engines that follow it; elsewhere E is
@@ -35,7 +35,8 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
   const nestedComments = /^(postgres|duckdb|sqlserver|tsql|mssql|fabric|clickhouse|spark|databricks)$/i.test(dialect ?? "")
   // Where nesting is not known either way, a /* inside a comment makes its end uncertain.
   const nestingKnown = nestedComments || /^(mysql|mariadb|sqlite|oracle|trino)$/i.test(dialect ?? "")
-  const dollarQuotes = /^(postgres|redshift|duckdb|snowflake)$/i.test(dialect ?? "")
+  // ClickHouse heredocs ($tag$...$tag$) read like PostgreSQL dollar quotes.
+  const dollarQuotes = /^(postgres|redshift|duckdb|snowflake|clickhouse)$/i.test(dialect ?? "")
   // Quoting is dialect-specific: brackets can contain executable array expressions,
   // and PostgreSQL backticks can be operators. Unknown dialects keep both visible.
   const backtickIdentifiers = /^(mysql|mariadb|sqlite|bigquery|clickhouse|spark|databricks|hive)$/i.test(dialect ?? "")
@@ -44,6 +45,9 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
   // may still escape a quote with a backslash or start a comment with #, so both make the SQL
   // uncertain. Oracle, Trino and Fabric do neither.
   const unmodeled = !isModeledDialect(dialect)
+  const sparkHints = unmodeled || /^(spark|databricks|hive|mysql|mariadb)$/i.test(dialect ?? "")
+  // Oracle q'<x>...<x>' quotes end at a delimiter this lexer does not track.
+  const qQuotes = unmodeled || /^oracle$/i.test(dialect ?? "")
   const ambiguousBackslash = unmodeled || /^(postgres|postgresql|spark|databricks|hive)$/i.test(dialect ?? "")
   let complete = true
   let i = 0
@@ -56,7 +60,7 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
     if (
       (sql.startsWith("--", i) && (!mysqlComments || i + 2 === sql.length || /[\x00-\x20\x7f]/.test(sql[i + 2]))) ||
       (hashComments && sql[i] === "#") ||
-      (clickhouseComments && (sql.startsWith("//", i) || sql.startsWith("# ", i) || sql.startsWith("#!", i))) ||
+      (clickhouse && (sql.startsWith("//", i) || sql.startsWith("# ", i) || sql.startsWith("#!", i))) ||
       (slashComments && sql.startsWith("//", i))
     ) {
       kind = "comment"
@@ -66,7 +70,8 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
     } else if (sql.startsWith("/*", i)) {
       // MySQL/MariaDB execute these comments. Generic SQL parsers discard them,
       // so they cannot establish that a query is read-only or scratch-confined.
-      if (/^\/\*(?:!|[mM]!)/.test(sql.slice(i))) complete = false
+      if (/^\/\*(?:!|[mM]!)/.test(sql.slice(i, i + 4))) complete = false
+      const hint = sparkHints && sql.startsWith("/*+", i)
       kind = "comment"
       let depth = 1
       i += 2
@@ -81,6 +86,8 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
         } else i++
       }
       if (depth) complete = false
+      // Spark and MySQL 8 lex a /*+ hint's body as code, so a quote there can hold the */ that seems to end it.
+      if (hint && /['"`]/.test(sql.slice(start, i))) complete = false
     } else if (tripleQuotes && (sql.startsWith("'''", i) || sql.startsWith('"""', i))) {
       // Ends only at the same three quotes; a backslash keeps the next character even in a raw
       // r'''...''' string (BigQuery reads the escape but keeps it as text).
@@ -130,7 +137,14 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
           // After an odd run of backslashes a quote is escaped in some modes and closes the literal in
           // others, so where it ends cannot be known: in an unmodeled dialect, in Spark, Databricks
           // and Hive (escapedStringLiterals) and in PostgreSQL (standard_conforming_strings off).
-          if (ambiguousBackslash && !escaped && quote !== "[" && sql[i] === end) {
+          // Not PostgreSQL "..." or `...` identifiers, which take no backslash escapes.
+          if (
+            ambiguousBackslash &&
+            !escaped &&
+            quote !== "`" &&
+            !(quote === '"' && /^(postgres|postgresql)$/i.test(dialect ?? "")) &&
+            sql[i] === end
+          ) {
             let run = 0
             for (let j = i - 1; j > start && sql[j] === "\\"; j--) run++
             if (run % 2 === 1) complete = false
@@ -149,9 +163,19 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
         }
         if (!closed) complete = false
         i = Math.min(i, sql.length)
-      } else if (dollarQuotes && sql[i] === "$" && !IDENTIFIER_CHAR.test(sql[i - 1] ?? "")) {
+      } else if (dollarQuotes && sql[i] === "$" && !(clickhouse ? /[\w$]/ : IDENTIFIER_CHAR).test(sql[i - 1] ?? "")) {
         DOLLAR_TAG.lastIndex = i
-        const tag = /^snowflake$/i.test(dialect ?? "") ? (sql.startsWith("$$", i) ? "$$" : undefined) : DOLLAR_TAG.exec(sql)?.[0]
+        // ClickHouse's heredoc tag is whatever lies between the two $ (ASCII words only before it).
+        const close = clickhouse ? sql.indexOf("$", i + 1) : -1
+        const tag = /^snowflake$/i.test(dialect ?? "")
+          ? sql.startsWith("$$", i)
+            ? "$$"
+            : undefined
+          : clickhouse
+            ? close < 0
+              ? undefined
+              : sql.slice(i, close + 1)
+            : DOLLAR_TAG.exec(sql)?.[0]
         if (tag) {
           kind = "literal"
           const end = sql.indexOf(tag, i + tag.length)
@@ -165,6 +189,8 @@ export function tokenizeSql(sql: string, dialect?: string): { tokens: SqlToken[]
     if (!kind) {
       // Unmodeled dialect: # starts a comment in MySQL and BigQuery, so what follows cannot be read.
       if (unmodeled && sql[start] === "#") complete = false
+      if (qQuotes && /[nNqQ]/.test(sql[start]) && /^[nN]?[qQ]'/.test(sql.slice(start, start + 3)) && !/[\w$]/.test(sql[start - 1] ?? ""))
+        complete = false
       i++
       continue
     }
diff --git a/packages/opencode/test/altimate/native/dbt-adapter-type.test.ts b/packages/opencode/test/altimate/native/dbt-adapter-type.test.ts
index 0dade10d7c..09340eff1c 100644
--- a/packages/opencode/test/altimate/native/dbt-adapter-type.test.ts
+++ b/packages/opencode/test/altimate/native/dbt-adapter-type.test.ts
@@ -118,4 +118,10 @@ test("once dbt's adapter is resolved, its project is the one read; a saved confi
   mkdirSync(path.dirname(configPath()), { recursive: true })
   writeFileSync(configPath(), "null")
   expect((await dbtProjectAdapterType()).project).toBe(false)
+  writeFileSync(configPath(), JSON.stringify({ projectRoot: 42 }))
+  expect((await dbtProjectAdapterType()).project).toBe(false)
+  // An adapter already created keeps running after its dbt_project.yml is removed.
+  rmSync(path.join(other, "dbt_project.yml"))
+  dbtRoute.root = other
+  expect((await dbtProjectAdapterType()).project).toBe(true)
 })
diff --git a/packages/opencode/test/altimate/sql-validation-adversarial.test.ts b/packages/opencode/test/altimate/sql-validation-adversarial.test.ts
index ef373e8469..e536bc4f38 100644
--- a/packages/opencode/test/altimate/sql-validation-adversarial.test.ts
+++ b/packages/opencode/test/altimate/sql-validation-adversarial.test.ts
@@ -113,10 +113,12 @@ describe("Adversarial: sql-classify bypass attempts", () => {
     expect(r.blocked).toBe(false)
   })
 
-  test("TRUNCATE inside string literal is not blocked", () => {
+  test("TRUNCATE inside a write's string literal is blocked", () => {
+    // A write can run its literals as SQL (dynamic SQL, routine bodies), so a write's literal that
+    // names TRUNCATE is refused even with --allow-write: an accepted limit of the best-effort block.
     const r = classifyAndCheck("INSERT INTO logs VALUES ('TRUNCATE TABLE users happened')")
     expect(r.queryType).toBe("write")
-    expect(r.blocked).toBe(false) // string literal, not real TRUNCATE
+    expect(r.blocked).toBe(true)
   })
 
   test("DROP SCHEMA inside double-quoted identifier is not blocked", () => {
diff --git a/packages/opencode/test/altimate/tool-cli/runner.test.ts b/packages/opencode/test/altimate/tool-cli/runner.test.ts
index 7e43e7e665..7fcf0bd60a 100644
--- a/packages/opencode/test/altimate/tool-cli/runner.test.ts
+++ b/packages/opencode/test/altimate/tool-cli/runner.test.ts
@@ -10,6 +10,7 @@ import { Effect, Exit } from "effect"
 import { Global } from "../../../src/global"
 import * as Registry from "../../../src/altimate/native/connections/registry"
 import { registerDefaultNativeHandlers } from "../../lib/native-handlers"
+import * as Dispatcher from "../../../src/altimate/native/dispatcher"
 
 let dir: string
 let runner: Runner
@@ -401,6 +402,52 @@ describe("data_diff write-carrying inputs", () => {
   })
 })
 
+describe("sql_explain", () => {
+  test("explains one statement only, since some drivers run every statement in the text", async () => {
+    Registry.setConfigs({ rs: { type: "redshift" } as any })
+    const sent: string[] = []
+    const get = spyOn(Registry, "get").mockResolvedValue({
+      execute: async (sql: string) => {
+        sent.push(sql)
+        return { columns: ["QUERY PLAN"], rows: [["Seq Scan"]], row_count: 1, truncated: false }
+      },
+    } as any)
+    try {
+      for (const sql of ["SELECT 1; DELETE FROM t", "SELECT 1; TRUNCATE t", "ANALYZE DELETE FROM t", "/* x */ ANALYZE DELETE FROM t", "(ANALYZE) DELETE FROM t", '("analyze") DELETE FROM t']) {
+        const result = await runner.run("sql_explain", { sql, warehouse: "rs" })
+        expect(JSON.stringify(result)).toMatch(/single statement|blocked for safety/)
+      }
+      expect(sent).toEqual([])
+      // A long run of whitespace after "(" is refused quickly.
+      const started = performance.now()
+      await runner.run("sql_explain", { sql: "(" + " ".repeat(200_000) + "x)", warehouse: "rs" })
+      expect(performance.now() - started).toBeLessThan(5_000)
+      // A parenthesized query is a query, not an EXPLAIN option list.
+      const plan = await runner.run("sql_explain", { sql: "(SELECT 1) UNION (SELECT 2)", warehouse: "rs" })
+      expect(JSON.stringify(plan)).not.toMatch(/single statement/)
+    } finally {
+      get.mockRestore()
+      Registry.reset()
+    }
+  })
+})
+
+describe("data_diff empty target_warehouse", () => {
+  test("dispatches the source warehouse", async () => {
+    Registry.setConfigs({ src: { type: "postgres" } as any, other: { type: "mysql" } as any })
+    const call = spyOn(Dispatcher, "call").mockResolvedValue({ success: true, outcome: { mode: "diff" } } as any)
+    try {
+      await runner.run("data_diff", { source: "orders", target: "orders_v2", key_columns: ["id"], source_warehouse: "src", target_warehouse: "" })
+      const params = call.mock.calls.find(([method]) => method === "data.diff")?.[1] as Record<string, unknown>
+      expect(params.source_warehouse).toBe("src")
+      expect(params.target_warehouse).toBeUndefined()
+    } finally {
+      call.mockRestore()
+      Registry.reset()
+    }
+  })
+})
+
 describe("data_diff hard block on long SQL", () => {
   test("is checked on the whole SQL, not the 120-character preview", async () => {
     Registry.setConfigs({ pg: { type: "postgres" } as any })
diff --git a/packages/opencode/test/altimate/tools/sql-classify.test.ts b/packages/opencode/test/altimate/tools/sql-classify.test.ts
index b39374c221..6d726db42c 100644
--- a/packages/opencode/test/altimate/tools/sql-classify.test.ts
+++ b/packages/opencode/test/altimate/tools/sql-classify.test.ts
@@ -120,7 +120,7 @@ describe("classify", () => {
 })
 
 describe("classifyMulti", () => {
-  for (const dialect of ["mysql", "sqlite", "bigquery", "clickhouse", "unknown", undefined]) {
+  for (const dialect of ["mysql", "sqlite", "bigquery", "unknown", undefined]) {
     test(`dollar-delimited writes fail closed for ${dialect ?? "omitted"} dialect`, () => {
       const sql = "SELECT $x$; DELETE FROM t; $x$"
       expect(classifyMulti(sql, dialect)).toBe("write")
@@ -440,6 +440,8 @@ describe("each engine's own literal and statement rules", () => {
     ["SELECT 1 USE master", "sqlserver", "write", false],
     ["SELECT enable, disable, rename, throw FROM t", "sqlserver", "read", false],
     ["SELECT 1 THROW 50000, 'x', 1", "sqlserver", "write", false],
+    ["SELECT throw FROM t ORDER BY throw", "sqlserver", "read", false],
+    ["SELECT 1 AS throw", "fabric", "read", false],
     // Spark continues a -- comment past a backslash before the line end.
     ["-- x\\\nSELECT '\nDELETE FROM t -- '", "databricks", "write", false],
     ["-- x\\\nSELECT '\nDROP SCHEMA s CASCADE -- '", "spark", "write", true],
@@ -453,7 +455,174 @@ describe("each engine's own literal and statement rules", () => {
     ["EXECUTE IMMEDIATE \"DROP SCHEMA ds CASCADE\"", "bigquery", "write", true],
     ["DO $$BEGIN EXECUTE 'DROP SCHEMA s CASCADE'; END$$", "postgres", "write", true],
     ["PREPARE s FROM 'DROP DATABASE prod'", "mysql", "write", true],
+    // ...with the literal's escapes decoded.
+    ["EXECUTE IMMEDIATE '\\x44ROP SCHEMA ds CASCADE'", "bigquery", "write", true],
+    ["EXECUTE IMMEDIATE '\\x44ROP DATABASE victim'", "snowflake", "write", true],
+    ["EXECUTE IMMEDIATE '\\104ROP SCHEMA ds'", "bigquery", "write", true],
+    ["DO $$BEGIN EXECUTE E'\\x44ROP SCHEMA victim CASCADE'; END$$", "postgres", "write", true],
+    ["DO $$BEGIN EXECUTE U&'\\0044ROP SCHEMA victim CASCADE'; END$$", "postgres", "write", true],
+    ["DO $$BEGIN EXECUTE U&'!0044ROP SCHEMA v' UESCAPE '!'; END$$", "postgres", "write", true],
     ["EXEC dbo.refresh @note = 'nightly'", "sqlserver", "write", false],
+    // A write runs its literals as SQL (dynamic SQL, routine bodies): each is read with escapes decoded.
+    ["sp_executesql N'DROP DATABASE prod'", "sqlserver", "write", true],
+    ["CREATE FUNCTION f() RETURNS void LANGUAGE sql AS $$ DROP SCHEMA s CASCADE $$; SELECT f()", "postgres", "write", true],
+    ["CREATE PROCEDURE p() AS $$ BEGIN TRUNCATE t; END; $$ LANGUAGE plpgsql; CALL p()", "redshift", "write", true],
+    ["SELECT dblink_exec('dbname=postgres', 'DROP DATABASE prod')", "postgres", "write", true],
+    ["WITH p AS PROCEDURE() RETURNS VARCHAR LANGUAGE SQL AS $$ BEGIN DROP DATABASE prod; RETURN 'x'; END $$ CALL p()", "snowflake", "write", true],
+    // Read word by word: escapes, nested quotes and comments in the literal cannot separate the words,
+    // so a write's text that merely mentions them is blocked too.
+    ["EXECUTE IMMEDIATE 'DROP\\nSCHEMA ds CASCADE'", "bigquery", "write", true],
+    ["DO $$BEGIN EXECUTE E'DROP\\tSCHEMA s CASCADE'; END$$", "postgres", "write", true],
+    ["EXEC(N'SELECT ''--''; DROP DATABASE victim')", "sqlserver", "write", true],
+    ["DO $$BEGIN EXECUTE $q$SELECT '--'; DROP SCHEMA s CASCADE$q$; END$$", "postgres", "write", true],
+    ["INSERT INTO log VALUES ('nightly truncate of staging')", "postgres", "write", true],
+    ["INSERT INTO log VALUES ('dropped rows', 'schema v2')", "postgres", "write", false],
+    ["SELECT * FROM audit WHERE action = 'DROP SCHEMA'", "postgres", "read", false],
+    // An escaped uppercase letter stands for itself; MySQL reads an unknown lowercase escape the same way.
+    ["DO E'BEGIN DROP SCH\\EMA s CASCADE; END'", "postgres", "write", true],
+    ["CREATE PROCEDURE p() AS 'BEGIN TRUNC\\ATE t; END' LANGUAGE plpgsql; CALL p()", "redshift", "write", true],
+    ["EXECUTE IMMEDIATE 'DROP SCH\\EMA s'", "snowflake", "write", true],
+    ["PREPARE s FROM 'D\\ROP DATABASE x'", "mysql", "write", true],
+    ["PREPARE s FROM 'DROP sch\\ema x'", "mysql", "write", true],
+    // TRUNCATE starting a block or a handler is the statement.
+    ["BEGIN REPEAT TRUNCATE t; UNTIL (TRUE) END REPEAT; END", "snowflake", "write", true],
+    ["CREATE PROCEDURE p() BEGIN DECLARE CONTINUE HANDLER FOR SQLEXCEPTION TRUNCATE t; END", "mysql", "write", true],
+    ["SELECT truncate, x FROM t ORDER BY truncate DESC", "postgres", "read", false],
+    // An alias written without AS after a column named truncate is refused (best effort, fails closed).
+    ["SELECT truncate result FROM (VALUES (1)) AS t(truncate)", "postgres", "write", true],
+    ["SELECT truncate AS result FROM (VALUES (1)) AS t(truncate)", "postgres", "read", false],
+    // TRUNCATE followed by a target, wherever it sits.
+    ["CREATE TASK k AS TRUNCATE t", "snowflake", "write", true],
+    ["CREATE OR REPLACE TASK k WAREHOUSE = w AS TRUNCATE t", "snowflake", "write", true],
+    ["CREATE PROCEDURE p() TRUNCATE t", "mysql", "write", true],
+    ["CREATE PROCEDURE p() SQL SECURITY INVOKER TRUNCATE t", "mysql", "write", true],
+    ["CREATE PROCEDURE p() LANGUAGE sql BEGIN ATOMIC TRUNCATE t; END; CALL p()", "postgres", "write", true],
+    ["ALTER TABLE t TRUNCATE PARTITION p", "oracle", "write", true],
+    ["ALTER TABLE t TRUNCATE SUBPARTITION p", "oracle", "write", true],
+    ["ALTER TABLE t TRUNCATE PARTITION p", "mysql", "write", true],
+    ["ALTER TABLE t TRUNCATE PARTITION (d = 1)", "databricks", "write", true],
+    ["SELECT t.truncate x FROM t", "postgres", "read", false],
+    ["SELECT t./* column comment */truncate x FROM t", "postgres", "read", false],
+    ["SELECT truncate ISNULL, truncate NOTNULL FROM t", "postgres", "read", false],
+    ["SELECT truncate GLOB 'a*' FROM t", "sqlite", "read", false],
+    ["SELECT sum(x) OVER (ORDER BY truncate ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) FROM t", "postgres", "read", false],
+    ["SELECT truncate, COUNT(*) FROM t GROUP BY truncate WITH ROLLUP", "mysql", "read", false],
+    ["SELECT id FROM t CONNECT BY PRIOR id = truncate START WITH id = 1", "oracle", "read", false],
+    ["SELECT truncate OPERATOR(pg_catalog.+) 1 FROM t", "postgres", "read", false],
+    ["SELECT * FROM t ORDER BY truncate ASCENDING", "clickhouse", "read", false],
+    ["SELECT sum(x) OVER (ORDER BY truncate RANGE BETWEEN 1 PRECEDING AND CURRENT ROW) FROM t", "postgres", "read", false],
+    ["SELECT * FROM t ORDER BY truncate SETTINGS max_threads = 1", "clickhouse", "read", false],
+    ["BEGIN NOT ATOMIC TRUNCATE $t; END", "mysql", "write", true],
+    ["BEGIN NOT ATOMIC TRUNCATE \u00fcber; END", "mysql", "write", true],
+    ["BEGIN NOT ATOMIC DECLARE CONTINUE HANDLER FOR 1062 TRUNCATE nulls; SELECT 1; END", "mysql", "write", true],
+    ["BEGIN NOT ATOMIC DECLARE EXIT HANDLER FOR SQLSTATE '23000' TRUNCATE t; SELECT 1; END", "mysql", "write", true],
+    ["BEGIN NOT ATOMIC DECLARE CONTINUE HANDLER FOR 1146" + " ".repeat(300) + "TRUNCATE t; SELECT * FROM missing; END", "mysql", "write", true],
+    ["IF 1 THEN TRUNCATE t; END IF", "mysql", "write", true],
+    ['BEGIN IF (x) THEN TRUNCATE "db"."t"; END IF; END', "snowflake", "write", true],
+    ['BEGIN IF (x) THEN TRUNCATE db."t"; END IF; END', "snowflake", "write", true],
+    ["BEGIN IF (x) THEN TRUNCATE IF EXISTS t; END IF; END", "snowflake", "write", true],
+    ["BEGIN IF (x) THEN TRUNCATE IDENTIFIER('t'); END IF; END", "snowflake", "write", true],
+    ["BEGIN NOT ATOMIC IF 1 THEN TRUNCATE `db`.`t`; END IF; END", "mysql", "write", true],
+    ["SELECT CASE WHEN a THEN truncate.x END FROM t", "postgres", "read", false],
+    ["EXPLAIN (ANALYZE) TRUNCATE t", "duckdb", "write", true],
+    ["EXPLAIN (ANALYZE, FORMAT JSON) TRUNCATE t", "duckdb", "write", true],
+    ["SELECT DEFAULT(truncate) FROM t", "mysql", "read", false],
+    ["SELECT atomic truncate FROM (VALUES (1)) AS t(atomic)", "postgres", "read", false],
+    ["SELECT found truncate, x FROM t", "postgres", "read", false],
+    ["SELECT seq.NEXTVAL", "mysql", "write", false],
+    ["SELECT s./*comment*/NEXTVAL FROM dual WHERE ROWNUM = 1", "oracle", "write", false],
+    ['SELECT s . /* c */ "NEXTVAL" FROM dual', "snowflake", "write", false],
+    ["BEGIN NOT ATOMIC IF 1 THEN TRUNCATE t NOWAIT; END IF; END", "mysql", "write", true],
+    ["BEGIN NOT ATOMIC IF 1 THEN TRUNCATE t WAIT 5; END IF; END", "mysql", "write", true],
+    ["IF 1 THEN TRUNCATE t NOWAIT; END IF", "mysql", "write", true],
+    ["CASE WHEN 1 THEN TRUNCATE t NOWAIT; END CASE", "mysql", "write", true],
+    ["IF 1 THEN TRUNCATE `d`.`t` WAIT 5; END IF", "mysql", "write", true],
+    // ClickHouse heredocs are strings.
+    ["SELECT $x$'$x$ AS a FROM jdbc($x$ds$x$, $x$INSERT INTO t VALUES (1)$x$) WHERE $x$'$x$ = 1", "clickhouse", "write", false],
+    ["SELECT $heredoc$it's$heredoc$ AS v", "clickhouse", "read", false],
+    ["SELECT $1$'$1$ AS a FROM jdbc($x$ds$x$, $x$INSERT INTO t VALUES (1)$x$) WHERE $1$'$1$ = 1", "clickhouse", "write", false],
+    ["SELECT $-$'$-$ AS a FROM jdbc($x$ds$x$, $x$INSERT INTO t VALUES (1)$x$) WHERE $-$'$-$ = 1", "clickhouse", "write", false],
+    ["SELECT\u00a0$x$'$x$ AS a FROM jdbc($x$ds$x$, $x$INSERT INTO t VALUES (1)$x$) WHERE\u00a0$x$'$x$ = $x$a$x$", "clickhouse", "write", false],
+    ["BEGIN IF (1=1) THEN TRUNCATE db..t; END IF; END", "snowflake", "write", true],
+    ["SELECT * FROM postgres_attach('dbname=x')", "duckdb", "write", false],
+    // Databricks functions that call outside systems.
+    ["SELECT http_request(conn => 'c', method => 'POST', path => '/x')", "databricks", "write", false],
+    ["SELECT * FROM remote_query('pg', database => 'db', query => 'DELETE FROM t RETURNING *')", "databricks", "write", false],
+    ["SELECT * FROM TABLE(GETNEXTVAL(s))", "snowflake", "write", false],
+    ["SELECT CASE WHEN true THEN truncate ELSE 0 END FROM metrics", "postgres", "read", false],
+    ["SELECT CASE WHEN a THEN 1 ELSE truncate END FROM metrics", "duckdb", "read", false],
+    // MySQL 8 reads quotes in a /*+ hint, so one there makes the SQL uncertain.
+    ["SELECT /*+ QB_NAME(`*/`) */ 1 INTO OUTFILE '/tmp/x' -- `", "mysql", "write", false],
+    ["SELECT /*+ SET_VAR(sql_mode='*/') */ 1 INTO OUTFILE '/tmp/x' -- '", "mysql", "write", false],
+    ["SELECT /*+ MAX_EXECUTION_TIME(1000) */ * FROM t", "mysql", "read", false],
+    // Snowflake: ->> chains statements; ASYNC ( and := ( hold one.
+    ["SELECT 1 ->> TRUNCATE t -- LIMIT", "snowflake", "write", true],
+    ["SELECT 1 ->> DELETE FROM t", "snowflake", "write", false],
+    ["BEGIN ASYNC (TRUNCATE t); AWAIT ALL; END", "snowflake", "write", true],
+    ["BEGIN LET r RESULTSET := (TRUNCATE t); RETURN TABLE(r); END", "snowflake", "write", true],
+    ["SELECT data ->> 'a' FROM t", "postgres", "read", false],
+    ["SELECT j ->> truncate FROM t", "postgres", "read", false],
+    ["BEGIN LET r RESULTSET DEFAULT (TRUNCATE t); RETURN TABLE(r); END", "snowflake", "write", true],
+    // DuckDB runs TRUNCATE under EXPLAIN ANALYZE and PREPARE ... AS.
+    ["EXPLAIN ANALYZE TRUNCATE t", "duckdb", "write", true],
+    ["PREPARE p AS TRUNCATE t; EXECUTE p", "duckdb", "write", true],
+    // Spark/Hive TRANSFORM ... USING runs a command; reflect() calls any Java method.
+    ["SELECT TRANSFORM(a) USING 'rm -rf /tmp/x' FROM t", "databricks", "write", false],
+    ["SELECT reflect('java.lang.System', 'exit', 0)", "spark", "write", false],
+    ["SELECT transform(arr, x -> x + 1) FROM t", "databricks", "read", false],
+    ["SELECT transform(arr, x -> x + 1) FROM t JOIN u USING (id)", "spark", "read", false],
+    ["SELECT TRANSFORM(a) USING /*c*/ 'touch /tmp/proof' FROM t", "spark", "write", false],
+    ['SELECT TRANSFORM(a) USING "touch /tmp/proof" FROM t', "databricks", "write", false],
+    // UESCAPE without a U& literal is a word, not an escape character.
+    ["UPDATE t SET uescape = 1", "postgres", "write", false],
+    ["UPDATE t SET uescape = menu&1", "postgres", "write", false],
+    // Sequence advances through a quoted name, and DuckDB extension functions that write.
+    ['SELECT s."NEXTVAL"', "snowflake", "write", false],
+    ['SELECT "NEXT VALUE FOR" FROM metrics', "postgres", "read", false],
+    ["SELECT s.nextval FROM metrics s", "postgres", "read", false],
+    ["SELECT CASE WHEN true THEN truncate SIMILAR TO 'a%' ELSE false END FROM metrics", "postgres", "read", false],
+    ["SELECT CASE WHEN a THEN truncate AT TIME ZONE 'UTC' END FROM metrics", "postgres", "read", false],
+    ["SELECT CASE WHEN a THEN 1 ELSE truncate END;", "postgres", "read", false],
+    ["SELECT * FROM bigquery_execute('p', 'DELETE FROM d.t WHERE true')", "duckdb", "write", false],
+    ["SELECT * FROM ducklake_expire_snapshots('lake', older_than => now())", "duckdb", "write", false],
+    // ClickHouse jdbc() can run a mutation through the bridge.
+    ["SELECT * FROM jdbc('ch-server?mutation', 'INSERT INTO audit VALUES (1)')", "clickhouse", "write", false],
+    // Identifiers that are never strings are not read as SQL.
+    ['CREATE TABLE t ("truncate" int)', "postgres", "write", false],
+    ["INSERT INTO t ([truncate]) VALUES (1)", "sqlserver", "write", false],
+    ['EXEC("DROP DATABASE prod")', "sqlserver", "write", true],
+    // Each engine's escapes: whitespace escapes are spaces, letter escapes the letter, and SQL Server
+    // joins a line ending in a backslash.
+    ["DO E'BEGIN DROP\\nSCH\\eMA s CASCADE; END'", "postgres", "write", true],
+    ["EXECUTE IMMEDIATE 'DROP\\nSCH\\eMA s'", "snowflake", "write", true],
+    ["PREPARE s FROM 'DROP\\nDAT\\aBASE x'", "mysql", "write", true],
+    ["EXEC('DROP DATA\\\nBASE x')", "sqlserver", "write", true],
+    ["EXEC sp_executesql N'TRUN\\\r\nCATE TABLE t'", "fabric", "write", true],
+    // A Spark hint's body is code: a quote in it makes the SQL uncertain.
+    ["WITH c AS (SELECT /*+ X('*/') */ ' AS a) SELECT 1 AS z /*' AS a) INSERT INTO t SELECT * FROM c -- */", "databricks", "write", false],
+    ["SELECT /*+ BROADCAST(t) */ * FROM t", "databricks", "read", false],
+    // Oracle q-quotes are uncertain.
+    ["DECLARE v VARCHAR2(9); w VARCHAR2(9); BEGIN v := q'{ ' }'; w := ' -- '; EXECUTE IMMEDIATE 'TRUNCATE TABLE t'; END;", "oracle", "write", true],
+    // A quoted target is still a target; Spark "..." is a string dynamic SQL can run.
+    ["BEGIN NOT ATOMIC TRUNCATE `t`; END", "mysql", "write", true],
+    ["SELECT \"truncate\" FROM t", "postgres", "read", false],
+    ['EXECUTE IMMEDIATE "TRUNCATE TABLE t"', "databricks", "write", true],
+    ['EXECUTE IMMEDIATE "DROP SCHEMA s CASCADE"', "spark", "write", true],
+    // DROP NAMESPACE is Spark's DROP SCHEMA, nested too.
+    ["EXECUTE IMMEDIATE 'DROP NAMESPACE s CASCADE'", "databricks", "write", true],
+    ["BEGIN DROP NAMESPACE s CASCADE; END", "databricks", "write", true],
+    ["SELECT dbo.truncate(x, 2) FROM t", "sqlserver", "read", false],
+    // Other names for dropping a database or schema.
+    ["DROP CATALOG prod CASCADE", "databricks", "write", true],
+    ["DROP EXTERNAL SCHEMA ds", "bigquery", "write", true],
+    ["DROP PLUGGABLE DATABASE pdb1 INCLUDING DATAFILES", "oracle", "write", true],
+    // A column or alias named truncate is not the statement.
+    ["SELECT truncate FROM t", "postgres", "read", false],
+    ["SELECT 1 AS truncate", "postgres", "read", false],
+    ["BEGIN TRUNCATE t; END", "snowflake", "write", true],
+    // Identifiers take no backslash escapes.
+    ['SELECT "a\\" FROM t', "postgres", "read", false],
+    ["SELECT `a\\` FROM t", "databricks", "read", false],
     // Sequence advances are writes in every dialect.
     ["SELECT NEXT VALUE FOR dbo.s", "sqlserver", "write", false],
     ["SELECT s.NEXTVAL FROM dual", "oracle", "write", false],
@@ -487,11 +656,28 @@ describe("each engine's own literal and statement rules", () => {
       expect(result.blocked).toBe(blocked)
     })
 
-  test("stays linear on repeated TRUNCATE inside a long comment", () => {
-    const started = performance.now()
-    classifyAndCheck("TRUNCATE /* ".repeat(40_000) + "*/ ( '", "mysql")
-    // Every TRUNCATE reaching the same comment close and whitespace after it.
-    classifyAndCheck("SELECT " + "TRUNCATE /*".repeat(40_000) + "*/" + " ".repeat(400_000) + "(", "postgres")
-    expect(performance.now() - started).toBeLessThan(2_000)
+  test("stays linear on large crafted inputs", () => {
+    // Each input on its own: a quadratic scan takes seconds to minutes here, a linear one well under a second.
+    const inputs: [string, string][] = [
+      ["TRUNCATE /* ".repeat(40_000) + "*/ ( '", "mysql"],
+      // An unclosed EXPLAIN option list, and a long dotted name after THEN truncate.
+      ["EXPLAIN (" + "ANALYZE ".repeat(40_000), "duckdb"],
+      ["EXPLAIN (".repeat(40_000), "duckdb"],
+      ["IF 1 THEN TRUNCATE IDENTIFIER(".repeat(20_000), "snowflake"],
+      // Long whitespace in SQL read with certainty (the TRUNCATE rules look behind only after a match).
+      ["SELECT 1" + " ".repeat(400_000) + "FROM t", "postgres"],
+      ["SELECT t." + " ".repeat(400_000) + "truncate x FROM t", "postgres"],
+      ["SELECT " + "t. truncate x, ".repeat(40_000) + "1 FROM t", "postgres"],
+      ["SELECT CASE WHEN true THEN truncate -" + "a.".repeat(2_000) + "a + 0 ELSE 0 END FROM t", "duckdb"],
+      // A write's literal full of comment openers.
+      ["INSERT INTO t VALUES ($$" + "/*x*/".repeat(100_000) + "x$$)", "postgres"],
+      // Every TRUNCATE reaching the same comment close and whitespace after it.
+      ["SELECT " + "TRUNCATE /*".repeat(40_000) + "*/" + " ".repeat(400_000) + "(", "postgres"],
+    ]
+    for (const [sql, dialect] of inputs) {
+      const started = performance.now()
+      classifyAndCheck(sql, dialect)
+      expect(performance.now() - started).toBeLessThan(2_000)
+    }
   })
 })
diff --git a/packages/opencode/test/altimate/tools/sql-comment-safety.test.ts b/packages/opencode/test/altimate/tools/sql-comment-safety.test.ts
index 2ba8be9012..96eca55685 100644
--- a/packages/opencode/test/altimate/tools/sql-comment-safety.test.ts
+++ b/packages/opencode/test/altimate/tools/sql-comment-safety.test.ts
@@ -233,3 +233,50 @@ test("omitted SQL warehouse cannot assume the registry dialect before dbt execut
     dbt.mockRestore()
   }
 })
+
+test("on the dbt route, Jinja in a comment or literal is a write, and a hard-denied run_query is blocked", async () => {
+  Registry.setConfigs({ fallback: { type: "postgres" } })
+  const dbt = spyOn(DbtAdapterType, "dbtProjectAdapterType").mockResolvedValue({ project: true, type: "postgres", targetTypes: ["postgres"] })
+  const dispatch = spyOn(Dispatcher, "call").mockResolvedValue({ columns: ["x"], rows: [[1]], row_count: 1, truncated: false } as never)
+  const run = async (query: string) => {
+    const asked: string[] = []
+    const tool = await initTool(SqlExecuteTool)
+    const result = tool.execute(
+      { query, limit: 10 },
+      {
+        sessionID: SessionID.make("ses_dbt_jinja"),
+        messageID: MessageID.make("msg_dbt_jinja"),
+        callID: "",
+        agent: "analyst",
+        abort: AbortSignal.any([]),
+        messages: [],
+        metadata: () => {},
+        ask: async (request: { permission: string }) => {
+          asked.push(request.permission)
+          throw new Error("write permission denied")
+        },
+      },
+    )
+    return { result, asked }
+  }
+  try {
+    const write = await run("SELECT 1 AS x /* {% do run_query('DELETE FROM t') %} */")
+    await expect(write.result).rejects.toThrow("write permission denied")
+    expect(write.asked).toEqual(["sql_execute_write"])
+    const drop = await run("SELECT 1 AS x -- {% do run_query('DROP SCHEMA prod CASCADE') %}")
+    await expect(drop.result).rejects.toThrow("blocked for safety")
+    // A Jinja comment removes text, which can expose a statement the SQL lexing reads as a literal.
+    const hidden = await run("SELECT 'x' /* {# */ || ' #} */) AS q ; DROP SCHEMA s CASCADE ; SELECT 1 -- '")
+    await expect(hidden.result).rejects.toThrow("blocked for safety")
+    // A Mongo default's command goes to dbt first as well.
+    Registry.setConfigs({ fallback: { type: "mongodb" } })
+    const mongo = await run(`{"command":"find","collection":"c","filter":{"x":"{% do run_query('DELETE FROM t') %}"}}`)
+    await expect(mongo.result).rejects.toThrow("write permission denied")
+    expect(mongo.asked).toEqual(["sql_execute_write"])
+    expect(dispatch).not.toHaveBeenCalled()
+  } finally {
+    dispatch.mockRestore()
+    dbt.mockRestore()
+    Registry.reset()
+  }
+})
diff --git a/packages/opencode/test/altimate/tools/sql-explain.test.ts b/packages/opencode/test/altimate/tools/sql-explain.test.ts
index b3bdccbc8b..09c9b0bc14 100644
--- a/packages/opencode/test/altimate/tools/sql-explain.test.ts
+++ b/packages/opencode/test/altimate/tools/sql-explain.test.ts
@@ -501,7 +501,7 @@ describe("SqlExplainTool.execute", () => {
 
       expect(result.metadata.success).toBe(false)
       expect(result.metadata.error_class).toBe("input_validation")
-      expect(String(result.output)).toContain("only a single read-only query can be analyzed")
+      expect(String(result.output)).toMatch(/only a single read-only query can be analyzed|single statement can be explained/)
       expect(spy).not.toHaveBeenCalled()
       spy.mockRestore()
     })
diff --git a/packages/opencode/test/altimate/tools/sql-read-safety.test.ts b/packages/opencode/test/altimate/tools/sql-read-safety.test.ts
index 024d8f1be5..f0cfe21491 100644
--- a/packages/opencode/test/altimate/tools/sql-read-safety.test.ts
+++ b/packages/opencode/test/altimate/tools/sql-read-safety.test.ts
@@ -106,7 +106,7 @@ describe("read-only SQL parser boundary", () => {
     expect(maskSql(tokenizeSql("SELECT $$-- ; DELETE$$ AS a -- actual comment", "postgres").tokens)).toMatch(/^SELECT\s+AS a\s+$/)
   })
 
-  for (const dialect of ["mysql", "sqlite", "bigquery", "clickhouse", "unknown", undefined]) {
+  for (const dialect of ["mysql", "sqlite", "bigquery", "unknown", undefined]) {
     test(`dollar delimiters leave statements visible for ${dialect ?? "omitted"} dialect`, () => {
       const sql = "SELECT $x$; DELETE FROM t; $x$"
       const { tokens, complete } = tokenizeSql(sql, dialect)