@trilogy-data/trilogy-studio-components 0.1.27 → 0.1.28

package/dist/{editorStore-CAlMJftt.js → editorStore-CiWH_nGN.js}

@@ -1,5 +1,5 @@
 import { s as e, u as t } from "./base-DuKnVD-A.js";
-import { $ as n, $t as r, B as i, Bt as a, Ft as o, H as s, J as c, Lt as l, Q as u, Qt as d, R as f, Rt as p, St as m, W as h, a as g, an as ee, at as te, c as ne, cn as _, ct as re, dt as ie, en as ae, ft as oe, gn as se, hn as ce, ht as le, i as ue, in as de, k as v, l as fe, ln as pe, mt as me, n as he, nn as ge, o as _e, on as ve, ot as ye, pt as be, r as xe, rn as Se, s as Ce, st as we, t as Te, tn as Ee, tt as De, ut as Oe, xt as ke, y, z as Ae, zt as je } from "./llm-ClU-2rgk.js";
+import { $ as n, $t as r, B as i, Bt as a, Ft as o, H as s, J as c, Lt as l, Q as u, Qt as d, R as f, Rt as p, St as m, W as h, a as g, an as ee, at as te, c as ne, cn as _, ct as re, dt as ie, en as ae, ft as oe, gn as se, hn as ce, ht as le, i as ue, in as de, k as v, l as fe, ln as pe, mt as me, n as he, nn as ge, o as _e, on as ve, ot as ye, pt as be, r as xe, rn as Se, s as Ce, st as we, t as Te, tn as Ee, tt as De, ut as Oe, xt as ke, y, z as Ae, zt as je } from "./llm-Wf4Etiu9.js";
 import { a as Me, c as Ne, f as Pe, l as Fe, n as Ie, o as Le, u as Re } from "./connections-CWH3zwH2.js";
 import { reactive as ze } from "vue";
 import { defineStore as b } from "pinia";

package/dist/{globalChatRuntime-CyS7vH51.js → globalChatRuntime-DAERmOdu.js}

@@ -1,8 +1,8 @@
 import { t as e } from "./screenBridge-BLlMdfhn.js";
-import { K as t, X as n, Z as r, ct as i, et as a } from "./llm-ClU-2rgk.js";
-import { a as o, d as s } from "./stores-LlyiwSbK.js";
-import { a as c, r as l } from "./editorStore-CAlMJftt.js";
-import { t as u } from "./useGlobalChatPanel-DuP1q7_f.js";
+import { K as t, X as n, Z as r, ct as i, et as a } from "./llm-Wf4Etiu9.js";
+import { a as o, d as s } from "./stores-DTGQZ4_B.js";
+import { a as c, r as l } from "./editorStore-CiWH_nGN.js";
+import { t as u } from "./useGlobalChatPanel-D9kMZLeF.js";
 import { watch as d } from "vue";
 //#region llm/registry/renderToolList.ts
 function f(e) {

package/dist/{helpers-BbVlA3Pf.js → helpers-C2pARJoZ.js}

@@ -1,8 +1,8 @@
 import { s as e } from "./base-DuKnVD-A.js";
-import { It as t, gn as n } from "./llm-ClU-2rgk.js";
-import "./stores-LlyiwSbK.js";
-import { M as r, Y as i, at as a, tt as o } from "./editorStore-CAlMJftt.js";
-import { n as s } from "./reconcile-DXIvGwBk.js";
+import { It as t, gn as n } from "./llm-Wf4Etiu9.js";
+import "./stores-DTGQZ4_B.js";
+import { M as r, Y as i, at as a, tt as o } from "./editorStore-CiWH_nGN.js";
+import { n as s } from "./reconcile-BH_ReGDR.js";
 //#region models/helpers.ts
 var c = class {
 	editorStore;

package/dist/llm/data/prompts.d.ts

@@ -1,6 +1,6 @@
 import { ModelConceptInput } from './models';
 export declare function filterDatePartConcepts(concepts: ModelConceptInput[]): ModelConceptInput[];
-export declare const leadIn = "You are a world-class assistant for generating queries in Trilogy, a SQL inspired language with similar syntax, Use the the syntax description below and field information to answer the user request. Return your answer with a short reasoning and a valid Trilogy query in triple double quotes.\n\n# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` \u2014 replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; \u2014 a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a \u2286 b; b authoritative for the key) or `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` declares EXACT domain equivalence. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order \u2014 all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection \u2014 grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation \u2014 it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`, and joined back like any concept (`subset join foo.key = other.key`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL \u2014 what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT \u2014 never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select \u2014 hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain \u2014 a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** \u2014 a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column) \u2014 never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` \u2261 `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages \u2014 give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship \u2014 the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology \u2014 use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE \u2014 its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` \u2192 leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows \u2014 use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain \u2014 on a select item (`~customer.id`) or in `merge a into ~b` (a \u2286 b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary \u2014 wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end \u2212 start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.\n\n\nFunctions (the `aggregate` family doubles as the Aggregate Functions list):\naggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)\nstring: concat|concat_strict|concat_ws(<arg1>, <arg2>, ..., <argN>) ; split|strpos|contains|trim|regexp_contains|format_time|parse_time(<arg1>, <arg2>) ; len|lower|upper|ltrim|rtrim|hex(<arg1>) ; substring|replace|regexp_extract|regexp_replace(<arg1>, <arg2>, <arg3>)\ndate/time: date|datetime|timestamp|second|minute|hour|day|day_of_week|week|month|quarter|year|month_name|day_name|unix_to_timestamp(<arg1>) ; date_spine(<arg1>, <arg2>) ; current_date|current_datetime()\n  date_part(<arg1>, <arg2>) e.g. date_part('2020-01-01'::date, year)\n  date_truncate(<arg1>, <arg2>) e.g. date_trunc('2020-01-01'::date, month)\n  date_add(<arg1>, <arg2>, <arg3>) e.g. date_add('2020-01-01'::date, month, 1)\n  date_sub(<arg1>, <arg2>, <arg3>) e.g. date_sub('2020-01-01'::date, day, 1)\n  date_diff(<arg1>, <arg2>, <arg3>) e.g. date_diff('2020-01-01'::date, '2020-01-02'::date, day)\n  current_timestamp() e.g. now()\narray/map/struct: unnest|array_distinct|array_sum|map_keys|map_values(<arg1>) ; struct|array(<arg1>, <arg2>, ..., <argN>) ; array_sort|array_to_string(<arg1>, <arg2>) ; array_transform|array_filter|generate_array(<arg1>, <arg2>, <arg3>)\nmath: mod|round|log|power|hash(<arg1>, <arg2>) ; abs|sqrt|random|floor|ceil(<arg1>)\nconditional/cast: case|simple_case|coalesce|greatest|least(<arg1>, <arg2>, ..., <argN>) ; cast|nullif(<arg1>, <arg2>) ; isnull(<arg1>)\ngeo: geo_from_text|geo_x|geo_y|geo_centroid(<arg1>) ; geo_point|geo_distance(<arg1>, <arg2>) ; geo_transform(<arg1>, <arg2>, <arg3>)\nother: is_not_distinct(<arg1>, <arg2>)\n\nSome types carry trait metadata that adds meaning: 'latitude', 'longitude' and 'currency' are all 'float' with extra semantics.\n\nAdditional syntax examples:\nThese less-common patterns have complete, copy-pasteable examples. Do NOT guess\nthe syntax \u2014 print the full example on demand with\n`trilogy agent-info syntax example <name>`:\n- `python-datasource` - run a local Python script as a datasource: wrap a function in `trilogy.io.run`, which writes the Arrow IPC stream to stdout for you from a table, dataframe, or list of dicts; declare concepts, map script columns in `datasource (...)`, use `grain (...) file `path.py`;`, then reference locally declared concepts WITHOUT the datasource name as a prefix\n- `query-structure` - the exact clause order of a query, and how to define a rowset\n- `chart` - drawing a result rather than tabulating it: `chart layer <type> ( <role> <- <expr> as <name>, ... )` is a STATEMENT that wraps its own query - the bindings ARE the select; layers, roles, scales, reference lines and `from <select>`\n- `filtered-aggregate` - aggregating only the rows matching a condition, and counting rows correctly\n- `row-level-where-vs-having` - choosing between WHERE and HAVING for a threshold condition\n- `nested-aggregate-group-average` - the \"above 1.2x the group norm\" ask: inner `sum(x) by entity, group`, outer `avg(inner) by group`; use WHERE for the comparison when a separate output-only filter must not narrow the benchmark\n- `pivot-columns` - turn distinct values of a dimension into columns with a reusable `def` macro + filtered aggregate\n- `case-when` - `case when <cond> then <val> ... else <val> end` to relabel codes or bucket a measure into a derived value, then group by it\n- `window-period-over-period` - `lead`/`lag` over (order by ...) to compare a row to an offset period; build a scoped set with `<col> ? <cond>` and filter with `in`\n- `union-stack-channels` - stacking several sources/channels together as rows\n- `except-intersect-setops` - set difference or co-occurrence across sources\n- `scoped-join` - blending a second fact model or rowset into one query\n- `existence-anti-join` - keeping or excluding rows by whether a key matches another model\n- `set-intersect-difference` - count tuples present in / absent from other groups via per-group presence flags - the multi-column INTERSECT/EXCEPT idiom\n- `dedup-before-aggregate` - a `rowset <- select <cols>` collapses identical tuples to one row (SELECT-grain group by) before a downstream sum - the `distinct`/UNION substitute\n- `rollup` - subtotals and a grand total in one pass, and labeling the levels\n- `rank-over-rollup` - rank rollup subtotals/leaves with a SINGLE `rank(a,b) over (partition by level, parent ...)` - not separate ranks per level\n- `staged-membership` - filtering a fine-grain query by a set computed at a coarser grain\n- `correlated-exists-via-grouped-counts` - translate `EXISTS other` / `NOT EXISTS other matching` over the same model into two `count(...) by <grain>` compared in `where` (`> 1` = another exists, `= 1` = no other matches) - pin the correlation grain with `by`\n\nFor any response to the user, use this format -> put your actual response within triple double quotes with thinking and justification before it, in this format (replace placeholders with relevant content): Reasoning: {{reasoning}} \"\"\"{{ response }}\"\"\"\n";
+export declare const leadIn = "You are a world-class assistant for generating queries in Trilogy, a SQL inspired language with similar syntax, Use the the syntax description below and field information to answer the user request. Return your answer with a short reasoning and a valid Trilogy query in triple double quotes.\n\n# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` \u2014 replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; \u2014 a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union\\|equal join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a \u2286 b; b authoritative for the key), `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept) or `equal join a = b` (ONE domain: a is an alias of b, every value on either side is on the other; the join renders inner).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` is the persistent equivalent of `equal join a = b`. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order \u2014 all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection \u2014 grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union|equal join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation \u2014 it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`. Pairing an output with a concept outside the rowset REQUIRES a declared join on the rowset's key (`subset join foo.key = key`); without one the query is disconnected. Alternatively project the concept inside the rowset and filter or select it through the handle (`where foo.cat = 'a'`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL \u2014 what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT \u2014 never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select \u2014 hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain \u2014 a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** \u2014 a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column; a scalar must also be one row: an aggregate with no `by`, or `limit 1`) \u2014 never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` \u2261 `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages \u2014 give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship \u2014 the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology \u2014 use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE \u2014 its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` \u2192 leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows \u2014 use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain \u2014 on a select item (`~customer.id`) or in `merge a into ~b` (a \u2286 b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary \u2014 wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end \u2212 start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.\n\n\nFunctions (the `aggregate` family doubles as the Aggregate Functions list):\naggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)\nstring: concat|concat_strict|concat_ws(<arg1>, <arg2>, ..., <argN>) ; split|strpos|contains|trim|regexp_contains|format_time|parse_time(<arg1>, <arg2>) ; len|lower|upper|ltrim|rtrim|hex(<arg1>) ; substring|replace|regexp_extract|regexp_replace(<arg1>, <arg2>, <arg3>)\ndate/time: date|datetime|timestamp|second|minute|hour|day|day_of_week|week|month|quarter|year|month_name|day_name|unix_to_timestamp(<arg1>) ; date_spine(<arg1>, <arg2>) ; current_date|current_datetime()\n  date_part(<arg1>, <arg2>) e.g. date_part('2020-01-01'::date, year)\n  date_truncate(<arg1>, <arg2>) e.g. date_trunc('2020-01-01'::date, month)\n  date_add(<arg1>, <arg2>, <arg3>) e.g. date_add('2020-01-01'::date, month, 1)\n  date_sub(<arg1>, <arg2>, <arg3>) e.g. date_sub('2020-01-01'::date, day, 1)\n  date_diff(<arg1>, <arg2>, <arg3>) e.g. date_diff('2020-01-01'::date, '2020-01-02'::date, day)\n  current_timestamp() e.g. now()\narray/map/struct: unnest|array_distinct|array_sum|map_keys|map_values(<arg1>) ; struct|array(<arg1>, <arg2>, ..., <argN>) ; array_sort|array_to_string(<arg1>, <arg2>) ; array_transform|array_filter|generate_array(<arg1>, <arg2>, <arg3>)\nmath: mod|round|log|power|hash(<arg1>, <arg2>) ; abs|sqrt|random|floor|ceil(<arg1>)\nconditional/cast: case|simple_case|coalesce|greatest|least(<arg1>, <arg2>, ..., <argN>) ; cast|nullif(<arg1>, <arg2>) ; isnull(<arg1>)\ngeo: geo_from_text|geo_x|geo_y|geo_centroid(<arg1>) ; geo_point|geo_distance(<arg1>, <arg2>) ; geo_transform(<arg1>, <arg2>, <arg3>)\nother: is_not_distinct(<arg1>, <arg2>) ; prior(<arg1>) ; grain_pin(<arg1>, <arg2>, ..., <argN>)\n\nSome types carry trait metadata that adds meaning: 'latitude', 'longitude' and 'currency' are all 'float' with extra semantics.\n\nAdditional syntax examples:\nThese less-common patterns have complete, copy-pasteable examples. Do NOT guess\nthe syntax \u2014 print the full example on demand with\n`trilogy agent-info syntax example <name>`:\n- `python-datasource` - run a local Python script as a datasource: wrap a function in `trilogy.io.run`, which writes the Arrow IPC stream to stdout for you from a table, dataframe, or list of dicts; declare concepts, map script columns in `datasource (...)`, use `grain (...) file `path.py`;`, then reference locally declared concepts WITHOUT the datasource name as a prefix\n- `query-structure` - the exact clause order of a query, and how to define a rowset\n- `chart` - drawing a result rather than tabulating it: `chart layer <type> ( <role> <- <expr> as <name>, ... )` is a STATEMENT that wraps its own query - the bindings ARE the select; layers, roles, scales, reference lines and `from <select>`\n- `filtered-aggregate` - aggregating only the rows matching a condition, and counting rows correctly\n- `row-level-where-vs-having` - choosing between WHERE and HAVING for a threshold condition\n- `nested-aggregate-group-average` - the \"above 1.2x the group norm\" ask: inner `sum(x) by entity, group`, outer `avg(inner) by group`; use WHERE for the comparison when a separate output-only filter must not narrow the benchmark\n- `pivot-columns` - turn distinct values of a dimension into columns with a reusable `def` macro + filtered aggregate\n- `case-when` - `case when <cond> then <val> ... else <val> end` to relabel codes or bucket a measure into a derived value, then group by it\n- `window-period-over-period` - `lead`/`lag` over (order by ...) to compare a row to an offset period; build a scoped set with `<col> ? <cond>` and filter with `in`\n- `union-stack-channels` - stacking several sources/channels together as rows\n- `except-intersect-setops` - set difference or co-occurrence across sources\n- `scoped-join` - blending a second fact model or rowset into one query\n- `existence-anti-join` - keeping or excluding rows by whether a key matches another model\n- `set-intersect-difference` - count tuples present in / absent from other groups via per-group presence flags - the multi-column INTERSECT/EXCEPT idiom\n- `dedup-before-aggregate` - a `rowset <- select <cols>` collapses identical tuples to one row (SELECT-grain group by) before a downstream sum - the `distinct`/UNION substitute\n- `rollup` - subtotals and a grand total in one pass, and labeling the levels\n- `rank-over-rollup` - rank rollup subtotals/leaves with a SINGLE `rank(a,b) over (partition by level, parent ...)` - not separate ranks per level\n- `staged-membership` - filtering a fine-grain query by a set computed at a coarser grain\n- `correlated-exists-via-grouped-counts` - translate `EXISTS other` / `NOT EXISTS other matching` over the same model into two `count(...) by <grain>` compared in `where` (`> 1` = another exists, `= 1` = no other matches) - pin the correlation grain with `by`\n\nFor any response to the user, use this format -> put your actual response within triple double quotes with thinking and justification before it, in this format (replace placeholders with relevant content): Reasoning: {{reasoning}} \"\"\"{{ response }}\"\"\"\n";
 export declare function conceptsToFieldPrompt(conceptInputs: ModelConceptInput[]): string;
 export declare function createPrompt(query: string, conceptInputs: ModelConceptInput[]): string;
 export declare function createFilterPrompt(query: string, conceptInputs: ModelConceptInput[]): string;

package/dist/llm/data/prompts.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"prompts.d.ts","sourceRoot":"","sources":["../../../llm/data/prompts.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,UAAU,CAAA;AAIjD,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,iBAAiB,EAAE,GAAG,iBAAiB,EAAE,CAEzF;AAED,eAAO,MAAM,MAAM,imiBAKlB,CAAA;AAED,wBAAgB,qBAAqB,CAAC,aAAa,EAAE,iBAAiB,EAAE,UAOvE;AAED,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,aAAa,EAAE,iBAAiB,EAAE,UAG7E;AAED,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,MAAM,EAAE,aAAa,EAAE,iBAAiB,EAAE,UAGnF;AAED,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,MAAM,EAAE,aAAa,EAAE,iBAAiB,EAAE,UAGtF"}
+{"version":3,"file":"prompts.d.ts","sourceRoot":"","sources":["../../../llm/data/prompts.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,UAAU,CAAA;AAIjD,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,iBAAiB,EAAE,GAAG,iBAAiB,EAAE,CAEzF;AAED,eAAO,MAAM,MAAM,mljBAKlB,CAAA;AAED,wBAAgB,qBAAqB,CAAC,aAAa,EAAE,iBAAiB,EAAE,UAOvE;AAED,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,aAAa,EAAE,iBAAiB,EAAE,UAG7E;AAED,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,MAAM,EAAE,aAAa,EAAE,iBAAiB,EAAE,UAGnF;AAED,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,MAAM,EAAE,aAAa,EAAE,iBAAiB,EAAE,UAGtF"}

package/dist/llm/data/trilogySyntax.generated.d.ts

@@ -1,6 +1,6 @@
-export declare const pytrilogySyntaxVersion = "0.3.370";
-export declare const trilogySyntaxReference = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` \u2014 replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; \u2014 a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a \u2286 b; b authoritative for the key) or `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` declares EXACT domain equivalence. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order \u2014 all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection \u2014 grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation \u2014 it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`, and joined back like any concept (`subset join foo.key = other.key`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL \u2014 what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT \u2014 never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select \u2014 hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain \u2014 a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** \u2014 a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column) \u2014 never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` \u2261 `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages \u2014 give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship \u2014 the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology \u2014 use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE \u2014 its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` \u2192 leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows \u2014 use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain \u2014 on a select item (`~customer.id`) or in `merge a into ~b` (a \u2286 b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary \u2014 wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end \u2212 start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.\n\n\nFunctions (the `aggregate` family doubles as the Aggregate Functions list):\naggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)\nstring: concat|concat_strict|concat_ws(<arg1>, <arg2>, ..., <argN>) ; split|strpos|contains|trim|regexp_contains|format_time|parse_time(<arg1>, <arg2>) ; len|lower|upper|ltrim|rtrim|hex(<arg1>) ; substring|replace|regexp_extract|regexp_replace(<arg1>, <arg2>, <arg3>)\ndate/time: date|datetime|timestamp|second|minute|hour|day|day_of_week|week|month|quarter|year|month_name|day_name|unix_to_timestamp(<arg1>) ; date_spine(<arg1>, <arg2>) ; current_date|current_datetime()\n  date_part(<arg1>, <arg2>) e.g. date_part('2020-01-01'::date, year)\n  date_truncate(<arg1>, <arg2>) e.g. date_trunc('2020-01-01'::date, month)\n  date_add(<arg1>, <arg2>, <arg3>) e.g. date_add('2020-01-01'::date, month, 1)\n  date_sub(<arg1>, <arg2>, <arg3>) e.g. date_sub('2020-01-01'::date, day, 1)\n  date_diff(<arg1>, <arg2>, <arg3>) e.g. date_diff('2020-01-01'::date, '2020-01-02'::date, day)\n  current_timestamp() e.g. now()\narray/map/struct: unnest|array_distinct|array_sum|map_keys|map_values(<arg1>) ; struct|array(<arg1>, <arg2>, ..., <argN>) ; array_sort|array_to_string(<arg1>, <arg2>) ; array_transform|array_filter|generate_array(<arg1>, <arg2>, <arg3>)\nmath: mod|round|log|power|hash(<arg1>, <arg2>) ; abs|sqrt|random|floor|ceil(<arg1>)\nconditional/cast: case|simple_case|coalesce|greatest|least(<arg1>, <arg2>, ..., <argN>) ; cast|nullif(<arg1>, <arg2>) ; isnull(<arg1>)\ngeo: geo_from_text|geo_x|geo_y|geo_centroid(<arg1>) ; geo_point|geo_distance(<arg1>, <arg2>) ; geo_transform(<arg1>, <arg2>, <arg3>)\nother: is_not_distinct(<arg1>, <arg2>)\n\nSome types carry trait metadata that adds meaning: 'latitude', 'longitude' and 'currency' are all 'float' with extra semantics.\n\nAdditional syntax examples:\nThese less-common patterns have complete, copy-pasteable examples. Do NOT guess\nthe syntax \u2014 print the full example on demand with\n`trilogy agent-info syntax example <name>`:\n- `python-datasource` - run a local Python script as a datasource: wrap a function in `trilogy.io.run`, which writes the Arrow IPC stream to stdout for you from a table, dataframe, or list of dicts; declare concepts, map script columns in `datasource (...)`, use `grain (...) file `path.py`;`, then reference locally declared concepts WITHOUT the datasource name as a prefix\n- `query-structure` - the exact clause order of a query, and how to define a rowset\n- `chart` - drawing a result rather than tabulating it: `chart layer <type> ( <role> <- <expr> as <name>, ... )` is a STATEMENT that wraps its own query - the bindings ARE the select; layers, roles, scales, reference lines and `from <select>`\n- `filtered-aggregate` - aggregating only the rows matching a condition, and counting rows correctly\n- `row-level-where-vs-having` - choosing between WHERE and HAVING for a threshold condition\n- `nested-aggregate-group-average` - the \"above 1.2x the group norm\" ask: inner `sum(x) by entity, group`, outer `avg(inner) by group`; use WHERE for the comparison when a separate output-only filter must not narrow the benchmark\n- `pivot-columns` - turn distinct values of a dimension into columns with a reusable `def` macro + filtered aggregate\n- `case-when` - `case when <cond> then <val> ... else <val> end` to relabel codes or bucket a measure into a derived value, then group by it\n- `window-period-over-period` - `lead`/`lag` over (order by ...) to compare a row to an offset period; build a scoped set with `<col> ? <cond>` and filter with `in`\n- `union-stack-channels` - stacking several sources/channels together as rows\n- `except-intersect-setops` - set difference or co-occurrence across sources\n- `scoped-join` - blending a second fact model or rowset into one query\n- `existence-anti-join` - keeping or excluding rows by whether a key matches another model\n- `set-intersect-difference` - count tuples present in / absent from other groups via per-group presence flags - the multi-column INTERSECT/EXCEPT idiom\n- `dedup-before-aggregate` - a `rowset <- select <cols>` collapses identical tuples to one row (SELECT-grain group by) before a downstream sum - the `distinct`/UNION substitute\n- `rollup` - subtotals and a grand total in one pass, and labeling the levels\n- `rank-over-rollup` - rank rollup subtotals/leaves with a SINGLE `rank(a,b) over (partition by level, parent ...)` - not separate ranks per level\n- `staged-membership` - filtering a fine-grain query by a set computed at a coarser grain\n- `correlated-exists-via-grouped-counts` - translate `EXISTS other` / `NOT EXISTS other matching` over the same model into two `count(...) by <grain>` compared in `where` (`> 1` = another exists, `= 1` = no other matches) - pin the correlation grain with `by`";
-export declare const rulesInput = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` \u2014 replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; \u2014 a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a \u2286 b; b authoritative for the key) or `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` declares EXACT domain equivalence. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order \u2014 all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection \u2014 grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation \u2014 it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`, and joined back like any concept (`subset join foo.key = other.key`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL \u2014 what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT \u2014 never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select \u2014 hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain \u2014 a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** \u2014 a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column) \u2014 never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` \u2261 `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages \u2014 give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship \u2014 the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology \u2014 use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE \u2014 its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` \u2192 leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows \u2014 use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain \u2014 on a select item (`~customer.id`) or in `merge a into ~b` (a \u2286 b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary \u2014 wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end \u2212 start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.";
+export declare const pytrilogySyntaxVersion = "0.3.391";
+export declare const trilogySyntaxReference = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` \u2014 replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; \u2014 a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union\\|equal join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a \u2286 b; b authoritative for the key), `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept) or `equal join a = b` (ONE domain: a is an alias of b, every value on either side is on the other; the join renders inner).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` is the persistent equivalent of `equal join a = b`. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order \u2014 all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection \u2014 grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union|equal join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation \u2014 it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`. Pairing an output with a concept outside the rowset REQUIRES a declared join on the rowset's key (`subset join foo.key = key`); without one the query is disconnected. Alternatively project the concept inside the rowset and filter or select it through the handle (`where foo.cat = 'a'`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL \u2014 what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT \u2014 never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select \u2014 hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain \u2014 a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** \u2014 a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column; a scalar must also be one row: an aggregate with no `by`, or `limit 1`) \u2014 never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` \u2261 `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages \u2014 give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship \u2014 the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology \u2014 use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE \u2014 its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` \u2192 leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows \u2014 use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain \u2014 on a select item (`~customer.id`) or in `merge a into ~b` (a \u2286 b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary \u2014 wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end \u2212 start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.\n\n\nFunctions (the `aggregate` family doubles as the Aggregate Functions list):\naggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)\nstring: concat|concat_strict|concat_ws(<arg1>, <arg2>, ..., <argN>) ; split|strpos|contains|trim|regexp_contains|format_time|parse_time(<arg1>, <arg2>) ; len|lower|upper|ltrim|rtrim|hex(<arg1>) ; substring|replace|regexp_extract|regexp_replace(<arg1>, <arg2>, <arg3>)\ndate/time: date|datetime|timestamp|second|minute|hour|day|day_of_week|week|month|quarter|year|month_name|day_name|unix_to_timestamp(<arg1>) ; date_spine(<arg1>, <arg2>) ; current_date|current_datetime()\n  date_part(<arg1>, <arg2>) e.g. date_part('2020-01-01'::date, year)\n  date_truncate(<arg1>, <arg2>) e.g. date_trunc('2020-01-01'::date, month)\n  date_add(<arg1>, <arg2>, <arg3>) e.g. date_add('2020-01-01'::date, month, 1)\n  date_sub(<arg1>, <arg2>, <arg3>) e.g. date_sub('2020-01-01'::date, day, 1)\n  date_diff(<arg1>, <arg2>, <arg3>) e.g. date_diff('2020-01-01'::date, '2020-01-02'::date, day)\n  current_timestamp() e.g. now()\narray/map/struct: unnest|array_distinct|array_sum|map_keys|map_values(<arg1>) ; struct|array(<arg1>, <arg2>, ..., <argN>) ; array_sort|array_to_string(<arg1>, <arg2>) ; array_transform|array_filter|generate_array(<arg1>, <arg2>, <arg3>)\nmath: mod|round|log|power|hash(<arg1>, <arg2>) ; abs|sqrt|random|floor|ceil(<arg1>)\nconditional/cast: case|simple_case|coalesce|greatest|least(<arg1>, <arg2>, ..., <argN>) ; cast|nullif(<arg1>, <arg2>) ; isnull(<arg1>)\ngeo: geo_from_text|geo_x|geo_y|geo_centroid(<arg1>) ; geo_point|geo_distance(<arg1>, <arg2>) ; geo_transform(<arg1>, <arg2>, <arg3>)\nother: is_not_distinct(<arg1>, <arg2>) ; prior(<arg1>) ; grain_pin(<arg1>, <arg2>, ..., <argN>)\n\nSome types carry trait metadata that adds meaning: 'latitude', 'longitude' and 'currency' are all 'float' with extra semantics.\n\nAdditional syntax examples:\nThese less-common patterns have complete, copy-pasteable examples. Do NOT guess\nthe syntax \u2014 print the full example on demand with\n`trilogy agent-info syntax example <name>`:\n- `python-datasource` - run a local Python script as a datasource: wrap a function in `trilogy.io.run`, which writes the Arrow IPC stream to stdout for you from a table, dataframe, or list of dicts; declare concepts, map script columns in `datasource (...)`, use `grain (...) file `path.py`;`, then reference locally declared concepts WITHOUT the datasource name as a prefix\n- `query-structure` - the exact clause order of a query, and how to define a rowset\n- `chart` - drawing a result rather than tabulating it: `chart layer <type> ( <role> <- <expr> as <name>, ... )` is a STATEMENT that wraps its own query - the bindings ARE the select; layers, roles, scales, reference lines and `from <select>`\n- `filtered-aggregate` - aggregating only the rows matching a condition, and counting rows correctly\n- `row-level-where-vs-having` - choosing between WHERE and HAVING for a threshold condition\n- `nested-aggregate-group-average` - the \"above 1.2x the group norm\" ask: inner `sum(x) by entity, group`, outer `avg(inner) by group`; use WHERE for the comparison when a separate output-only filter must not narrow the benchmark\n- `pivot-columns` - turn distinct values of a dimension into columns with a reusable `def` macro + filtered aggregate\n- `case-when` - `case when <cond> then <val> ... else <val> end` to relabel codes or bucket a measure into a derived value, then group by it\n- `window-period-over-period` - `lead`/`lag` over (order by ...) to compare a row to an offset period; build a scoped set with `<col> ? <cond>` and filter with `in`\n- `union-stack-channels` - stacking several sources/channels together as rows\n- `except-intersect-setops` - set difference or co-occurrence across sources\n- `scoped-join` - blending a second fact model or rowset into one query\n- `existence-anti-join` - keeping or excluding rows by whether a key matches another model\n- `set-intersect-difference` - count tuples present in / absent from other groups via per-group presence flags - the multi-column INTERSECT/EXCEPT idiom\n- `dedup-before-aggregate` - a `rowset <- select <cols>` collapses identical tuples to one row (SELECT-grain group by) before a downstream sum - the `distinct`/UNION substitute\n- `rollup` - subtotals and a grand total in one pass, and labeling the levels\n- `rank-over-rollup` - rank rollup subtotals/leaves with a SINGLE `rank(a,b) over (partition by level, parent ...)` - not separate ranks per level\n- `staged-membership` - filtering a fine-grain query by a set computed at a coarser grain\n- `correlated-exists-via-grouped-counts` - translate `EXISTS other` / `NOT EXISTS other matching` over the same model into two `count(...) by <grain>` compared in `where` (`> 1` = another exists, `= 1` = no other matches) - pin the correlation grain with `by`";
+export declare const rulesInput = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` \u2014 replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; \u2014 a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union\\|equal join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a \u2286 b; b authoritative for the key), `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept) or `equal join a = b` (ONE domain: a is an alias of b, every value on either side is on the other; the join renders inner).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` is the persistent equivalent of `equal join a = b`. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order \u2014 all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection \u2014 grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union|equal join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation \u2014 it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`. Pairing an output with a concept outside the rowset REQUIRES a declared join on the rowset's key (`subset join foo.key = key`); without one the query is disconnected. Alternatively project the concept inside the rowset and filter or select it through the handle (`where foo.cat = 'a'`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL \u2014 what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT \u2014 never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select \u2014 hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain \u2014 a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** \u2014 a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column; a scalar must also be one row: an aggregate with no `by`, or `limit 1`) \u2014 never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` \u2261 `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages \u2014 give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship \u2014 the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology \u2014 use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE \u2014 its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` \u2192 leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows \u2014 use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain \u2014 on a select item (`~customer.id`) or in `merge a into ~b` (a \u2286 b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary \u2014 wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end \u2212 start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.";
 export declare const aggFunctions: string[];
 export declare const functions: string[];
 export declare const datatypes: string[];

package/dist/llm/data/trilogySyntax.generated.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"trilogySyntax.generated.d.ts","sourceRoot":"","sources":["../../../llm/data/trilogySyntax.generated.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,sBAAsB,YAAY,CAAA;AAC/C,eAAO,MAAM,sBAAsB,mjhBAA27gB,CAAA;AAC99gB,eAAO,MAAM,UAAU,+vXAA4oX,CAAA;AACnqX,eAAO,MAAM,YAAY,UAExB,CAAA;AACD,eAAO,MAAM,SAAS,UAerB,CAAA;AACD,eAAO,MAAM,SAAS,UAsBrB,CAAA;AACD,eAAO,MAAM,cAAc;;;;;GAmH1B,CAAA"}
+{"version":3,"file":"trilogySyntax.generated.d.ts","sourceRoot":"","sources":["../../../llm/data/trilogySyntax.generated.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,sBAAsB,YAAY,CAAA;AAC/C,eAAO,MAAM,sBAAsB,qiiBAA66hB,CAAA;AACh9hB,eAAO,MAAM,UAAU,wrYAAqkY,CAAA;AAC5lY,eAAO,MAAM,YAAY,UAExB,CAAA;AACD,eAAO,MAAM,SAAS,UAerB,CAAA;AACD,eAAO,MAAM,SAAS,UAsBrB,CAAA;AACD,eAAO,MAAM,cAAc;;;;;GAmH1B,CAAA"}

package/dist/{llm-ClU-2rgk.js → llm-Wf4Etiu9.js}

@@ -15448,7 +15448,7 @@ var jg = class e extends nt {
 		let t = [...e].sort(Ag);
 		return t.find((e) => kg(e)?.tier === "flash") || t[0] || "";
 	}
-}, Mg = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` — replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; — a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a ⊆ b; b authoritative for the key) or `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` declares EXACT domain equivalence. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order — all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection — grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation — it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`, and joined back like any concept (`subset join foo.key = other.key`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL — what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT — never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select — hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain — a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** — a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column) — never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` ≡ `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages — give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship — the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology — use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE — its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` → leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows — use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain — on a select item (`~customer.id`) or in `merge a into ~b` (a ⊆ b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary — wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end − start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.\n\n\nFunctions (the `aggregate` family doubles as the Aggregate Functions list):\naggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)\nstring: concat|concat_strict|concat_ws(<arg1>, <arg2>, ..., <argN>) ; split|strpos|contains|trim|regexp_contains|format_time|parse_time(<arg1>, <arg2>) ; len|lower|upper|ltrim|rtrim|hex(<arg1>) ; substring|replace|regexp_extract|regexp_replace(<arg1>, <arg2>, <arg3>)\ndate/time: date|datetime|timestamp|second|minute|hour|day|day_of_week|week|month|quarter|year|month_name|day_name|unix_to_timestamp(<arg1>) ; date_spine(<arg1>, <arg2>) ; current_date|current_datetime()\n  date_part(<arg1>, <arg2>) e.g. date_part('2020-01-01'::date, year)\n  date_truncate(<arg1>, <arg2>) e.g. date_trunc('2020-01-01'::date, month)\n  date_add(<arg1>, <arg2>, <arg3>) e.g. date_add('2020-01-01'::date, month, 1)\n  date_sub(<arg1>, <arg2>, <arg3>) e.g. date_sub('2020-01-01'::date, day, 1)\n  date_diff(<arg1>, <arg2>, <arg3>) e.g. date_diff('2020-01-01'::date, '2020-01-02'::date, day)\n  current_timestamp() e.g. now()\narray/map/struct: unnest|array_distinct|array_sum|map_keys|map_values(<arg1>) ; struct|array(<arg1>, <arg2>, ..., <argN>) ; array_sort|array_to_string(<arg1>, <arg2>) ; array_transform|array_filter|generate_array(<arg1>, <arg2>, <arg3>)\nmath: mod|round|log|power|hash(<arg1>, <arg2>) ; abs|sqrt|random|floor|ceil(<arg1>)\nconditional/cast: case|simple_case|coalesce|greatest|least(<arg1>, <arg2>, ..., <argN>) ; cast|nullif(<arg1>, <arg2>) ; isnull(<arg1>)\ngeo: geo_from_text|geo_x|geo_y|geo_centroid(<arg1>) ; geo_point|geo_distance(<arg1>, <arg2>) ; geo_transform(<arg1>, <arg2>, <arg3>)\nother: is_not_distinct(<arg1>, <arg2>)\n\nSome types carry trait metadata that adds meaning: 'latitude', 'longitude' and 'currency' are all 'float' with extra semantics.\n\nAdditional syntax examples:\nThese less-common patterns have complete, copy-pasteable examples. Do NOT guess\nthe syntax — print the full example on demand with\n`trilogy agent-info syntax example <name>`:\n- `python-datasource` - run a local Python script as a datasource: wrap a function in `trilogy.io.run`, which writes the Arrow IPC stream to stdout for you from a table, dataframe, or list of dicts; declare concepts, map script columns in `datasource (...)`, use `grain (...) file `path.py`;`, then reference locally declared concepts WITHOUT the datasource name as a prefix\n- `query-structure` - the exact clause order of a query, and how to define a rowset\n- `chart` - drawing a result rather than tabulating it: `chart layer <type> ( <role> <- <expr> as <name>, ... )` is a STATEMENT that wraps its own query - the bindings ARE the select; layers, roles, scales, reference lines and `from <select>`\n- `filtered-aggregate` - aggregating only the rows matching a condition, and counting rows correctly\n- `row-level-where-vs-having` - choosing between WHERE and HAVING for a threshold condition\n- `nested-aggregate-group-average` - the \"above 1.2x the group norm\" ask: inner `sum(x) by entity, group`, outer `avg(inner) by group`; use WHERE for the comparison when a separate output-only filter must not narrow the benchmark\n- `pivot-columns` - turn distinct values of a dimension into columns with a reusable `def` macro + filtered aggregate\n- `case-when` - `case when <cond> then <val> ... else <val> end` to relabel codes or bucket a measure into a derived value, then group by it\n- `window-period-over-period` - `lead`/`lag` over (order by ...) to compare a row to an offset period; build a scoped set with `<col> ? <cond>` and filter with `in`\n- `union-stack-channels` - stacking several sources/channels together as rows\n- `except-intersect-setops` - set difference or co-occurrence across sources\n- `scoped-join` - blending a second fact model or rowset into one query\n- `existence-anti-join` - keeping or excluding rows by whether a key matches another model\n- `set-intersect-difference` - count tuples present in / absent from other groups via per-group presence flags - the multi-column INTERSECT/EXCEPT idiom\n- `dedup-before-aggregate` - a `rowset <- select <cols>` collapses identical tuples to one row (SELECT-grain group by) before a downstream sum - the `distinct`/UNION substitute\n- `rollup` - subtotals and a grand total in one pass, and labeling the levels\n- `rank-over-rollup` - rank rollup subtotals/leaves with a SINGLE `rank(a,b) over (partition by level, parent ...)` - not separate ranks per level\n- `staged-membership` - filtering a fine-grain query by a set computed at a coarser grain\n- `correlated-exists-via-grouped-counts` - translate `EXISTS other` / `NOT EXISTS other matching` over the same model into two `count(...) by <grain>` compared in `where` (`> 1` = another exists, `= 1` = no other matches) - pin the correlation grain with `by`", Ng = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` — replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; — a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a ⊆ b; b authoritative for the key) or `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` declares EXACT domain equivalence. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order — all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection — grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation — it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`, and joined back like any concept (`subset join foo.key = other.key`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL — what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT — never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select — hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain — a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** — a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column) — never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` ≡ `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages — give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship — the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology — use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE — its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` → leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows — use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain — on a select item (`~customer.id`) or in `merge a into ~b` (a ⊆ b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary — wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end − start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.", Pg = ["aggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)"], Fg = [
+}, Mg = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` — replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; — a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union\\|equal join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a ⊆ b; b authoritative for the key), `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept) or `equal join a = b` (ONE domain: a is an alias of b, every value on either side is on the other; the join renders inner).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` is the persistent equivalent of `equal join a = b`. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order — all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection — grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union|equal join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation — it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`. Pairing an output with a concept outside the rowset REQUIRES a declared join on the rowset's key (`subset join foo.key = key`); without one the query is disconnected. Alternatively project the concept inside the rowset and filter or select it through the handle (`where foo.cat = 'a'`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL — what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT — never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select — hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain — a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** — a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column; a scalar must also be one row: an aggregate with no `by`, or `limit 1`) — never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` ≡ `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages — give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship — the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology — use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE — its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` → leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows — use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain — on a select item (`~customer.id`) or in `merge a into ~b` (a ⊆ b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary — wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end − start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.\n\n\nFunctions (the `aggregate` family doubles as the Aggregate Functions list):\naggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)\nstring: concat|concat_strict|concat_ws(<arg1>, <arg2>, ..., <argN>) ; split|strpos|contains|trim|regexp_contains|format_time|parse_time(<arg1>, <arg2>) ; len|lower|upper|ltrim|rtrim|hex(<arg1>) ; substring|replace|regexp_extract|regexp_replace(<arg1>, <arg2>, <arg3>)\ndate/time: date|datetime|timestamp|second|minute|hour|day|day_of_week|week|month|quarter|year|month_name|day_name|unix_to_timestamp(<arg1>) ; date_spine(<arg1>, <arg2>) ; current_date|current_datetime()\n  date_part(<arg1>, <arg2>) e.g. date_part('2020-01-01'::date, year)\n  date_truncate(<arg1>, <arg2>) e.g. date_trunc('2020-01-01'::date, month)\n  date_add(<arg1>, <arg2>, <arg3>) e.g. date_add('2020-01-01'::date, month, 1)\n  date_sub(<arg1>, <arg2>, <arg3>) e.g. date_sub('2020-01-01'::date, day, 1)\n  date_diff(<arg1>, <arg2>, <arg3>) e.g. date_diff('2020-01-01'::date, '2020-01-02'::date, day)\n  current_timestamp() e.g. now()\narray/map/struct: unnest|array_distinct|array_sum|map_keys|map_values(<arg1>) ; struct|array(<arg1>, <arg2>, ..., <argN>) ; array_sort|array_to_string(<arg1>, <arg2>) ; array_transform|array_filter|generate_array(<arg1>, <arg2>, <arg3>)\nmath: mod|round|log|power|hash(<arg1>, <arg2>) ; abs|sqrt|random|floor|ceil(<arg1>)\nconditional/cast: case|simple_case|coalesce|greatest|least(<arg1>, <arg2>, ..., <argN>) ; cast|nullif(<arg1>, <arg2>) ; isnull(<arg1>)\ngeo: geo_from_text|geo_x|geo_y|geo_centroid(<arg1>) ; geo_point|geo_distance(<arg1>, <arg2>) ; geo_transform(<arg1>, <arg2>, <arg3>)\nother: is_not_distinct(<arg1>, <arg2>) ; prior(<arg1>) ; grain_pin(<arg1>, <arg2>, ..., <argN>)\n\nSome types carry trait metadata that adds meaning: 'latitude', 'longitude' and 'currency' are all 'float' with extra semantics.\n\nAdditional syntax examples:\nThese less-common patterns have complete, copy-pasteable examples. Do NOT guess\nthe syntax — print the full example on demand with\n`trilogy agent-info syntax example <name>`:\n- `python-datasource` - run a local Python script as a datasource: wrap a function in `trilogy.io.run`, which writes the Arrow IPC stream to stdout for you from a table, dataframe, or list of dicts; declare concepts, map script columns in `datasource (...)`, use `grain (...) file `path.py`;`, then reference locally declared concepts WITHOUT the datasource name as a prefix\n- `query-structure` - the exact clause order of a query, and how to define a rowset\n- `chart` - drawing a result rather than tabulating it: `chart layer <type> ( <role> <- <expr> as <name>, ... )` is a STATEMENT that wraps its own query - the bindings ARE the select; layers, roles, scales, reference lines and `from <select>`\n- `filtered-aggregate` - aggregating only the rows matching a condition, and counting rows correctly\n- `row-level-where-vs-having` - choosing between WHERE and HAVING for a threshold condition\n- `nested-aggregate-group-average` - the \"above 1.2x the group norm\" ask: inner `sum(x) by entity, group`, outer `avg(inner) by group`; use WHERE for the comparison when a separate output-only filter must not narrow the benchmark\n- `pivot-columns` - turn distinct values of a dimension into columns with a reusable `def` macro + filtered aggregate\n- `case-when` - `case when <cond> then <val> ... else <val> end` to relabel codes or bucket a measure into a derived value, then group by it\n- `window-period-over-period` - `lead`/`lag` over (order by ...) to compare a row to an offset period; build a scoped set with `<col> ? <cond>` and filter with `in`\n- `union-stack-channels` - stacking several sources/channels together as rows\n- `except-intersect-setops` - set difference or co-occurrence across sources\n- `scoped-join` - blending a second fact model or rowset into one query\n- `existence-anti-join` - keeping or excluding rows by whether a key matches another model\n- `set-intersect-difference` - count tuples present in / absent from other groups via per-group presence flags - the multi-column INTERSECT/EXCEPT idiom\n- `dedup-before-aggregate` - a `rowset <- select <cols>` collapses identical tuples to one row (SELECT-grain group by) before a downstream sum - the `distinct`/UNION substitute\n- `rollup` - subtotals and a grand total in one pass, and labeling the levels\n- `rank-over-rollup` - rank rollup subtotals/leaves with a SINGLE `rank(a,b) over (partition by level, parent ...)` - not separate ranks per level\n- `staged-membership` - filtering a fine-grain query by a set computed at a coarser grain\n- `correlated-exists-via-grouped-counts` - translate `EXISTS other` / `NOT EXISTS other matching` over the same model into two `count(...) by <grain>` compared in `where` (`> 1` = another exists, `= 1` = no other matches) - pin the correlation grain with `by`", Ng = "# Trilogy Syntax\n\nTrilogy statements define a semantic model or a query; collections of statements\nare written as `.preql` files. A statement returns data (SELECT) or defines\nconcepts (key | property | auto | metric), datasources, or parameters.\n\nimport <model> as <alias>; makes a model's fields available. <model> is a DOT-SEPARATED MODULE PATH relative to the project root (where trilogy.toml lives), never a file path: the file `raw/store_sales.preql` is imported as `import raw.store_sales as ss;` — replace each `/` with `.`, drop the `.preql`, keep the folder segment. When an imported model imports others (a fact with imported dimensions), those are dot paths: `enroll.student.name`. Typical usage imports ONE fact and reaches dimensions through it; nested dimensions carry ALL values (order.customers includes customers without orders).\n\nkey | property | auto | metric; define new concepts. `auto x <- ...` is a DEFINITION, not a precomputed value: each reference expands and re-evaluates in the consuming query's scope.\n\nparameter NAME TYPE [default <literal>]; — a runtime value supplied via `--param NAME=VALUE` (repeatable); append `?` to the type to allow null.\n\n## Combining models\n\n| Goal | Use |\n|---|---|\n| Typical query | no merge, no join: access all fields through dot-paths |\n| Blend two models on shared keys inside one query | scoped `subset\\|union\\|equal join` |\n| Make a connection universal to a whole file | `merge` |\n| Stack subsets/channels as rows | `union(...)` |\n| Rows in A but never in B (set difference) | `except(...)` |\n| Rows present in every source (set intersection) | `intersect(...)` |\n\n### Query-scoped join\n\nA typical fact already has its dimensions merged in, so no join is needed. Blending fact models or rowset outputs takes a scoped join, placed right after the select list, which DECLARES how the key domains relate: `subset join a = b` (a's values are contained in b's, a ⊆ b; b authoritative for the key), `union join a = b` (neither contains the other; the key is the coalesce of both sides and unmatched rows from BOTH sides are kept) or `equal join a = b` (ONE domain: a is an alias of b, every value on either side is on the other; the join renders inner).\n\nTwo rules decide whether the numbers are right, so apply them before reaching for the example:\n\n- Joins NEVER drop a row and NULL keys match null-safely; there is NO inner join. An intersection is an explicit `where <optional-side attr> is not null`.\n- JOIN ON THE FULL GRAIN: `explore` prints each fact's grain as @<k1, k2>; write one clause PER key. Matching only one key of a multi-key grain causes duplication.\n\nChaining, expression keys, union/subset mixing, multi-key and self-pair shapes: trilogy agent-info syntax example scoped-join.\n\nmerge <a> into ~<b>; is the persistent (whole-file) equivalent of `subset join a = b`; plain `merge a into b;` is the persistent equivalent of `equal join a = b`. Standalone statements; prefer a scoped join unless the connection is universal.\n\n### union / except / intersect (row set operations)\n\n`union(...)` row-stacks self-contained select arms positionally (SQL UNION ALL) into one named rowset (`with combined as union((armA), (armB), ...) -> (out1, out2, ...)`); `except(...)` / `intersect(...)` share that arm shape but are SQL SET operators (DISTINCT rows, null-safe whole-row compare, and except subtracts later arms from the FIRST, so arm order matters). Prefer except over a multi-column `not in` for \"in A but never in B\". Each arm dedups to its own output grain BEFORE stacking, so aggregate INSIDE each arm or the total silently undersums. Examples: trilogy agent-info syntax example union-stack-channels / except-intersect-setops.\n\n## Query structure\n\nTop-level statements first (each `;`-terminated), then a query with clauses in this EXACT order — all optional except select:\n\n```\n<top-level statements>;            # import / key / auto / rowset / merge / def / parameter\nwith <name> as                     # optional: name the select as a reusable rowset (CTE)\nwhere   <row condition>            # 1. filters INPUT rows, BEFORE aggregation\nselect  <col>, <agg> as name,      # 2. projection — grouping is AUTOMATIC by the\n                                   #    non-aggregated columns; never write GROUP BY\n  subset|union|equal join a = b (= c)?   # 3. blend models; one clause per key, right after the select list\nby rollup|cube|grouping sets (...) # 4. optional multi-level grouping for the whole select\nhaving  <result condition>         # 5. filters aggregated/joined RESULTS\norder by <col> asc|desc            # 6. sort\nlimit   <n>;                       # 7. cap rows\n```\n\nA rowset (`with <name> as where ... select ...;`) is a standalone statement, evaluated in isolation — it does NOT respond to the consuming query's context. All its outputs are namespaced: output `abc.def` of rowset `foo` is referenced as `foo.abc.def`. Pairing an output with a concept outside the rowset REQUIRES a declared join on the rowset's key (`subset join foo.key = key`); without one the query is disconnected. Alternatively project the concept inside the rowset and filter or select it through the handle (`where foo.cat = 'a'`). Alias every reused expression with `as`.\n\nFull annotated example: trilogy agent-info syntax example query-structure.\n\n### Not SQL — what to never write\n\n- **No FROM, GROUP BY, DISTINCT, SELECT \\*, or SQL-style set operators between selects.** Stack rows with `union(...)`; set difference/intersection with `except(...)`/`intersect(...)`; blend fact models with a scoped join.\n- **Grouping is automatic** by the non-aggregated fields in the SELECT — never write GROUP BY. Aggregates inherit the select's grain wherever they appear (where/select/having). Override per-aggregate with `agg(x) by <dims>`; `by *` aggregates across all data (single row). To weight a related dimension property once per fact row, project it to the fact grain first: `auto row_birth_year <- group(enroll.student.birth_year) by enroll.student_id, enroll.course_id;` then aggregate that.\n- **Output rows are deduplicated to the select grain.** To preserve legitimate duplicates (one row per matching fact), include the fact's grain keys in the select — hidden with the PREFIX `--` if they shouldn't appear in the output.\n- **Never write `distinct`.** `count(<key>)` is already distinct because keys are unique; `count_distinct(<property>)` counts distinct values of a non-key property.\n- **Count a COMPOSITE grain with `grain(...)`.** `count(grain(order_id, item.id))` counts order+item combinations; `grain(...)` is NEVER NULL. Never count one column of a multi-key grain — a coarser key counts its own distinct values and undercounts. `count_distinct(a, b)` is accepted as sugar for `count(grain(a, b))` (distinct combinations, NULL members included).\n- **An aggregate over a group with no qualifying rows is NULL, never 0.** `sum(x ? cond) by k` yields NULL for groups where nothing matches, so `= 0` matches nothing. Test emptiness with `count(key ? cond) by k = 0` or wrap: `coalesce(sum(x ? cond), 0)`.\n- **`count(1)` is invalid** — a constant does not identify rows. Use the row key, `count(grain(...))`, or `sum(1) by <grain>`.\n- **One-column expression subqueries only.** `(select ...)` is allowed where a scalar or membership set is expected (exactly one projected column; a scalar must also be one row: an aggregate with no `by`, or `limit 1`) — never as a SQL-style FROM-subquery. For related-entity filters prefer the dot-path: `where enrollments.student.state = 'TN'`, not a subselect.\n- **`--` is a HIDDEN-field prefix, not a comment** (it still changes query structure). Comments use `#` only.\n- **Aliases**: always use the full dot-path (`enroll.student.id`); alias every new expression with `as`; a select alias is usable in `having`/`order by` but NOT inside other select expressions or `where`, and must not rename a field back to an existing concept name.\n\n## Filtering\n\n`where` filters input rows before aggregation; `having` filters results after aggregation/windows. The inline filter `x ? cond` filters the immediate prior expression: `sum(credits ? enrolled = true)`.\n\nAggregates/windows in WHERE do not filter each other's inputs. Use inline filters, or a staged `then where` chain: each `then where` stage's aggregates compute over only the rows passing earlier stages (`where x = 5 then where sum(y) > 10` ≡ `where x = 5 and sum(y ? x = 5) > 10`). One rule: the SAME cross-row expression cannot gate two different stages — give one an inline filter.\n\nWindows need all rows in their range: filter AFTER the window (in `having`) rather than in `where` when the window must see the full set.\n\n## Membership (semijoins)\n\n`in` / `not in` cross models without any declared relationship — the semijoin IS a scoped intersection. The right side is a concept or expression (never a subselect); build computed sets with an inline filter: `auto big_zip <- student.zip ? (count(student.id ? student.honors = true) by student.zip) > 10; ... where school.zip in big_zip`.\n\nMembership is IDENTITY matching, not SQL three-valued logic: NULL matches NULL; the result is always TRUE or FALSE (safe as a projected flag); `not in` is the exact complement (a NULL in the set never empties the result). For strict SQL behavior, add `x is not null`.\n\nA fact model contains ALL dimensional members, so `where students.id not in enroll.student_id` is typically a tautology — use `where enroll.id is null` instead. Tuple membership `(a, b) in (m.a, m.b)` requires both right-side concepts to resolve to ONE source (model or rowset); to test against a cross-model pair, stage it through a rowset first. Examples: existence-anti-join, staged-membership.\n\n## Aggregation\n\nAn aggregate without an explicit `by` is RESPONSIVE — its grain is set where it is consumed, not where it is declared (`auto avg_credits <- avg(enroll.credits);` becomes per-department when selected with `enroll.department`). Pin a grain inline: `sum(metric) by dim1, dim2 as name`; `by` accepts bare identifiers or parenthesized expressions: `avg(price) by (substring(phone, 1, 2))`.\n\nMulti-level grouping (`by rollup|cube|grouping sets (...)`) is a clause of the WHOLE select, placed before `having`: it computes every no-explicit-grain aggregate at every level in one pass (`by rollup (d1, d2)` → leaf, per-d1 subtotal, grand total). `grouping(<field>)` = 1 on rolled-up rows — use it to label levels rather than testing for NULL. Full detail: trilogy agent-info syntax example rollup.\n\n## Window functions\n\nDefault to windows for self-referential queries (period-over-period, running totals, rank, share of group). SQL-style syntax: `rank(<key>) over (partition by <g> order by <expr> desc)` (`dense_rank`/`row_number` same shape; multi-key `rank(a, b)`); `sum(x) over (partition by g order by t)` for running totals; `lag(<field>, <offset>) over (...)` / `lead(...)` for offsets (default 1). `partition by` is optional and accepts expressions.\n\n## Data types\n\nScalars: `string`, `bytes`, `bool`, `int`, `bigint`, `float`, `double`, `number`, `numeric(p,s)` (alias `decimal`; use for exact money), `date`, `datetime`, `timestamp`, `geography`, `any`. Composites: `array<T>`, `map<K, V>`, `struct<name: T, ...>`, `enum<T>[v1, v2, ...]`. Cast with `::type` or `cast(x as type)`: `\"2020-01-01\"::date`.\n\n`?` after a type marks it nullable (`property user.middle_name string?;`). `~` prefixing a concept marks the reference as a PARTIAL subset of the concept's domain — on a select item (`~customer.id`) or in `merge a into ~b` (a ⊆ b); it tells the planner this side is not authoritative.\n\n## Expression gotchas\n\n- `x ? cond` binds to a primary — wrap arithmetic: `(a - b) ? cond`, NOT `a - b ? cond`.\n- `a || b` is NULL if either side is NULL; `concat(...)` skips NULL arguments; `concat_ws(sep, ...)` joins with sep, skipping NULLs.\n- Date parts are unquoted: `date_part(d, year)`, never `'year'`; prefer `year(d)`. `date_diff(start, end, unit)` = end − start (a shipping lag is `date_diff(sold_date, ship_date, day)`).\n- Zero-argument functions still take parentheses: `current_date()`.\n- Factor repeated calculations into a `def` macro (invoked `@name(...)`) or reusable `auto` concepts.", Pg = ["aggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)"], Fg = [
 	"aggregate: count|count_distinct|sum|max|min|avg|stddev|variance|array_agg|bool_or|bool_and|any|grouping(<arg1>) ; grouping_id(<arg1>, <arg2>, ..., <argN>)",
 	"string: concat|concat_strict|concat_ws(<arg1>, <arg2>, ..., <argN>) ; split|strpos|contains|trim|regexp_contains|format_time|parse_time(<arg1>, <arg2>) ; len|lower|upper|ltrim|rtrim|hex(<arg1>) ; substring|replace|regexp_extract|regexp_replace(<arg1>, <arg2>, <arg3>)",
 	"date/time: date|datetime|timestamp|second|minute|hour|day|day_of_week|week|month|quarter|year|month_name|day_name|unix_to_timestamp(<arg1>) ; date_spine(<arg1>, <arg2>) ; current_date|current_datetime()",
@@ -15462,7 +15462,7 @@ var jg = class e extends nt {
 	"math: mod|round|log|power|hash(<arg1>, <arg2>) ; abs|sqrt|random|floor|ceil(<arg1>)",
 	"conditional/cast: case|simple_case|coalesce|greatest|least(<arg1>, <arg2>, ..., <argN>) ; cast|nullif(<arg1>, <arg2>) ; isnull(<arg1>)",
 	"geo: geo_from_text|geo_x|geo_y|geo_centroid(<arg1>) ; geo_point|geo_distance(<arg1>, <arg2>) ; geo_transform(<arg1>, <arg2>, <arg3>)",
-	"other: is_not_distinct(<arg1>, <arg2>)"
+	"other: is_not_distinct(<arg1>, <arg2>) ; prior(<arg1>) ; grain_pin(<arg1>, <arg2>, ..., <argN>)"
 ], Ig = [
 	"string",
 	"bytes",
@@ -15550,7 +15550,7 @@ var jg = class e extends nt {
 		name: "query-structure",
 		title: "Full query structure - every clause and where it goes (+ rowsets)",
 		summary: "the exact clause order of a query, and how to define a rowset",
-		body: "# Every Trilogy query has the SAME fixed skeleton. Top-level statements come\n# first (each ends with `;`), then ONE query whose clauses appear in this exact\n# order - each clause is OPTIONAL except `select`:\n#\n#   <top-level statements>;        # import / key / property / auto / metric /\n#                                  #   rowset / merge / def / datasource / parameter\n#   where   <row condition>        # 1. filter INPUT rows (BEFORE aggregation)\n#   select  <col>, <agg> as name,  # 2. the projection - grouping is AUTOMATIC by the\n#                                  #      non-aggregated columns (never write GROUP BY)\n#   <subset|union join a = b (=c)?>* # 3. blend models (zero or more joins; one per concept RIGHT AFTER the\n#                                  #      select list - the SQL-like spot)\n#   having  <result condition>     # 4. filter on an AGGREGATED / joined RESULT\n#   order by <col> asc|desc        # 5. sort\n#   limit   <n>                    # 6. cap rows\nimport enrollments as enroll;\nimport students as students;\n\n# --- TOP-LEVEL DEFINITIONS (reusable; all live ABOVE the query) -------------\n# An `auto`/`metric`/`property`/`key` defines a field. It expands inline wherever\n# referenced (a macro), re-evaluated in the referencing query's scope.\nauto completed_credits <- sum(enroll.credits ? enroll.completed = true);\n\n# A ROWSET is a NAMED select: `rowset <name> <- <a full select>;`. It runs as its\n# own scoped query (its own where/select/having) and exposes its outputs as\n# `<name>.<col>`. ALIAS every column you reuse downstream with `as`.\nrowset dept_totals <- select\n    enroll.department as department,\n    sum(enroll.credits) as total_credits,\n;\n\n# --- THE QUERY - clauses in the fixed order above ---------------------------\nwhere enroll.year >= 2015                   # 1. WHERE: per-row filter, before aggregation\nselect                                      # 2. SELECT: grouped automatically by students.major\n    students.major,\n    count(enroll.id) as enrollments,\n    --completed_credits,                    #    a leading `--` HIDES a column from the output\nsubset join enroll.student_id = students.id  # 3. JOIN: blend students onto enrollments (after the select list)\nhaving completed_credits > 0                # 4. HAVING: condition on an aggregated RESULT\norder by enrollments desc nulls first       # 5. ORDER BY\nlimit 100;                                  # 6. LIMIT\n\n# Reference a rowset's outputs as `<rowset>.<output>` in a later query.\nselect\n    dept_totals.department,\n    dept_totals.total_credits,\norder by dept_totals.total_credits desc nulls first\nlimit 100;\n\n# ---------------------------------------------------------------------------\n# NOTES:\n#  - Joins sit RIGHT AFTER the select list (the SQL-like, preferred spot). They\n#    may also appear before `select` - both parse - but keep them after the\n#    select list for consistency.\n#  - `where` filters INPUT rows (before aggregation). To filter on something only\n#    known AFTER the join / aggregation (e.g. comparing two joined rowsets'\n#    counts), select that value (hide it with `--`) and test it in `having` -\n#    `having` is the post-aggregation/output filter. There is only ONE `where`.\n#  - Output rows dedupe to the select grain. To keep legitimate duplicates (one\n#    row per matching fact), select the fact's grain keys - hide with `--` if\n#    they shouldn't appear in the output.\n#  - No FROM, GROUP BY, DISTINCT, SELECT *, subqueries, or SQL-style set/JOIN\n#    operators. Grouping is automatic; blend with the scoped `join` above.\n"
+		body: "# Every Trilogy query has the SAME fixed skeleton. Top-level statements come\n# first (each ends with `;`), then ONE query whose clauses appear in this exact\n# order - each clause is OPTIONAL except `select`:\n#\n#   <top-level statements>;        # import / key / property / auto / metric /\n#                                  #   rowset / merge / def / datasource / parameter\n#   where   <row condition>        # 1. filter INPUT rows (BEFORE aggregation)\n#   select  <col>, <agg> as name,  # 2. the projection - grouping is AUTOMATIC by the\n#                                  #      non-aggregated columns (never write GROUP BY)\n#   <subset|union|equal join a = b (=c)?>* # 3. blend models (zero or more joins; one per concept RIGHT AFTER the\n#                                  #      select list - the SQL-like spot)\n#   having  <result condition>     # 4. filter on an AGGREGATED / joined RESULT\n#   order by <col> asc|desc        # 5. sort\n#   limit   <n>                    # 6. cap rows\nimport enrollments as enroll;\nimport students as students;\n\n# --- TOP-LEVEL DEFINITIONS (reusable; all live ABOVE the query) -------------\n# An `auto`/`metric`/`property`/`key` defines a field. It expands inline wherever\n# referenced (a macro), re-evaluated in the referencing query's scope.\nauto completed_credits <- sum(enroll.credits ? enroll.completed = true);\n\n# A ROWSET is a NAMED select: `rowset <name> <- <a full select>;`. It runs as its\n# own scoped query (its own where/select/having) and exposes its outputs as\n# `<name>.<col>`. ALIAS every column you reuse downstream with `as`.\nrowset dept_totals <- select\n    enroll.department as department,\n    sum(enroll.credits) as total_credits,\n;\n\n# --- THE QUERY - clauses in the fixed order above ---------------------------\nwhere enroll.year >= 2015                   # 1. WHERE: per-row filter, before aggregation\nselect                                      # 2. SELECT: grouped automatically by students.major\n    students.major,\n    count(enroll.id) as enrollments,\n    --completed_credits,                    #    a leading `--` HIDES a column from the output\nsubset join enroll.student_id = students.id  # 3. JOIN: blend students onto enrollments (after the select list)\nhaving completed_credits > 0                # 4. HAVING: condition on an aggregated RESULT\norder by enrollments desc nulls first       # 5. ORDER BY\nlimit 100;                                  # 6. LIMIT\n\n# Reference a rowset's outputs as `<rowset>.<output>` in a later query.\nselect\n    dept_totals.department,\n    dept_totals.total_credits,\norder by dept_totals.total_credits desc nulls first\nlimit 100;\n\n# ---------------------------------------------------------------------------\n# NOTES:\n#  - Joins sit RIGHT AFTER the select list (the SQL-like, preferred spot). They\n#    may also appear before `select` - both parse - but keep them after the\n#    select list for consistency.\n#  - `where` filters INPUT rows (before aggregation). To filter on something only\n#    known AFTER the join / aggregation (e.g. comparing two joined rowsets'\n#    counts), select that value (hide it with `--`) and test it in `having` -\n#    `having` is the post-aggregation/output filter. There is only ONE `where`.\n#  - Output rows dedupe to the select grain. To keep legitimate duplicates (one\n#    row per matching fact), select the fact's grain keys - hide with `--` if\n#    they shouldn't appear in the output.\n#  - No FROM, GROUP BY, DISTINCT, SELECT *, subqueries, or SQL-style set/JOIN\n#    operators. Grouping is automatic; blend with the scoped `join` above.\n"
 	},
 	{
 		name: "rank-over-rollup",
@@ -15572,9 +15572,9 @@ var jg = class e extends nt {
 	},
 	{
 		name: "scoped-join",
-		title: "Blend two models in one query with a scoped subset/union join",
+		title: "Blend two models in one query with a scoped subset/union/equal join",
 		summary: "blending a second fact model or rowset into one query",
-		body: "# A query-scoped `join` blends a second model into ONE select without editing the\n# model files (the query-local equivalent of `merge`). A join DECLARES how the\n# two keys' value domains relate - it never drops rows (rendering is always\n# row-preserving; restrict rows with explicit `is not null` filters):\n#   subset join a = b  -> declares a's values are contained in b's (a is the\n#                         narrow side, b superset, all values in a in b)\n#   union  join a = b  -> neither domain contains the other; the key coalesces\n#                         and unmatched rows from EITHER side are kept\n# Place the clause(s) RIGHT AFTER the `select` list (the SQL-like spot).\nimport enrollments as enroll;\nimport students as students;\n\n# (1) SINGLE-KEY blend: bring students onto enrollments by the shared student\n# key. Enrollment student ids are a SUBSET of all student ids, so declare that;\n# majors with no enrollments are preserved with a NULL count.\nselect\n    students.major,\n    count(enroll.id) as enrollments,\nsubset join enroll.student_id = students.id\norder by enrollments desc nulls first\nlimit 100;\n\n# The select plus its join clause(s) is ONE statement, so it must end with `;`\n# even when no `order by`/`limit` follows the join - a quick probe that stops\n# right after the join clause still needs the terminator:\nselect\n    students.major,\n    count(enroll.id) as enrollments,\nsubset join enroll.student_id = students.id;\n\n# ---------------------------------------------------------------------------\n# (2) MULTI-KEY blend - JOIN ON THE FULL GRAIN. When facts share a COMPOSITE\n# grain, write ONE join clause per key. `trilogy explore` shows each fact's grain\n# as `@<k1, k2>` (e.g. a sales/returns fact keyed `@<department, course>`);\n# matching only ONE key fans out and DOUBLE-COUNTS. Two rowsets, joined on both\n# of their shared keys:\nrowset completed_by <- where enroll.completed = true\n    select enroll.department as dept, enroll.course as course, count(enroll.id) as completed_cnt;\nrowset incomplete_by <- where enroll.completed = false\n    select enroll.department as dept, enroll.course as course, count(enroll.id) as incomplete_cnt;\nrowset null_completed <- where enroll.completed is null\n    select enroll.department as dept, enroll.course as course, count(enroll.id) as null_completed_cnt;\n\nselect\n    completed_by.dept,\n    completed_by.course,\n    completed_by.completed_cnt,\n    incomplete_by.incomplete_cnt,\n    null_completed.null_completed_cnt,\nunion join completed_by.dept = incomplete_by.dept = null_completed.dept\nunion join completed_by.course = incomplete_by.course = null_completed.course\norder by completed_by.dept asc nulls first\nlimit 100;\n\n# ---------------------------------------------------------------------------\n# (3) SELF-PAIR across two periods/sets (period-over-period) - a useful shape\n# for \"this year vs last year\": build one rowset per period (each aggregated to\n# the SAME keys), filter to restrict to output to one side or the other\n# in having clause.\n\nrowset y2020 <- where enroll.year = 2020 select enroll.department as dept, count(enroll.id) as cnt;\nrowset y2021 <- where enroll.year = 2021 select enroll.department as dept, count(enroll.id) as cnt;\n\nselect\n    y2020.dept,\n    y2020.cnt as cnt_2020,\n    y2021.cnt as cnt_2021,\n    y2021.cnt - y2020.cnt as yoy_diff,\nunion join y2020.dept = y2021.dept\nhaving y2021.cnt is not null\norder by yoy_diff desc nulls first\nlimit 100;\n\n# ---------------------------------------------------------------------------\n# (4) EXPRESSION join keys - a key may be ANY expression (arithmetic, an\n# aggregate, a window), not just a bare field. Here each year is matched to the\n# PRIOR year by joining on an OFFSET key (`prev_year.yr + 1`), so the prior\n# year's count lands on the current row without a self-merge. Only `=` equality\n# is supported (a key below comparison level - wrap a comparison in parens).\nrowset cur_year  <- select enroll.year as yr, count(enroll.id) as cnt;\nrowset prev_year <- select enroll.year as yr, count(enroll.id) as cnt;\n\nselect\n    cur_year.yr,\n    cur_year.cnt as this_year,\n    prev_year.cnt as prior_year,\nunion join cur_year.yr = prev_year.yr + 1\norder by cur_year.yr asc\nlimit 100;\n\n# NOTES:\n#  - JOIN ON THE FULL GRAIN: one `join` clause per key in the two facts' shared\n#    `@<...>` grain. Under-joining (one key of a multi-key grain) can cause\n#    duplication. Chain `= c` to also pull a base key into the join\n#    group so its properties stay reachable (`union join a.k = b.k = base.k`).\n#  - Stacked clauses of the SAME type can share one prefix via `and`:\n#    `union join a.k1 = b.k1 and a.k2 = b.k2` == two stacked `union join` clauses.\n#    `and` joins distinct KEY-EQUALITY groups (not filters); `= c` chains keys into\n#    ONE group - both compose: `union join a.k = b.k = base.k and a.k2 = b.k2`.\n#    A COMMA never separates join groups (it is a select-list separator only):\n#    `union join a.k1 = b.k1, a.k2 = b.k2` is an error - use `and`.\n#  - Terminate the statement with `;` after the LAST clause, whatever that clause\n#    is (join, `having`, `order by`, `limit`).\n#  - A join key may be ANY expression (a computed/offset key, an aggregate, a\n#    window), not only a bare field - see (4). Only `=` equality is supported.\n#  - Joins NEVER drop rows: for an intersection (rows present in BOTH sides),\n#    add explicit conditions - `where <side_a attr> is not null and <side_b\n#    attr> is not null`. A one-sided `is not null` keeps the other side's\n#    exclusive rows. Either spelling reads as \"did this side match?\": the\n#    side's own JOIN KEY (`where a.k is not null`, tested per side even though\n#    the projected key is the coalesced group axis) or any of its measures /\n#    attributes (`where a.cnt is not null`). Same in `having`.\n#  - `left`/`full`/`inner`/`right`/`cross join` are NOT valid query-scoped joins\n#    (they error) - only `subset`/`union`. Map an old `left join a = b` to\n#    `subset join b = a` (swap operands) and `full join a = b` to\n#    `union join a = b`. A union key-group must be ALL union (can't mix with\n#    subset on the SAME key; `union join a = b = c` chains one union group);\n"
+		body: "# A query-scoped `join` blends a second model into ONE select without editing the\n# model files (the query-local equivalent of `merge`). A join DECLARES how the\n# two keys' value domains relate - it never drops rows (rendering is always\n# row-preserving; restrict rows with explicit `is not null` filters):\n#   subset join a = b  -> declares a's values are contained in b's (a is the\n#                         narrow side, b superset, all values in a in b)\n#   union  join a = b  -> neither domain contains the other; the key coalesces\n#                         and unmatched rows from EITHER side are kept\n#   equal  join a = b  -> ONE domain: a is an alias of b, every value on either\n#                         side is on the other (`merge a into b` scoped to this\n#                         query); the join renders inner. Two rowsets ranking\n#                         the same rows two ways pair this way.\n# Place the clause(s) RIGHT AFTER the `select` list (the SQL-like spot).\nimport enrollments as enroll;\nimport students as students;\n\n# (1) SINGLE-KEY blend: bring students onto enrollments by the shared student\n# key. Enrollment student ids are a SUBSET of all student ids, so declare that;\n# majors with no enrollments are preserved with a NULL count.\nselect\n    students.major,\n    count(enroll.id) as enrollments,\nsubset join enroll.student_id = students.id\norder by enrollments desc nulls first\nlimit 100;\n\n# The select plus its join clause(s) is ONE statement, so it must end with `;`\n# even when no `order by`/`limit` follows the join - a quick probe that stops\n# right after the join clause still needs the terminator:\nselect\n    students.major,\n    count(enroll.id) as enrollments,\nsubset join enroll.student_id = students.id;\n\n# ---------------------------------------------------------------------------\n# (2) MULTI-KEY blend - JOIN ON THE FULL GRAIN. When facts share a COMPOSITE\n# grain, write ONE join clause per key. `trilogy explore` shows each fact's grain\n# as `@<k1, k2>` (e.g. a sales/returns fact keyed `@<department, course>`);\n# matching only ONE key fans out and DOUBLE-COUNTS. Two rowsets, joined on both\n# of their shared keys:\nrowset completed_by <- where enroll.completed = true\n    select enroll.department as dept, enroll.course as course, count(enroll.id) as completed_cnt;\nrowset incomplete_by <- where enroll.completed = false\n    select enroll.department as dept, enroll.course as course, count(enroll.id) as incomplete_cnt;\nrowset null_completed <- where enroll.completed is null\n    select enroll.department as dept, enroll.course as course, count(enroll.id) as null_completed_cnt;\n\nselect\n    completed_by.dept,\n    completed_by.course,\n    completed_by.completed_cnt,\n    incomplete_by.incomplete_cnt,\n    null_completed.null_completed_cnt,\nunion join completed_by.dept = incomplete_by.dept = null_completed.dept\nunion join completed_by.course = incomplete_by.course = null_completed.course\norder by completed_by.dept asc nulls first\nlimit 100;\n\n# ---------------------------------------------------------------------------\n# (3) SELF-PAIR across two periods/sets (period-over-period) - a useful shape\n# for \"this year vs last year\": build one rowset per period (each aggregated to\n# the SAME keys), filter to restrict to output to one side or the other\n# in having clause.\n\nrowset y2020 <- where enroll.year = 2020 select enroll.department as dept, count(enroll.id) as cnt;\nrowset y2021 <- where enroll.year = 2021 select enroll.department as dept, count(enroll.id) as cnt;\n\nselect\n    y2020.dept,\n    y2020.cnt as cnt_2020,\n    y2021.cnt as cnt_2021,\n    y2021.cnt - y2020.cnt as yoy_diff,\nunion join y2020.dept = y2021.dept\nhaving y2021.cnt is not null\norder by yoy_diff desc nulls first\nlimit 100;\n\n# ---------------------------------------------------------------------------\n# (4) EXPRESSION join keys - a key may be ANY expression (arithmetic, an\n# aggregate, a window), not just a bare field. Here each year is matched to the\n# PRIOR year by joining on an OFFSET key (`prev_year.yr + 1`), so the prior\n# year's count lands on the current row without a self-merge. Only `=` equality\n# is supported (a key below comparison level - wrap a comparison in parens).\nrowset cur_year  <- select enroll.year as yr, count(enroll.id) as cnt;\nrowset prev_year <- select enroll.year as yr, count(enroll.id) as cnt;\n\nselect\n    cur_year.yr,\n    cur_year.cnt as this_year,\n    prev_year.cnt as prior_year,\nunion join cur_year.yr = prev_year.yr + 1\norder by cur_year.yr asc\nlimit 100;\n\n# NOTES:\n#  - JOIN ON THE FULL GRAIN: one `join` clause per key in the two facts' shared\n#    `@<...>` grain. Under-joining (one key of a multi-key grain) can cause\n#    duplication. Chain `= c` to also pull a base key into the join\n#    group so its properties stay reachable (`union join a.k = b.k = base.k`).\n#  - Stacked clauses of the SAME type can share one prefix via `and`:\n#    `union join a.k1 = b.k1 and a.k2 = b.k2` == two stacked `union join` clauses.\n#    `and` joins distinct KEY-EQUALITY groups (not filters); `= c` chains keys into\n#    ONE group - both compose: `union join a.k = b.k = base.k and a.k2 = b.k2`.\n#    A COMMA never separates join groups (it is a select-list separator only):\n#    `union join a.k1 = b.k1, a.k2 = b.k2` is an error - use `and`.\n#  - Terminate the statement with `;` after the LAST clause, whatever that clause\n#    is (join, `having`, `order by`, `limit`).\n#  - A join key may be ANY expression (a computed/offset key, an aggregate, a\n#    window), not only a bare field - see (4). Only `=` equality is supported.\n#  - Joins NEVER drop rows: for an intersection (rows present in BOTH sides),\n#    add explicit conditions - `where <side_a attr> is not null and <side_b\n#    attr> is not null`. A one-sided `is not null` keeps the other side's\n#    exclusive rows. Either spelling reads as \"did this side match?\": the\n#    side's own JOIN KEY (`where a.k is not null`, tested per side even though\n#    the projected key is the coalesced group axis) or any of its measures /\n#    attributes (`where a.cnt is not null`). Same in `having`.\n#  - `left`/`full`/`inner`/`right`/`cross join` are NOT valid query-scoped joins\n#    (they error) - only `subset`/`union`/`equal`. Map an old `left join a = b` to\n#    `subset join b = a` (swap operands) and `full join a = b` to\n#    `union join a = b`. A union key-group must be ALL union (can't mix with\n#    subset on the SAME key; `union join a = b = c` chains one union group);\n"
 	},
 	{
 		name: "set-intersect-difference",

package/dist/llm.js

@@ -1,3 +1,3 @@
-import { $ as e, At as t, Ct as n, Dt as r, Et as i, G as a, H as o, J as s, K as c, Mt as l, Nt as u, Ot as d, Pt as f, Tt as p, V as m, Y as h, _t as g, ct as _, gt as v, it as y, jt as b, lt as x, nt as S, ot as C, q as w, rt as T, st as E, t as D, vt as O, wt as k } from "./llm-ClU-2rgk.js";
+import { $ as e, At as t, Ct as n, Dt as r, Et as i, G as a, H as o, J as s, K as c, Mt as l, Nt as u, Ot as d, Pt as f, Tt as p, V as m, Y as h, _t as g, ct as _, gt as v, it as y, jt as b, lt as x, nt as S, ot as C, q as w, rt as T, st as E, t as D, vt as O, wt as k } from "./llm-Wf4Etiu9.js";
 import { n as A, t as j } from "./chatHelpers-F_woEbYF.js";
 export { p as AnthropicProvider, C as CHAT_TOOLS, b as DEFAULT_COMPACTION_THRESHOLD_TOKENS, a as DEFAULT_KEEP_RECENT_TOKENS, n as DeepSeekProvider, r as DemoProvider, m as EDITOR_REFINEMENT_TOOLS, D as EditorRefinementToolExecutor, k as GoogleProvider, l as KEY_PLACEHOLDERS, t as LLMProvider, i as OpenAIProvider, d as OpenRouterProvider, u as PROVIDERS, f as PROVIDER_LABELS, E as RETURN_TO_USER_TOOL, _ as buildChatAgentSystemPrompt, x as buildCustomTrilogyPrompt, o as buildEditorRefinementPrompt, c as compactChat, w as compactHistory, j as createChatNamePrompt, v as createDashboardPrompt, g as createFilterPrompt, O as createPrompt, S as estimateHistoryTokens, T as estimateMessageTokens, y as estimateTextTokens, A as extractChatName, s as maybeCompactChat, e as runToolLoop, h as shouldCompact };

package/dist/{localStorage-CGQ1iJXt.js → localStorage-C3xms-ab.js}

@@ -1,6 +1,6 @@
-import { Ct as e, Dt as t, Et as n, It as r, Ot as i, R as a, Tt as o, hn as s, wt as c } from "./llm-ClU-2rgk.js";
-import { _ as l } from "./stores-LlyiwSbK.js";
-import { Z as u, et as d } from "./editorStore-CAlMJftt.js";
+import { Ct as e, Dt as t, Et as n, It as r, Ot as i, R as a, Tt as o, hn as s, wt as c } from "./llm-Wf4Etiu9.js";
+import { _ as l } from "./stores-DTGQZ4_B.js";
+import { Z as u, et as d } from "./editorStore-CiWH_nGN.js";
 import { c as f, f as p, l as m, n as h, o as g, u as _ } from "./connections-CWH3zwH2.js";
 import { reactive as v } from "vue";
 //#region data/idbKv.ts

package/dist/monaco.js

@@ -1,5 +1,5 @@
-import { t as e } from "./editorStore-CAlMJftt.js";
-import { a as t, c as n, i as r, l as i, o as a, r as o, s } from "./vocabulary-CKLGILtR.js";
+import { t as e } from "./editorStore-CiWH_nGN.js";
+import { a as t, c as n, i as r, l as i, o as a, r as o, s } from "./vocabulary-DQK5jB7m.js";
 import { l as c, p as l } from "./workers-B9Eybl-M.js";
 import "./editor.main-BQh6DxDB.js";
 //#region language/constants.ts

package/dist/{reconcile-DXIvGwBk.js → reconcile-BH_ReGDR.js}

@@ -1,4 +1,4 @@
-import { et as e } from "./editorStore-CAlMJftt.js";
+import { et as e } from "./editorStore-CiWH_nGN.js";
 import { reactive as t } from "vue";
 //#region editors/reconcile.ts
 var n = (e, t, n) => Object.values(e).filter((e) => e.storage === "remote" && e.remoteStoreId === t && e.remotePath === n), r = (r, i) => {

package/dist/{remoteStoreSync-CuQBWpWf.js → remoteStoreSync-B0K8QScO.js}

@@ -1,6 +1,6 @@
 import { t as e } from "./_plugin-vue_export-helper-f9Rw-Il4.js";
-import { p as t } from "./useGlobalChatPanel-DuP1q7_f.js";
-import { n } from "./reconcile-DXIvGwBk.js";
+import { p as t } from "./useGlobalChatPanel-D9kMZLeF.js";
+import { n } from "./reconcile-BH_ReGDR.js";
 import { createBlock as r, createElementVNode as i, defineComponent as a, openBlock as o, toDisplayString as s, withCtx as c } from "vue";
 //#region components/ConfirmDialog.vue?vue&type=script&setup=true&lang.ts
 var l = { class: "confirm-message" }, u = ["data-testid"], d = ["data-testid", "disabled"], f = /*#__PURE__*/ e(/* @__PURE__ */ a({