[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7C2MMFX40K0WTDNW05VCQEJ":3,"skill-same:01M7C2MMFX40K0WTDNW05VCQEJ":49},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":13,"repositoryHtmlUrl":26,"repositoryHomepageUrl":6,"ownerHtmlUrl":27,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":28,"install":29,"resources":42},"01M7C2MMFX40K0WTDNW05VCQEJ","semantics",null,"Use when the views of an existing Denodo 9.5 database have to be understood by people and by AI consumers — the Denodo MCP Server, Assisted Query in the Data Marketplace, the AI SDK. Auditing a database for views and fields without descriptions, primary keys that are missing or not unique, missing associations and the VDP tag the MCP Server shows views by; writing view, field, association and tag descriptions from the data, and primary keys. Also for \"describe \u002F document these views\", \"prepare this database for AI, MCP or Assisted Query\", \"why does the agent not see this view through MCP\", a description that says the wrong thing. Not for building or changing what a view computes — \u002Fdenodo:views; not for marketplace-only metadata (logical names, property groups, marketplace tags) — \u002Fdenodo:marketplace.","kirivandubai\u002Fdenodo_wipe_coading_skills","kirivandubai","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F221521416?v=4","01M7C2HC9KETA166XKR09YMKZQ","01M7C2HCAJ787NNYKMCG12S3DY",0,"MIT","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fsemantics\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: semantics\ndescription: Use when the views of an existing Denodo 9.5 database have to be understood by people and by AI consumers — the Denodo MCP Server, Assisted Query in the Data Marketplace, the AI SDK. Auditing a database for views and fields without descriptions, primary keys that are missing or not unique, missing associations and the VDP tag the MCP Server shows views by; writing view, field, association and tag descriptions from the data, and primary keys. Also for \"describe \u002F document these views\", \"prepare this database for AI, MCP or Assisted Query\", \"why does the agent not see this view through MCP\", a description that says the wrong thing. Not for building or changing what a view computes — \u002Fdenodo:views; not for marketplace-only metadata (logical names, property groups, marketplace tags) — \u002Fdenodo:marketplace.\n---\n\n# Make views understandable: descriptions, keys, associations, visibility\n\nAI consumers of Denodo build their prompts from the metadata of a view and a few of its\nvalues (the MCP Server always; Assisted Query and the AI SDK when sample data is on), not\nfrom reading the view. The Denodo MCP Server, Assisted Query and the AI SDK all read the\nview's description, each field's description, type, primary-key and nullable flags, the\nnames of its tags (Assisted Query and the Denodo Assistant also their descriptions), and its\ndirect associations, the only joins they are told about. The MCP Server also shows a view at\nall only when it carries the tag it is configured with. A view with a cryptic name and no\ndescription is answered with a guess; a wrong description or a primary key that is not\nunique is answered with confidence.\n\n**This skill audits a database and fills that metadata in:** view, field, association and\ntag descriptions, primary keys, the associations that are missing, and the MCP visibility\ntag. Logical names, property groups and descriptions edited in the marketplace live only in\nthe Data Marketplace (`\u002Fdenodo:marketplace`, or its UI); configuring the MCP Server, the AI\nSDK or Assisted Query is the administrator's (the last two are Enterprise Plus features).\nApplying files is `\u002Fdenodo:execute`; the working loop and the safety rule are `\u002Fdenodo:vql`.\n\n**A view you are building now** (`\u002Fdenodo:views`) gets the same metadata in its own file:\nprofile it (section 2), write the texts (section 3), put them in its `CREATE OR REPLACE VIEW`\n(section 5), apply and verify — no approval step for the texts, the view is yours. The MCP tag\ngoes in too only when the human asked for the view to be visible to agents (MCP visibility,\nbelow). What the data cannot settle stays out of the text and goes to the human as a question.\n\n## The rule: from the data, approved before it is written\n\n1. **Every sentence comes from the data, the view's definition or the human.** A column name\n   is a question, not an answer: a `count` column holding `-1` for a fifth of the rows, a\n   `code` column that turns out to be a region, a `GROUP BY` on a label that two keys share\n   — none of it shows in a name. What the data cannot settle (what `-1` means, the currency,\n   whether an amount is monthly or yearly) is a question in your proposal, never \"probably\"\n   in a description.\n2. **Nothing is written to a view you did not create in this session until the human says\n   yes to the texts** — whatever the statement: `ALTER VIEW`, `ALTER TAG`, or the view\n   re-declared with `CREATE OR REPLACE VIEW`. Whose view it is, is the `own` of `vql plan` on\n   the file (`\u002Fdenodo:vql`).\n3. **When you cannot ask** — the human is away, the go-live is close — the answer is the\n   audit, the proposal and the statement file, not the statements applied. Say which file\n   to apply and what applying it changes. The one exception is a view the human named to be\n   made visible to an agent: that request is the yes for its tag (MCP visibility, below).\n\n| Rationalization | Reality |\n|---|---|\n| \"`CREATE OR REPLACE` is mine to do; only `ALTER` needs a yes\" | Re-declaring someone's view to change its description is a write to their view. The rule is about the object, not the keyword. |\n| \"The request was exactly to describe these views\" | They asked for descriptions. The yes is to the texts, after they read them. |\n| \"A wrong description or key misleads the AI, so fixing it cannot wait\" | Show it with the evidence and the counts first — a proposal, not an `ALTER` nobody reviewed; the owner may know why it says what it says. |\n| \"The view is not in the MCP tag yet, so tagging it is harmless\" | The tag publishes the view (its sample values and its rows) to every agent whose Denodo user may read it. Which views get it is the owner's decision, view by view. |\n\n## 1. Audit — read-only\n\n```sql\n-- verified: 9.5.1 (live, 2026-10-07)\nSELECT e.name, e.subtype, e.folder,\n       CASE WHEN e.description IS NULL OR TRIM(e.description) = '' THEN 'none' ELSE 'yes' END AS view_description,\n       c.fields, c.undescribed_fields, c.pk_fields,\n       CASE WHEN t.view_name IS NULL THEN 'no' ELSE 'yes' END AS mcp_tag\n  FROM GET_ELEMENTS() AS e\n       LEFT OUTER JOIN (\n           SELECT view_name, COUNT(*) AS fields,\n                  SUM(CASE WHEN column_remarks IS NULL OR TRIM(column_remarks) = '' THEN 1 ELSE 0 END) AS undescribed_fields,\n                  SUM(CASE WHEN column_is_primary_key THEN 1 ELSE 0 END) AS pk_fields\n             FROM GET_VIEW_COLUMNS()\n            WHERE input_database_name = 'sales_analytics'\n            GROUP BY view_name) AS c\n       ON e.name = c.view_name\n       LEFT OUTER JOIN (\n           SELECT DISTINCT view_name\n             FROM GET_VIEW_TAGS()\n            WHERE input_database_name = 'sales_analytics' AND tag_name = 'mcp' AND column_name IS NULL) AS t\n       ON e.name = t.view_name\n WHERE e.input_database_name = 'sales_analytics' AND e.input_type = 'views'\n ORDER BY subtype, name;\n\nSELECT view_name, column_name, column_vdp_type, column_is_primary_key, column_is_nullable, column_remarks\n  FROM GET_VIEW_COLUMNS()\n WHERE input_database_name = 'sales_analytics';\n\nSELECT association_name, association_description, mappings, valid\n  FROM GET_ASSOCIATIONS()\n WHERE input_database_name = 'sales_analytics' AND input_type = 'views';\n```\n\nOne row per view, then every field, then every association — the whole database in three\ncalls. `'mcp'` in the first query is the tag the MCP Server is configured with (MCP\nvisibility, below) — put in the one the human names; until you know it, drop the second\n`LEFT OUTER JOIN` and the `mcp_tag` column rather than guess. Then, per view the consumers\nwill read:\n\n- **An existing description is checked, not trusted.** \"One row per household\" is a claim\n  about the grain; the profile (next section) confirms it or not. Descriptions copied from\n  another view are common.\n- **A declared primary key is checked**: `SELECT COUNT(*), COUNT(\u003Ckey>), COUNT(DISTINCT \u003Ckey>)\n  FROM \u003Cview>` — all three equal. For a composite key: `SELECT COUNT(*) FROM (SELECT \u003Ck1>,\n  \u003Ck2> FROM \u003Cview> GROUP BY \u003Ck1>, \u003Ck2> HAVING COUNT(*) > 1) d` is `0`, and `COUNT(\u003Ck1>)`,\n  `COUNT(\u003Ck2>)` equal `COUNT(*)`. Denodo does not enforce keys. A declared key also marks\n  its columns `NOT NULL` in the metadata, so a wrong key is two wrong facts.\n- **A view with no key to give it** — an aggregate grouped by a label two keys share, or\n  with a `NULL` group — gets none; its description says what one row is instead.\n- **Missing associations**: a column that carries another view's key (same values, 0\n  orphans — the check is in `\u002Fdenodo:views`) with no association between the two. Assisted\n  Query cannot join them (it is sent only the view's direct associations); an MCP agent is not\n  told they join and has to guess the condition from column names. Creating one is the\n  association template in `\u002Fdenodo:views`; between views you did not create it goes into\n  the proposal like the texts — it becomes a dependant of both views, and their owner's\n  `DROP VIEW` then needs `CASCADE`.\n- **Tags**: `DESC TAG \u003Ctag>` gives its name and description only, no assignments — Assisted\n  Query reads that description; a tag without one tells it nothing.\n- `column_remarks` is the description a consumer sees, **inherited ones included**: a field\n  that passes a column through unchanged shows the description of the column below. `DESC\n  VQL` shows only the view's own. Base views: their field descriptions reach every view\n  above that passes the column through, so one description there can cover many views.\n- **A metric view** (`subtype = 'metric'`) is described like any view, field by field —\n  its dimensions and metrics. A view built over it inherits the descriptions of the\n  dimensions it passes through and none of the metrics: its `evaluate_metric` columns need\n  their own (`\u002Fdenodo:metrics`).\n\n## 2. Profile — what the data says\n\n```sql\n-- verified: 9.5.1 (live, 2026-10-07)\nCONNECT DATABASE sales_analytics;\n\n-- The grain: the key the description will name. Equal counts, no NULL key.\nSELECT COUNT(*) AS row_count, COUNT(household_sk) AS keys_not_null,\n       COUNT(DISTINCT household_sk) AS distinct_keys\n  FROM iv_household_income;\n\n-- Every column in one pass: NULLs, distinct values, range, padding.\nSELECT COUNT(buy_potential) AS buy_potential_not_null,\n       COUNT(DISTINCT buy_potential) AS buy_potential_distinct,\n       SUM(CASE WHEN buy_potential LIKE '% ' THEN 1 ELSE 0 END) AS buy_potential_padded,\n       COUNT(vehicles) AS vehicles_not_null,\n       MIN(vehicles) AS vehicles_min, MAX(vehicles) AS vehicles_max,\n       COUNT(DISTINCT vehicles) AS vehicles_distinct\n  FROM iv_household_income;\n\n-- The values themselves, for each column with a short list.\nSELECT vehicles, COUNT(*) AS row_count\n  FROM iv_household_income GROUP BY vehicles ORDER BY vehicles;\n```\n\nWhat to read out of it, and where it goes:\n\n| You see | It means | The description says |\n|---|---|---|\n| `distinct_keys` = `row_count`, no `NULL` key | the grain | \"One row per household (`household_sk`)\" |\n| a value far outside the rest (`-1`, `0`, `9999`, `'Unknown'`) | a sentinel | what it stands for — if the data or the human says; otherwise a question |\n| `…_padded` above 0 | padded text — counted with `LIKE '% '`: `LEN` pushed down to SQL Server ignores trailing spaces (*verified: 9.5.1 (live, 2026-10-06)*) | \"compare with `TRIM(x) = '…'`\" — Denodo's own equality misses padded values, a delegated one may not; only the `TRIM` form holds everywhere |\n| a short value list | codes | each code with its meaning, when known |\n| `NULL`s in a column | missing values | what a `NULL` stands for, when known |\n| an aggregate grouped by a label | one row per label, not per key | two keys with the same text are one row — compare the label's distinct count with the key's |\n| a join in the definition (`DESC VQL`) | rows dropped (`INNER`), doubled, or — an outer join under a `GROUP BY` — gathered into one `NULL` row | which rows are in, and what the `NULL` row holds |\n\nEach query reads the whole view. Over a view with statistics gathered, `SELECT * FROM\nGET_VIEW_STATISTICS() WHERE input_database_name = '\u003Cdb>' AND input_name = '\u003Cview>'` gives\nrows, distinct values, `NULL`s and ranges without reading it (0 rows: statistics not gathered,\na misspelt view, or no `EXECUTE` on it) — one view per call: without `input_name` it answers\n`The following fields are obligatory: input_name` — *verified: 9.5.1 (live, 2026-10-06)*.\n\n## 3. The texts\n\n**A view description** says what one row is and its key; what the view leaves out or adds\n(a filter, a join that drops rows, a group of `NULL`s); how to query it where that is not\nobvious (padding, a sentinel to exclude). **A field description** says what the value means,\nits unit, its values and what each code or sentinel stands for, and what `NULL` means. **An\nassociation description** says what the relationship is in business words. **A tag\ndescription** says what carrying the tag means.\n\n- No row counts or `NULL` counts: the next load changes them, and a consumer reads them as\n  today's. A count belongs in a description only when it is the point, with its date.\n- No restating the name (`dep_count` → \"The dependent count\"). No \"probably\", \"likely\",\n  \"appears to\".\n- Keep a correct description; propose a replacement for a contradicted one, with the\n  evidence. A description of up to 4,000 characters; a few sentences are enough.\n- `''` inside a text is a quote: `'the band''s households'`.\n\n## 4. The proposal — what the human approves\n\n```\nDatabase claims_analytics — 12 views audited.\nWrong now:\n  policies: PRIMARY KEY holder_id — 4,210 distinct in 9,800 rows. Proposed: policy_id (9,800 distinct, no NULL).\n  claims_by_region: its description is the one of claims_by_month. Proposed: \"Claims per region: one row per …\"\nNew texts:\n  | object            | proposed text                                    | from the data            |\n  | policies (view)   | Insurance policies. One row per policy (…)       | 9,800 rows = 9,800 keys  |\n  | policies.status   | Policy status: A active, L lapsed, X \u003Cyour answer> | values A, L, X          |\nQuestions: what does status X mean? Is premium monthly or yearly?\nMCP: which views should the agent see? (tag \u003Cname>, from the MCP Server's configuration)\nStatements: semantics\u002Fclaims_analytics.vql — applied after your yes.\n```\n\n### A database of many views: batches\n\nThe whole database is still audited and every view profiled — batches cut how much the human\nreads at once, not how much of the data is read (`\u002Fdenodo:vql`, **Many objects at once**):\n\n- The audit's field query returns a row per field: run it with `--max-rows 5000` and check\n  `truncated`; the profile is a file of reads, one all-columns statement per view, run with\n  `--continue-on-error`. A view whose columns call the server's LLM is profiled from its cache\n  or not read at all: every row read is a request to the provider (`\u002Fdenodo:ai`).\n- A batch is a statement file and its part of the proposal, `semantics\u002F\u003Cdatabase>\u002Fbatch_NN_\u003Cgroup>.vql`,\n  sized by its texts, not its views — field texts are most of the reading. Base views go in\n  the first batches: their field texts are inherited above (section 1).\n- Reads saved into the project keep counts and value lists of codes, not the rows of columns\n  that hold people's data: a sample with names and emails in a profile file is that data in git.\n- A question stays with the batch it blocks, so a yes to one batch is not held up by another's.\n- After a batch is applied, the audit query over its views is its check: `view_description =\n  'yes'` and `undescribed_fields` down to the fields left to inherit — per view, into the plan.\n\n## 5. Write it where the view is defined\n\n| Where the view's definition lives | Write the metadata |\n|---|---|\n| a `.vql` file in the project that is applied (git, CI, `\u002Fdenodo:views`) | **in that file**: `DESCRIPTION`, `PRIMARY KEY`, `TAGS ( … )` and `( \u003Cfield> ( description = '…' ) )` in the view's own `CREATE OR REPLACE VIEW`, in the clause order of `\u002Fdenodo:views`; then re-apply it |\n| only on the server (built in Design Studio) | an `ALTER` file of its own, `semantics\u002F\u003Cdatabase>.vql` — the template below. Do not re-declare the view to change its metadata; a change to what it computes gives it a file first (`\u002Fdenodo:views`, **Before a column changes**) |\n| a base view | `ALTER TABLE \u003Cview> …`, the same forms as below |\n\nStatements still waiting for a yes go into a file of their own, next to the one being\napplied — each file is applied whole, so what has a yes never goes out with what has not.\n\n**The file wins.** Re-applying a `CREATE OR REPLACE VIEW` without these clauses removes the\ndescription, the field descriptions, the primary key and every tag assignment of that view,\nsilently — including the MCP tag, so the view disappears from the agent. An `ALTER` over a\nview that has a file lasts until the file is next applied.\n\n```sql\n-- verified: 9.5.1 (live, 2026-10-07)\nCONNECT DATABASE sales_analytics;\n\nALTER VIEW household_income_by_band\n    DESCRIPTION = 'Households per income band. One row per income band (income_band_sk); a band with no households has no row.';\n\nALTER VIEW household_income_by_band (\n    ALTER COLUMN household_count ADD ( DESCRIPTION = 'Number of households in the band.' )\n    ALTER COLUMN avg_dependents ADD ( DESCRIPTION = 'Average number of dependents of the band''s households, rounded to two decimals.' )\n);\n\nALTER VIEW household_income_by_band ADD PRIMARY KEY ( 'income_band_sk' );\n\nALTER ASSOCIATION a_income_band_household\n    DESCRIPTION = 'The income band a household belongs to. Every household has one band; a band has many households.';\n\n-- only for a view the human named to be made visible to agents (MCP visibility, below)\nALTER TAG mcp\n    ADD_TO ( VIEWS ( sales_analytics.household_income_by_band ) COLUMNS () )\n    REMOVE_FROM ( VIEWS () COLUMNS () );\n```\n\n- Each statement changes that metadata and nothing else: the cache, the views built on it,\n  their status and the privileges granted on it stay as they were — *verified: 9.5.1 (live,\n  2026-09-30)*. The tool marks each one `destructive: alter`.\n- `ALTER TAG … ADD_TO` keeps the tag's other assignments; the tag must exist\n  (`\u002Fdenodo:catalog`). Replacing a key, `ALTER TAGS` and inheritance: Silent failures, below;\n  `DROP PRIMARY KEY`, `DROP DESCRIPTION` and the base-view forms: `references\u002Fmetadata.md`.\n\n## MCP visibility — \"the agent does not see this view\"\n\nThe MCP Server shows only views tagged with one of the tags in `mcp.visibility.tags` of its\n`config\u002Fapplication.properties`; the file it ships with sets `mcp`. When the property is not\nset, it shows every view the client's Denodo user may read, in every database, base views\nincluded. That file is on the MCP Server's host, not in Denodo: ask the human which tag, or\nread their copy of it. \"We run the shipped configuration\" settles it as `mcp`, and the tag\ncarried by the views the agent does see confirms it. Otherwise do not guess a tag name, and\ndo not create one to make a view appear.\n\nVDP tags are part of the Enterprise Plus subscription bundle (documentation;\n`env check --env dev` → `features.enterprise_plus`). On a server without it expect `ALTER TAG`\nto fail, and no view can be made visible by tag: say so, leave out the tag lines of the audit\nand of section 5, and treat which views the MCP Server shows as the administrator's\nconfiguration (`mcp.visibility.tags`).\n\nIn order:\n\n1. **The tag is on the view, not only on a column**:\n   `SELECT view_name, column_name FROM GET_VIEW_TAGS() WHERE input_database_name = '\u003Cdb>' AND input_view_name = '\u003Cview>' AND tag_name = '\u003Ctag>'`\n   — a row with an empty `column_name` (*verified: 9.5.1 (live, 2026-10-05)*). A tag on a column leaves the view hidden\n   (documentation).\n2. **The agent's user may read it.** The MCP Server works as the Denodo user its client\n   authenticates with (Basic or OAuth), so the user is the agent's own account (ask the human\n   which). `SELECT dbconnect, elementname, elementexecute FROM GET_CATALOG_EFFECTIVE_PERMISSIONS()\n   WHERE input_user_name = '\u003Cuser>' AND input_database_name = '\u003Cdb>'` — the database row\n   needs `dbconnect = true`, the view's row `elementexecute = true` (*verified: 9.5.1 (live, 2026-10-05)*).\n   For a user that exists only in the identity provider it fails with an error that the user\n   does not exist, unless you run it as that user (documentation). Granting is a change of\n   who reads what: `\u002Fdenodo:security`, after the human's yes.\n3. **The server has picked the change up**: it refreshes its schema while\n   `mcp.schema-refresh.enabled` is on (the default), but not at once: a change can wait until\n   a subsequent request prompts the server to update. Off, the server does not detect the\n   change; the manual names no way to force it, so ask the administrator.\n4. The view's file does not carry the tag → the next apply of that file hides the view\n   again (section 5).\n\nAdding the tag is a write to a shared object that publishes the view (its sample values and\nits rows) to every agent whose Denodo user may read it: the human names the views. \"Make\n`\u003Cview>` visible to the agent\" names one — that is the yes for the tag on that view, and\nonly for it; its missing description, and every other view, go into the proposal. The\nstatement is `ALTER TAG … ADD_TO`, plus `TAGS ( \u003Ctag> )` in the view's file if it has one —\nnever a re-declared view.\n\n## After writing: when each consumer sees it\n\n- **The MCP Server** reads Denodo directly — the next schema refresh.\n- **Assisted Query and the AI SDK** read through the Data Marketplace: nothing changes for\n  them until the marketplace is synchronised with Denodo (`\u002Fdenodo:marketplace`,\n  **Who sends it** — a server-wide call — and whether the database is there at all, on every\n  registered server), and the AI SDK until its `getMetadata` runs again for that database or\n  tag — whoever runs the SDK does that.\n- The AI SDK pointed at a **database** reads every view in it, base views included; pointed\n  at a **tag**, only the tagged ones. Which it is decides whether base views need describing\n  for it.\n\n## Verify\n\n| Check | Query | Expect |\n|---|---|---|\n| View descriptions | the audit query, section 1 | `view_description = 'yes'` for every view you wrote |\n| Field descriptions | `SELECT column_name, column_remarks FROM GET_VIEW_COLUMNS() WHERE input_database_name = '\u003Cdb>' AND input_view_name = '\u003Cview>'` | your texts; an unchanged field still shows the inherited one |\n| Primary key | the same query with `column_is_primary_key` | the key you proposed, and only it |\n| The key holds | `SELECT COUNT(*), COUNT(DISTINCT \u003Ckey>) FROM \u003Cview>` | equal |\n| Association descriptions | `SELECT association_name, association_description FROM GET_ASSOCIATIONS() WHERE input_database_name = '\u003Cdb>' AND input_type = 'views'` | your texts, `valid = true` |\n| MCP tag | `SELECT view_name FROM GET_VIEW_TAGS() WHERE input_database_name = '\u003Cdb>' AND tag_name = '\u003Ctag>' AND column_name IS NULL` | exactly the views the human named |\n| The file agrees | `vql desc --env dev --database \u003Cdb> \u003Cview> --vql` against the view's file | the same clauses — or the next apply undoes them |\n\n## Silent failures\n\nEach runs without an error — *verified: 9.5.1 (live, 2026-09-30)*.\n\n| You did | What consumers get | Instead |\n|---|---|---|\n| 1. `ALTER` on a view whose `.vql` file is applied later | description, field descriptions, key and tags gone after that apply; the view drops out of MCP | the metadata in the file |\n| 2. declared a key without counting | a key that is not unique, and its columns marked `NOT NULL` — the AI counts and joins as if both were true | the count first |\n| 3. described only the top view | fields computed by an expression or aggregate stay empty; pass-through fields inherit from below | describe computed fields where they are computed |\n| 4. `ALTER TAG … ADD_TO` a misspelt view | `ok`, no assignment | read `GET_VIEW_TAGS()` back |\n| 5. `ALTER TABLE \u003Cbase view> ( ALTER TAGS ( VIEW ( mcp ) COLUMNS () ) )` | every other tag of that view and of its columns removed | `ALTER TAG \u003Ctag> ADD_TO` |\n| 6. `ADD PRIMARY KEY` on a view that has one | the old key silently replaced | read the key first (audit, section 1) |\n\n## Common mistakes\n\n| You wrote | Server says | Fix |\n|---|---|---|\n| a view description over 4,000 characters | `Error storing view '\u003Cview>': Error accessing the metadata while loading\u002Fstoring objects` | shorten it |\n| `WHERE input_view_name IN ('a', 'b')` on a `GET_*` procedure | nothing — zero rows | one query per view, or the whole database and filter on `view_name` |\n| `DESCRIPTION 'x'` without `=` | `Syntax error: Exception parsing query near '''` | `DESCRIPTION = 'x'` |\n| `SELECT e.name FROM GET_ELEMENTS() AS e … ORDER BY e.name` — one procedure, aliased | `Error in ORDER BY clause: Invalid field name: e.name` | `ORDER BY name` — over a join of procedures both work |\n\n## Reference\n\n- `references\u002Fmetadata.md` — what each AI consumer reads and from where, every statement\n  form for views, base views, interface views, associations and tags, what survives which\n  statement, inheritance in full, the procedures that read metadata, statistics, and the\n  Denodo Assistant procedures that draft descriptions.\n","# Make views understandable: descriptions, keys, associations, visibility\n\nAI consumers of Denodo build their prompts from the metadata of a view and a few of its\nvalues (the MCP Server always; Assisted Query and the AI SDK when sample data is on), not\nfrom reading the view. The Denodo MCP Server, Assisted Query and the AI SDK all read the\nview's description, each field's description, type, primary-key and nullable flags, the\nnames of its tags (Assisted Query and the Denodo Assistant also their descriptions), and its\ndirect associations, the only joins they are told about. The MCP Server also shows a view at\nall only when it carries the tag it is configured with. A view with a cryptic name and no\ndescription is answered with a guess; a wrong description or a primary key that is not\nunique is answered with confidence.\n\n**This skill audits a database and fills that metadata in:** view, field, association and\ntag descriptions, primary keys, the associations that are missing, and the MCP visibility\ntag. Logical names, property groups and descriptions edited in the marketplace live only in\nthe Data Marketplace (`\u002Fdenodo:marketplace`, or its UI); configuring the MCP Server, the AI\nSDK or Assisted Query is the administrator's (the last two are Enterprise Plus features).\nApplying files is `\u002Fdenodo:execute`; the working loop and the safety rule are `\u002Fdenodo:vql`.\n\n**A view you are building now** (`\u002Fdenodo:views`) gets the same metadata in its own file:\nprofile it (section 2), write the texts (section 3), put them in its `CREATE OR REPLACE VIEW`\n(section 5), apply and verify — no approval step for the texts, the view is yours. The MCP tag\ngoes in too only when the human asked for the view to be visible to agents (MCP visibility,\nbelow). What the data cannot settle stays out of the text and goes to the human as a question.\n\n## The rule: from the data, approved before it is written\n\n1. **Every sentence comes from the data, the view's definition or the human.** A column name\n   is a question, not an answer: a `count` column holding `-1` for a fifth of the rows, a\n   `code` column that turns out to be a region, a `GROUP BY` on a label that two keys share\n   — none of it shows in a name. What the data cannot settle (what `-1` means, the currency,\n   whether an amount is monthly or yearly) is a question in your proposal, never \"probably\"\n   in a description.\n2. **Nothing is written to a view you did not create in this session until the human says\n   yes to the texts** — whatever the statement: `ALTER VIEW`, `ALTER TAG`, or the view\n   re-declared with `CREATE OR REPLACE VIEW`. Whose view it is, is the `own` of `vql plan` on\n   the file (`\u002Fdenodo:vql`).\n3. **When you cannot ask** — the human is away, the go-live is close — the answer is the\n   audit, the proposal and the statement file, not the statements applied. Say which file\n   to apply and what applying it changes. The one exception is a view the human named to be\n   made visible to an agent: that request is the yes for its tag (MCP visibility, below).\n\n| Rationalization | Reality |\n|---|---|\n| \"`CREATE OR REPLACE` is mine to do; only `ALTER` needs a yes\" | Re-declaring someone's view to change its description is a write to their view. The rule is about the object, not the keyword. |\n| \"The request was exactly to describe these views\" | They asked for descriptions. The yes is to the texts, after they read them. |\n| \"A wrong description or key misleads the AI, so fixing it cannot wait\" | Show it with the evidence and the counts first — a proposal, not an `ALTER` nobody reviewed; the owner may know why it says what it says. |\n| \"The view is not in the MCP tag yet, so tagging it is harmless\" | The tag publishes the view (its sample values and its rows) to every agent whose Denodo user may read it. Which views get it is the owner's decision, view by view. |\n\n## 1. Audit — read-only\n\n```sql\n-- verified: 9.5.1 (live, 2026-10-07)\nSELECT e.name, e.subtype, e.folder,\n       CASE WHEN e.description IS NULL OR TRIM(e.description) = '' THEN 'none' ELSE 'yes' END AS view_description,\n       c.fields, c.undescribed_fields, c.pk_fields,\n       CASE WHEN t.view_name IS NULL THEN 'no' ELSE 'yes' END AS mcp_tag\n  FROM GET_ELEMENTS() AS e\n       LEFT OUTER JOIN (\n           SELECT view_name, COUNT(*) AS fields,\n                  SUM(CASE WHEN column_remarks IS NULL OR TRIM(column_remarks) = '' THEN 1 ELSE 0 END) AS undescribed_fields,\n                  SUM(CASE WHEN column_is_primary_key THEN 1 ELSE 0 END) AS pk_fields\n             FROM GET_VIEW_COLUMNS()\n            WHERE input_database_name = 'sales_analytics'\n            GROUP BY view_name) AS c\n       ON e.name = c.view_name\n       LEFT OUTER JOIN (\n           SELECT DISTINCT view_name\n             FROM GET_VIEW_TAGS()\n            WHERE input_database_name = 'sales_analytics' AND tag_name = 'mcp' AND column_name IS NULL) AS t\n       ON e.name = t.view_name\n WHERE e.input_database_name = 'sales_analytics' AND e.input_type = 'views'\n ORDER BY subtype, name;\n\nSELECT view_name, column_name, column_vdp_type, column_is_primary_key, column_is_nullable, column_remarks\n  FROM GET_VIEW_COLUMNS()\n WHERE input_database_name = 'sales_analytics';\n\nSELECT association_name, association_description, mappings, valid\n  FROM GET_ASSOCIATIONS()\n WHERE input_database_name = 'sales_analytics' AND input_type = 'views';\n```\n\nOne row per view, then every field, then every association — the whole database in three\ncalls. `'mcp'` in the first query is the tag the MCP Server is configured with (MCP\nvisibility, below) — put in the one the human names; until you know it, drop the second\n`LEFT OUTER JOIN` and the `mcp_tag` column rather than guess. Then, per view the consumers\nwill read:\n\n- **An existing description is checked, not trusted.** \"One row per household\" is a claim\n  about the grain; the profile (next section) confirms it or not. Descriptions copied from\n  another view are common.\n- **A declared primary key is checked**: `SELECT COUNT(*), COUNT(\u003Ckey>), COUNT(DISTINCT \u003Ckey>)\n  FROM \u003Cview>` — all three equal. For a composite key: `SELECT COUNT(*) FROM (SELECT \u003Ck1>,\n  \u003Ck2> FROM \u003Cview> GROUP BY \u003Ck1>, \u003Ck2> HAVING COUNT(*) > 1) d` is `0`, and `COUNT(\u003Ck1>)`,\n  `COUNT(\u003Ck2>)` equal `COUNT(*)`. Denodo does not enforce keys. A declared key also marks\n  its columns `NOT NULL` in the metadata, so a wrong key is two wrong facts.\n- **A view with no key to give it** — an aggregate grouped by a label two keys share, or\n  with a `NULL` group — gets none; its description says what one row is instead.\n- **Missing associations**: a column that carries another view's key (same values, 0\n  orphans — the check is in `\u002Fdenodo:views`) with no association between the two. Assisted\n  Query cannot join them (it is sent only the view's direct associations); an MCP agent is not\n  told they join and has to guess the condition from column names. Creating one is the\n  association template in `\u002Fdenodo:views`; between views you did not create it goes into\n  the proposal like the texts — it becomes a dependant of both views, and their owner's\n  `DROP VIEW` then needs `CASCADE`.\n- **Tags**: `DESC TAG \u003Ctag>` gives its name and description only, no assignments — Assisted\n  Query reads that description; a tag without one tells it nothing.\n- `column_remarks` is the description a consumer sees, **inherited ones included**: a field\n  that passes a column through unchanged shows the description of the column below. `DESC\n  VQL` shows only the view's own. Base views: their field descriptions reach every view\n  above that passes the column through, so one description there can cover many views.\n- **A metric view** (`subtype = 'metric'`) is described like any view, field by field —\n  its dimensions and metrics. A view built over it inherits the descriptions of the\n  dimensions it passes through and none of the metrics: its `evaluate_metric` columns need\n  their own (`\u002Fdenodo:metrics`).\n\n## 2. Profile — what the data says\n\n```sql\n-- verified: 9.5.1 (live, 2026-10-07)\nCONNECT DATABASE sales_analytics;\n\n-- The grain: the key the description will name. Equal counts, no NULL key.\nSELECT COUNT(*) AS row_count, COUNT(household_sk) AS keys_not_null,\n       COUNT(DISTINCT household_sk) AS distinct_keys\n  FROM iv_household_income;\n\n-- Every column in one pass: NULLs, distinct values, range, padding.\nSELECT COUNT(buy_potential) AS buy_potential_not_null,\n       COUNT(DISTINCT buy_potential) AS buy_potential_distinct,\n       SUM(CASE WHEN buy_potential LIKE '% ' THEN 1 ELSE 0 END) AS buy_potential_padded,\n       COUNT(vehicles) AS vehicles_not_null,\n       MIN(vehicles) AS vehicles_min, MAX(vehicles) AS vehicles_max,\n       COUNT(DISTINCT vehicles) AS vehicles_distinct\n  FROM iv_household_income;\n\n-- The values themselves, for each column with a short list.\nSELECT vehicles, COUNT(*) AS row_count\n  FROM iv_household_income GROUP BY vehicles ORDER BY vehicles;\n```\n\nWhat to read out of it, and where it goes:\n\n| You see | It means | The description says |\n|---|---|---|\n| `distinct_keys` = `row_count`, no `NULL` key | the grain | \"One row per household (`household_sk`)\" |\n| a value far outside the rest (`-1`, `0`, `9999`, `'Unknown'`) | a sentinel | what it stands for — if the data or the human says; otherwise a question |\n| `…_padded` above 0 | padded text — counted with `LIKE '% '`: `LEN` pushed down to SQL Server ignores trailing spaces (*verified: 9.5.1 (live, 2026-10-06)*) | \"compare with `TRIM(x) = '…'`\" — Denodo's own equality misses padded values, a delegated one may not; only the `TRIM` form holds everywhere |\n| a short value list | codes | each code with its meaning, when known |\n| `NULL`s in a column | missing values | what a `NULL` stands for, when known |\n| an aggregate grouped by a label | one row per label, not per key | two keys with the same text are one row — compare the label's distinct count with the key's |\n| a join in the definition (`DESC VQL`) | rows dropped (`INNER`), doubled, or — an outer join under a `GROUP BY` — gathered into one `NULL` row | which rows are in, and what the `NULL` row holds |\n\nEach query reads the whole view. Over a view with statistics gathered, `SELECT * FROM\nGET_VIEW_STATISTICS() WHERE input_database_name = '\u003Cdb>' AND input_name = '\u003Cview>'` gives\nrows, distinct values, `NULL`s and ranges without reading it (0 rows: statistics not gathered,\na misspelt view, or no `EXECUTE` on it) — one view per call: without `input_name` it answers\n`The following fields are obligatory: input_name` — *verified: 9.5.1 (live, 2026-10-06)*.\n\n## 3. The texts\n\n**A view description** says what one row is and its key; what the view leaves out or adds\n(a filter, a join that drops rows, a group of `NULL`s); how to query it where that is not\nobvious (padding, a sentinel to exclude). **A field description** says what the value means,\nits unit, its values and what each code or sentinel stands for, and what `NULL` means. **An\nassociation description** says what the relationship is in business words. **A tag\ndescription** says what carrying the tag means.\n\n- No row counts or `NULL` counts: the next load changes them, and a consumer reads them as\n  today's. A count belongs in a description only when it is the point, with its date.\n- No restating the name (`dep_count` → \"The dependent count\"). No \"probably\", \"likely\",\n  \"appears to\".\n- Keep a correct description; propose a replacement for a contradicted one, with the\n  evidence. A description of up to 4,000 characters; a few sentences are enough.\n- `''` inside a text is a quote: `'the band''s households'`.\n\n## 4. The proposal — what the human approves\n\n```\nDatabase claims_analytics — 12 views audited.\nWrong now:\n  policies: PRIMARY KEY holder_id — 4,210 distinct in 9,800 rows. Proposed: policy_id (9,800 distinct, no NULL).\n  claims_by_region: its description is the one of claims_by_month. Proposed: \"Claims per region: one row per …\"\nNew texts:\n  | object            | proposed text                                    | from the data            |\n  | policies (view)   | Insurance policies. One row per policy (…)       | 9,800 rows = 9,800 keys  |\n  | policies.status   | Policy status: A active, L lapsed, X \u003Cyour answer> | values A, L, X          |\nQuestions: what does status X mean? Is premium monthly or yearly?\nMCP: which views should the agent see? (tag \u003Cname>, from the MCP Server's configuration)\nStatements: semantics\u002Fclaims_analytics.vql — applied after your yes.\n```\n\n### A database of many views: batches\n\nThe whole database is still audited and every view profiled — batches cut how much the human\nreads at once, not how much of the data is read (`\u002Fdenodo:vql`, **Many objects at once**):\n\n- The audit's field query returns a row per field: run it with `--max-rows 5000` and check\n  `truncated`; the profile is a file of reads, one all-columns statement per view, run with\n  `--continue-on-error`. A view whose columns call the server's LLM is profiled from its cache\n  or not read at all: every row read is a request to the provider (`\u002Fdenodo:ai`).\n- A batch is a statement file and its part of the proposal, `semantics\u002F\u003Cdatabase>\u002Fbatch_NN_\u003Cgroup>.vql`,\n  sized by its texts, not its views — field texts are most of the reading. Base views go in\n  the first batches: their field texts are inherited above (section 1).\n- Reads saved into the project keep counts and value lists of codes, not the rows of columns\n  that hold people's data: a sample with names and emails in a profile file is that data in git.\n- A question stays with the batch it blocks, so a yes to one batch is not held up by another's.\n- After a batch is applied, the audit query over its views is its check: `view_description =\n  'yes'` and `undescribed_fields` down to the fields left to inherit — per view, into the plan.\n\n## 5. Write it where the view is defined\n\n| Where the view's definition lives | Write the metadata |\n|---|---|\n| a `.vql` file in the project that is applied (git, CI, `\u002Fdenodo:views`) | **in that file**: `DESCRIPTION`, `PRIMARY KEY`, `TAGS ( … )` and `( \u003Cfield> ( description = '…' ) )` in the view's own `CREATE OR REPLACE VIEW`, in the clause order of `\u002Fdenodo:views`; then re-apply it |\n| only on the server (built in Design Studio) | an `ALTER` file of its own, `semantics\u002F\u003Cdatabase>.vql` — the template below. Do not re-declare the view to change its metadata; a change to what it computes gives it a file first (`\u002Fdenodo:views`, **Before a column changes**) |\n| a base view | `ALTER TABLE \u003Cview> …`, the same forms as below |\n\nStatements still waiting for a yes go into a file of their own, next to the one being\napplied — each file is applied whole, so what has a yes never goes out with what has not.\n\n**The file wins.** Re-applying a `CREATE OR REPLACE VIEW` without these clauses removes the\ndescription, the field descriptions, the primary key and every tag assignment of that view,\nsilently — including the MCP tag, so the view disappears from the agent. An `ALTER` over a\nview that has a file lasts until the file is next applied.\n\n```sql\n-- verified: 9.5.1 (live, 2026-10-07)\nCONNECT DATABASE sales_analytics;\n\nALTER VIEW household_income_by_band\n    DESCRIPTION = 'Households per income band. One row per income band (income_band_sk); a band with no households has no row.';\n\nALTER VIEW household_income_by_band (\n    ALTER COLUMN household_count ADD ( DESCRIPTION = 'Number of households in the band.' )\n    ALTER COLUMN avg_dependents ADD ( DESCRIPTION = 'Average number of dependents of the band''s households, rounded to two decimals.' )\n);\n\nALTER VIEW household_income_by_band ADD PRIMARY KEY ( 'income_band_sk' );\n\nALTER ASSOCIATION a_income_band_household\n    DESCRIPTION = 'The income band a household belongs to. Every household has one band; a band has many households.';\n\n-- only for a view the human named to be made visible to agents (MCP visibility, below)\nALTER TAG mcp\n    ADD_TO ( VIEWS ( sales_analytics.household_income_by_band ) COLUMNS () )\n    REMOVE_FROM ( VIEWS () COLUMNS () );\n```\n\n- Each statement changes that metadata and nothing else: the cache, the views built on it,\n  their status and the privileges granted on it stay as they were — *verified: 9.5.1 (live,\n  2026-09-30)*. The tool marks each one `destructive: alter`.\n- `ALTER TAG … ADD_TO` keeps the tag's other assignments; the tag must exist\n  (`\u002Fdenodo:catalog`). Replacing a key, `ALTER TAGS` and inheritance: Silent failures, below;\n  `DROP PRIMARY KEY`, `DROP DESCRIPTION` and the base-view forms: `references\u002Fmetadata.md`.\n\n## MCP visibility — \"the agent does not see this view\"\n\nThe MCP Server shows only views tagged with one of the tags in `mcp.visibility.tags` of its\n`config\u002Fapplication.properties`; the file it ships with sets `mcp`. When the property is not\nset, it shows every view the client's Denodo user may read, in every database, base views\nincluded. That file is on the MCP Server's host, not in Denodo: ask the human which tag, or\nread their copy of it. \"We run the shipped configuration\" settles it as `mcp`, and the tag\ncarried by the views the agent does see confirms it. Otherwise do not guess a tag name, and\ndo not create one to make a view appear.\n\nVDP tags are part of the Enterprise Plus subscription bundle (documentation;\n`env check --env dev` → `features.enterprise_plus`). On a server without it expect `ALTER TAG`\nto fail, and no view can be made visible by tag: say so, leave out the tag lines of the audit\nand of section 5, and treat which views the MCP Server shows as the administrator's\nconfiguration (`mcp.visibility.tags`).\n\nIn order:\n\n1. **The tag is on the view, not only on a column**:\n   `SELECT view_name, column_name FROM GET_VIEW_TAGS() WHERE input_database_name = '\u003Cdb>' AND input_view_name = '\u003Cview>' AND tag_name = '\u003Ctag>'`\n   — a row with an empty `column_name` (*verified: 9.5.1 (live, 2026-10-05)*). A tag on a column leaves the view hidden\n   (documentation).\n2. **The agent's user may read it.** The MCP Server works as the Denodo user its client\n   authenticates with (Basic or OAuth), so the user is the agent's own account (ask the human\n   which). `SELECT dbconnect, elementname, elementexecute FROM GET_CATALOG_EFFECTIVE_PERMISSIONS()\n   WHERE input_user_name = '\u003Cuser>' AND input_database_name = '\u003Cdb>'` — the database row\n   needs `dbconnect = true`, the view's row `elementexecute = true` (*verified: 9.5.1 (live, 2026-10-05)*).\n   For a user that exists only in the identity provider it fails with an error that the user\n   does not exist, unless you run it as that user (documentation). Granting is a change of\n   who reads what: `\u002Fdenodo:security`, after the human's yes.\n3. **The server has picked the change up**: it refreshes its schema while\n   `mcp.schema-refresh.enabled` is on (the default), but not at once: a change can wait until\n   a subsequent request prompts the server to update. Off, the server does not detect the\n   change; the manual names no way to force it, so ask the administrator.\n4. The view's file does not carry the tag → the next apply of that file hides the view\n   again (section 5).\n\nAdding the tag is a write to a shared object that publishes the view (its sample values and\nits rows) to every agent whose Denodo user may read it: the human names the views. \"Make\n`\u003Cview>` visible to the agent\" names one — that is the yes for the tag on that view, and\nonly for it; its missing description, and every other view, go into the proposal. The\nstatement is `ALTER TAG … ADD_TO`, plus `TAGS ( \u003Ctag> )` in the view's file if it has one —\nnever a re-declared view.\n\n## After writing: when each consumer sees it\n\n- **The MCP Server** reads Denodo directly — the next schema refresh.\n- **Assisted Query and the AI SDK** read through the Data Marketplace: nothing changes for\n  them until the marketplace is synchronised with Denodo (`\u002Fdenodo:marketplace`,\n  **Who sends it** — a server-wide call — and whether the database is there at all, on every\n  registered server), and the AI SDK until its `getMetadata` runs again for that database or\n  tag — whoever runs the SDK does that.\n- The AI SDK pointed at a **database** reads every view in it, base views included; pointed\n  at a **tag**, only the tagged ones. Which it is decides whether base views need describing\n  for it.\n\n## Verify\n\n| Check | Query | Expect |\n|---|---|---|\n| View descriptions | the audit query, section 1 | `view_description = 'yes'` for every view you wrote |\n| Field descriptions | `SELECT column_name, column_remarks FROM GET_VIEW_COLUMNS() WHERE input_database_name = '\u003Cdb>' AND input_view_name = '\u003Cview>'` | your texts; an unchanged field still shows the inherited one |\n| Primary key | the same query with `column_is_primary_key` | the key you proposed, and only it |\n| The key holds | `SELECT COUNT(*), COUNT(DISTINCT \u003Ckey>) FROM \u003Cview>` | equal |\n| Association descriptions | `SELECT association_name, association_description FROM GET_ASSOCIATIONS() WHERE input_database_name = '\u003Cdb>' AND input_type = 'views'` | your texts, `valid = true` |\n| MCP tag | `SELECT view_name FROM GET_VIEW_TAGS() WHERE input_database_name = '\u003Cdb>' AND tag_name = '\u003Ctag>' AND column_name IS NULL` | exactly the views the human named |\n| The file agrees | `vql desc --env dev --database \u003Cdb> \u003Cview> --vql` against the view's file | the same clauses — or the next apply undoes them |\n\n## Silent failures\n\nEach runs without an error — *verified: 9.5.1 (live, 2026-09-30)*.\n\n| You did | What consumers get | Instead |\n|---|---|---|\n| 1. `ALTER` on a view whose `.vql` file is applied later | description, field descriptions, key and tags gone after that apply; the view drops out of MCP | the metadata in the file |\n| 2. declared a key without counting | a key that is not unique, and its columns marked `NOT NULL` — the AI counts and joins as if both were true | the count first |\n| 3. described only the top view | fields computed by an expression or aggregate stay empty; pass-through fields inherit from below | describe computed fields where they are computed |\n| 4. `ALTER TAG … ADD_TO` a misspelt view | `ok`, no assignment | read `GET_VIEW_TAGS()` back |\n| 5. `ALTER TABLE \u003Cbase view> ( ALTER TAGS ( VIEW ( mcp ) COLUMNS () ) )` | every other tag of that view and of its columns removed | `ALTER TAG \u003Ctag> ADD_TO` |\n| 6. `ADD PRIMARY KEY` on a view that has one | the old key silently replaced | read the key first (audit, section 1) |\n\n## Common mistakes\n\n| You wrote | Server says | Fix |\n|---|---|---|\n| a view description over 4,000 characters | `Error storing view '\u003Cview>': Error accessing the metadata while loading\u002Fstoring objects` | shorten it |\n| `WHERE input_view_name IN ('a', 'b')` on a `GET_*` procedure | nothing — zero rows | one query per view, or the whole database and filter on `view_name` |\n| `DESCRIPTION 'x'` without `=` | `Syntax error: Exception parsing query near '''` | `DESCRIPTION = 'x'` |\n| `SELECT e.name FROM GET_ELEMENTS() AS e … ORDER BY e.name` — one procedure, aliased | `Error in ORDER BY clause: Invalid field name: e.name` | `ORDER BY name` — over a join of procedures both work |\n\n## Reference\n\n- `references\u002Fmetadata.md` — what each AI consumer reads and from where, every statement\n  form for views, base views, interface views, associations and tags, what survives which\n  statement, inheritance in full, the procedures that read metadata, statistics, and the\n  Denodo Assistant procedures that draft descriptions.\n",{"name":5,"description":7},"main","fd41fb386b25141c459264d10a1c49bd3b4f4c4b","skills\u002Fsemantics","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills","https:\u002F\u002Fgithub.com\u002Fkirivandubai","2026-10-29 13:49:22",{"source":8,"skillName":5,"commands":30,"prompt":40,"zipAvailable":41},[31,34,37],{"tool":32,"command":33},"npx","npx skills add kirivandubai\u002Fdenodo_wipe_coading_skills --skill semantics",{"tool":35,"command":36},"bunx","bunx skills add kirivandubai\u002Fdenodo_wipe_coading_skills --skill semantics",{"tool":38,"command":39},"pnpm","pnpm dlx skills add kirivandubai\u002Fdenodo_wipe_coading_skills --skill semantics","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fsemantics\u002FSKILL.md",true,[43],{"path":44,"resourceType":45,"resourceTypeDesc":46,"sizeBytes":47,"sourceHtmlUrl":48},"references\u002Fmetadata.md","REFERENCE","参考文档",8660,"https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fsemantics\u002Freferences\u002Fmetadata.md",[50,57,64,71,78,85],{"skillNo":51,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":18},"01M7C2NGAJEKKG90BG9Q1Z2M96","vql","Use when doing anything with Denodo — creating or changing a database, folder, tag, data source, wrapper, base, derived, interface or metric view, association, cache, role or security policy, remote table or summary, a Data Marketplace object, a regression test or a Scheduler job; writing or fixing VQL, or an expression in a SELECT — text, dates, casts; deciding where an object lives and what to name it; \"where do I start\", \"build a mart\", \"connect a source\", \"vibe-code on Denodo\". Start here when the request is ambiguous or names no single object.","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fvql\u002FSKILL.md",[],[],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":18},"01M7C2N5W4203SQ2TRYT6E0VSR","views","Use when building on top of views that already exist in Denodo 9.5 — a derived view (a join, an aggregate, a mart, a filtered or parameterised view, a UNION ALL of views of one entity, FLATTEN of a JSON array into rows, NEST of rows into an array), an interface view, or an association between two views. Also before changing a view others depend on — \"what uses this view\", \"can I drop this column\", \"where does this field come from\" — for whether a mart over a database runs there or pulls every row into Denodo, for \"combine these sources into one view\", \"expose this as a data product\", and for a view created without an error that fails on SELECT or turns up INVALID. Not for sources and base views — \u002Fdenodo:datasources; databases, folders, VDP tags — \u002Fdenodo:catalog; a view's cache — \u002Fdenodo:cache; a copy stored as a table — \u002Fdenodo:materialize; a metric view — \u002Fdenodo:metrics.","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fviews\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":18},"01M7C2MY48XXHR87G39R0F0VTK","testing","Use when a Denodo 9.5 project needs regression tests, or its tests need running or fixing — .denodotest files that the Denodo Testing Tool runs in CI (%EXECUTION, %RESULTS[data|query|csv|exception], configuration.properties, denodo-test.sh) — \"add tests for this mart\", \"a safety net before I rewrite these views\", \"make sure the dashboards see no difference\", \"what should CI check on our data products\", \"prove the mart is right\", \"CI is red on the Denodo tests\", a test that fails on row order, on a column it did not ask for or on a value padded with spaces. Not for checking a new view once, right after creating it — \u002Fdenodo:views.","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Ftesting\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":18},"01M7C2M9XVTRVXYXMGTEC62EFB","security","Use when deciding who may read what in Denodo 9.5 — masking or hiding columns (PII, contact details, salaries, bank accounts) from some users or roles, letting a role see only some rows (a country, a tenant) or each person only their own, taking someone's access away (ALTER USER … REVOKE), denying a view to a group, a global security policy (CREATE GLOBAL_SECURITY_POLICY) and the VDP tags it reads, giving a user a role or giving a role access to a database or views (CREATE ROLE … GRANT EXECUTE, ALTER USER … GRANT ROLE), checking what a given user really gets back, or a user who sees data a restriction should hide (\"he still gets every row\", \"the masking stopped working\"). Not for a tag as a label without a policy (\u002Fdenodo:catalog) or for descriptions and the MCP visibility tag (\u002Fdenodo:semantics).","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fsecurity\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":18},"01M7C2KYWB8B6ZC026CT3Z1MZA","scheduler","Use when something in Denodo 9.5 has to happen on a schedule, or a scheduled run misbehaves — refreshing the cache of a view every night or every hour, running REFRESH of a remote table or a procedure at a set time, exporting a view to a CSV file every week or month, running a Denodo Scheduler job now, stopping, pausing, enabling or deleting it, reading what last night's run did or why it failed, \"the cache doubles after every refresh\", \"the job fires at the wrong hour\", \"the export file is missing or has no header\". Denodo Scheduler projects, jobs, cron triggers and reports, through the REST API of the Scheduler administration tool. Not for switching a cache on or loading it once (\u002Fdenodo:cache), not for creating the remote table or summary itself (\u002Fdenodo:materialize).","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fscheduler\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":18},"01M7C2KQBX7XS4JYW68BBRPXQN","procedures","Use when a Denodo 9.5 stored procedure is in play — calling one the server already ships (SELECT … FROM GET_VIEWS() \u002F USED_BY() … WHERE input_… = …, the CALL statement, finding which predefined procedure answers a question and what parameters it takes), writing your own procedural logic in VQL (CREATE OR REPLACE VQL PROCEDURE … BEGIN … END, local variables, IF\u002FLOOP\u002FCASE, cursors, exceptions, EXECUTE of DDL), or importing a Java one from a JAR (CREATE PROCEDURE … CLASSNAME … JARS). Also for \"run the stats procedure\", \"loop over rows and create views\", \"call that procedure from a view\". Not for introspecting a JDBC source into base views — that is \u002Fdenodo:datasources.","https:\u002F\u002Fgithub.com\u002Fkirivandubai\u002Fdenodo_wipe_coading_skills\u002Fblob\u002Fmain\u002Fskills\u002Fprocedures\u002FSKILL.md",[],[]]