myriade_audit_list_map
ChatGPTThe organisation's audit value map in one read, as its page shows it: each domain with its document and headline, then its active zones — starred first, then by gain — with verdict, gain, data, confidence, last verification, document, one-line summary, next step and lever count per class (quick_win, quick_action, waiting, project, long_term). Opens with the domain keys zones use without a domain document and the zone or domain documents whose block could not be projected (with the error to fix); a map too long for one read names the domains (or, zoomed on one, the zones) it leaves out. Read zones and levers in full with myriade_audit_read_zones / myriade_audit_read_levers, a zone's narrative with myriade_read_report (the uuid inside <DOCUMENT:uuid>).
myriade_audit_read_levers
ChatGPTLevers as their page shows them: every field, the class, the human follow-up (status, assignee, note) and the dated figures behind each. An id that is no lever on the map comes back under errors. A query is shown only on connections you may read.
myriade_audit_read_zones
ChatGPTZones as their page shows them: the record from the ``zone block, the dated figures recorded on the zone, its levers in short (id, name, stage, class, follow-up, estimate) and its document. A slug with no active zone comes back under errors. A figure's query is shown only on connections you may read (restricted` otherwise).
myriade_audit_record_evidence
ChatGPTRecord a dated figure backing a zone (or one of its levers), linked to the query that computed it, on a connection you may read. Recording the same label again on the same zone or lever replaces the figure.
myriade_audit_record_lever
ChatGPTCreate a lever in a zone of the audit map, or update one by its id (an argument you omit keeps its value). Cite the returned id as <LEVER:id> in the zone document. Its class is computed from effort × gain_delay × timing; the follow-up (status, assignee, note) is set on its page. A lever is an action to take on the business. What happened about an alert the brief showed belongs on that alert (myriade_triage_alert), not here.
myriade_audit_write_overview
ChatGPTWrite the audit's overview page, whole: the business synthesis (title, key figures, trend, economic model, main constraint, risks), then the levers worth reading now, above the map the page draws itself. Written for a manager: no tool, table or team jargon; ranges like "0,8–1,6 M€/an"; every link reads « Examiner » or « Diagnostiquer ». A key figure or a lever line cites the lever by its uuid: the page reads its figure, dots and certainty from it, so record the lever's gain_low / gain_high first. A refused write lists every problem to fix.
myriade_brief
ChatGPTThe brief — the page this person lands on in Myriade — and the workspace behind it, across every connection they may open, with no database id. Returns `brief (the page as last written: text and sections, or null), freshness (written_at, age_hours, stale — true when the page would rewrite it on the person's next open, i.e. it predates their day — and status: ready, or revalidating / writing while a run is rewriting it; read again in a minute), tasks (its open to-do list) and workspace: standing alerts, proposed checks, the monitoring topics they follow with their latest reading, recent reports, open catalog questions and suggestions, recent conversations, unread activity — each entry with cite_as (the id the other tools take: <ALERT:id> feeds myriade_triage_alert) and actionable when this person may act on it, every section capped with its real total`. Say when the brief is stale and offer myriade_refresh_brief; earlier briefs are in myriade_brief_history.
myriade_brief_history
ChatGPTThe person's earlier briefs: every rewrite keeps a dated publication. Without `brief_id, list them (newest first, excerpt, paged by cursor), narrowed by query and dates — "what did my brief say on Monday about X". With brief_id`, read one in full: its text, sections and recommendations with their current outcome (an action after a brief is not proof the brief caused it). Governed by the person's current access: a publication they could no longer read in full is left out.
myriade_create_metric
ChatGPTGive a business concept its calculation: the term named name carries it, so the term becomes a governed metric. Use when the user wants a number to become a reusable, validated metric. A term and a metric on the same concept are one object: when no term has this name, a DRAFT term carrying the calculation is created; a draft term without calculation receives it in place; any other term without calculation gets it as a suggestion for the data team. Catalog editors are notified; publishing is their call — nothing is org-visible until then. After it succeeds, include <METRIC:{id}> in your answer so the user sees the metric card. When a semantic model covers the data, base the metric on it: set base_model_id to the model's id instead of writing base_sql — the model's SQL becomes the rows, and the metric follows when the model changes. Write base_sql only over a clean table no model covers. A per-row value the model already computes (e.g. nps_points) is aggregated as is, never recomputed in the measure. The definition has two halves. base_sql is a SELECT you write freely — joins of any depth, CTEs, CASE, window functions — producing the ROWS the metric aggregates: one row per countable event/entity, with every column the envelope needs (measure inputs, the time column, dimension columns — join in business labels, not ids, when a label table exists). The envelope (measure, filters, time_dimension, grain, dimensions) references OUTPUT COLUMNS of that SELECT and is what powers grain switching and one-click breakdowns. Row shaping goes in base_sql, aggregation in the measure — NEVER pre-aggregate in base_sql (a GROUP BY there freezes the metric and makes re-slicing wrong: avg/ratio/count_distinct cannot be re-aggregated). LEGIBILITY RULE — the panel's business reading (formula + filter chips) is rendered from the ENVELOPE, so what you push into base_sql disappears from it. Keep the metric's defining arithmetic in measure.expr (expr: "unit_price quantity (1 - discount)", not a base_sql alias like line_revenue — alias only what cannot be a fragment, e.g. a windowed column) and business exclusions in filters (a WHERE in base_sql is for structural conditions). In the description, do NOT restate tables, formula or grain — the panel already shows them from the definition, which cannot drift; say what the metric measures, for whom, and the caveats. STILL OUT OF REACH: percentiles and medians as the measure (the agg list is fixed), a metric built from another metric, and period-over-period as the measure (the trend chart shows evolution; the metric itself is one number per bucket). A window function can never BE the measure, but is fine inside base_sql. When the user needs one of those, do NOT bend the definition and do NOT phrase the intent in the description as if it were computed — name the limit you hit, then offer the way out: answer with a saved SQL query, or have them define it in dbt where it gets mirrored into this catalog.
myriade_create_report
ChatGPTCreate a report owned by the caller, private until shared. Share it with myriade_set_report_access(visibility="everyone"): an audit zone or domain document joins the map only once organisation-wide.
myriade_define_model
ChatGPTDefine a semantic model as a DRAFT: one business entity (customer, order, return…) — the fields the business needs and the SQL that reaches them — that you and other agents read before writing queries. One model per business entity, even over a clean table: naming the entity, choosing its fields and mapping them is the point. Never a model for a technical or staging table. Start from the need: declare the fields the business wants even when no data has them (they show as missing); give each field its role and its source — the column it comes from ("public.customers.customer_id") or where it lives when the SQL cannot reach it ("Salesforce Account.Segment"). The SQL, when given, is one read-only SELECT over the connection's tables with business rules as columns (e.g. "o.canal = 'web' AS is_web"); it is run once with no row to check it and read its columns. Each fact lives in one place: grain, rules and pitfalls in the description; a per-row value several metrics aggregate (e.g. nps_points) in the model; an aggregation formula only in a metric, never here. Flag pre-aggregated or ratio columns in the description ("do not sum", "do not average"). A catalog editor publishes it. After it succeeds, link it in your answer as [name](/semantic-layer/definitions/{id}).
myriade_edit_report
ChatGPTReplace text in a report, as the caller (a version is recorded). Read it first with myriade_read_report. A published report is editable only by its owner or an org admin.
myriade_get_asset_activities
ChatGPTGet the full activity feed for a catalog object (asset or glossary term), including content and changes. Assets and terms share one id space — any catalog id works. Use this when you need detailed history beyond the summary a read returns.
myriade_get_check
ChatGPTGet a monitoring check (name, SQL, severity, cadence, status) and its most recent run results, newest first. On governed databases the SQL and evidence rows are redacted unless you manage the database or authored the proposal.
myriade_get_database_info
ChatGPTGet detailed information about a specific database connection. Returns information about the database including its name, engine type, the SQL dialect myriade_query takes, description, write mode, and any AI memory notes.
myriade_get_query_results
ChatGPTRe-execute a saved Myriade query and return fresh results. The query's SQL always re-runs under the caller's governance policies (table permissions, column masking) — results are never served from Myriade's cached snapshot. Prefer this over myriade_query for recurring metrics (dashboards, cockpits): the SQL stays governed in Myriade instead of travelling through prompts. Write statements are blocked. Returns up to 50 rows with column metadata, same shape as myriade_query, and next_offset while rows remain. Each page re-runs the query, so rows can move between pages.
myriade_guide
ChatGPTHow Myriade works and how to use these tools: the method for querying a database and documenting it in the catalog as you go, then — for an organisation that uses them — the monitoring, brief and audit objects with their French names, which tool reads or writes each, where a follow-up on a matter is recorded, and what is a gesture on the page rather than a tool. Read it once before acting on someone's data.
myriade_list_databases
ChatGPTList databases configured in Myriade (paginated, sorted by name). Returns paginated database connections with their IDs, names, engines, and descriptions. Use the database ID with other tools; the alias is how myriade_search_workspace and myriade_read name the database.
myriade_list_reports
ChatGPTList or search reports (documents) written in Myriade. Without `query returns the most recently updated reports first; with query` returns the best matches first. Only reports the caller may see are returned. Read one with myriade_read_report.
myriade_mark_no_doc_needed
ChatGPTMark assets as not needing documentation, in bulk (status='not_needed'). Use for temp/backup tables, dated copies, dead schemas and anything that doesn't deserve documentation. They leave the completion denominator, and any certification they carried is dropped — nothing vouches for what nobody documents. A SCHEMA id cascades to all its child tables. Reversible via myriade_update_asset with another status.
myriade_model_map
ChatGPTThe map of the semantic models: each business entity with its domain, status, how far its data reaches the need, freshness, usage, metrics and links to other models. Use it to see what the semantic layer covers before building on it, or to place a new model. Read one model for its SQL, fields and lineage.
myriade_monitoring_overview
ChatGPTSnapshot of ONE database's monitoring board: check count, standing alerts (open / acknowledged / known) with their headlines, recent notifications and triage notes, followed by the open alerts as JSON — each with its investigation and evidence while they fit, then as listing lines, with `shown / total / detailed` saying which. To find an alert without knowing its database, or to see the whole workspace, start with myriade_brief instead. Alert ids feed myriade_triage_alert; check ids feed myriade_get_check. Investigation details are redacted unless you manage the database.
myriade_note_alert
ChatGPTWrite a durable note on an alert (requires manage access on its connection), signed by you: what whoever reads this connection's alerts should know — "pilot paused on purpose until December", "the customer was asked to update on 14/10". Shown on the alert's timeline and re-read before any future notification on that connection. The alert resolves its database by itself. Its state and headline change with myriade_triage_alert, not here.
myriade_note_topic
ChatGPTWrite a note on a monitoring topic, signed by you: a durable fact about the whole subject — a decision taken, a tolerated situation, a planned change ("Dickson relancé le 14/10 pour passer en 2.327 avant le 28/10, Julian en appui"). Shown on the topic's page, where the person can remove it, and re-read at every reading of the topic. For what was done about ONE alert, prefer myriade_triage_alert (its state and headline) or myriade_note_alert (a note on that alert).
myriade_post_message
ChatGPTReply on a catalog object's activity feed (asset or glossary term). This is a reply channel, not a notebook. Post only when a person is waiting for it on that feed: answering a comment that mentioned you, or the closing report of a documentation pass. Nothing posted here is tracked or chased. Do NOT post: notes for a future reader, observations, corrections to one of your own questions, or anything that needs an answer. Durable knowledge goes in the description (myriade_update_asset / myriade_request_catalog_change); an open point goes through ask_question; a wrong question is fixed with myriade_update_question or myriade_withdraw_question. Feed etiquette: - Keep it short — it lands in a timeline - Mention a user as <USER:id|Display Name> — both halves. A feed entry gives you both: actor_id and actor. Without the name the reader sees the raw id in the sentence.
myriade_prepare_database_connection
ChatGPTPrepare a new database connection for the person to finish on Myriade. Takes only what is not secret (engine, host, port, user, database name…) and returns a link to the connection form filled in with it; the person types the password or key there. Nothing is saved until they do.
myriade_propose_check
ChatGPTPropose a monitoring check for the user to review in Myriade. This creates a PROPOSAL, never an active check: nothing runs until a database owner or org admin accepts it in the app. The SQL is test-run once under your governance context to validate it. Contract for `check_sql: it MUST return rows ONLY when something is wrong (zero rows = healthy) and use relative time (now() - interval`), never absolute dates. Minimum cadence is 5 minutes. Names are unique per database — read myriade_monitoring_overview before proposing a near-duplicate.
myriade_query
ChatGPTExecute a read-only SQL query against a database. Write the SQL in the target database's dialect — the dialect field returned by myriade_get_database_info. Write operations (INSERT, UPDATE, DELETE, DROP, etc.) are blocked. Returns up to 50 rows of results with column metadata. Successful queries are saved in Myriade; the response's query_id can be replayed later with myriade_get_query_results without resending SQL.
myriade_read
ChatGPTOpen up to five workspace objects at once — the tables, terms or reports a question needs, in ONE call. A path is what myriade_search_workspace prints: a table as typed in SQL ("finance.fact_revenue_daily", a column as "finance.fact_revenue_daily.amount"), "terms/<name>", "metrics/<name>", "reports/<id>", "checks/<id>", "alerts/<id>", "cockpits/<id>", "reviews/<id>" (a monitoring review, as a <REVIEW:id> attachment names it), "topics/<id>" (a monitoring topic: charter, checks, latest reading, notes); a catalog UUID works too. A table comes back with its columns (name, type, description), sample rows and documentation; a report with its content. Long content is paged: the footer gives the offset of the next window. A path that matches nothing costs only its own line.
myriade_read_report
ChatGPTRead a report's full markdown content. Applies the same visibility rule as the web app; a report the caller may not see returns a "not found" error. Edit it with myriade_edit_report.
myriade_read_skill
ChatGPTRead one of Myriade's skills: the method Myriade's agent follows in a domain (documenting the catalog, the audit map, forecasting…), or one this organisation or person wrote. myriade_guide lists them. Read the skill before acting in its domain.
myriade_record_model_check
ChatGPTRecord your diagnosis of a semantic model's data, after investigating it with your own myriade_query calls. Investigate first, as an analyst would, on the model's SQL and its source tables: how many rows; how filled each field is overall and on recent rows (a field turning empty lately is a broken source); whether it is up to date, on the date key that fits the entity (created_at, updated_at…) and on each source table when the model joins several; whether the key is unique and joins find their rows; whether values look right; whether the data can be used analytically (size, indexes, an operational database under load). Then record one verdict per field you checked and one for the model. This replaces the previous check; it is shown to every reader of the model, with the queries you name as evidence.
myriade_refresh_brief
ChatGPTHave Myriade's agent rewrite this person's brief now, as opening the page does on the first open of their day. Use it when myriade_brief says the page is stale, or when the person asks for a fresh one. The run takes about a minute in the background and keeps a dated publication of the previous page. Returns `status (writing for a first brief, revalidating when a page already exists) and already_running when a run holds the guard; then read myriade_brief until its freshness.status is ready. subscription_required` means the organisation has no subscription for Myriade's agents.
myriade_request_catalog_change
ChatGPTChange a catalog description, tags or a term's definition — written directly where you may, otherwise filed for the data team to review. The server decides (your edit rights, the object's status): a draft or a table you may document is written, validated content becomes a reviewable suggestion. The response says which; when a suggestion was filed, include <SUGGESTION:{id}> in your answer so the user sees it. A term that carries a calculation (a metric) takes its wording here too, in words; its calculation changes through myriade_update_metric, never here. dbt-managed assets (managed_by_dbt when read) are refused: their description/tags live in the dbt project — offer the user a pull request on it instead of a suggestion.
myriade_resolve_question
ChatGPTClose a question once its answer is reflected in the documentation. Only call this after folding the answer into the description or definition, once that write has come back: it can be refused. Resolving a question whose answer you have not written down loses the answer. If the reply does not settle the question, leave it open and ask a follow-up. If the question itself turned out wrong, use myriade_withdraw_question instead.
myriade_search_workspace
ChatGPTSearch the governed workspace (catalog, reports, and enabled monitoring/topic/cockpit families) for one or MORE focused concepts in a single call — or browse it by type/parent instead. Pass every concept the question needs at once — short queries of 1-2 words each. Each printed path is the exact `myriade_read argument that opens that hit. Unshown relevant hits become counters: narrow the query, pass families, or add database rather than paginating. Browsing replaces the queries: pass asset_type and/or parent with queries=[]` to list the objects of one type, or the children of one object (the tables of a schema, the columns of a table). A query alongside them filters that listing by name.
myriade_semantic_work
ChatGPTWhat to focus on in the semantic layer (terms, metrics, models): what waits on this person, what is left to build, what they asked for. Use it when the person asks what to work on, what waits on them, or where their request stands. Lists names, steps and pages, no calculation: read an entry before working on it, and judge from a term's definition whether it deserves a calculation at all. Deciding stays a person's act in the app: publishing, certifying, accepting a proposal. Give the entry's page for those.
myriade_set_report_access
ChatGPTChange who can read a report: visibility (its owner only) and web publication (its owner or an org admin).
myriade_triage_alert
ChatGPTTriage an alert (requires manage access: database owner or org admin). The alert resolves its database by itself. This is where what was done about a matter is recorded — a mail sent, a decision taken, who is on it: the `statement becomes the alert's headline, which the brief and the topic's reading show. Nothing else needs writing. - resolve: the underlying problem is handled or gone. - acknowledge: keep it standing with a statement of what is happening — only for a situation that may worsen. - dismiss: false positive / expected; refused while the check still trips (it would just respawn). - snooze: silence for hours with a statement; reopens by itself if the check still trips afterwards. - reopen: send a parked alert back to "needs action" with a statement` of what is no longer covered.
myriade_update_asset
ChatGPTUpdate catalog asset documentation. Descriptions are in Markdown. Read myriade_read_skill('catalog-documentation') first for the rules on confidence, draft vs published, and knowledge management. Say what should change: it is written directly when you may, and filed as a suggestion for the data team otherwise — the response says which.
myriade_update_metric
ChatGPTChange the calculation and/or the definition of a term that carries a calculation (a metric) — draft OR published. A draft you may edit is edited in place (include <METRIC:{id}> in your answer). Anything else is never touched directly: the change is filed as one reviewable suggestion for the data team — this is THE way to propose an evolution of a published metric's calculation. The server decides; the response says which. Never describe a formula change in the description text; pass the new definition object instead. Changing the calculation drops the term's certification. The definition is a FULL REPLACEMENT: always read the metric's current definition immediately before calling this and build on it — a copy from earlier in the conversation may be stale and would silently revert changes reviewers approved in between. Same contract as myriade_create_metric: base_sql shapes the rows (free SELECT — joins, CTEs, windows), the envelope aggregates and slices them via base_sql's OUTPUT columns; never pre-aggregate in base_sql. Same LEGIBILITY RULE too: defining arithmetic in measure.expr, business exclusions in filters (both render in the panel; base_sql does not), description carries meaning, not tables/formula/grain. STILL OUT OF REACH: percentiles/medians as the measure, a metric built from another metric, period-over-period as the measure. When the user asks for one of those, do NOT bend the definition and do NOT reword the description as if the change had been computed — name the limit, then offer a saved SQL query or dbt as the way out.
myriade_update_model
ChatGPTChange a semantic model's SQL, fields, description, domain or declared links — draft or published. A draft you may edit changes in place; anything else is filed as a suggestion for the data team (the response says which). Omitted arguments keep their value; fields, depends_on and relations, when given, replace the whole list. A changed SQL is run once with no row to check it (and a sketch gains its first SQL this way). Changing a model the response names as used by other models means their SQL may repeat the old logic: read them and update them too when it applies.
myriade_update_question
ChatGPTRewrite one of your open questions when its wording or a premise was wrong. Use it instead of posting a correction next to the question: the question keeps its id, its place in the feed and its reviewers, and the reader sees the current text with the earlier wording kept underneath. Only open questions can change.
myriade_update_report_title
ChatGPTRename a report (an audit zone takes its document's title).
myriade_withdraw_question
ChatGPTRetract one of your open questions that should not have been asked. For a question built on a wrong premise, or one the data answered after all. Nothing is folded into the catalog — that is what myriade_resolve_question is for. The question stays visible in the feed with the reason, so nobody wonders where it went.
myriade_ask
Claudemyriade_audit_list_map
Claudemyriade_audit_read_levers
Claudemyriade_audit_read_zones
Claudemyriade_audit_record_evidence
Claudemyriade_audit_record_lever
Claudemyriade_audit_write_overview
Claudemyriade_brief
Claudemyriade_brief_history
Claudemyriade_create_report
Claudemyriade_edit_report
Claudemyriade_get_answer
Claudemyriade_get_check
Claudemyriade_get_database_info
Claudemyriade_get_query_results
Claudemyriade_get_table_schema
Claudemyriade_guide
Claudemyriade_list_databases
Claudemyriade_list_reports
Claudemyriade_list_tables
Claudemyriade_monitoring_overview
Claudemyriade_note_alert
Claudemyriade_note_topic
Claudemyriade_prepare_database_connection
Claudemyriade_propose_check
Claudemyriade_query
Claudemyriade_read
Claudemyriade_read_asset
Claudemyriade_read_report
Claudemyriade_refresh_brief
Claudemyriade_request_catalog_change
Claudemyriade_sample_data
Claudemyriade_search_catalog
Claudemyriade_search_workspace
Claudemyriade_set_report_access
Claudemyriade_triage_alert
Claudemyriade_update_asset
Claudemyriade_update_report_title
Claude