Metricflow Ingest Agent Skill

MetricFlow to ktx Semantic Layer

A MetricFlow semantic_model maps to an SL source; MetricFlow measures map to ktx measures; MetricFlow entities map to ktx joins; MetricFlow metrics (top-level) map to ktx measures OR to cross-model derived measures. Files in one WorkUnit are ALWAYS part of the same logical entity (a connected component, possibly spanning extends: + cross-model metric refs). Flatten inheritance and cross-file references at write time.

Mapping table

| MetricFlow | ktx form | Notes |

|---|---|---|

| semantic_model: X { model: ref('t') } with measures + dimensions | Overlay named X with measures, computed-only columns, column_overrides, joins | The model: ref resolves to a manifest table. |

| semantic_model: X { model: source('s','t') } | Overlay named X over table t. | Same shape; source() still resolves to a physical table. |

| semantic_model: X { model: <literal> } with no manifest entry | Standalone with explicit sql:, grain:, columns: | Happens when the dbt manifest isn't available. |

| semantic_model: Y { extends: X } | Merge Y's measures/dimensions/entities into X's overlay, or write a single overlay named for the most-derived child (Y) containing both X's and Y's primitives | Do not emit a second overlay for X - flatten. |

| measures: [{ name, agg, expr }] | measures: [{ name, expr: "<agg>(<expr>)" }] | Aggregation inlined. agg: count_distinct → count(distinct ...). |

| entities: [{ name, type: primary }] | grain: [<entity_name-or-expr>] on the overlay/standalone | Primary/unique entities drive grain. |

| entities: [{ name, type: foreign }] | joins: entry joining to the primary-entity's semantic_model | Only when a matching primary is discoverable. |

| metrics: [{ type: simple, type_params: { measure: X } }] | If the base measure is labeled/described by the metric: in-place edit to the existing measure. Otherwise leave as-is. | Same-name metrics can absorb metadata. |

| metrics: [{ type: simple, filter: <jinja> }] | New measure on the same source, with the filter translated to SQL and attached via filter: | Translate Jinja {{ Dimension('x__y') }} to the column name y. |

| metrics: [{ type: derived, type_params: { expr, metrics } }] | Derived measure on whichever source owns the referenced measures, with expr: referencing measure names | If the metric spans models, still write it once on the source owning the "primary" measure (the one the agent judges most central). Mention the cross-model chain in the description. |

| metrics: [{ type: ratio, type_params: { numerator, denominator } }] | Same as derived; expr: "numerator / NULLIF(denominator, 0)" if no explicit expr | Safe-division by default. |

| metrics: [{ type: cumulative, type_params: { window, grain_to_date } }] | Standalone source with a window-function SQL; reference the resulting column as a normal measure | ktx SL has no first-class cumulative primitive (spec Non-goals). |

| metrics: [{ type: conversion }] | Flag for human - do NOT write. Emit a wiki note describing the intended semantics. | No ktx equivalent in v1. |

| Metric not mappable | Wiki page <metric_name>-definition.md with the full YAML body quoted | Capture the intent even if we can't emit SL. |

Type map: MetricFlow time to ktx time; categorical to string; number to number; boolean to boolean. Follow expr over name when both differ - expr is the physical column.

Verify each MetricFlow model source table with entity_details before producing

the corresponding sl_write_source.

Identifier Verification Protocol

Before writing a wiki page or SL source on any topic:

• discover_data({query: "<topic>"}) - see what wikis, SL sources, and raw

tables already exist. Prefer updating existing pages over creating new ones.

Before emitting any schema.table or schema.table.column into a wiki body,

SL source, tables: frontmatter, sl_refs, or emit_unmapped_fallback:

• entity_details({connectionId, targets: [{display: "<identifier>"}]}) -

confirm the identifier resolves; inspect native types, FK/PK, and

sampleValues.

• For literal values from the source, such as status codes or plan tiers,

check whether they appear in entity_details sampleValues for the relevant

column. If sampleValues is short or the sample may have missed real values,

run a sql_execution probe with the same warehouse connection id:

sql_execution({connectionId, sql: "SELECT DISTINCT <col> FROM <ref> LIMIT 50"}).

• If the candidate identifier still does not resolve, do one of:
• Use sql_execution({connectionId, sql: "SELECT 1 FROM <ref> LIMIT 0"}).

If it errors, the identifier is fictional.

• Wrap the identifier in [unverified - from <rawPath>] in the wiki body,

citing the exact raw path that mentioned it.

• When recording emit_unmapped_fallback with no_physical_table, include

the failing probe error in clarification.

• Never copy <schema>.<table> placeholder strings from these instructions

into output.

Flattening extends:

Within one WorkUnit, multiple semantic_models linked by extends: are guaranteed to be present (the chunker groups them). Resolve inheritance before writing:

• Start with the most-derived child (the one that no other semantic_model extends).
• Walk the extends: chain upward, accumulating measures, dimensions, entities.
• Write ONE overlay/standalone, named for the most-derived child's SL-appropriate name (not the base).
• Parents that lack their own distinctive content should NOT get a separate overlay. If a parent has unique measures a child doesn't inherit, consider whether the base is used elsewhere - if yes, write both; if no, still one overlay.
• Measure/dimension name collisions: child wins, but note the overridden parent in the overlay's description or in a sibling wiki page.

The spec's worked example has orders, orders_ext (extends orders), and metrics/orders_final.yml (defines revenue referencing both). The right output is ONE overlay named orders_ext (or orders if the team's naming favors the base) containing order_count, gross_amount, refund_amount, and a derived revenue measure. Provenance tags point to all three source files.

model: ref resolution

The model: field on a semantic_model is a string like ref('table_name'), source('src','table_name'), or a literal. Resolve:

• ref('x') → table name x. Verify via sl_discover(x).
• source('s','t') → table name t. Verify via sl_discover(t).
• Literal (no ref(...) / source(...)) → treat as the table name directly.

If sl_discover errors because no such table exists, use discover_data and

entity_details to find the warehouse target. If a SQL probe is still needed,

call sql_execution with the same warehouse connection id, for example:

sql_execution({connectionId: "warehouse", sql: "SELECT 1 FROM analytics.orders LIMIT 0"}).

Never invent column names - every column in computed columns:, column_overrides:, grain:, and

sql: must be sourced from raw files, entity_details, or a successful SQL

probe.

After every sl_write_source, call sl_validate. The warehouse will reject invented columns with Unrecognized name: <name> - treat as a hard failure and re-read the schema.

Cumulative metrics - sql-standalone fallback

ktx SL has no first-class window: or grain_to_date: primitive in v1 (spec Non-goals). Translate a MetricFlow cumulative metric to a standalone SL source with a window-function SQL:

# MetricFlow input:
metrics:
  - name: cum_revenue_7d
    type: cumulative
    type_params:
      measure: gross_amount
      window: 7 days

# ktx standalone output:
name: cum_revenue_7d
source_type: sql
sql: |
  SELECT
    ordered_at,
    SUM(amount) OVER (ORDER BY ordered_at RANGE BETWEEN INTERVAL '7' DAY PRECEDING AND CURRENT ROW) AS cum_revenue_7d,
    order_id
  FROM analytics.orders
grain: [order_id]
columns:
  - {name: ordered_at, type: time, role: time}
  - {name: cum_revenue_7d, type: number}
  - {name: order_id, type: string}
measures:
  - {name: cum_revenue_7d, expr: "max(cum_revenue_7d)"}

Pick the time column based on the semantic_model's defaults.agg_time_dimension (e.g. ordered_at). If the MetricFlow config omits it, probe the base table for time-typed columns and choose the most obvious. After writing the standalone SQL source, call emit_unmapped_fallback with rawPath set to the MetricFlow file path, reason: "cumulative_metric_unsupported", and fallback: "sql_standalone".

Conversion metrics - flag for human

metrics:
  - name: signup_to_first_order
    type: conversion
    type_params:
      conversion_type_params:
        entity: customer
        base_measure: signup_count
        conversion_measure: first_order_count
        window: 30 days

Do NOT emit SL for this. Instead:

• Write a wiki page at wiki/global/<metric_name>-intent.md quoting the full YAML body and a one-line explanation of the intended semantics (base event → conversion event within window).
• Call emit_unmapped_fallback with rawPath set to the MetricFlow file path, reason: "conversion_metric_unsupported", and fallback: "flagged".

When ktx SL gains conversion primitives, re-ingesting will find the prior wiki note (via priorProvenance) and replace it with an SL source.

Provenance markers

Every overlay/standalone/wiki page emitted from a MetricFlow source carries HTML-comment provenance tags. When one overlay derives from multiple files (e.g. an extends chain), emit one tag per contributing file:

# <!-- from: raw-sources/conn-1/metricflow/<syncId>/models/orders.yml#L1-20 -->
# <!-- from: raw-sources/conn-1/metricflow/<syncId>/models/orders_ext.yml#L1-12 -->
# <!-- from: raw-sources/conn-1/metricflow/<syncId>/metrics/orders_final.yml#L1-10 -->
name: orders_ext
...

Line ranges (#L<start>-<end>) point to the exact YAML span within the file (the semantic_models: entry for its own name). Use read_raw_span to identify those ranges before writing.

Example 1 - single semantic_model to overlay

# MetricFlow:
semantic_models:
  - name: orders
    model: ref('orders')
    entities:
      - {name: order_id, type: primary}
    measures:
      - {name: order_count, agg: count, expr: order_id}
      - {name: gross_amount, agg: sum, expr: amount}

# ktx overlay at <connId>/orders.yaml:
# <!-- from: raw-sources/.../models/orders.yml#L1-10 -->
name: orders
descriptions:
  user: Order fact table.
measures:
  - {name: order_count, expr: "count(order_id)"}
  - {name: gross_amount, expr: "sum(amount)"}
grain: [order_id]

Example 2 - extends chain → one flattened overlay

# MetricFlow:
# models/orders.yml
semantic_models:
  - name: orders
    model: ref('orders')
    measures:
      - {name: order_count, agg: count, expr: order_id}
      - {name: gross_amount, agg: sum, expr: amount}

# models/orders_ext.yml
semantic_models:
  - name: orders_ext
    model: ref('orders_ext')
    extends: orders
    measures:
      - {name: refund_amount, agg: sum, expr: refund_amt}

# metrics/orders_final.yml
metrics:
  - name: revenue
    type: derived
    type_params:
      expr: gross_amount - refund_amount
      metrics:
        - {name: gross_amount}
        - {name: refund_amount}

# ktx overlay at <connId>/orders_ext.yaml (one file; inheritance flattened):
# <!-- from: raw-sources/.../models/orders.yml#L1-10 -->
# <!-- from: raw-sources/.../models/orders_ext.yml#L1-8 -->
# <!-- from: raw-sources/.../metrics/orders_final.yml#L1-10 -->
name: orders_ext
descriptions:
  user: Extended order fact including refund handling; `revenue` = gross - refund.
measures:
  - {name: order_count, expr: "count(order_id)"}
  - {name: gross_amount, expr: "sum(amount)"}
  - {name: refund_amount, expr: "sum(refund_amt)"}
  - {name: revenue, expr: "gross_amount - refund_amount"}
grain: [order_id]

Example 3 - derived metric spanning two semantic_models

# models/sales.yml
semantic_models:
  - name: sales
    model: ref('sales')
    measures:
      - {name: revenue, agg: sum, expr: revenue_cents}
# models/costs.yml
semantic_models:
  - name: costs
    model: ref('costs')
    measures:
      - {name: cost, agg: sum, expr: cost_cents}
# metrics/margin.yml
metrics:
  - name: margin
    type: derived
    type_params:
      expr: revenue - cost
      metrics: [{name: revenue}, {name: cost}]

Because the WorkUnit bundles all three files (cross-component union via the metric), write the derived measure on ONE of the two sources - pick the source whose domain "owns" the metric (here, sales - margin is inherently a sales metric). Cross-source references aren't native in ktx SL; treat the metric's operands as already-resolvable in the target source's query context OR emit a standalone SQL that joins the two tables:

# <connId>/sales.yaml
# <!-- from: .../models/sales.yml#L1-8 -->
# <!-- from: .../models/costs.yml#L1-8 -->
# <!-- from: .../metrics/margin.yml#L1-8 -->
name: sales
measures:
  - {name: revenue, expr: "sum(revenue_cents)"}

# <connId>/margin.yaml - standalone because it spans two tables
# <!-- from: .../models/sales.yml#L1-8 -->
# <!-- from: .../models/costs.yml#L1-8 -->
# <!-- from: .../metrics/margin.yml#L1-8 -->
name: margin
source_type: sql
sql: |
  SELECT s.period_id, s.revenue_cents, COALESCE(c.cost_cents, 0) AS cost_cents
  FROM analytics.sales s
  LEFT JOIN analytics.costs c ON c.period_id = s.period_id
grain: [period_id]
columns:
  - {name: period_id, type: string}
  - {name: revenue_cents, type: number}
  - {name: cost_cents, type: number}
measures:
  - {name: revenue, expr: "sum(revenue_cents)"}
  - {name: cost, expr: "sum(cost_cents)"}
  - {name: margin, expr: "sum(revenue_cents) - sum(cost_cents)"}

Also write a wiki page at wiki/global/margin-metric.md explaining the cross-source origin.

Example 4 - filtered metric creates a new measure

metrics:
  - name: paid_order_count
    type: simple
    type_params:
      measure: order_count
    filter: "{{ Dimension('orders__status') }} = 'paid'"

# <connId>/orders.yaml
measures:
  - {name: order_count, expr: "count(order_id)"}
  - {name: paid_order_count, expr: "count(order_id)", filter: "status = 'paid'"}

Translate {{ Dimension('orders__status') }} to the bare column name status (the table alias prefix is implicit within the SL source's scope).