Maintain

Keep the repository correct as the world underneath it moves, on both of its axes: the dbt project and the semantic layer. Maintenance is the recurring half of the loop: warehouses drift, loads half-fail, models go stale, keys stop being unique, and business definitions change. This skill compares a known-good baseline against current reality, classifies what drifted, and proposes the reconciling edit. It is manual and on-demand here; continuous drift detection and automated PRs are the commercial product.

The model: baseline, detect, reconcile

Drift is measured against a baseline (the .dex/snapshot.json fingerprint of the warehouse map and the repository's per-layer definitions). Detection is read-only; only reconcile proposes edits.

The two project layers are fingerprinted independently. The transform layer comes from the dbt project; the semantic layer comes from whichever vendor semantic.vendor names, which may be dbt's own or a native format such as Apache Ossie. A repository with a semantic layer and no dbt project at all still gets a baseline and still runs every free axis: transform_layer comes back null, and the warning that names why no project was fingerprinted is reserved for the case where neither layer answered, since that is the one you could otherwise mistake for a clean read.

Snapshot discipline matters. A snapshot is only as trustworthy as the moment it froze. Take one right after a known-good build (maintain snapshot), and commit .dex/snapshot.json like a lockfile so the whole team diffs against the same reference. Snapshot a state that is already drifted and check will mask the very drift you care about. When you accept a change as the new normal (re-run explore map first, then maintain snapshot); check warns when the baseline looks stale.

On a warehouse past the rank cutoff, use explore map --full before snapshotting. Past 50 objects explore map profiles the top 25 by rank and enters the rest as metadata alone, and the baseline can only compare columns for objects it has columns for. Snapshotting a partial map is still valid, and the envelope reports column_detail_count against dataset_count plus a warning naming what it could not cover, so the gap is visible rather than silently mistaken for a clean bill.

How to drive it

uv run --no-project --script "${CLAUDE_SKILL_DIR}/scripts/run.py" <subcommand> [flags]

dex runs its engine through uv, which is a prerequisite and is not installed by Claude Code. If the shell reports uv: command not found, stop and tell the user to install it (curl -LsSf https://astral.sh/uv/install.sh | sh, or brew install uv, or pipx install uv), then re-run. Never fall back to diffing the warehouse against the project by hand instead: the drift axes and the baseline comparison live in the engine, so any other path is guesswork.

The first command in a fresh environment installs the engine, so it can take tens of seconds where later ones take well under a second. --warm pays that install up front and exits without running anything:

uv run --no-project --script "${CLAUDE_SKILL_DIR}/scripts/run.py" --warm

Offer it once at setup. It is not something to run before an ordinary command.

• maintain snapshot captures or refreshes the baseline. Run it after a clean explore or transform session so later runs have a known-good reference. It pins the current .dex/cache.json (so the grain baseline is the exact-distinct verdicts explore map already computed) plus per-layer fingerprints of the dbt project and of the semantic layer. A native semantic layer contributes its definitions per dataset and per metric, each with a content hash, the relation behind it, the column each field resolves to, its declared keys in the arity they were written, and its relationships with every ordered column pair; whether that side was captured is itself recorded, so a baseline written before it reports the relationship axis as unchecked rather than clean. Without a cache it captures a metadata-only baseline and says so. It also warns when the cache it pinned is thin (objects without column detail) or older than the profile freshness window, because either makes an "accept current state" only partly true.

• maintain snapshot --project-only is for a project-only refactor, such as moved model files or a dbt project rename. It refreshes transform and semantic fingerprints without opening the warehouse, carrying the previous warehouse evidence and original capture time forward instead. It requires an existing snapshot and refuses connection-target flags.

• maintain check is the everyday entry point: it sweeps every axis and returns a report ranked by blast radius. Read-only.

• maintain schema [<objects>] detects structural drift: source columns and tables added, dropped, retyped, or renamed; nullability changes; declared sources the warehouse no longer honors.

• maintain volume [<objects>] detects freshness drift: row counts that collapsed, spiked, or went to zero. This is the "is the data still flowing correctly?" axis, distinct from "did the shape change?".

• maintain grain [<objects>] detects grain drift: a key that now has duplicates, a changed row-per-entity cardinality, or an increased join fanout. It also re-verifies the grains the repository declares, which measurement on its own can miss: a dbt model-level unique_combination_of_columns, and a semantic layer's own key declarations. A multi-column declaration is measured as one complete composite and never one column at a time. Uses aggregates, never raw rows.

  A native semantic layer's keys reach this axis and nothing else reaches it for them, since such a layer is never the transformation project. They go through the identical billed handshake on a metered warehouse: nothing here is cheaper or less gated because the declaration came from a document rather than from dbt.

  Two findings come out of the uniqueness checks and the difference is the baseline. key_lost_uniqueness is a key that was proven unique and is not any more: something changed in the data. declared_grain_not_unique is a declared combination that does not hold, and nothing changed at all: the project asserts a grain the data never had, so the fix is to the declaration (widen it, dedup upstream, or drop the claim) rather than to the data.

• maintain semantic [<objects>] detects definition drift: definitions that changed, were added, or were removed against the baseline; a source relation that is gone; a dimension, entity, measure, or declared key naming a column that is gone; a relationship whose endpoint or column pairs no longer resolve, which is high because a join nothing can resolve is a broken layer rather than a stale one; and categorical dimensions whose set of values widened or narrowed underneath their metrics.

  Read unavailable on the layer before hunting for an element kind. A native Ossie layer has no measures and no entities at all, so their absence is the format rather than drift. Its cardinality half also never fires, because that check needs a semantic model naming a transformation model and Ossie names none: on such a layer this command is free and offers no scan.

• maintain verify [<selector>] answers a different question from every command above it: not "what changed since the baseline" but "is this project right now", and it needs no baseline at all, so it works on a project that was never correct and on one somebody else just built. Two classes of finding. Build status: nodes that failed, nodes skipped because a parent failed (naming the one that actually failed), and models the project declares that built no relation. Row population: row_loss where a model holds materially fewer rows than its driving parent (the relation in its FROM clause, followed through the CTE chain, as distinct from anything it joins) and nothing in its SQL accounts for the shortfall, and row_fanout where it holds materially more, each naming the join and its key and stating both counts.

  Row population is conservative on purpose. A model with a WHERE, GROUP BY, DISTINCT, QUALIFY, LIMIT, a semi or anti join, or a set operation was written to hold a different number of rows and is never reported for loss; an incremental model is skipped outright. So a quiet answer here is weaker evidence than a finding, and the warnings say which models could not be lined up at all.

  A project that does not compile is reported first and suppresses everything else, since a manifest a broken project could not have produced is not evidence. Read data.suppressed before reading an empty data.findings as a clean bill of health.

• maintain reconcile [<class>] proposes the dbt edits that bring the project back in sync, as reviewable diffs. Optionally scope it to one class (schema, volume, grain, or semantic). It composes every layer's declarations first, so a grain the semantic layer already declares is not proposed as though nothing declared it. Where there is no editable dbt project it has nothing to author: every proposal is advisory and no plan is stored. Authoring into a native semantic layer is semantic ossie in the transform skill, never this command.

The usual flow: check to triage, a focused detector to understand one axis in depth, then reconcile to get the proposed fix. With no baseline, or on a project whose numbers were never right, start at verify instead: it is the one command here that does not need a snapshot, and it answers "is this right" rather than "what moved".

Per-axis cost: what is free and what scans

Detection is read-only, but read-only is not the same as free on a metered connector (BigQuery, Snowflake, Databricks, Postgres, Redshift, ClickHouse). The axes split:

• Schema, volume, and the reference/definition half of semantic are free everywhere: they read metadata and the snapshot, and run immediately.
• Grain and the dimension-cardinality half of semantic scan the warehouse, so on a metered connector they run the two-step handshake. Asked for directly, maintain grain returns needs_confirmation with an estimate in cost.estimate (and a per-table breakdown). Surface it to the user in human units, get an explicit budget, and re-issue the same command with --confirm --budget <magnitude> in the paradigm's unit (bytes on BigQuery, warehouse-seconds on Snowflake and Databricks, compute-seconds on Redshift, database-seconds on Postgres and ClickHouse). Never invent a budget the user did not agree to, and never retry with a raised budget on an over-ceiling refusal without asking. An over-ceiling refusal carries a calibration line from .dex/spend.jsonl (what this connector's recent commands billed as a fraction of estimate, or a sentence saying there is too little history to say): relay it, and note that the ceiling binds on the estimate, so a budget set at that fraction of the estimate is refused again.
• verify is free except for the counts a warehouse does not keep. Its build-status findings read artifacts on disk, and its row counts come from object metadata. A view has no stored row count anywhere, and a view is dbt's default materialization, so on a metered connector those counts are batched into one aggregate-only statement, priced, and returned in data.offer beside findings that are already final. On DuckDB there is no gate, so every count is measured rather than estimated and the findings come back exact.
• check, semantic and verify answer first and offer second. Their free axes complete on every call, so the envelope is ok and the findings in it are final. The price of the scanning axes sits in data.offer, with axes naming what it would add; data.axes_run names what already ran. Confirming is a choice, not a required next step: quote the estimate, say which axes are still dark, and let the user decide. A triage pass that stops at the free axes is a complete piece of work, not an abandoned one.
• Read warnings on these responses, always. They carry the reasons the baseline may no longer describe the warehouse (a cache newer than the snapshot, a baseline pinned from a stale cache), which bound every finding above them. A stale baseline is often the most important line in the response and it is never in findings.

On DuckDB everything is free and local, so nothing prompts.

A needs_confirmation envelope carrying suggested_session_ceiling is the project's one-time ask for a cumulative daily cap, separate from the per-command --budget. Surface it, get the user's answer, and add --session-ceiling <value> or --no-session-ceiling to the same re-issue; it is written to .dex/config.yml once and never asked again. Never answer it for them.

Reconcile proposals are mechanical or advisory

Reconcile tags every proposal by kind, because the fix differs sharply by axis:

• mechanical: schema drift reconciles in one of two shapes. On a dex-scaffolded staging model it re-scaffolds the model from the drifted source; on a project format that places a declaration but authors no staging model, it edits the drifted columns into that declaration and says so. High-confidence, but still a reviewable diff: read it for hand-written logic the scaffold cannot know about.
• advisory: grain, volume, and semantic drift are decisions, not auto-fixes (dex cannot dedup your warehouse or decide whether a new 'refunded' status belongs in a metric). The proposal is the decision surfaced, at most backed by a test edit that makes the break visible in builds. It declines that test where the test would be wrong: if your model declares a composite grain covering the column, no column-level unique is proposed on it, and the warning names the combination so you can tell "re-baseline, this is still the grain" from "something relied on that column alone".

A type change is advisory on every format. Nothing dex writes declares a type, and the type it holds is the connector's own spelling rather than a canonical one (Snowflake reports NUMBER(38,0) and NUMBER(10,2) both as FIXED), so the proposal names both spellings and the edit is yours. One consequence to know: on ClickHouse nullability is part of the type, so a column that starts accepting nulls is reported as a retype and gets advice where other connectors get an edit.

When reconcile produces edits it stores them as a plan and prints a plan_id. Apply them with transform apply <plan-id> (the one apply door): a human edit made since detection surfaces as a conflict, never a silent overwrite.

Guardrails (enforced in the engine, not here)

• Read-only against data. Schema, volume, and semantic references are computed from metadata and the snapshot; grain and dimension-cardinality use aggregates only. Raw rows and dimension values never cross the envelope.
• Propose, don't impose. Reconciliation is always a reviewable diff, applied through transform apply. Human edits to the project and to the semantic layer are authoritative; on conflict the engine surfaces the divergence and asks rather than overwriting.
• The repository is the source of truth, on both axes; the .dex/ snapshot is a non-canonical fingerprint used only to detect change.