docs: Kimi-K2.6 ADE-Bench behavioral analysis + dbt skill improvements

.opencode/skills/dbt-develop/SKILL.md

@@ -1,6 +1,27 @@
 ---
 name: dbt-develop
-description: Create and modify dbt models — staging, intermediate, marts, incremental, medallion architecture. Use when building new SQL models, extending existing ones, scaffolding YAML configs, or reorganizing project structure. Powered by altimate-dbt.
+description: |
+  REQUIRED before writing or modifying ANY dbt model. Invoke this skill FIRST
+  whenever a task says "create", "build", "add", "modify", "update", "fix", or
+  "refactor" a dbt model, staging file, mart, incremental, or snapshot.
+
+  Skipping this skill is the leading cause of silent-correctness bugs —
+  models that compile and `dbt build` cleanly but produce wrong values. It
+  contains the patterns that prevent the most common such bugs encountered
+  in real dbt projects:
+
+  • Incremental high-water marks (`>=` vs `>` ties → silent row dropout)
+  • Snapshot strategy selection (timestamp vs check, `unique_key` choice)
+  • `LEFT JOIN + COUNT(*)` phantom rows from unmatched parents
+  • Type harmonization in `COALESCE` / `CASE` / `UNION` legs
+  • Date-spine completeness (every period present, even empty ones)
+  • Off-by-one window boundaries (`BETWEEN d - (N-1) AND d` for N-wide)
+  • Uniqueness enforcement when schema implies a key
+  • Window-function `LIMIT` with deterministic tiebreaker
+  • Verifying transformation correctness with dbt unit tests, not just `dbt build`
+  • Enumerating every requested deliverable and checking each exists on disk
+
+  Do not start writing SQL until this skill is loaded. Powered by altimate-dbt.
 ---
 
 # dbt Model Development
@@ -31,6 +52,12 @@ description: Create and modify dbt models — staging, intermediate, marts, incr
 
 Before writing any SQL:
 - Read the task requirements carefully
+- **Enumerate every concrete deliverable the task asks for** — write down each
+  model name, every column/test/config change mentioned, and any "create N
+  models" count. This list becomes the checklist you verify against in
+  step 4. A task asking for four models is not done if only three exist on
+  disk. If the task references a `schema.yml`, `_models.yml`, or similar
+  spec file, every entry there is a deliverable.
 - Identify which layer this model belongs to (staging, intermediate, mart)
 - Check existing models for naming conventions and patterns
 - **Check dependencies:** If `packages.yml` exists, check for `dbt_packages/` or `package-lock.yml`. Only run `dbt deps` if packages are declared but not yet installed.
@@ -98,6 +125,44 @@ altimate-dbt compile --model <name>                        # catch Jinja errors
 altimate-dbt build --model <name>                          # materialize + run tests
 ```
 
+**Verify transformation correctness with unit tests:**
+
+For models with non-trivial transformation logic — aggregations, JOINs, CASE/WHEN,
+window functions, ratio / rate / NPS calculations, COALESCE / NULL coalescing, date
+spines, incremental merge keys — generate and run dbt unit tests before declaring
+the model done. Schema checks ("table exists with the right columns") only verify
+mechanics; value-level correctness needs unit tests.
+
+Invoke the **dbt-unit-tests** skill, which will:
+- Analyze your SQL for the constructs above
+- Build typed mock input rows from the manifest
+- Compute expected outputs by running the SQL against the mocks
+- Write a `unit_tests:` block in the model's `_models.yml`
+
+Then run them:
+```bash
+altimate-dbt test --model <name>     # runs unit tests + schema tests
+```
+
+If a unit test fails, the transformation logic is wrong — **fix the SQL, do not
+weaken the test**. Skip unit tests only for genuinely trivial models: pure renames,
+simple `SELECT *` passthrough, materialization / config-only changes, format-only
+edits.
+
+**Verify every requested deliverable exists:**
+
+Walk the checklist you wrote in the Plan step. For each model the task asked
+for, confirm: (1) the `.sql` file exists in the project, (2) it appears in
+`altimate-dbt info` / the manifest, (3) `altimate-dbt columns --model <name>`
+returns the expected columns, (4) the materialization config matches the
+spec. A task that asked for N models is not complete with N-1 files on disk,
+even if those N-1 build cleanly. Use:
+
+```bash
+ls models/                                                   # confirm every requested file exists
+altimate-dbt info                                            # confirm every requested model is in the project
+```
+
 **Verify the output:**
 ```bash
 altimate-dbt columns --model <name>                        # confirm expected columns exist
@@ -127,6 +192,81 @@ Use `altimate-dbt children` and `altimate-dbt parents` to verify the DAG is inta
 3. **Match existing patterns.** Read 2-3 existing models in the same directory before writing.
 4. **One model, one purpose.** A staging model should not contain business logic. An intermediate model should not be materialized as a table unless it has consumers.
 5. **Fix ALL errors, not just yours.** After creating/modifying models, run a full `dbt build`. If ANY model fails — even pre-existing ones you didn't touch — fix them. Your job is to leave the project in a fully working state.
+6. **Verify transformation correctness, not just mechanics.** For non-trivial models, generate and run dbt unit tests as part of the validate step (use the `dbt-unit-tests` skill). Passing `dbt build` only proves the SQL is syntactically valid — it doesn't prove the *values* are right.
+7. **Enumerate deliverables, then check them off.** The task is not done until every model, column, test, and config change explicitly requested exists on disk and in the manifest. Re-read the prompt at the end and verify each requested item — don't trust your own intermediate "done" feeling.
+
+## Common Pitfalls in Transformation Logic
+
+When the model involves any of the following SQL constructs, watch for these
+generic bugs that mostly compile cleanly but produce wrong values:
+
+### Incremental models and snapshots
+
+- **High-water mark boundary**: in the `{% if is_incremental() %}` filter, use
+  `>=` (not `>`) when the upstream timestamp can repeat or land exactly on the
+  prior max — a strict `>` silently drops every event that ties with the most
+  recent prior load.
+- **`unique_key` choice**: must be the *natural* unique key of the row. Picking
+  a column that is not actually unique (e.g. a foreign-key like `customer_id`
+  instead of `order_id`) causes silent merges and lost rows.
+- **`on_schema_change`**: set `append_new_columns` (or `sync_all_columns` if
+  upstream evolves) so a new source column doesn't NULL-out existing data.
+- **Snapshots — strategy selection**: use `strategy='timestamp'` only when the
+  source has a reliable `updated_at` that monotonically increases on every
+  change. If `updated_at` can be NULL, be reset, or move backwards, switch to
+  `strategy='check'` with an explicit `check_cols` list. Verify by querying
+  the source for `MAX(updated_at)` and looking for repeats or NULLs.
+- **Backfilling**: `--full-refresh` rebuilds incremental tables from scratch.
+  Use it whenever you change the incremental SQL, the merge key, or
+  `on_schema_change`.
+
+### Date and time arithmetic
+
+- **"current age", "days since", "elapsed", "tenure"** — if the column is not
+  pre-computed in the source, compute it. For year-based age, account for
+  month/day so the change happens on the birthday, not on Jan 1:
+  ```sql
+  date_part('year', age(birth_date))                              -- in postgres-family
+  EXTRACT(YEAR FROM CURRENT_DATE) - EXTRACT(YEAR FROM birth_date)
+    - CASE WHEN (EXTRACT(MONTH FROM CURRENT_DATE), EXTRACT(DAY FROM CURRENT_DATE))
+              < (EXTRACT(MONTH FROM birth_date), EXTRACT(DAY FROM birth_date))
+           THEN 1 ELSE 0 END                                       -- portable form
+  ```
+- **Date spines**: when a daily/weekly/monthly model must have a row for
+  every period (even periods with zero events), build a spine first with
+  `dbt_utils.date_spine` or a recursive CTE, then LEFT JOIN the events onto
+  it. Never compute date series by `DISTINCT date_col FROM events` — that
+  silently drops empty periods.
+- **Date boundaries for windowed sums**: rolling-N-day windows expressed as
+  `BETWEEN d - (N-1) AND d` (inclusive both ends) give a width of exactly N.
+  `BETWEEN d - N AND d` gives N+1 — a classic off-by-one.
+
+### Type harmonization in `COALESCE` / `CASE` / `UNION`
+
+`COALESCE(timestamp_col, integer_col)` and `CASE WHEN ... THEN '0' ELSE 0 END`
+fail at compile or coerce silently to whatever type the engine guesses.
+Cast every branch / argument to the same explicit type:
+```sql
+COALESCE(CAST(timestamp_col AS TIMESTAMP), CAST(integer_col AS TIMESTAMP))
+CASE WHEN cond THEN CAST('0' AS NUMERIC) ELSE CAST(0 AS NUMERIC) END
+```
+Same applies to `UNION` / `UNION ALL` — column types must match across legs.
+
+### Uniqueness when the schema implies it
+
+If the model is named `dim_*`, has a `unique` test in `schema.yml`, or the
+task says "one row per X", the model must enforce that grain. Source data
+often has duplicates. Use one of:
+- `SELECT DISTINCT ...`
+- `QUALIFY ROW_NUMBER() OVER (PARTITION BY <key> ORDER BY <tiebreaker>) = 1`
+- `GROUP BY <key>` with explicit aggregation of all other columns
+
+### Window functions with `LIMIT` and ties
+
+`ORDER BY metric DESC LIMIT N` over a column with ties returns a
+non-deterministic set — it may include any N of the tied rows. If the
+business wants a stable top-N, add a deterministic tiebreaker to the
+`ORDER BY` (e.g. an `id` column) so repeated runs return the same rows.
 
 ## Common Mistakes
 
@@ -138,6 +278,7 @@ Use `altimate-dbt children` and `altimate-dbt parents` to verify the DAG is inta
 | Creating a staging model with JOINs | Staging = 1:1 with source. JOINs belong in intermediate or mart |
 | Not checking existing naming conventions | Read existing models in the same directory first |
 | Using `SELECT *` in final models | Explicitly list columns for clarity and contract stability |
+| `COUNT(*)` over a `LEFT JOIN` — counts unmatched parent rows as if they had one child (e.g. a `dim_listings LEFT JOIN fct_reviews` with no matching reviews still yields one row, so `COUNT(*) = 1` instead of `0`) | Use `COUNT(<child_key>)` or `COUNT(CASE WHEN <child_key> IS NOT NULL THEN 1 END)`. If you intended to exclude unmatched parents, switch to `INNER JOIN`. Same trap applies to `SUM`, `AVG`, etc. when the unmatched side contributes a "ghost" `NULL` row |
 
 ## Reference Guides
 

.opencode/skills/dbt-unit-tests/SKILL.md

@@ -32,6 +32,19 @@ description: Generate dbt unit tests automatically for any model. Analyzes SQL l
 3. **Use sql format for ephemeral models.** Dict format fails silently for ephemeral upstreams.
 4. **Never weaken a test to make it pass.** If the test fails, the model logic may be wrong. Investigate before changing expected values.
 5. **Compile before committing.** Always run `altimate-dbt test --model <name>` to verify tests compile and execute.
+6. **Mock data MUST exercise the failure modes of every SQL construct in the model.** A unit test that only covers the happy path validates that the model handles easy inputs — it does not validate correctness. Before writing `given:` rows, list every SQL construct in the model and the boundary case it can mishandle, then ensure at least one mock row triggers each. Universal cases to always cover when the construct appears:
+   - **`LEFT JOIN` / `LEFT OUTER JOIN`** → at least one parent row with **no matching child** (catches `COUNT(*)` phantom rows, `SUM` over `NULL`, fan-out / dropout)
+   - **`INNER JOIN`** → at least one parent row whose child is filtered out by the JOIN condition (catches missing rows)
+   - **`COUNT(*)` / `COUNT(<col>)`** → row where the counted column is `NULL` (catches `COUNT(*)` vs `COUNT(col)` divergence)
+   - **`NULLIF(x, y)`** → row where `x = y` (so the result is `NULL`, exercising downstream `NULL`-handling)
+   - **`/` division** → row where the denominator is `0` or `NULL`
+   - **`CASE WHEN`** → at least one row matching each branch, including the implicit `ELSE NULL` if no explicit `ELSE` is set
+   - **`COALESCE` / `IFNULL`** → row where every argument is `NULL`
+   - **Window functions (`OVER`)** → an empty partition, a partition of size 1, and a row at the partition boundary
+   - **Date arithmetic / date spines** → a row at the start of range, end of range, and a gap day with no events
+   - **Aggregations with `GROUP BY`** → at least one group of size 1 (often masks fan-out bugs) and one group whose key is `NULL`
+   - **Incremental merge keys** → both an "insert" row and an "update" row matching an existing key
+   If you can't think of a failure mode for a construct, you don't yet understand it well enough to test it — read the SQL again before guessing inputs.
 
 ## Core Workflow: Analyze -> Generate -> Refine -> Validate -> Write
 

research/kimi-k26-ade-bench-2026-05-10/README.md

@@ -0,0 +1,11 @@
+# Kimi-K2.6 on ADE-Bench — 2026-05-10
+
+Behavioral analysis of the Moonshot Kimi-K2.6 model running inside altimate-code's agent loop against the ADE-Bench analytics/data-engineering benchmark.
+
+- **Headline:** 61 / 75 = 81.3% pass rate (canonical re-tally across all per-trial directories: 59 / 78 = 75.6%)
+- **Total cost:** $14.91 across ~9.6 hours of wall clock
+- **Source:** [`findings.md`](./findings.md)
+
+Read [`findings.md`](./findings.md) for the full writeup — tool usage distribution, wall-clock anatomy (~89% of time is the model thinking), prompt-cache amplification (85.8% cache hit), per-failure-class taxonomy, and what would be needed to recover the remaining 14–19 failures.
+
+Trace data referenced throughout lives under `experiments/ade-bench-upstream/experiments/2026-05-10__*__none/`. The post is blog-ready; cite or extract sections as needed.