taiyou-dw

Data warehouse / ETL — ClickHouse + Dagster + Python. Closest of the four to Step 2 thanks to nested agent docs. Cherny harness audit — 2026-08-04.

Step 1

Assisted (current)

~30%

Toward Step 2 (best of 4)

Strong

Agent docs / patterns

No CI

No lint / typecheck

Verdict

Strong nested AGENTS.md and copyable ETL/schema patterns make assisted pairing excellent. Still not Step 2 ready: fragmented local verify, no CI/lint/types, shared Docker volumes, sparse asset checks, no auto-review.

Capability scores (0–5)

Agent context

Copyable patterns

Self-verify

Lint / types

Auto review

Parallel isolation

Eval / quality

Deploy autonomy

Stack snapshot

Layer	Reality	Evidence
Warehouse	ClickHouse single-node	config/docker-compose.yml · clickhouse-server:26.2.3
Orchestration	Dagster · 19 workflows · UI :3366	etl/definitions.py
Python	3.14 + uv	service/etl_agent/ · service/migrations/
Schema	CH 19 migrations · PG 11 (runner TBD)	schema/clickhouse/ · schema/postgres/
ETL shape	raw → dim / fact / snap / agg + ML	etl/workflows/
Ops	EC2-oriented · Alloy→Loki · Caddy	config/ · bin/

Gate checklist

Gate	Status	Evidence
Agent docs	Strong	Root + nested: `AGENTS.md`, `schema/AGENTS.md`, `etl/AGENTS.md`, `etl/workflows/AGENTS.md` + `GUIDELINES.md` / `DEVELOPMENT.md`. No CLAUDE.md / .cursor.
Copyable patterns	Ready	`etl/workflows/AGENTS.md` full templates for assets/jobs/SQL; concrete refs (raw_students, dim_students…).
Self-verification	Weak	`bin/test_etl_agent.sh` + testcontainers; migrations/notif tests; smoke curls in DEVELOPMENT.md — fragmented, not in root AGENTS.md, no single verify entrypoint.
Lint / types / format	Gap	No ruff/mypy/black/pyright in either package `pyproject.toml`.
CI / auto review	Gap	No `.github/` — no Actions, CODEOWNERS, or merge gates.
Parallel isolation	Partial	Docker Compose exists; shared host volumes (`clickhouse-data/`, `dagster-data/`) and fixed ports 8123/3366. No worktree guidance.
ASSET_CHECKS	Sparse	17/19 workflows empty. Real checks: `dim_students`, `notification_prediction`.
Secrets hygiene	OK	`.env` gitignored; `.env.example` tracked.
Migrations	Partial	CH runner via `bin/migrate.sh` + migrations compose. PG migrations files-only (runner TBD).

What makes this the strongest harness

Nested agent guides	Root → schema → etl → workflows — actionable, not boilerplate
Conventions	Naming, migrations, roles encoded in GUIDELINES
Local Docker stack	ClickHouse + Dagster + Alloy + Caddy via `bin/dev_env.sh`
Some real tests	etl_agent (~17), migrations, notification_prediction ML checks
Metadata / lineage	Rich ClickHouse metadata MD alongside migrations

Top 3 actions to unlock Step 2

Unified verify + CI — One documented command (lint + typecheck + bin/test_etl_agent.sh + migrations tests) and a GitHub Actions PR workflow.
Document verify + isolation in root AGENTS.md — How agents self-check; git worktree / compose project-name / port overrides so parallel agents don’t collide on shared volumes.
PR auto-review / same quality bar — Required checks + review bot / CODEOWNERS so agent PRs aren’t mergeable without the human bar.

One-liner

taiyou-dw = Step 1 (Assisted): excellent nested agent docs and copyable ETL/schema patterns; blocked from Step 2 by missing CI/lint/types, fragmented self-verify, and no parallel isolation or auto-review gates.