taiyou-dw — before and current

Compare the 2026-08-04 baseline with the expected state after feat/ai-adoption-finalize merges.

Step 1 → 2

Assisted → early Parallel

30% → 70%

Progress toward Step 2

2.1 → 4.0

Average capability score

+1.9

Average gain across 7 areas

Adjusted scope

This comparison excludes automated PR review, required CI pass rules, and branch-protection rules from both scoring and remaining gaps. The assessment assumes all pending adoption work merges soon.

Verdict

Before · 2026-08-04

Step 1 — Assisted. Agent guidance and copyable patterns were strong. Verification was fragmented. The repository had no lint, type checks, CI, or safe parallel stack setup.

Current · expected after merge

Early Step 2 — Parallel-capable. Agents now have one fast gate, focused test suites, broad seed+verify coverage, and isolated local environments. Production autonomy remains weak.

Capability score comparison (0–5)

Capability	Before	Current	Change	Why the score changed
Agent context	4	5	+1	`CLAUDE.md`, seven skills, and focused agent docs complete the strong nested guides.
Copyable patterns	5	5	—	Schema and workflow templates remain excellent. Skills now package common procedures.
Self-verification	2	4	+2	`bin/check.sh` provides one fast gate. `bin/test/*.sh` provides focused suites.
Lint / types / format	0	4	+4	Root `pyproject.toml` configures ruff and ty. CI runs the same checks.
Parallel isolation	1	4	+3	Worktree guidance, port offsets, project names, and local compose changes prevent collisions.
Eval / quality	2	4	+2	24 of 25 non-smoke workflows have seed+verify tests. Schema migration suites add broad coverage.
Operational / deploy support	1	2	+1	Local and operational scripts improved. Production deployment remains mostly manual.
Average	2.1	4.0	+1.9	Most repository-level Step 2 mechanics are now present.

Major evidence: before and current

Area	2026-08-04 baseline	Expected current state
Agent entry	Strong root and nested `AGENTS.md` guides.	`CLAUDE.md` links agents to the guides and the seven skills under `.agents/skills/`.
Verification	Tests and smoke commands were fragmented.	`bin/check.sh` is the fast gate. Focused suites live under `bin/test/`.
Static analysis	No configured lint or type check.	Root `pyproject.toml` configures ruff and ty.
CI execution	No `.github/` workflow existed.	`.github/workflows/ci.yml` runs the repository checks. Merge enforcement is outside this assessment.
Parallel work	Fixed ports and shared stack names caused collisions.	`.agents/docs/worktrees.md`, `infra/local/docker-compose.yml`, and `bin/dev/env.sh` support isolated worktrees.
Workflow evaluation	Only a small number of workflows had meaningful tests.	24 of 25 non-smoke workflows have seed+verify tests.
Test infrastructure	Useful tests existed without one shared warehouse harness.	`testkit/warehouse_testkit/` supports workflow and migration suites for ClickHouse and Postgres.
Operations	Production work required manual EC2 steps.	Local and operational tooling improved. Production deployment still requires manual action.

What produced the gain

Clear agent procedures 4 → 5

CLAUDE.md connects compatible agents to repository guidance. Seven skills encode repeatable warehouse tasks.

Unified verification 2 → 4

One fast command covers static checks. Focused scripts run the matching Docker and Testcontainers suites.

Safe parallel work 1 → 4

Worktree instructions and configurable local ports let agents work in separate checkouts without stack collisions.

Broad evaluation 2 → 4

Seed+verify cases cover nearly every workflow. Shared test infrastructure also verifies ClickHouse and Postgres migrations.

Remaining gaps in the adjusted assessment

Production deployment remains manual. The deploy workflow is a placeholder. Production changes still need human action.
Postgres has no production migration CLI. The test suite applies migrations, but production has no equivalent runner.
Runtime asset checks remain sparse. ASSET_CHECKS is non-empty in only 2 of 26 workflows.
Type coverage has localized exceptions. A few justified ignores reduce coverage in specific areas.

One-liner

taiyou-dw moves from Step 1 at ~30% toward Step 2 to early Step 2 at ~70%, with strong parallel mechanics and broad self-verification.