---

name: repo-development description: "Guides DataChain repository maintenance, packaging, focused tests, contributor workflows, and backend-sensitive source changes." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Repo Development

Use this sub-skill when the user is working inside a DataChain checkout and asks how to edit, test, package, document, or review the repository itself. It is for maintainer workflows, not ordinary DataChain SDK usage.

Trigger Phrases

Load this sub-skill for prompts containing or implying:

• contributor setup, nox, pytest, lint, docs build, package build, optional extras, Python version support, or repository test layout;
• changes to signal schemas, DataModel, flatten/unflatten, Python↔SQL type conversion, physical storage, warehouse SQL, or nested field exports;
• backend divergence between local SQLite, Studio ClickHouse, and future BigQuery/Snowflake/Postgres support;
• source edits under src/datachain/lib, src/datachain/func, src/datachain/sql, src/datachain/query, src/datachain/cli, or src/datachain/skill;
• dirty/stale checkout checks, generated repo skill refresh decisions, or repository provenance.

First Decision

1. Need setup, tests, nox sessions, optional extras, or focused test choice → read development-and-testing and use select_tests.py for suggestions.
2. Changing schema, nested signal mapping, SQL types, export/read-back, or backend-sensitive behavior → read schema-backend-change-matrix before proposing completion criteria.
3. Need stale-skill or checkout provenance evidence → run snapshot_repo_state.py in the current checkout and compare with the root skill provenance.
4. Debugging maintainer failures → read troubleshooting, then pick the narrowest reproducible test before running broad sessions.

Durable Maintainer Rules

• A signal is a typed column. A nested DataModel flattens into DB columns; the logical SignalSchema travels with every dataset version.
• Chain operations return new chains and must not mutate the receiver in place.
• Logical schema and physical columns are mapped in many paths: ingestion, export, query operations, object hydration, flat tabular export, and each backend. There is no single chokepoint.
• Local SQLite is the convenient default, not proof of backend parity. ClickHouse has stricter nullability and array/map behavior; future backends add more divergence. Verify target behavior on the actual backend when the change depends on backend semantics.
• For signal→column work, done means a literal matrix of affected paths × backend(s) × composition axes with permanent tests and read-back assertions. A green smoke test is not enough.
• Keep comments sparse and timeless. Public APIs need docstrings; internal helpers usually need clear names instead of explanatory comments.

Boundaries and Reroutes

• This sub-skill owns DataChain source maintenance, packaging, nox/pytest, test selection, backend-sensitive review, and stale-skill provenance.
• For user-facing SDK code patterns, UDFs, saves, exports, File/DataModel usage, and LLM operations, read sdk-pipelines.
• For native query operations, function expressions, and schema/backend behavior as an SDK user, read query-engine.
• For datachain CLI command usage or Studio job/pipeline operation, read cli-and-studio.
• For DataChain bundled coding-agent skills or dc-knowledge, read agent-harness.

Safety Rules

• Do not run full nox or broad example/cloud/Studio tests when a focused test can isolate the issue first.
• Treat tests that need cloud credentials, Studio tokens, model downloads, optional GPU/torch/HF/video extras, or shared remote services as optional or credential-gated unless the user explicitly supplies access.
• Do not claim backend parity from a local CPU/SQLite run.
• Do not overwrite generated repo skill directories or live managed skills from maintainer scripts; refresh/import workflows are owned by DisCo meta skills.
• Do not include local environment paths, secrets, or machine-specific checkout paths in public docs or skill content.

Native Verification Candidates Owned Here

• Optional-dependency import gate: base import datachain must not require torch, while datachain.torch should explain the missing torch extra.
• CLI/skill/knowledge tests for changes under src/datachain/cli or src/datachain/skill.
• Signal/schema/type conversion tests for changes under lib/convert, lib/signal_schema, data_storage, sql, and query operations.
• nox -s tests -- <focused pytest path> for a selected subset after narrow pytest commands pass.