Agent-assisted development¶
This guide turns repository standards into a repeatable multi-agent workflow.
AGENTS.md files contain concise invariants; this page contains the detailed
coordination protocol.
For governance, security, and repository-administration controls, use:
- Agent governance for owners, control lifecycle, and required reviews;
- Agent task contracts for machine-checkable writer boundaries in parallel worktrees;
- Agent risk register for known AI-agent failure modes and mitigation status;
- Agent security mapping for OWASP, NIST, SLSA, and SSDF control mapping;
- GitHub branch protection for the repository ruleset checklist.
Workflow¶
1. Classify¶
Classify the task as one or more of:
- public feature/contract;
- internal implementation;
- manifest/schema;
- connector/route/strategy;
- state/identity/evidence;
- CLI/Python UX;
- Airflow/dbt/GitOps integration;
- docs only;
- CI/release/security.
2. Read-only fan-out¶
For a non-trivial task, run independent analysis in parallel:
dpone_explorer: actual code paths, tests, docs, and shared files;dpone_architect: boundaries, contracts, state, and alternatives;dpone_test_certifier: risk-based test and evidence matrix;dpone_docs_ux_reviewer: first-time-user journey and docs impact.
The parent/integrator reconciles disagreement. Do not let four agents each implement their own architecture.
3. Specify and approve¶
For material changes use docs/agent-templates/feature-design-spec.md and the
design-dpone-feature skill. Implementation starts only at APPROVED.
4. Partition writes¶
Create one worktree per writer. Copy
docs/agent-templates/agent-task-contract.yml into the task artifact directory and
fill:
owned_paths;read_only_paths;forbidden_paths;- acceptance criteria;
- required checks;
- public-contract impact;
- shared-file owner.
Writers do not expand ownership themselves. They return a conflict/request to the integrator.
Validate a concrete task contract before handing it to a writer:
5. Integrate and review¶
The integrator owns shared schemas, registries, dependency files, workflows, MkDocs navigation, changelog, and common fixtures. After integration, a fresh-context reviewer examines correctness, compatibility, silent-data-risk, tests, docs, and evidence.
6. Validate¶
Generate the recommended plan:
Run focused checks first, then the required broad gate. Record every check as
PASS, FAIL, SKIP, N/A, or UNVERIFIED.
Change-to-context map¶
| Change | Read first | Additional proof |
|---|---|---|
| CLI | docs/cli-reference.md, docs/ux.md |
exit/output/file/negative matrix |
| Manifest/schema | docs/manifests-variant-c.md, docs/schema-evolution.md |
old manifests, schema/examples, migration |
| Runtime/state | docs/architecture.md, docs/state.md |
retry, replay, crash boundary, evidence order |
| Nested data | docs/nested-normalization.md, docs/schema-identity.md |
deterministic lineage and no orphans |
| Connector/strategy | docs/connector-sdk.md, docs/source-sink-matrix.md |
capability matrix and route evidence |
| Airflow | docs/airflow-pack-provider.md |
import/parse/serialization/runtime compatibility |
| dbt | docs/dbt.md |
project discovery, invocation, artifacts, compatibility |
| Docs | docs/documentation-standard.md |
strict build, examples, rendered CJM |
| Release | docs/release.md, docs/agent-release-protocol.md |
frozen-commit R1-R9 report |
Shared-file policy¶
Default integrator-owned paths:
pyproject.toml
uv.lock
CHANGELOG.md
mkdocs.yml
.github/workflows/**
src/dpone/schema/**
shared registries and factories
tests/conftest.py
release and certification indexes
A task may override this list explicitly, but two simultaneous writers must not own the same semantic contract.
Agent evaluation¶
Repository agent changes are product changes. Evaluate them using representative tasks and track:
- first-pass CI success;
- accepted-without-rework rate;
- unrelated changed files;
- correct test selection;
- public-contract violations;
- escaped defects;
- merge-conflict rate;
- tokens and wall time per accepted change.
Use evals/agent/result.schema.json for structured run results. Do not optimize
only for generated code volume or a green CI result.