Verification Pipeline
The verification pipeline runs lint, type-checking, and tests against a changeset to confirm structural correctness. It is called explicitly via dk_verify — typically by the harness orchestrator during the landing pipeline (submit → verify → review → approve → merge → push).
Verification vs. Review
Verification and code review are complementary quality gates:
Verification (dk_verify) | Review (dk_review) | |
|---|---|---|
| What it checks | Correctness — lint, type-check, tests | Quality — patterns, security, conventions |
| When it runs | Explicitly, when called by the agent or harness | Local review runs automatically on every dk_submit; deep review runs async with BYOK |
| Output | Pass/fail per gate with structured errors | Score (1–5) with findings by severity |
| Purpose | "Does the code compile and pass tests?" | "Is the code well-written and safe?" |
Both must pass before the harness approves a changeset. In the landing pipeline, the orchestrator calls dk_verify first (correctness), then dk_review (quality), and only proceeds to dk_approve if both pass.
How It Works
When an agent or the harness calls dk_verify, dkod runs the project's lint, type-check, and test commands against the changeset in an isolated environment. Results are streamed back as each gate completes — the agent receives structured pass/fail data with error details, file paths, and suggested fixes.
{
"status": "passed",
"gates": [
{ "name": "typecheck", "status": "passed", "duration_ms": 1200 },
{ "name": "lint", "status": "passed", "duration_ms": 400 },
{ "name": "affected_tests", "status": "passed", "duration_ms": 8500 }
]
}If any required gate fails, the agent gets structured failure data and can immediately fix and resubmit — no waiting for CI, no parsing log output.
In the Harness
The harness orchestrator calls dk_verify as part of the landing pipeline:
- Generator agent calls
dk_submit(local review runs automatically, returnsreview_summary) - Orchestrator calls
dk_verify— runs lint, type-check, tests - Orchestrator calls
dk_review— gets full review findings - If both pass, orchestrator calls
dk_approve→dk_merge→dk_push - If either fails, the generator is re-dispatched with the failure details to fix and resubmit
Next Steps
- Code Review Intelligence — code quality review that complements verification
- Agent Protocol — the six protocol operations
- Session Isolation — how agents work in isolated sessions
- Multi-Agent Workflows — orchestrate multiple agents