LANES.md

OVK Check Types

The five check types OVK ships today and their input contracts.

Default path: ovk check --changed-files <diff> infers and evaluates all affected checks from a unified diff. Focused CLI commands below remain for debugging and manifest-driven workflows.

Overview

Check type	Intent ID	Focused CLI	Via ovk check
Self-protection	agent-cannot-disable-own-ci-gate	ovk ci	ovk check
Authorization	no-admin-route-bypass	ovk auth-obligation	ovk check
Infrastructure	no-public-sensitive-resource	ovk infra-exposure	ovk check
CI secrets	no-secrets-in-untrusted-context	ovk ci-secrets	ovk check
Deployment state	no-skipped-approval-state	ovk deployment-state	ovk check

Multi-check manifest: ovk verify --manifest <manifest.json>

Self-protection

Checks that AI-authored changes do not weaken verification controls governing merge.

ovk ci \
  --metadata examples/no_agent_self_approval/metadata_gate_preserved.json \
  --changed-files examples/no_agent_self_approval/changed_files_workflow.txt \
  --advisory

Condition	Recommendation
Required check removed	block
Missing required-check metadata	require_human_review
Controls preserved	allow

Backend: deterministic evaluator (default), optional OPA via --backend-strategy.

Authorization

Checks that non-admin principals cannot reach admin-only routes.

ovk auth-obligation examples/auth_regression/input_admin_bypass.json --advisory

Input schema: schemas/authorization.input.schema.json

Required field: non-empty routes list. Each route has path, admin_only_before, admin_only_after, and reachable_after.

Solver result	Recommendation
Violation found	block
No violation	allow
Solver unavailable	require_human_review

Optional Z3 backend: pip install -e '.[solvers]'

Infrastructure exposure

Checks that sensitive resources are not publicly exposed.

ovk infra-exposure examples/infrastructure_exposure/input_public_sensitive_resource.json --advisory

Input schema: schemas/infrastructure.input.schema.json

Input formats

ovk infra-exposure accepts:

Format	Flag
Native OVK infra	--input-format infra (default)
Terraform-plan subset	--input-format terraform
Kubernetes Service subset	--input-format kubernetes
Graph-style	--input-format graph

Normalizer: ovk.adapters.infra.normalize

Policy

Optional policy file: schemas/infrastructure.policy.schema.json

{
  "blocked_public_sensitivities": ["confidential", "restricted"]
}

Pass via --policy <path>.

Condition	Recommendation
Sensitive resource publicly exposed	block
Resource remains private	allow
Invalid abstraction	require_human_review

CI secrets

Checks that secrets are not referenced in untrusted workflow contexts.

ovk ci-secrets examples/ci_secrets/input_secrets_exposed.json --advisory
ovk extract-workflow .github/workflows/deploy.yml

Input schema: schemas/ci_secrets.input.schema.json

Workflow YAML extraction handles PyYAML 1.1 on: → True quirk. Unified diffs reconstruct workflow content via ovk plan / ovk infer.

Condition	Recommendation
Secrets on untrusted trigger (pull_request, etc.)	block
pull_request_target with PR head checkout + secrets	block
Safe context	allow

Deployment approval state

Checks that deployment state machines do not skip required approval states.

ovk deployment-state examples/deployment_state/input_skipped_approval.json --advisory

Input schema: schemas/deployment_state.input.schema.json

Uses graph reachability over deployment states. Skipped required approval → block.

Conservative rule

Invalid input must not produce allow. Unknown and error states require human review for every check type.

Examples: examples/ per check type. Benchmark scorer: benchmarks/formal_pr_bench/score_all_lanes.py (internal script name).

Older wrapper scripts

Deprecated scripts/run_*.py runners mirror the focused CLI commands above (ovk ci, ovk auth-obligation, ovk infra-exposure, ovk ci-secrets, ovk deployment-state) with identical flags. Prefer ovk for new integrations; see MIGRATION.md.