Dvalin security runtime

Dvalin is an independent security verifier for code written by humans and AI agents. The product boundary is not “which model writes the best code”; it is the stable contract between code production and merge:

Agent writes. Dvalin verifies.

Architecture

flowchart LR
  H["Human developer"] --> C["Code change"]
  A["Coding agent"] --> C
  C --> S["Dvalin scan contract"]
  S --> F["Normalized findings"]
  F --> G["Baseline + policy gate"]
  G -->|pass| E["Evidence / CI / merge"]
  G -->|needs work| R["External agent or Dvalin remediation"]
  R --> V["Independent re-scan + check exits"]
  V --> G

The generic DvalinCode Agent Loop remains responsible for a single model turn (RESTORE → COMPACT → COMMAND → BUILD → RUN → SAVE → RESPOND). Security orchestration is a separate persisted state machine. This keeps model control flow generic while making security transitions deterministic and resumable.

Security workflow states

The version 1 workflow supports:

created → scanning → ready/needs_work/passed → external_handoff/internal_fix → verifying → passed/needs_work → evidence_ready → publication_pending → published

Blocked work can resume at scanning or verification. Workflows live under ~/.dvalincode/security/workflows/ (or DVALINCODE_HOME) and contain compact finding snapshots, not source snippets or prompts.

Repository contract

Run:

dvalin init
dvalin baseline
dvalin scan

dvalin.security.json is a versioned repository policy:

{
  "version": 1,
  "scanners": ["builtin"],
  "gate": { "severity": "high", "mode": "new" },
  "baseline": ".dvalin/baseline.json",
  "checks": ["test"],
  "suppressions": []
}

Unconfigured repositories remain compatible: the built-in scanner runs and the gate is non-blocking. Initializing opts the repository into a high-severity, new-findings gate. A suppression needs a fingerprint or rule, a reason, and may also carry a path, owner, and expiry date. Expired suppressions are reported and no longer hide findings.

Finding fingerprints are stable over path spelling normalization. Verification also uses a conservative target fingerprint (scanner + rule + file), so moving the same vulnerability within a file does not make it appear fixed.

Interfaces

• Human CLI: dvalin scan, baseline, verify, doctor, and scanners.
• Compatibility CLI: dvalincode security ...; the earlier dvalincode dvalin ... repair command remains available.
• Agent MCP: dvalin_scan, dvalin_get_finding, dvalin_verify_findings, dvalin_list_scanners, and optional governed coding and evidence tools. Structured tools return MCP structuredContent.
• CI: the GitHub Action produces SARIF; a GitLab example is in docs/examples/gitlab-ci.yml.
• Editors: the VS Code extension and Dvalin workspace remain presentation layers over the same scanner suite.

Verification boundary

Scanner results and actual process exit codes decide the gate. In automated remediation, focused checks must run through run_check; Dvalin reads their exit codes from the tamper-evident audit trail and then independently re-runs the original scanners. The model's DVALIN_VERIFICATION_PASSED marker remains only as a compatibility fallback for API callers that do not provide audit evidence—it is not used by the automated publication path.

dvalin verify <workflow-id> runs the checks selected in dvalin.security.json and reports scan-and-checks assurance. The MCP verification tool reports scan-only because the calling agent owns its check execution; the governed automated-remediation path separately requires recorded run_check evidence before publication.

Scanner installation

Dvalin Built-in always exists. Optional engines are detected from PATH. dvalin scanners install <id> prints a fixed, reviewable command and does not execute it. Adding --yes runs the fixed executable/argument vector under the resolved Dvalin command and egress policy. Scanner discovery never triggers an installation by itself. The Dvalin status UI uses the same fixed plan: its download icon shows a native confirmation containing the exact command, calls a workspace-restricted local endpoint, verifies the executable on PATH, and does not require a configured model.

Honest limits

• A passing scan is not proof that a program is secure.
• The health score is a triage heuristic, not a certification.
• Business-logic flaws may require threat modeling and human review.
• External scanners may access their own registries and advisory databases; availability and errors are included in the result.
• Dvalin can orchestrate a repair, but publication stays explicit and it never auto-merges a change.