neohaskell/docs/archive/2026-07-ai-artifacts/claude-skills/integration-pipeline-preview/04-security-design

ARCHIVAL KNOWLEDGE — DO NOT USE OR REFERENCE. Superseded by codemap/ + root AGENTS.md. Scheduled for deletion once the pipeline is verified (Phase 6). Manifest: docs/archive/2026-07-ai-artifacts/MANIFEST.md

name: 04-security-design description: Runs the design-level security review for the outbound integration (threat-model → ground → record). kind: process executor: haiku model: claude-haiku-4-5-20251001

Security review (design)

Runs a threat-model pass, then the mandatory grounding pass against the integration classification, then writes the structured findings file. Replaces a single monolithic "security review" with three small steps so the grounding loop cannot be skipped.

Steps

1. Threat model — spawn an Agent (model: opus) and instruct it to read ./01-threat-model/SKILL.md and follow it on the design draft at .integration-pipeline/integration-design.md (path from pipeline.py get design_path). Verify: stdout returns a JSON array of raw findings with severity and rule fields.
2. Ground — spawn an Agent (model: sonnet) and instruct it to read ./02-ground/SKILL.md and follow it on the threat-model output, the classification at .integration-pipeline/classification.json, and the references in ../references/grounding-loop.md. Verify: stdout returns a JSON array where every finding has a grounding_outcome and a severity_after_grounding.
3. Record — read ./03-record/SKILL.md and follow it. Verify: .integration-pipeline/findings-04.json is written and pipeline.py status shows the findings registered for phase 4.

Walk these steps in order. The grounding step MUST run, even on a trivial integration with zero raw findings — the resulting record is the audit trail. The record leaf (step 3) calls pipeline.py complete 4 internally via scripts/record-findings.py; the orchestrator must not call it again.

Shared invariants

• The threat-model step does not filter findings — that is the grounding step's job.
• The grounding step does not introduce new findings — it only re-classifies the ones produced by step 1.
• Findings file path is fixed at .integration-pipeline/findings-04.json so the parent process can verify deterministically.
• Outbound integrations almost always cross a trust boundary; treat HTTP / queue / file-storage egress as security-relevant even when the design says "no secrets" — the threat model still produces findings the grounding loop can demote.