✨ Set up Copilot instructions

[igor-holt]

Copilot Coding Agent — Repository Instructions (Onboard this repo)
Repo: CVE Matter-Analysis OS (defense-only)
Language/stack: Python 3.11, CUDA (optional), Docker, K8s (GKE/gVisor), Argo, Terraform, GitHub Actions
Mission: Build and maintain the CVE matter-analysis pipeline with Bayesian evidence, positional alignment, reverse-adversarial refractors, and a black-box arbiter.
Non-negotiables: Private repo; defense-only; no cryptographic-breaking, no offensive payloads.

1. Agent Operating Rules
   Act from files, not guesswork. Prefer repo artifacts over improvisation.

Defense-only guardrails. Refuse crypto-breaking or exploit authoring.

Idempotent PRs. Every change must be reproducible via scripts/CI.

Citations & logs. Reference edited files; emit structured logs where applicable.

Never commit secrets. Read keys via env/Secrets. Fail closed if missing.

2. File Anchors (read these first)
   .copilot/AGENT_GUIDE.md — operating guide

.copilot/tasks/ — atomic tasks with acceptance criteria

config/matter.yaml — mode=cve, thresholds, alignment, refractors, arbiter

src/ — modules: ingest/, alignment/, models/, refractors/, evaluation/, orchestrate/

k8s/ — CRD+webhook, runtime classes, GPU job specs, policies

argo/ — WorkflowTemplate, CronWorkflow, Events

terraform/ — GKE cluster, GPU node pool, GCS capsules bucket

prompts/ & capsules/ — LID-LIFT prompt/capsule artifacts (link-pack)

3. Default Task Order (finish each before moving on)
   NVD ingest

Edit: src/ingest/nvd_client.py, tests/test_ingest.py

Criteria: Delta sync, ETag/backoff, JSONL out, tests pass

Positional alignment

Edit: src/alignment/procrustes.py, src/alignment/cca.py, tests/test_alignment.py

Criteria: alignment_R2 ≥ 0.85, re-register on fail

Arbiter (stacked) + Pareto knee

Edit: src/models/arbiter.py

Criteria: choose knee over {accuracy,f1,robust_auc_ε} vs {latency_ms,energy_J,cost_$}

Refractor grid (ε & axes)

Edit: src/refractors/epsilon.py, src/refractors/shifts.py

Criteria: grid runs; outputs robust AUC & CVaR@95

Evidence & decision

Edit: src/evaluation/evidence.py

Criteria: BIC/WAIC log-evidence; Bayes factor + Jeffreys class

Legendary prompt sync (Notion)

Run: scripts/register_prompt_notion.sh

Criteria: SHA in Notion matches prompts/legendary_lidlift_v14.md

Capsule publish

Run: scripts/publish_capsules.sh (on tag)

Criteria: capsules/*.json reachable at configured bucket/domain

GPU enablement (optional)

Files: terraform/gpu.tf, k8s/gpu-job.yaml, docker/Dockerfile (CUDA)

Criteria: GPU pod requests succeed; tests marked Marco Hutter (@gpu) pass

Admission webhook & triggers

Files: k8s/crd-policytrigger.yaml, k8s/deploy-webhook.yaml, k8s/webhook-config.yaml

Criteria: Pods labeled macrosegment: code receive sedation init; suspicious binaries rejected

Argo ε-sweep

Files: argo/workflowtemplate-tensor-macrosegments.yaml, argo/cronworkflow-nightly.yaml

Criteria: nightly sweep executes and stores artifacts

4. CI/CD & Quality Gates (must pass on every PR)
   CI: lint (flake8), tests (pytest), Python 3.11

Security: CodeQL (Python), Trivy image scan (fail on HIGH/CRITICAL)

Branch protection: required status checks, at least 1 review, linear history

No-secret check: block PRs containing keys or tokens

Container: build Docker image; rootless, read-only filesystem, no privilege escalation

5. Environment & Secrets
   Use GitHub Secrets / Environments only:

NVD_API_KEY, GCP_PROJECT_ID, GCP_SA_JSON (if needed), NOTION_API_TOKEN, NOTION_PROMPTS_DB_ID

Local/dev runs must read from env; CI injects via env: or secrets:.

6. K8s/GKE Runtime Rules
   Workload Identity: map KSA→GSA; no node-level keys

RuntimeClass (gVisor) for default pods; GPUs limited to macrosegment: code

AdmissionWebhook applies: inject-sleep-init, reject-suspicious-binary, freeze-image-pull as configured by PolicyTrigger CRDs

7. Argo Orchestration
   Run ε-grid with tensor-macrosegments WorkflowTemplate; CronWorkflow nightly at 03:00 UTC

Label all compute pods macrosegment: code for webhook policies

8. Definition of Done (per task/PR)
   Tests + linters pass; CI security gates green

No secrets in diff; changelog & docs updated

For K8s/Argo changes: kubectl diff or dry-run manifests attached

For Terraform: terraform plan output attached (redacted)

9. Failure Handling (what Copilot must do)
   If a task fails due to missing tools/perm: print exact simulated steps and mark SIMULATED in the PR body.

If prompt/spec is unclear: re-template the failing prompt and propose a single follow-up question as code comments, not as commits.

If latency/energy budgets breach: propose rollback to last stable config and open a separate “perf-budget” issue.

10. One-shot bootstrap prompt (for Copilot PR generator)
    Build/maintain a private, defense-only CVE Matter-Analysis OS: NVD ingest → positional alignment (Procrustes/CCA) → stacked arbiter (Pareto knee) → ε-refractors → Bayesian evidence. Provide tests, Docker (+CUDA optional), K8s (gVisor + AdmissionWebhook + PolicyTrigger CRD), Argo ε-sweep, Terraform (GKE/GPU), CI (lint, pytest, CodeQL, Trivy), SECURITY.md (CVD), .copilot/tasks/*, and config/matter.yaml. No offensive/crypto-breaking. Use env secrets.