vegastack

When to dispatch this skill

1. Query mentions a specific resource/argument lookup → run vegastack tf "<query>" (single call)
2. Query mentions importing an existing resource → load references/import.md first
3. Query mentions migrating between provider versions → load references/migrations.md first
4. Query spans multiple providers (zero-trust, GitOps, observability) → load references/recipes.md first
5. Query mentions a recent rename / deprecation / "is X still valid?" → load references/recent-changes.md first
6. Query is "what arguments does X accept?" → run vegastack tf directly; envelope's manifest_entry answers it
7. Otherwise → run vegastack tf directly; the response envelope is sufficient

The decision tree above costs ~150 tokens. Loading the matching reference adds ~200-400 more. The historical eager-load pattern was ~4,400 tokens per turn — avoid.

Terraform Providers Kit — v0.1

You have local, deterministic, read-only access to documentation for 31 Terraform providers (1Password, Ansible, Auth0, AWS, Azure, ClickHouse, Cloudflare, CrowdStrike, Datadog, DigitalOcean, External, GCP, GitHub, GitLab, Grafana, Helm, Kubernetes, Local, MongoDB Atlas, Netlify, Okta, PagerDuty, Pinecone, Random, Redis Cloud, Snowflake, Splunk, Time, TLS, Vault, Vercel). The single command is:

vegastack tf "<the user's request, in natural language>"

vegastack (npm: @vegastack/cli) auto-resolves the docs bundle at ~/.config/vegastack/bundle/, auto-detects the provider, and returns one JSON envelope on stdout with four channels read in this order:

1. knowledge[] — date-stamped recent-change cards; if overrides_training: true, trust the card over your training memory.
2. recipes[] — multi-provider topology scaffolds (HCL fragments + pitfalls).
3. files[] — ranked resource/data-source pages, each with manifest_entry (required/optional args, blocks, enum values, import syntax, recommended companions, deprecation flags) and example_usage inline.
4. concept_aliases_used[] — natural-language → resource mapping, surface in your reply for transparency.

The response is enriched: in the common case one tool call answers the task because manifest_entry and example_usage are inlined into files[]. Cite citations[] verbatim at the end of your reply.

One worked example

vegastack tf "S3 backend state locking dynamodb"

Read in order: knowledge[] first surfaces aws-s3-native-state-locking (overrides_training: true) — as of AWS provider 5.55 / Terraform 1.10 the s3 backend supports use_lockfile = true and DynamoDB is optional. Cite the card's authoritative_source. Then files[0] gives the terraform_remote_state / s3 backend manifest. Total: one CLI call, one citation block.

For multi-provider queries ("EKS plus Cloudflare DNS"), the envelope returns recipes[] if a topology matches; otherwise call vegastack tf --provider <name> once per provider.

Guardrails

• Never invent resource names. If a name doesn't appear in files[].manifest_entry, it does not exist. Hallucinations like aws_lb_v2 / aws_load_balancer are common — re-phrase and re-run instead of mutating names.
• Never quote arguments from memory. If an arg isn't in manifest_entry.required_args / optional_args / computed_attrs / blocks.*.{required,optional}_args, it does not exist on this provider version.
• Never fabricate import IDs. Use manifest_entry.import_syntax.command verbatim — composite IDs (zone_id/record_id, region:name) are easy to get wrong.
• Respect deprecation. If manifest_entry.deprecated: true, tell the user before writing code; surface suggested_alternative.
• Honor knowledge-card timestamps over training. When overrides_training: true, the card represents post-training-cutoff facts the maintainers explicitly verified.

References (load on demand per the decision tree)

• references/import.md — import-ID formats, composite-ID gotchas, terraform import workflow
• references/migrations.md — provider-version migration playbook (CF v5, azurerm v4, k8s v2, helm v3, mongodbatlas)
• references/recipes.md — recipe schema + full inventory of 10 cross-provider topologies
• references/recent-changes.md — full inventory of 16 knowledge cards by date_authored
• references/discover-cli.md — full vegastack tf flag reference, output schema, scoring weights
• references/manifest-schema.md — every field in per-provider and root MANIFEST.json
• references/concept-aliases.md — per-provider table of 28 NL → resource aliases
• references/eval-baseline.md — methodology for the published lift numbers
• references/troubleshooting.md — vegastack doctor matrix and common failure modes

This skill bundles all 3 progressive-disclosure phases (load_skill / read_skill_resource / run_skill_script) into one CLI call — the dispatch tree above is the equivalent of the decision Microsoft Agent Framework recommends agents make before invoking load_skill.