pod-storage

name	pod-storage
description	Resolve durable Pod writes through shared descriptors, Consensus, pod_schema, and pod_storage instead of inventing paths or Turtle.
allowed-tools	Read, Write, Edit, Bash, Grep, Glob, pod_read, pod_write

Pod Storage

Use this skill when the user asks AI Secretary or another agent to save credentials, configuration, grants, preferences, or other durable Pod-backed data.

Core Rule

Do not invent Pod paths, RDF predicates, subject templates, or Turtle.

Durable storage semantics belong to @undefineds.co/models:

• descriptors,
• resource base paths,
• subject templates,
• field predicates,
• validation,
• repository-backed commits.

Workflow

1. Restate the storage request in structured fields when possible.
2. Use the injected udfs tool for schema lookup and semantic consensus. LinX/linx-lite should provide runtime context for remote Consensus; without that context, udfs can still run local deterministic schema commands.
3. For direct schema lookup, call udfs schema classes, udfs schema predicates, or udfs schema describe.
4. For ambiguous natural-language storage requests, call udfs consensus with a JSON input payload. The future service-backed path should use the Responses protocol (/v1/conversations + /v1/responses).
5. Ask the user any returned clarification questions.
6. Continue the same Responses conversation through the conversation_id request field.
7. Build a domain-shaped DTO from the returned schemaUri and field mapping.
8. Optionally call udfs storage validate --input <json> before committing.
9. Persist through the shared models ORM/repository, not through hand-written Turtle or a resource-specific udfs shortcut.
10. Report the schema URI, resource summary, and follow-up status.

Current MVP Check

The local verification path is the injected udfs tool. Do not call back into linx; udfs must also work when injected into non-LinX coding agents:

udfs consensus --input '{"session_id":"sess_123","request":"我要保存这个 Cloudflare token","answers":{"token_type":"tunnel-token"}}' --json
udfs schema classes --uri 'https://vocab.xpod.dev/credential#Credential'
udfs schema predicates --uri 'https://vocab.xpod.dev/credential#Credential'
udfs schema predicates --uri 'https://vocab.xpod.dev/credential#Credential' --field apiKey

Inside LinX or another coding-agent shell, AI should use the injected udfs tool:

udfs consensus --input '{"session_id":"sess_123","request":"我要保存这个 Cloudflare token","answers":{"token_type":"tunnel-token"}}' --json
udfs schema describe 'https://vocab.xpod.dev/credential#Credential'
udfs schema predicates --uri 'https://vocab.xpod.dev/credential#Credential'
udfs storage validate --input '{"schemaUri":"https://vocab.xpod.dev/credential#Credential","operation":"upsert","match":{"service":"infra","providerId":"cloudflare","secretType":"tunnel-token"},"set":{"label":"Cloudflare Tunnel Token","status":"active"}}'

Expected outcome:

• clarification question: API token or tunnel token,
• schema URI: https://vocab.xpod.dev/credential#Credential,
• predicate mapping includes apiKey -> https://vocab.xpod.dev/credential#apiKey,
• descriptor storage: /settings/credentials.ttl#{id},
• validation plan: /settings/credentials.ttl#infra-cloudflare-tunnel-token,
• no commit in udfs; actual writes go through shared ORM/repository code.
• remote Consensus, when available, is reached through injected UDFS_CONSENSUS_BASE_URL and UDFS_CONSENSUS_TOKEN; do not ask the user for a command-line API key.

Secretary Rules

• Never store secrets in free-form notes.
• Never log full secret values.
• Use validation before risky writes, but persist through model-owned repositories/ORM.
• Treat unmodeled durable data as a model proposal flow, not a permanent catch-all note.
• User/developer-created descriptors must be validated before activation.

Low-Level Pod Tools

pod_read and pod_write are protocol tools, not modeling tools. Use them only when:

• no shared business model exists yet,
• the operation is explicitly protocol-level,
• or you are inspecting raw Pod state for debugging.

If a shared model exists, use model-owned helpers instead of hand-editing Turtle.