Polygres Agent Skills 0.4.0

Added

• Added polygres-data-pipeline for guided ingestion and retrieval setup from
  conversations, databases, files, APIs, event streams, and existing Polygres
  data.
• Added read-only device inspection and user choice for local Ollama, Sentence
  Transformers, llama.cpp, and application-owned ONNX embedding options.
• Added a versioned pipeline-plan validator, setup-pack scaffolder, safe .env
  status checker, local embedding adapters, and an idempotent SQLite checkpoint
  ledger.
• Added separate documented store and retrieve interface decisions for CLI,
  SDK, Runtime API, and approved direct Postgres fallback workflows.
• Added a versioned local and hosted embedding model catalog plus deterministic,
  device-aware recommendation tooling.

Changed

• Expanded the plugin from four skills to five.
• Added one-line setup prompts such as Look at my data and use $polygres-data-pipeline to set up a Polygres data pipeline.
• Required generated local setup packs to ignore .env, keep credential values
  out of agent context, and give the user exact local paste instructions.
• Changed pipeline setup to a bounded build-first fast path. The skill creates
  and tests source-specific runtime code before full source inspection,
  backfill, or broad Polygres capability discovery.
• Limited questions to genuine blockers and consolidated upload and remote
  mutation approval into one review.
• Updated CLI and SDK guidance for the single-row Runtime write surface,
  including exact no-retry behavior and ambiguous outcomes.
• Made Context-backed capture one user-facing row operation that also completes
  or durably starts point reconciliation.
• Defined per-surface checkpoint state for pending or partially failed Context
  reconciliation, with exact idempotent replay instead of a manual point step.
• Preserved generic row-only behavior unless a Context collection is explicitly
  selected or resolved safely.
• Moved an unknown local-versus-hosted embedding choice into the single setup
  review, so selecting either fully disclosed option does not require another
  approval.
• Clarified that workflow order, numeric values, retrieval timing, capture,
  graph, embeddings, and agent integration are adaptable guidance rather than
  required pipeline components.
• Made generated setup packs omit explicitly disabled components and allowed
  managed agent instructions to configure capture, recall, or both.
• Made fully vague setup requests ask one short source-and-outcome question
  before inspection, while contextual and detailed requests keep the fast path.
• Added read-only personalized project recommendations for questions such as
  What can I do with Polygres? without starting setup automatically.
• Added a direct reply that carries an accepted recommendation into setup
  without repeating discovery or bypassing the mutation review.