Polygres Data Pipeline

Turn setup intent into the smallest complete working result. Adapt to what the user actually needs. Do not force every setup through schema creation, embeddings, graph, backfill, continuous capture, retrieval code, or agent instructions.

Use guidance at the right strength

Treat this skill as a decision guide, not a mandatory architecture or ordered checklist. Adapt, reorder, replace, or omit recommended components and numeric defaults when inspected evidence or the user's outcome supports a better design. Numeric defaults are starting points, including sample sizes, result counts, token budgets, recall timing, capture timing, and model rankings.

Keep only safety and correctness boundaries firm: preserve authorization and provenance, keep secrets out of plans and payloads, disclose and approve material effects, use compatible embedding contracts, avoid unsupported public interfaces and invented graph relationships, and test the important selected path before calling it operational.

Choose the execution surface

When Polygres MCP tools are available, read references/mcp-tool-contract.md, call whoami, and prefer discovered compatible MCP tools for interactive inspection, setup, retrieval, and operation tracking. Treat the connection's project boundary, feature selection, read-only setting, and visible catalog as authoritative.

Use the existing CLI, SDK, Dashboard, or PostgreSQL workflow when MCP is absent, incompatible, or does not expose the required operation. Keep the SDK for persistent application code, the Dashboard for source secret entry and import start, and the CLI for migrations, Runtime keys, and other documented operator workflows.

Recognize setup intent

Activate for broad prompts such as "Help me set up Polygres," contextual prompts such as "look at my conversations and set up Polygres," and detailed technical specifications. Treat examples as illustrations, not required wording.

For a question such as "What can I do with Polygres?", use a personalized recommendation branch. Inspect the accessible current workspace and any uniquely resolved Polygres project with bounded, read-only checks. Look only at the data shape, existing retrieval configuration, and application or agent entry points needed to identify useful opportunities. Then give a short answer that says what was found, leads with the most useful Polygres outcome for this project, and mentions only relevant alternatives. Do not return a generic feature list, create a plan, scaffold files, or mutate anything. End with a direct next step such as:

To proceed, reply: Set up the recommended Polygres pipeline.

Treat that reply or an equivalent acceptance as setup intent. Carry the inspected source, project, outcome, and recommendation into the setup flow without repeating discovery unless the evidence is stale. This acceptance starts setup; it is not mutation approval. Prepare the implementation and show the normal consolidated review before making covered changes.

If the prompt identifies neither a source or inspectable context nor a desired outcome, do not inspect, design, scaffold, or configure yet. Ask one short direction question first:

What would you like Polygres to do, and where is the relevant data? For
example: conversations or agent memory, an existing database, files or an API,
search and retrieval, or connected-data exploration.

Treat a response that identifies a source, an outcome, or both as enough to begin. Infer the remaining reversible details from inspection instead of turning the opening question into a form. Contextual prompts such as "look at my conversations and set up Polygres" and detailed specifications skip this question and proceed immediately.

If the user asks for a generic explanation or design comparison that does not request personalized inspection, use $polygres-retrieval-design. If they want a usable project, pipeline, memory, search, ingestion, synchronization, or agent integration, continue here.

Move immediately once direction is known

1. Parse the prompt and current workspace for source, target project, desired

outcome, ownership boundary, freshness, embedding choice, and existing code.

1. Inspect one bounded source sample and the narrow project capabilities needed

for the likely setup. Do not inventory every Polygres surface.

1. Resolve target.project_mode before selecting any write, import, migration,

database, or retrieval surface. For an existing PostgreSQL source, evaluate managed sync with references/synced-projects.md before designing custom capture.

1. Infer reversible defaults. Ask one concise batch of questions only for

critical facts that inspection cannot resolve safely. Do not ask about optional components that are unnecessary.

1. Keep a small internal setup plan when work has multiple actions, then create

the local source-specific adapter, privacy filter, writer, retrieval entry point, tests, and operator files that the selected design requires.

1. Test locally with a small safe sample appropriate to the source. Present one

consolidated review before the first upload, remote mutation, modification of active agent instructions, or installation of a runtime integration.

1. After approval, apply all actions covered by that exact review without

repeated prompts, verify a bounded end-to-end slice, then continue any approved backfill or integration.

Do not spend setup time explaining Polygres unless the user asks. Do not read an entire source before a privacy filter exists. Do not stop at a plan or generic scaffold when runnable source-specific code can be produced.

Ask only for critical unknowns

Critical means the answer changes safety or makes the implementation invalid:

• no unique target project can be resolved;
• source access or source scope is ambiguous;
• ownership/authorization cannot be inferred;
• semantic retrieval is required but inspection finds no compatible local,

hosted, or existing-vector path that can be fully disclosed in the review;

• a destructive, externally visible, paid, or difficult-to-reverse choice has

no safe default.

Group critical unknowns into one short request. Prefer a reversible default and state it in the review. Use references/guided-interview.md for the initial vague-prompt question or when genuinely blocked; it is not the normal flow for contextual or detailed requests.

Select only useful components

• Reuse a suitable table. Create or alter schema only when needed for stable

IDs, ownership, provenance, content, timestamps, deletion state, metadata, or selected retrieval inputs.

• Start with relational or text retrieval when it satisfies the outcome. Add

pgContext for meaning, similarity, natural-language recall, or agent memory.

• Follow an established embedding deployment preference. If it is unknown,

silently rank one compatible local recommendation and one hosted alternative and put both in the existing consolidated review. Device feasibility does not imply a local preference. Do not create a separate model questionnaire or a second approval after the user selects a fully reviewed option. Polygres does not generate embeddings.

• Recommend pgGraph when validated relationships improve the requested

retrieval. A single memory table does not by itself justify graph, but self-references or reliably derived relationships may. Omit graph when it adds no value.

• Add backfill, checkpointing, deletion propagation, and ongoing capture only

when the source or freshness requirement needs them.

• Add retrieval code when the user needs application or agent recall. Choose

timing, result bounds, token budget, and fallback for that application while preserving provenance and authorization.

• Update agent instructions only when an agent should capture or recall. Scope

the managed block to the relevant repository or agent, and do not claim guaranteed capture without a tested runtime hook.

Read references/schema-and-graph.md for schema and graph decisions and references/context-and-retrieval.md for text, embedding, Context, and recall. For semantic retrieval, follow references/embedding-model-selection.md and use scripts/recommendembeddingmodels.py after bounded inspection.

Generate a working implementation

Read only the source reference that matches the inspected input:

Read the focused MCP playbook that matches the outcome:

Follow references/pipeline-runtime.md. The generated implementation must use stable source identities, filter before persistence or embedding, write idempotently, checkpoint only after durable success, expose exact capture and retrieval commands when selected, and include focused tests.

Use the public interface appropriate to each workload:

• synced project: keep the source database authoritative, hand sync creation

and configuration to the dashboard, and use the Runtime API key only for supported retrieval and retrieval configuration;

• dataset or bounded backfill: reviewed CLI import is normally sufficient;
• one JSON object or runtime event: use the rows surface when the target and

workload pass its read-only validation and deployed limits;

• runtime record capture: public rows API, SDK, or CLI only when installed

client and deployed Runtime compatibility evidence confirm the surface;

• retrieval: documented SDK or Runtime API;
• deletion or unsupported/high-throughput writes: use another documented public

operation when available, otherwise approved direct Postgres;

• direct Postgres: only when no public operation satisfies the approved need.

Never apply the standard-project mutation routes to a synced project. Do not probe rows validation, request target database information, or infer that a project API key can call sync control-plane operations.

The single-row contract is available in CLI/SDK 0.3.0 and includes insert, upsert, ignore, and validate. If the installed client or project does not contain that endpoint version, mark capture upgrade-required, give the exact upgrade requirement, and continue all unaffected setup work. Never infer the endpoint or disguise a bulk import as per-turn capture.

When a bulk import feeds a selected Context collection, reconcile the imported source rows into Context before declaring semantic retrieval operational. The rows API does not delete records; route deletion through an approved source-row deletion path and remove the corresponding Context, text, and graph evidence.

Keep one execution record

For multi-step work, use references/pipeline-plan-contract.md and quietly lint the internal plan with scripts/validatepipelineplan.py. Automatically fix blockers from available evidence; resolve or accept warnings without turning them into user questions. Never make the user read or edit the plan. Use scripts/scaffoldpipeline.py only as a base for selected local files, then add the source-specific runtime. Render the single review with scripts/renderpipeline_review.py.

Follow references/security-and-approvals.md. One approval covers the reviewed setup while project, source scope, data egress, destructive effects, and paid processing remain unchanged. Implementation details and harmless local files do not invalidate it. Credentials are always local environment-variable values; inspect presence with scripts/check_env.py, never their contents.

Connect agent capture and recall when selected

Use scripts/updateagentinstructions.py to add an idempotent managed block to the relevant AGENTS.md or equivalent file. Preserve all user-authored text. Prepare a preview before approval. Modify the active instruction file only after the consolidated review is approved. The block may contain capture, recall, or both according to the selected integration. It must name each selected command, say what is safe to store, describe the selected timing, and state the real guarantee.

Agent instructions are guidance. For guaranteed or retryable capture, also wire and test an application hook, wrapper, outbox, worker, or equivalent runtime. Never store system instructions, retrieved context, credentials, attachments, or tool/environment output unless the user explicitly selected and approved that content.

Verify truthfully

First prove the selected vertical slice: safe normalization, rejection before egress, idempotent write when applicable, ownership filtering, provenance, and one useful retrieval result. Verify graph or Context readiness only when enabled. Then test update, deletion, retry, resume, reconciliation, and agent integration only when those capabilities were selected.

Report operational, partial, or blocked from observed evidence. Name omitted components as intentionally not selected, not missing. Do not claim remote setup, continuous capture, or retrieval works when only local files were generated.