The user picked the advanced path. Two non-negotiable rules at this tier:

1. Oracle AI DB is the only state store. No Redis, no Postgres, no SQLite, no Chroma/FAISS/Qdrant/Pinecone, no JSON or pickle on disk for runtime state. verify.py greps src/ for forbidden imports and fails the build.
2. You compose, you don't write boilerplate. The Oracle layer comes from skills/oracle-aidb-docker-setup, skills/langchain-oracledb-helper, skills/oracle-mcp-server-helper. Your job is to invoke those, then write the application logic (the agent loop, the memory adapters, the tools, the adapter, the notebook). ~500-700 LOC of project code per idea — not 1500.

Step 0 — Read these references first

• skills/README.md — the composition pattern.
• skills/oracle-aidb-docker-setup/SKILL.md
• skills/langchain-oracledb-helper/SKILL.md
• skills/oracle-mcp-server-helper/SKILL.md
• shared/references/oamp.md — load-bearing for the conversational + per-user durable memory layer (ideas 1 + 2).
• All of shared/references/ — yes all of them. The advanced tier can touch every feature.
• advanced/project-ideas.md.

Step 1 — Interview

Run shared/interview.md plus the advanced-only questions.

• Q3 (DB target) — local Docker default.
• Q4 (Inference) — not optional. OCI GenAI for chat (xai.grok-4 via the OpenAI-compat bearer-token endpoint at us-phoenix-1; OCI_GENAI_API_KEY only — no ~/.oci/config needed). In-DB ONNX for embeddings (MY_MINILM_V1, 384 dim) registered via onnx2oracle CLI.
• Q5 (Topic) — one of three from advanced/project-ideas.md. Map free-text pitches; default to idea 1 (NL2SQL + doc-RAG hybrid analyst).
• Q6 (Notebook) — yes, mandatory. Reject "no" — advanced is where notebook payoff lives.
• Q7 (advanced-only) — sql_mode for MCP? — read_only (default; ideas 1 + 2). Idea 3 (conversational schema designer) requires read_write and an explicit confirmation captured in writing — surface in the README as a callout.
• Q8 (advanced-only) — Demo focus? "Polished UI demo" / "Notebook deep-dive" / "Both" — affects how much Open WebUI vs notebook narrative the skill produces.

Print confirmation block. Wait for y.

Step 2 — Resolve choices

Per-idea collections + tools:

Step 3 — Scaffold

3a — Foundation via building-block skills

1. Refuse if target_dir is non-empty.
2. Invoke skills/oracle-aidb-docker-setup. Block until OK.
3. Append the Open WebUI service to the generated compose file (same as beginner / intermediate).
4. Register the in-DB ONNX model via onnx2oracle CLI. Same as intermediate Step 3a-4: pip install onnx2oracle, then onnx2oracle load all-MiniLM-L6-v2 --name MY_MINILM_V1 --dsn "$DB_USER/$DB_PASSWORD@$DB_DSN" --force. Smoke with SELECT VECTOR_EMBEDDING(MY_MINILM_V1 USING 'test' AS data) FROM dual. Required GRANTs (CREATE MINING MODEL, EXECUTE ON SYS.DBMS_VECTOR) are issued by oracle-aidb-docker-setup Step 6.
5. Invoke skills/langchain-oracledb-helper. Pass embedder=in-db-onnx, the per-idea collections, has_chat_history=True (idea 3 only — ideas 1+2 don't need OracleChatHistory; OAMP threads replace it). Block until OK.
6. Invoke skills/oracle-mcp-server-helper. Pass the per-idea sql_mode and allowed_tools. For idea 3, the helper will refuse to proceed silently — capture the explicit user y before invoking.
7. Configure OAMP (ideas 1 + 2 only). Copy shared/snippets/oamp_helpers.py into the project as src/<package_slug>/oamp_helpers.py and import make_oamp_client(conn) from there. Schema is auto-created (schema_policy="create_if_necessary"). Auto-extraction enables iff OCI_GENAI_API_KEY is set; otherwise the helper degrades to manual-add mode and the rest of OAMP still works. See shared/references/oamp.md for the full decision tree and the OAMP/OracleVS/OracleChatHistory/SQL split.

3b — Per-idea memory & app code

Write only the files specific to the chosen idea. Order: migrations → memory adapters → app code → adapter → notebook.

Idea 1 — Hybrid analyst

7. migrations/100_seed_dummy.sql — same fake schema as intermediate idea 1 (10 tables, ~50K rows via Faker).
8. src/<package_slug>/router.py — turn classifier. Two-step: small Grok call ("is this a data question, a knowledge question, or both?") → routes to SQL family / vector family / both. ~80 LOC.
9. src/<package_slug>/ingest.py — walks data/ subdirs (runbooks/, glossary/, decisions/), embeds via in-DB VECTOR_EMBEDDING, inserts into the matching collection.
10. src/<package_slug>/memory.py — thin wrapper around oamp_helpers. Exposes get_thread(user_id, agent_id="hybrid-analyst-v1") which returns an OAMP thread (creates it if missing, otherwise reuses the existing thread for the user). Routes per-user conversation + durable memory through OAMP — replaces the OracleChatHistory wiring entirely. ~40 LOC.
11. src/<package_slug>/agent.py — tool-calling agent. Tools = MCP tools. The router's classification becomes a system-prompt hint that shapes which tool the agent picks first. Pulls thread.get_context_card() from memory.py and prepends it to the user turn before each LLM call.
12. src/<package_slug>/adapter.py — FastAPI /v1/chat/completions, streams agent events. Resolves user_id from the request (Open WebUI surfaces a stable id per browser session); routes through memory.get_thread(user_id).

Idea 2 — Self-improving research agent

7. migrations/100_tool_registry.sql — relational TOOL_REGISTRY (tool_name, signature, last_used, success_count, fail_count).

8. src/<package_slug>/memory/toolbox.py — register_tool, mark_success, mark_fail, recommend_next (SQL queries).

9. src/<package_slug>/memory/log.py — append(tool, args, result, score) writes to TOOL_RUNS (OracleVS) with embedding via VECTOR_EMBEDDING. retrieve_similar(query) uses vector_search.

10. src/<package_slug>/memory/summary.py — OAMP-backed. Thin wrapper around oamp_helpers: get_or_create_thread(user_id, agent_id) returns an OAMP thread; record_turn(thread, role, content) calls add_turn(thread, role, content) from oamp_helpers (one add_messages() per turn — batching breaks extraction, see shared/references/oamp.md V4-OAMP-1); context_card(thread) returns thread.get_context_card() for the planner to splice into the prompt. The hand-rolled SESSION_SUMMARIES collection is GONE — OAMP's rolling thread summary plus retrieved memories is the summary store now. Cold-start retrieval at session boot becomes client.get_thread(saved_thread_id) — recovers the same conversation across processes for free.

11. src/<package_slug>/tools/web_fetch.py — copy shared/snippets/web_fetch_tool.py verbatim. It's a httpx.get + trafilatura.extract BaseTool with a (url, fallback_query) signature so the agent can pivot to a corpus search when a URL 4xx/5xx or times out. The corpus_search_fn constructor arg is wired to memory.log.retrieve_similar (or any retriever you prefer) at scaffold time. Tool calls get logged via memory.log.append.

12. src/<package_slug>/agent.py — planner-executor loop. The skeleton (note: memory.summary calls now go through OAMP):

    thread = memory.summary.get_or_create_thread(user_id, agent_id)
    memory.summary.record_turn(thread, "user", task)
    state = {"task": task, "context": thread.get_context_card(), "step": 0}
    while state["step"] < MAX_STEPS:
        relevant_runs = memory.log.retrieve_similar(state["task"])
        plan = grok.plan(state, relevant_runs)
        if plan.tool == "finish":
            break
        result = execute_tool(plan.tool, plan.args)
        memory.log.append(plan.tool, plan.args, result, score(result))
        memory.toolbox.mark_success_or_fail(plan.tool, ...)
        state = update(state, result)
    memory.summary.record_turn(thread, "assistant", state["final_answer"])
    # OAMP auto-extracts durable facts from the thread on the schedule defined
    # by memory_extraction_frequency; no explicit summary write needed.

    Planner system-prompt rules (load-bearing — friction P1-9):

    • Tool args MUST be a JSON object, never a list. Grok 4 will emit lists by default; correct this in the system prompt.
    • In read_only SQL mode, refuse run_sql calls whose statement starts with anything other than SELECT or WITH ... SELECT — the RunSQLTool already enforces this; the planner should also know.
    • Memory writes are AUTOMATIC at finish. The agent must not emit explicit save_to_memory(...) tool calls — the loop persists summaries and tool_runs for it.
    • finish substantively after 2-3 useful tool calls. Don't wander; MAX_STEPS=12 is a safety net, not a target.

    Planner LLM call must allow enough tokens for final_answer (friction v2-F-v2-1). The agent loop emits structured JSON containing the final_answer text alongside the tool plan. With max_tokens=600 (the OCI default), Grok 4 can truncate mid-JSON at the boundary between the plan envelope and the answer — the JSON is unparseable and the loop crashes. Use max_tokens=1500 minimum on the planner chat_complete call so final_answer (target 1000-1500 chars) fits with envelope overhead.

    Without these prompt rules the agent burns through MAX_STEPS and persists placeholder content into memory — observed in run #4.

13. src/<package_slug>/adapter.py — FastAPI /v1/chat/completions with streaming, exposes "task plan" + "tool calls" + "summary" as separate event types in the SSE stream so Open WebUI can render the agent's reasoning.

Idea 3 — Conversational schema designer

7. migrations/100_design_history.sql — DESIGN_HISTORY (id, ddl, rationale, run_at, success).
8. src/<package_slug>/migrations.py — wraps every DDL the agent emits into a DESIGN_HISTORY row + executes it via run_sql MCP tool. Replay-capable: rerun DESIGN_HISTORY rows in order on a fresh DB.
9. src/<package_slug>/duality.py — JSON Duality view generator. Takes a list of base tables + relationships, emits CREATE OR REPLACE JSON RELATIONAL DUALITY VIEW ... WITH INSERT UPDATE DELETE. Validator checks the generated view loads without error before committing it to history.
10. src/<package_slug>/seeder.py — NL "seed it with 5 fake customers, each with 1-3 commissions" → grok generates INSERT statements → run via MCP. Captured in history.
11. src/<package_slug>/agent.py — conversation loop with confirmation gating: every DDL the agent wants to run is surfaced as a structured event in the SSE stream; the adapter holds the SQL until the UI POSTs back to /confirm/<request_id>. Open WebUI's tool-output rendering is enough — no custom frontend needed.
12. src/<package_slug>/adapter.py — FastAPI; supports the confirm endpoint plus standard /v1/chat/completions.

3c — Verify, notebook, README (all ideas)

13. verify.py — fill template:
    • inference_enabled = True.
    • Round-trip ONNX dim (== 384).
    • Smoke MCP tools list.
    • Forbidden-imports grep: grep -RE 'import (redis|psycopg|psycopg2|sqlite3|chromadb|qdrant_client|pinecone|faiss)' src/. Non-empty output = fail.
    • OAMP cold→warm round-trip (ideas 1 + 2 only). Construct an OAMP client, write a memory (client.add_memory("verify-check fact", user_id="verify", agent_id="verify-bot")), close the connection, reopen a new connection + new client, retrieve via client.search("verify-check", user_id="verify"), assert the fact comes back. This proves the DB-as-only-store invariant for OAMP.
    • Per-idea smoke:
      • 1: ingest the seed schema, ask "what was Q3 revenue?"; assert SQL ran.
      • 2: empty-task run + OAMP thread write + cold→warm thread recovery via get_thread(saved_id).
      • 3: dry-run a CREATE TABLE customers (...) through the agent — assert it lands in DESIGN_HISTORY without executing.
14. notebook.ipynb — mandatory at advanced; clean execution is a Bar B requirement, not optional. The notebook is the demo payoff (the "what does this thing actually do?" artifact for an influencer demo).
    • polished_ui focus → 8 cells, last launches the adapter (and you open WebUI manually).
    • deep_dive focus → 12-15 cells per idea, walks every component (memory, MCP, agent loop, ONNX SQL).
    • both → 12-15 cells, last launches. Cells must execute clean via jupyter nbconvert --to notebook --execute notebook.ipynb. Save the executed copy alongside the source so reviewers can see the outputs without re-running.
15. README.md — the "Why Oracle" paragraph names: in-DB ONNX, vector + relational + JSON Duality + property graph (idea 1 references the agentic_rag-style 6-memory pattern; ideas 2 + 3 use specific subsets), MCP server. Include a "Skills composition" diagram (mermaid) showing the three skills feeding into the project. Include the "DB-as-only-store proof" callout pointing at the verify forbidden-imports grep.

Step 4 — Verify

1. DB up (skill 1 ensured).
2. ONNX registered (step 3a-4).
3. python -m pip install -e .
4. python verify.py. Expect verify: OK (db, vector, inference, mcp, memory, oamp, no_forbidden_imports) (idea 3 reports oamp: skipped — DDL-audit shape, not retrieved memory).
5. Run the notebook clean.
6. Boot adapter, hit /v1/models, kill it.
7. On any failure, follow shared/verify.md recovery loop, max 3 retries.

Resetting memory between dev runs (idea 2 specifically)

Running the agent twice in dev pollutes memory tables — by design. To reset between iterations:

TRUNCATE TABLE TOOL_RUNS;
TRUNCATE TABLE FINDINGS;
DELETE FROM TOOL_REGISTRY;  -- not TRUNCATE — IDENTITY column counter resets weird
COMMIT;

Plus reset OAMP-managed state (the OAMP client owns its own tables):

# scripts/reset_oamp.py
from <package_slug>.oamp_helpers import make_oamp_client
from <package_slug>.store import get_connection
client = make_oamp_client(get_connection())
# Delete every thread + memory for the dev user/agent pair:
for t in client.list_threads(user_id="dev", agent_id="dev-bot"):
    client.delete_thread(t.thread_id)
# `recreate` schema_policy on construction is an alternative — drops and
# rebuilds OAMP tables. Faster than per-row delete; nukes everything.

Add a scripts/reset_memory.py that runs the SQL block + the OAMP block. The notebook's last cell can call it for a clean re-run.

Step 5 — Polish for sharing

1. README placeholders filled.
2. docs/ — placeholders for: demo GIF, screenshot of Open WebUI, mermaid skills-composition diagram.
3. Final report:

   Done.
     project at:    <target_dir>
     features used: in-DB ONNX, OracleVS multi-collection, OAMP (ideas 1+2) | OracleChatHistory (idea 3), oracle-database-mcp-server, <idea-specific>
     skills used:   oracle-aidb-docker-setup, langchain-oracledb-helper, oracle-mcp-server-helper
     run with:      cd <target_dir>
                    docker compose up -d
                    python -m <pkg>.adapter   # blocks; Open WebUI on :3000
     verify:        OK (no forbidden imports)
     notebook:      <target_dir>/notebook.ipynb (executed clean)
     proof:         no Redis/Postgres/SQLite/Chroma/etc — verified by grep.
     next:          record 2-3 min demo, fill "What I built", architecture diagram, push.

Stop conditions

• User declines OCI GenAI (no tenancy / cost concern). Tell them this tier requires it.
• User declines the Oracle-as-only-store constraint. Suggest intermediate path instead.
• ONNX model registers but dim ≠ 384. Drop, surface error, stop.
• Idea 3 picked but user won't confirm read_write MCP. Default to read_only blocks idea 3 — pick a different idea.
• Verify fails 3 times.
• Notebook fails to execute 3 times.

What you must NOT do

• Don't add Redis. Don't add Postgres. Don't add SQLite. Don't add Chroma / FAISS / Qdrant / Pinecone. Don't write to a filesystem JSON file as state. verify.py will catch you.
• Don't make memory ephemeral (in-process dicts). All state in DB.
• Don't roll your own per-user durable memory layer. Use OAMP via oamp_helpers (ideas 1 + 2). The hand-rolled chat history / session summary / context-card plumbing from earlier versions of this tier is gone — OAMP owns it. See shared/references/oamp.md for the decision tree.
• Don't write OracleVS boilerplate yourself — langchain-oracledb-helper did it.
• Don't write oracle-database-mcp-server boilerplate yourself — oracle-mcp-server-helper did it.
• Don't write docker-compose.yml for Oracle yourself — oracle-aidb-docker-setup did it.
• Don't ship without the executed notebook. Mandatory.
• Don't claim done before verify and notebook execution are both green.
• Don't use recursive WITH for bidirectional graphs in idea 1 if you reach for the agentic_rag 6-memory pattern. Use Python BFS over an adjacency table.
• Don't try to register a SentencePiece-tokenized ONNX model. BertTokenizer family only.
• Don't expose run_sql in read_write mode without surfacing the safety warning in the project README.