Classify and Route Persistent Storage

You receive one or more pieces of content that might be worth persisting (a business fact, a validated SQL, a metric or model definition, a session preference, a project convention, a reusable workflow). For each piece you must:

1. Classify it into exactly one of the seven persistent stores — or decide it should NOT be persisted.
2. Route it the prescribed way for that store (delegate to a task subagent / another skill / a memory tool, or edit AGENTS.md directly).

You are a router and curator, not a generator. The goal is that every artifact lands in the one store that fits its shape, lifetime, and reuse pattern — and is written with the mechanism that owns that store.

Role & Boundary

• semantic_models / metrics MUST go through task(type="semantic_modeling", ...); reference_sql MUST go through task(type="gen_sql_summary", ...). Never hand-write their YAML files or vector-store rows yourself — the subagents own schema, validation, and indexing.
• knowledge goes through extract-knowledge in lite mode — summarize the atomic fact directly against the source; do NOT trigger the deep blind-SQL-iteration flow.
• memory goes through add_memory / edit_memory — the only writers; they enforce the 2000-byte cap.
• skills goes through create-skill.
• AGENTS.md is the only file you edit yourself (write_file / edit_file).
• When content does not clearly belong to any store, or routing would overwrite something destructively, call ask_user. If ask_user is unavailable, report the gap and do not write by default.
• This skill only routes — it does not own persistence mechanics. Whether a store is a flat file, a vector index, or both, and the durability of that write, is the responsibility of the target subagent/skill (e.g. gen_sql_summary must persist and index its reference_sql). Do not add file-vs-index verification logic here.

Context Handoff to Subagents

Every task() subagent and extract-knowledge run in an isolated context. They do NOT inherit the current session/conversation, the project goal, the active datasource, the inventory you built, or the facts you just gathered. A bare table name or a SQL string with no surrounding context yields a generic, poorly-indexed artifact (wrong dimensions, empty search_text, a summary that ignores a mandatory filter). So whoever dispatches the task MUST inline everything the generator needs. At minimum pass:

• Datasource name (and dialect) the artifact targets — so paths, identifiers, and YAML data_source resolve to the right place.
• Business intent / the question the artifact answers — what it is for, in one or two lines.
• Rules and encodings you already discovered that the generator must honor (a mandatory constant filter like rtype='D', a column encoding, a join-key trap), so it does not rediscover, ignore, or contradict them.
• For reference_sql / metrics: the knowledge atoms mined from the same source, so the summary / search_text / measure expression stay consistent with the rule that explains the SQL.

The dispatching skill owns this handoff — the prompt seeds in the decision tree below (<table names + intent>, <complete SQL + business context>, …) are minimums, not the whole prompt.

Decision Tree

Walk top to bottom; the first match wins. Each item routes to exactly one store.

1. Reusable atomic business fact or rule — field encoding / enum / status code, mandatory constant filter, join trap (must go through a mapping table), boundary trap (strict vs non-strict inequality), business-term-to-field mapping, same-name-field divergence. → knowledge — delegate to extract-knowledge (lite mode).
2. Structured semantic definition of a table — its identifiers, measures, dimensions (the structure, not a single metric). → semantic_models — task(type="semantic_modeling", prompt=<table name(s) + intent>).
3. A reusable metric built on a measure — a named business metric, its aggregation expr, base measures, available dimensions. → metrics — task(type="semantic_modeling", prompt=<metric description or SQL>).
4. A complete, validated SQL worth indexing for semantic search / future reuse (plus a human summary). → reference_sql — task(type="gen_sql_summary", prompt=<the complete SQL + business context>). One SQL = one call = one entry. gen_sql_summary produces exactly one reference_sql row per invocation, so dispatch a separate task(gen_sql_summary) call for every distinct query. NEVER bundle multiple SQLs into one prompt — that collapses them into a single useless entry whose sql/search_text mixes unrelated queries and breaks few-shot retrieval. When the query came with an original natural-language question, pass it along and instruct the generator to use it verbatim as search_text — a future user question matches a stored question far better than it matches SQL keywords.
5. A high-level project overview — architecture, directory map, services, data assets, or the knowledge index. → AGENTS.md — edit the file directly (see AGENTS.md Section Ownership).
6. A lightweight, cross-session preference or context bound to one agent (≤ 2000 bytes) — a user/team habit, a default the agent should remember next session. → memory — add_memory / edit_memory.
7. A reusable, multi-step workflow or operating procedure that will be invoked repeatedly. → skills — delegate to create-skill.
8. None of the above — one-shot, inferable from INFORMATION_SCHEMA / table comments / column names, or not reusable. → Do NOT persist. Say so and stop.

Per-Store Reference

When NOT to pick a store:

• Not semantic_models — a single metric (that's metrics); a one-table fact already in the column comment.
• Not metrics — structural table/column/measure definitions (that's semantic_models).
• Not reference_sql — a one-shot query never to be reused; or the reason a query is written a certain way (that atomic rule is knowledge).
• Not knowledge — anything inferable from INFORMATION_SCHEMA / comments, generic SQL knowledge, or a fact that mechanically composes from existing facts.
• Not memory — a team-level long-lived business fact (that's knowledge); anything that would exceed 2000 bytes.
• Not AGENTS.md — fine-grained atomic facts (those are knowledge; AGENTS.md only holds their index).
• Not skills — a declarative "what is true" fact (that's knowledge); skills capture executable "how to do" procedures.

AGENTS.md Section Ownership

Because you edit AGENTS.md yourself, keep its sections consistent and minimal. Sections and their intent:

AGENTS.md is the KB entry point. It is injected into every node's <project_context>, so it is the one reliable place that tells a downstream agent what KB exists and how to reach it. Index retrieval-backed stores (semantic_models/metrics/reference_sql) by count + which search tool; index knowledge by file links. Never inline the stores' contents.

Write only sections that have real content. Omit any section you have nothing concrete to put in — never write empty "none yet" / "not built yet" placeholder lines. A section is added later (per the handling order below) when content for it actually exists.

Handling order:

1. AGENTS.md missing → write_file a minimal skeleton (# <project dir name> plus the relevant section); leave the inventory + knowledge/memory sections for /init to fill in and the vector-index sections (## Semantic Models / ## Metrics / ## Reference SQL) for /build-kb. Do not block on the user running /init first.
2. Exists but missing the target section → insert it in the canonical order above (Knowledge goes last).
3. Exists with the section → prefer a minimal edit_file. Never rewrite the whole file when a scoped edit suffices, and never touch ## Knowledge entries owned by extract-knowledge.

Disambiguation (overlapping boundaries)

• knowledge vs memory — a team-level, long-lived business fact → knowledge; a session-level, agent-bound preference or working context → memory.
• reference_sql vs knowledge — not exclusive, they feed each other. The complete reusable SQL → reference_sql (an example that teaches answer shape, retrieved later for few-shot); the atomic business rule mined from why that SQL is written that way → knowledge (teaches why). A single (question, gold_sql) pair routes to BOTH: store the example, and run extract-knowledge to mine the rule.
• output / answer conventions are NOT knowledge. "Return only the ratio column", "match the exact column projection", "don't add helper columns", "count questions return a single scalar" — these are presentation conventions, not facts about the data. They live in two project-level places: as reference_sql examples (few-shot, the strongest teacher) and, when a recurring pattern can be induced from the corpus, as schema-free bullets in AGENTS.md ## SQL Conventions (rides in <project_context>). They are NOT knowledge atoms and NOT a global prompt rule — answer shape is project/context-specific (strict-eval wants minimal projection; interactive BI often wants extra context columns). Only durable data-model facts (field encodings, term→field mappings, mandatory filters) are knowledge. Never encode a benchmark/eval quirk as knowledge.
• semantic_models vs metrics — the structure (table/column/measure/dimension definitions) → semantic_models; a concrete metric built on a measure → metrics.
• AGENTS.md vs knowledge — a high-level overview or index → AGENTS.md; a fine-grained atomic fact → knowledge (AGENTS.md only indexes it).
• skills vs knowledge — an executable, repeatable procedure ("how to do") → skills; a declarative statement ("what is true") → knowledge.

Output / Behavior

For each piece of input:

1. State the verdict: { target store, the decision-tree branch that matched and why }.
2. Execute the routing for that store — call the prescribed task subagent / skill / memory tool, or directly edit AGENTS.md.
3. If the content does not clearly fit any store, or routing would destructively overwrite existing content, call ask_user with numbered options before acting.

Process multiple pieces one at a time. When ask_user is unavailable and a verdict is ambiguous, report what is missing and route nothing by default.

End with a short human-readable summary: how many pieces were classified, where each was routed (or why it was not persisted), and any items still awaiting user confirmation. Do not return a JSON envelope.

Forbidden

• Do not hand-write semantic_models / metrics / reference_sql YAML or vector-store rows — use semantic_modeling for Dosi semantic assets and gen_sql_summary for reference SQL.
• Do not bundle multiple SQLs into one gen_sql_summary call — one query per call, one entry per query.
• Do not run extract-knowledge in deep mode — use lite.
• Do not exceed the 2000-byte memory cap, and do not write memory with any tool other than add_memory / edit_memory.
• Do not overwrite ## Knowledge entries in AGENTS.md — those are owned by extract-knowledge.
• Do not persist one-shot content, generic SQL knowledge, or anything inferable from INFORMATION_SCHEMA / table comments / column names.
• Do not generate the artifact's content yourself (e.g. write the SQL, invent the metric) — your job is classification and routing, not authoring.
• Do not resolve a destructive overwrite silently — go through ask_user.