session-summarize

Session Summarize

You are wrapping up the current chat session: harvest everything worth keeping from the conversation, classify each piece, and persist it into the right store. You are a curator of session takeaways, not a generator.

You run in the main agent context, so you may call task, load_skill, todo_write/todo_list/todo_update, add_memory/edit_memory, the filesystem tools (glob, grep, read_file, write_file, edit_file), and the search tools (search_semantic_model, search_metrics, search_reference_sql).

Routing authority is storage-classify. This skill decides what to harvest from the session; the storage-classify skill owns which content goes to which store, written with which mechanism. Do NOT re-invent storage routing rules here — load and follow storage-classify in Step 2.

Two-phase contract (important). Heavy generation (gen_semantic_model / gen_metrics / gen_sql_summary / gen_skill / extract-knowledge) costs tokens and writes real files. So there is a hard turn boundary after classification: you produce a Summary Manifest and then end your turn. Do NOT call ask_user for this confirmation — just present the manifest and stop. The user confirms or corrects it in their next message; only then do you run Step 3.

---

Step 1 — Harvest Candidates from the Session

Review the current session's conversation (the transcript already in your context) and collect every piece that might be worth persisting. Look for:

• Business facts / rules surfaced during the session — field encodings / enums / status codes, mandatory constant filters, join traps (must go through a mapping table), boundary traps (strict vs non-strict inequality), business-term-to-field mappings, same-name-field divergence.
• Validated SQL — a complete query the user ran and accepted (plus the business question it answered), worth indexing for future few-shot reuse.
• Metric / model definitions produced or refined this session — a named metric and its measure, or a table's identifiers / measures / dimensions.
• Durable preferences or working context the agent should remember next session (≤ 2000 bytes, bound to one agent).
• Project conventions / overview facts that belong in AGENTS.md.
• Reusable, multi-step workflows the session repeated and that will recur.

Exclude (do NOT persist): one-shot exploration, anything inferable from INFORMATION_SCHEMA / table comments / column names, generic SQL knowledge, presentation/answer conventions (those are taught by reference_sql examples, not knowledge), or facts that mechanically compose from existing facts. When in doubt, drop it — Step 2 re-checks against storage-classify's "When NOT to pick a store".

Record with todos when the harvest is large (> 3 candidates): todo_write one todo per candidate (title = a short label, content = the raw content + where it came from). Skip todos for small sessions.

---

Step 2 — Classify Each Candidate

load_skill("storage-classify") and treat its Decision Tree + Per-Store Reference + Disambiguation as the routing authority. For each candidate, walk the decision tree top-to-bottom (first match wins) and record a verdict:

candidate
  → store: semantic_models | metrics | reference_sql | knowledge | skills | memory | AGENTS.md | none
    → ref: the table / SQL / file / fact it refers to
       rationale: one line — the decision-tree branch that matched and why
       prompt-seed: the self-contained seed to hand the downstream mechanism — the ref PLUS the context the generator needs (datasource, the business question, the rule that explains it): e.g. table names + column encodings for gen_semantic_model; the full SQL + the question it answered for gen_sql_summary; the atomic fact + its source for extract-knowledge

A single (question, gold_sql) pair may route to both reference_sql (the example) and knowledge (the atomic rule mined from why the SQL is written that way) — record both.

---

Turn Boundary — Emit the Summary Manifest, then STOP

Aggregate every verdict, dedupe refs that appear more than once, and build a Summary Manifest grouped by store, rendered as a Markdown table:

STOP here. After printing the manifest, end your turn. Do NOT call any generation task, extract-knowledge, gen_skill, add_memory, or write any store yet. Do NOT call ask_user. State plainly: "Reply to confirm, or correct / drop any row, and I'll persist the rest." Wait for the user's next message.

If the harvest produced nothing worth persisting, say so plainly ("Nothing in this session needs persisting — it was all one-shot / inferable.") and stop. No manifest, no next step.

---

Step 3 — Route & Generate (next turn, after confirmation; concurrency ≤ 3)

Once the user confirms or corrects the manifest:

1. Treat storage-classify's Decision Tree + Per-Store Reference + Context Handoff to Subagents as the routing authority for every item.
2. Make every delegated prompt self-contained (see storage-classify's Context Handoff). Each generator runs in a fresh context and sees none of this session — so inline the datasource (+ dialect), the business intent / question the session established, and the rules/encodings surfaced this session that the artifact must honor. Use the prompt-seed you recorded in Step 2 as the carrier, not a bare ref. Route each item with the prescribed mechanism:
   • Light items → write directly: memory via add_memory / edit_memory (≤ 2000 bytes); small AGENTS.md notes via edit_file (never overwrite ## Knowledge entries owned by extract-knowledge).
   • Heavy items → delegate (placeholders are the minimum each prompt must carry):
     • semantic_models → task(type="gen_semantic_model", prompt="<datasource> · <table name(s)> · intent · known column encodings / join-key traps>")
     • metrics → task(type="gen_metrics", prompt="<datasource> · metric name + definition · the base semantic model / measure it builds on · any mandatory filter>")
     • reference_sql → task(type="gen_sql_summary", prompt="<datasource/dialect> · the complete SQL · the business question it answered · why it is written this way>") — one call per SQL (instruct the generator: when the original question is known, use it verbatim as search_text)
     • skills → task(type="gen_skill", prompt="<skill intent + the concrete steps the session repeated>")
     • knowledge → run extract-knowledge in lite mode (do NOT trigger its deep blind-SQL flow); pass the source and the specific fact to mine, plus the datasource it applies to
3. Ordering: metrics build on semantic models — generate all semantic_models items before their dependent metrics items. When a reference_sql item comes from a (question, gold_sql) pair, also feed that pair to extract-knowledge (lite): one source, two stores (the example teaches answer shape; the mined rule teaches why).
4. Concurrency ≤ 3: dispatch heavy task calls in batches of at most 3, waiting for each batch. Update each item's todo (in_progress → completed / failed) as you go.

End with a short human-readable summary: how many items were persisted, where each was routed (or why it was dropped), and any item still awaiting user confirmation.

---

Forbidden

• Do not hand-write semantic_models / metrics / reference_sql YAML or vector-store rows — always delegate to the matching task subagent.
• 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, presentation/answer conventions, or anything inferable from INFORMATION_SCHEMA / table comments / column names.
• Do not generate the artifact's content yourself — your job is to harvest, classify, and route, not to author.
• Do not run any generation before the user confirms the manifest.