docs: trait filtering is client-side; add memory schema

workspace/AGENTS.md

@@ -55,17 +55,46 @@ Avoids re-running `nfts list-by-account` on every heartbeat. Invalidate a chain'
 
 So `events by-account` only streams the tail since last heartbeat.
 
-### `memory/scan_state.json` — once-per-day scan timestamps
+### `memory/scan_state.json` — once-per-day scan timestamps + resumable cursors
 
-Tracks the last run of HEARTBEAT.md steps 5 (drops) and 6 (trending), which run at most every ~20 hours.
+Tracks the last run of HEARTBEAT.md steps 5 (drops) and 6 (trending), which run at most every ~20 hours. Also holds resumable cursors for any per-collection trait scan in progress, so a heartbeat that hits 429 mid-scan can pick up where it left off.
 
 ```json
 {
   "last_drop_scan": "2026-04-17T02:00:00Z",
-  "last_trending_scan": "2026-04-17T02:00:00Z"
+  "last_trending_scan": "2026-04-17T02:00:00Z",
+  "trait_scans": {
+    "<slug>": {
+      "started_at": "2026-04-17T02:00:00Z",
+      "completed_at": null,
+      "next_cursor": "abc123",
+      "tokens_seen": 1450
+    }
+  }
 }
 ```
 
+When a trait scan completes, set `completed_at`, clear `next_cursor`, and write the result to `memory/trait_holders.<slug>.json` (below).
+
+### `memory/trait_holders.<slug>.json` — per-collection trait index
+
+Built by the whole-collection scan pattern in SOUL.md → *How You Work* → *Trait filtering*. Re-scan at most once per day. Index is `trait_type` → `value` → `token_id[]`.
+
+```json
+{
+  "slug": "tiny-dinos-eth",
+  "updated_at": "2026-04-17T02:30:00Z",
+  "by_trait": {
+    "feet": {
+      "hoverboard": ["6292", "5996"],
+      "skateboard": []
+    }
+  }
+}
+```
+
+Only create one of these for collections the user has explicitly asked to track by trait — they're expensive to maintain.
+
 ### `MEMORY.md` (in `workspace/`, not `memory/`)
 
 Free-text long-form observations that don't fit a schema: API quirks, volatile slugs, seller wallets that pattern wash-trade, drops that disappointed, specific things the user said about their taste that you want to remember word-for-word.

workspace/SOUL.md

@@ -62,6 +62,7 @@ Before submitting any transaction, run the gate in order. Any RED stops the flow
 - **Ordering:** run requests sequentially on the shared `OPENSEA_API_KEY`. Parallel fans out 429s.
 - **Memory:** `memory/floors.json` = latest watchlist snapshot (overwrite each heartbeat). `memory/actions.jsonl` = append-only log of anything with a side-effect. `memory/taste.json` = structured taste model you maintain per collection. See `AGENTS.md` for schemas.
 - **Stale-reject listings:** before proposing a buy, re-fetch the listing — OpenSea will 404 or return a new order hash if it expired.
+- **Trait filtering is client-side.** Enumerate the schema with `collections traits <slug>`, then fetch tokens and filter on `traits[]` locally. Two patterns: (a) listed-only — pull `listings all <slug>`, then `nfts get` per token (cheap, ~1 + N_listings calls); (b) whole-collection — paginate `nfts list-by-collection`, fetch traits per token, cache the trait → token-id index in a per-collection `memory/trait_holders.<slug>.json` (expensive, do at most once per day, throttle to ≤1 req / 0.3s, persist cursor in `memory/scan_state.json`, resume on 429). See `skills/opensea/SKILL.md` → *Reading NFT data* → *Filtering NFTs by trait*.
 
 ## Wallet