feat: evolve retrieval layer with content_type, memory FTS, forget(), and recap references

Extract schema DDL into scripts/schema.sql. Add content_type and is_meta columns to messages for transcript control-plane filtering. Introduce FTS5-backed memory recall with safe tokenization, memory soft-delete via forget() through the renamed --attune runtime, and anchors on memory records. Expand query helpers (includeMeta, thread opts, overview project-path awareness). Add per-card recap retrieval and writing references under references/recap/.
2026-06-15 02:25:21 +08:00
parent b52f57b538
commit 3822bda89e
24 changed files with 1290 additions and 157 deletions
@@ -0,0 +1,33 @@
+# Card 2 Thinking Retrieval
+
+Goal: find turning points. This card is not a project timeline and not an implementation log. It is the record of what changed in the user's mind.
+
+Retrieve 3-6 turns. A turn needs both sides:
+
+- the user question, friction, doubt, or request that started the turn;
+- the later decision, reframing, finding, or constraint that replaced the earlier
+  state.
+
+Useful searches:
+
+- user questions in the period: "为什么", "是不是", "怎么", "我觉得", "不应该";
+- places where the user corrected the direction and then approved a new frame;
+- summaries that name decisions, followed by `context()` or `thread()` for the
+  user's actual words;
+- memory records only as hints; raw session evidence must provide the prompt and
+  turn.
+
+Prefer helpers first. If you need custom SQL for message windows or user-turn
+counts, read `references/schema.md` before writing the SQL.
+
+Do not use workflow names, feature names, or agent task labels as prompts when
+the user had their own wording. Do not use counts, "5 rounds", "13 agents", or
+implementation effort as turns unless that count is the turn itself.
+
+Read this card's writing file immediately after the turns are chosen:
+`references/recap/writing2-thinking.md`. Then update the JSON `thinking_path`
+card and add evidence for each item.
+Do not read `pattern3-vibe.md` until this JSON update is done.
+
+Stop when each item has a source-language prompt label, a short changed-state
+prompt, turn, and an `evidence` anchor.