ygl-trees 0.79.1 → 0.80.0

package/CHANGELOG.md

@@ -1,5 +1,10 @@
 # Changelog
 
+## 0.80.0（2026-10-17）rel 织物并入 eval 上下文 + 会话内裁决载体
+- **eval 上下文包并入 rel 挂接摘要**（用户实测发现：裁决已 `decided-by` 挂链，P 仍 0.30，必须 `--state k=v` 手动携带事实才 0.92）：`evalContext` 新增 `attachments`——节点挂接的 relations 双向查询（≤12 条），claim 摘录 text+status+confidence、evidence 摘录 quote（截 120 字）、kb/source 标题、node key+name、其余 id-only；无挂接时字段缺省（与旧形状逐字节兼容）
+- **`research source add --adjudication "<裁决摘要>"`**：会话内裁决原生载体——url 自动 `session://adjudication/<yyyymmddhhmm>`，摘要兼作缺省标题；`--url` 校验放宽收 `session://`；裸 `--adjudication` 旗标缺值 rc=2
+- 测试：`packages/techtree/model/tests/evalctx.spec.ts` 5 例（无挂接兼容 / decided-by 摘要 / 反向 evidenced-by / 截断与悬空退化 / 12 条上限）；`scripts/readme-verify.sh` +1（--adjudication rc=0）
+
 ## 0.79.1（2026-10-17）锁身份 + serve --help 短路 + 批量判定 + 回测状态桥
 - **异常①锁归属**：`.ygllock` 记录持有者命令行；撞锁报错、`db unlock`（含 `holder_cmd` JSON 字段）、`doctor` 三处回显真实持有者身份——CLI 撞 CLI 不再被误标为 serve（serve 从不取写锁，报错明示这一点）
 - **异常②**：`webui serve --help` 短路打印用法即退出，不再直接启动服务（随机端口误触发实证）

package/README.md

@@ -103,7 +103,9 @@ ygl-trees node import --file tree.json --tree copy  # 严格校验、事务导
 
 **批量判定 + `--state-file`（0.79.1）**：`node can k1 k2 k3` 一次进程判定多节点（多键自动进批模式，`--json` 返回数组）；`--state-file <json>` 喂状态袋，两种形状——**扁平计数** `{"y_imported": 40}` 全节点共享，或**按节点** `{"a": {"y_imported": 40}, "b": {...}}` 各取各袋。探针脚本从此每树起一次进程而非每节点。
 
-**决策模型评估门（spec R8；0.14 布尔 → 0.26 三题型 → 0.27 混批）**：语义型必检条件用 eval 题型声明，`node can <key> --eval` 会把**节点元数据 + 直接父（关联）+ 祖先链（闭包索引）+ 条件组 + 已解锁集 + --state 袋**组装为上下文包，交给评估模型（TypeSafe jev，经 Vercel AI Gateway）判定：
+**决策模型评估门（spec R8；0.14 布尔 → 0.26 三题型 → 0.27 混批 → 0.80 织物并入）**：语义型必检条件用 eval 题型声明，`node can <key> --eval` 会把**节点元数据 + 直接父（关联）+ 祖先链（闭包索引）+ 条件组 + 已解锁集 + --state 袋 + rel 挂接摘要**组装为上下文包，交给评估模型（TypeSafe jev，经 Vercel AI Gateway）判定：
+
+- **rel 织物并入（0.80）**：节点挂接的 evidence（quote 摘要）/ claim（text+status+confidence）/ kb/source（标题）等 rel 关系（双向，≤12 条，文本截 120 字）自动进上下文包的 `attachments`——「证据挂链」纪律从此对语义门槛可见，不必再 `--state k=v` 手动携带裁决事实（实测 P 0.30 → 0.92 的差距由挂链补平）
 
 - **三题型**：`--op eval`（布尔 P(true)）/ `--op eval-choice`（多选一，判 **P(期望答案)**，`--options "k=desc,..."` + `--answer k`）/ `--op eval-score`（等级评分 `--levels "级0,..."` 2-10 级 + `--min-score n`，判概率加权均值 ±1 灰带）
 - P≥0.7 通过 · P≤0.4 未通过 · **灰带 pending**（过程态，禁作终态，单列不阻塞也不放行）；score 的灰带是 [min−1, min)
@@ -124,6 +126,7 @@ ygl-trees node import --file tree.json --tree copy  # 严格校验、事务导
 query add       登记检索式
 query attach    检索式 ↔ 来源关联（query list 显示召回数）
 source add      收来源，按 (project, url_hash) 去重；--query 顺带关联
+                --adjudication "<裁决摘要>"：会话内裁决原生载体（url 自动 session://adjudication/<时刻>，摘要兼作标题）；--url 也收 session://
 source fetch <url>  直接抓取网页入库：title/meta description/去标签正文；15s 超时、非 html 拒收、失败零副作用，正文过薄自动打 needs_review
 evidence add    从来源摘原文片段，带 locator（页码/章节）；长摘录用 --file <path> 免转义
 claim add       写下断言

package/dist/bundle/bin.cjs

@@ -2313,6 +2313,12 @@ var TreeStore = class {
 	* + unlocked set + the caller's state bag. The eval API demands a context
 	* package — bare state fell into the gray band 6/8 in the material's
 	* routing test, so we ship everything the tree knows about this node.
+	*
+	* 0.80: the rel fabric joins the package — evidence/claim/kb summaries
+	* attached to the node (either direction) land in `attachments`, so a
+	* decided-by chain is visible to the eval model without manually
+	* re-carrying facts via --state (user finding: P 0.30 → 0.92 gap).
+	* Bounded: ≤12 rows, text clipped to 120 chars.
 	*/
 	evalContext(treeId, keyOrId, stateBag, context = "default") {
 		const n = this.nodeGet(treeId, keyOrId);
@@ -2336,7 +2342,7 @@ var TreeStore = class {
 			effect: c.effect
 		}));
 		const unlocked = [...this.unlockedSet(treeId, context)].map((id) => idToKey.get(id) || id);
-		return {
+		const pkg = {
 			node: {
 				key: n.key,
 				name: n.name,
@@ -2350,6 +2356,62 @@ var TreeStore = class {
 			unlocked,
 			state: stateBag
 		};
+		const attachments = this.relAttachments(String(n.id));
+		if (attachments.length) pkg.attachments = attachments;
+		return pkg;
+	}
+	/** 0.80: rel-fabric summaries for the eval context package (both directions). */
+	relAttachments(nodeId) {
+		const rows = this.db.all(`SELECT rel, src_type, src_id, dst_type, dst_id FROM relations
+       WHERE (src_type='node' AND src_id=?) OR (dst_type='node' AND dst_id=?)
+       ORDER BY rowid LIMIT 12`, [nodeId, nodeId]);
+		const clip = (v, n = 120) => String(v ?? "").slice(0, n);
+		return rows.map((r) => {
+			const srcIsMe = r.src_type === "node" && r.src_id === nodeId;
+			const type = String(srcIsMe ? r.dst_type : r.src_type);
+			const id = String(srcIsMe ? r.dst_id : r.src_id);
+			const base = {
+				rel: String(r.rel),
+				type,
+				id
+			};
+			const q = (sql) => this.db.get(sql, [id]);
+			if (type === "claim") {
+				const c = q("SELECT text, status, confidence FROM claims WHERE id=?");
+				if (c) return {
+					...base,
+					text: clip(c.text),
+					status: c.status,
+					confidence: c.confidence
+				};
+			} else if (type === "evidence") {
+				const e = q("SELECT quote FROM evidence WHERE id=?");
+				if (e) return {
+					...base,
+					quote: clip(e.quote)
+				};
+			} else if (type === "kb") {
+				const k = q("SELECT title FROM kb_notes WHERE id=?");
+				if (k) return {
+					...base,
+					title: clip(k.title, 80)
+				};
+			} else if (type === "source") {
+				const s = q("SELECT title FROM sources WHERE id=?");
+				if (s) return {
+					...base,
+					title: clip(s.title, 80)
+				};
+			} else if (type === "node") {
+				const nd = q("SELECT key, name FROM nodes WHERE id=?");
+				if (nd) return {
+					...base,
+					key: nd.key,
+					name: clip(nd.name, 80)
+				};
+			}
+			return base;
+		});
 	}
 	/** Every node whose required gates are currently satisfied. */
 	ready(treeId, context = "default", state = {}) {
@@ -6492,6 +6554,10 @@ init_writes();
 init_entities$1();
 /** Run a write through the TypeORM transactional write service and print
 *  with JSON discipline. Keeps every write command 6 lines. */
+/** 0.80: compact timestamp for session:// adjudication urls (yyyymmddhhmm). */
+function nowCompact() {
+	return (/* @__PURE__ */ new Date()).toISOString().replace(/[-:T]/g, "").slice(0, 12);
+}
 function runWrite(env, fn, human) {
 	const { ctx } = env;
 	return getOrm(env.db.path).then(fn).then((r) => {
@@ -6616,14 +6682,24 @@ function run$8(env, action, rest, args) {
 				bad(ctx, "flag needs a value: --title \"...\" / --url \"...\"  (bare flag was treated as boolean)");
 				return 2;
 			}
-			const title = String(titleRaw || "");
-			const url = String(urlRaw || "");
+			let title = String(titleRaw || "");
+			let url = String(urlRaw || "");
+			const adjRaw = args.flags.adjudication;
+			if (adjRaw === true) {
+				bad(ctx, "flag needs a value: --adjudication \"<裁决摘要>\"");
+				return 2;
+			}
+			if (adjRaw !== void 0 && adjRaw !== "") {
+				const adj = String(adjRaw);
+				if (!url) url = `session://adjudication/${nowCompact()}`;
+				if (!title) title = adj.slice(0, 60);
+			}
 			if (!title && !url) {
 				bad(ctx, "need --title and/or --url");
 				return 2;
 			}
-			if (url && !/^https?:\/\/|^file:\/\//.test(url)) {
-				bad(ctx, `--url must be http(s):// or file:// (got: "${require_text.clip(url, 30)}")`);
+			if (url && !/^https?:\/\/|^file:\/\/|^session:\/\//.test(url)) {
+				bad(ctx, `--url must be http(s)://, file:// or session:// (got: "${require_text.clip(url, 30)}")`);
 				return 2;
 			}
 			const spec = {

package/dist/packages/research/pipeline/src/cmds.js

@@ -48,6 +48,10 @@ const config_1 = require("../../../ops/tools/src/config");
 const entities_1 = require("./entities");
 /** Run a write through the TypeORM transactional write service and print
  *  with JSON discipline. Keeps every write command 6 lines. */
+/** 0.80: compact timestamp for session:// adjudication urls (yyyymmddhhmm). */
+function nowCompact() {
+    return new Date().toISOString().replace(/[-:T]/g, "").slice(0, 12);
+}
 function runWrite(env, fn, human) {
     const { ctx } = env;
     return (0, orm_1.getOrm)(env.db.path)
@@ -177,14 +181,31 @@ function run(env, action, rest, args) {
                 (0, fmt_1.bad)(ctx, 'flag needs a value: --title "..." / --url "..."  (bare flag was treated as boolean)');
                 return 2;
             }
-            const title = String(titleRaw || "");
-            const url = String(urlRaw || "");
+            let title = String(titleRaw || "");
+            let url = String(urlRaw || "");
+            // 0.80: in-session adjudication carrier — the most common human
+            // arbitration form had no native --url shape (http(s)/file only),
+            // forcing temp-file workarounds. `--adjudication "<裁决摘要>"`
+            // fills url with a session:// pseudo-scheme and doubles as the
+            // title when --title is absent.
+            const adjRaw = args.flags.adjudication;
+            if (adjRaw === true) {
+                (0, fmt_1.bad)(ctx, 'flag needs a value: --adjudication "<裁决摘要>"');
+                return 2;
+            }
+            if (adjRaw !== undefined && adjRaw !== "") {
+                const adj = String(adjRaw);
+                if (!url)
+                    url = `session://adjudication/${nowCompact()}`;
+                if (!title)
+                    title = adj.slice(0, 60);
+            }
             if (!title && !url) {
                 (0, fmt_1.bad)(ctx, "need --title and/or --url");
                 return 2;
             }
-            if (url && !/^https?:\/\/|^file:\/\//.test(url)) {
-                (0, fmt_1.bad)(ctx, `--url must be http(s):// or file:// (got: "${(0, text_1.clip)(url, 30)}")`);
+            if (url && !/^https?:\/\/|^file:\/\/|^session:\/\//.test(url)) {
+                (0, fmt_1.bad)(ctx, `--url must be http(s)://, file:// or session:// (got: "${(0, text_1.clip)(url, 30)}")`);
                 return 2;
             }
             const spec = {

package/dist/packages/techtree/model/src/reads.js

@@ -386,6 +386,12 @@ class TreeStore {
      * + unlocked set + the caller's state bag. The eval API demands a context
      * package — bare state fell into the gray band 6/8 in the material's
      * routing test, so we ship everything the tree knows about this node.
+     *
+     * 0.80: the rel fabric joins the package — evidence/claim/kb summaries
+     * attached to the node (either direction) land in `attachments`, so a
+     * decided-by chain is visible to the eval model without manually
+     * re-carrying facts via --state (user finding: P 0.30 → 0.92 gap).
+     * Bounded: ≤12 rows, text clipped to 120 chars.
      */
     evalContext(treeId, keyOrId, stateBag, context = "default") {
         const n = this.nodeGet(treeId, keyOrId);
@@ -398,7 +404,7 @@ class TreeStore {
             name: c.name, kind: c.kind, op: c.op, value: c.value, required: !!c.required, effect: c.effect,
         }));
         const unlocked = [...this.unlockedSet(treeId, context)].map((id) => idToKey.get(id) || id);
-        return {
+        const pkg = {
             node: { key: n.key, name: n.name, kind: n.kind, tier: n.tier, cost: n.cost },
             parents,
             ancestors,
@@ -406,6 +412,50 @@ class TreeStore {
             unlocked,
             state: stateBag,
         };
+        const attachments = this.relAttachments(String(n.id));
+        if (attachments.length)
+            pkg.attachments = attachments;
+        return pkg;
+    }
+    /** 0.80: rel-fabric summaries for the eval context package (both directions). */
+    relAttachments(nodeId) {
+        const rows = this.db.all(`SELECT rel, src_type, src_id, dst_type, dst_id FROM relations
+       WHERE (src_type='node' AND src_id=?) OR (dst_type='node' AND dst_id=?)
+       ORDER BY rowid LIMIT 12`, [nodeId, nodeId]);
+        const clip = (v, n = 120) => String(v ?? "").slice(0, n);
+        return rows.map((r) => {
+            const srcIsMe = r.src_type === "node" && r.src_id === nodeId;
+            const type = String(srcIsMe ? r.dst_type : r.src_type);
+            const id = String(srcIsMe ? r.dst_id : r.src_id);
+            const base = { rel: String(r.rel), type, id };
+            const q = (sql) => this.db.get(sql, [id]);
+            if (type === "claim") {
+                const c = q("SELECT text, status, confidence FROM claims WHERE id=?");
+                if (c)
+                    return { ...base, text: clip(c.text), status: c.status, confidence: c.confidence };
+            }
+            else if (type === "evidence") {
+                const e = q("SELECT quote FROM evidence WHERE id=?");
+                if (e)
+                    return { ...base, quote: clip(e.quote) };
+            }
+            else if (type === "kb") {
+                const k = q("SELECT title FROM kb_notes WHERE id=?");
+                if (k)
+                    return { ...base, title: clip(k.title, 80) };
+            }
+            else if (type === "source") {
+                const s = q("SELECT title FROM sources WHERE id=?");
+                if (s)
+                    return { ...base, title: clip(s.title, 80) };
+            }
+            else if (type === "node") {
+                const nd = q("SELECT key, name FROM nodes WHERE id=?");
+                if (nd)
+                    return { ...base, key: nd.key, name: clip(nd.name, 80) };
+            }
+            return base; // query/artifact/skill or dangling ref: id only
+        });
     }
     /** Every node whose required gates are currently satisfied. */
     ready(treeId, context = "default", state = {}) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ygl-trees",
-  "version": "0.79.1",
+  "version": "0.80.0",
   "description": "猿光 · 科技树 — database-centric CLI for research → production → knowledge reuse, built for AI agents.",
   "type": "commonjs",
   "main": "dist/apps/cli/index.js",

package/skills/ygl-trees-eval-gates/SKILL.md

@@ -60,7 +60,12 @@ ygl-trees node can gate-node --eval --json     # evaluated[] 带 type 字段 + p
 ## 4. 上下文包（模型看到什么）
 
 节点元数据（key/name/kind/tier/cost）+ 直接父（node_edges）+ 祖先链（node_closure）+
-该节点全部条件 + 已解锁集 + `--state` 袋。**state 必带上下文包**——裸 state 大概率落灰带。
+该节点全部条件 + 已解锁集 + `--state` 袋 + **rel 挂接摘要（0.80）**。**state 必带上下文包**——裸 state 大概率落灰带。
+
+**attachments（0.80 织物并入）**：节点挂接的 rel 关系（双向 ≤12 条）自动摘录入包——
+claim（text+status+confidence）、evidence（quote 截 120 字）、kb/source（标题）、node（key+name）、
+其余类型 id-only。裁决链 `decided-by` 挂好后 eval 模型直接可见，不必 `--state k=v` 手动携带事实
+（用户实测：不挂链 P=0.30，挂链后 0.92）。
 
 ## 5. 凭证与端点（优先级：config > 环境变量/文件 > 默认网关）
 

package/skills/ygl-trees-relating/SKILL.md

@@ -17,6 +17,11 @@ tags: ygl-trees,methodology,relations,fabric
 **两条平行流水线**：研究管线（source→evidence→claim→report）和科技树各自发育、零缝合——
 检阅者看不出"这个节点依据什么证据、由什么断言决定、跟踪着什么任务"。`relations` 表就是缝合线。
 
+**缝合直接喂决策模型（0.80）**：节点挂接的 rel 关系（双向 ≤12 条）自动进 eval 上下文包
+`attachments`——claim 带 text+status+confidence、evidence 带 quote 摘要。裁决 `decided-by`
+挂链后 `node can --eval` 不必 `--state k=v` 手动携带事实（实测 P 0.30 → 0.92）。
+挂链纪律由此获得语义门槛层面的回报。
+
 ## 2. rel 词汇表（方向：src --rel→ dst）
 
 | rel | 语义 | 典型挂法 |

package/skills/ygl-trees-research/SKILL.md

@@ -26,9 +26,14 @@ ygl-trees research source add --title t --url https://... \
         --abstract "..." --relevance 4 --query "邻接表 vs 闭包表"   # 手工入库+关联检索式
 ygl-trees research source fetch https://example.com/page     # 自动抓取（title/meta/正文；
                                                               # 非 html 拒收零副作用；薄内容打 needs_review）
+ygl-trees research source add --adjudication "会话内裁决：三条件已逐条核对" \
+        --tags adjudication                                    # 会话内裁决原生载体（0.80）：
+                                                              # url 自动 session://adjudication/<时刻>，
+                                                              # 摘要兼作标题——不再需要临时文件绕行
 ```
 
 来源按 `(project, url_hash)` 去重；`--query` 顺带建立召回关联（query list 显示召回数）。
+`--url` 收 http(s):// · file:// · session://（0.80）。
 
 ## 2. 摘录与断言