@lmzhen/dsh-evolution-learning-graph 0.3.83 → 0.4.1

package/README.md

@@ -10,25 +10,22 @@ Memory node ids carry a trailing snapshot token (`memory:<source>:<index>:<snaps
 
 Memory→skill edges are word-level, not substring: the entry is tokenized on non-letter/digit/hyphen runs and a skill name links only when it is a whole token. This prevents a skill named `run` from linking the words `running`/`grunt`.
 
-`graph edit`/`graph delete` route through the evolution approval seam when it is mounted (soft-probed; the write executes directly when it is absent) — for BOTH skill nodes and memory nodes (P2-6, v15; memory staged args mirror the `memory` tool runner's replay shape). Before staging, the seam's `hasRunner` is checked (P2-7, v15): with approval enabled, a session policy that is not `never`, and the write actually staging (foreground stages per approval `stageForeground`; subagent origins always stage), a missing replay runner's row (tool-skill-manage / tool-memory) refuses the write; allow-direct combinations execute unchanged instead of creating a pending record no approver could replay. Each approved/executed edit bumps the skill's patch counter and a delete archives it, matching `skill_manage` — including the no-op gate: an edit whose content is byte-equivalent to the current file (`noop`) writes nothing and does not bump the patch counter, exactly as `skill_manage` treats an unchanged update/patch. The command invocation's session rides the approval request (v12 N1), so a `never`-policy session stages nothing instead of deriving every graph write as foreground.
+**`graph edit`/`graph delete` route through the evolution approval seam** when mounted (soft-probed: the write executes directly when absent) — for BOTH skill nodes and memory nodes. Each approved/executed edit bumps the skill's patch counter and a delete archives it, matching `skill_manage` — including the no-op gate: an edit whose content is byte-equivalent to the current file (`noop`) writes nothing and does not bump the patch counter, exactly as `skill_manage` treats an unchanged update/patch.
 
-## Model Experience
+## Model surface
 
-### Indirect model surface
+- **Model-visible:** nothing of its own: no prompt section and no tool schema; `/graph` is a human command, and consumers own the model-visible effects.
+- **Prompt prefix / KV cache:** independent of request-prefix construction — it does not alter the assembled prompt or tool list; family-level rules: `packages/README.md` §"Model-visible prompt prefix and the KV cache".
+- **Mount it?** yes — the `evolution-learning-graph` row, carried by the `evolution-host`, `evolution-all` and one-click `evolution-preset` bundles.
 
-`@lmzhen/dsh-evolution-learning-graph` registers no direct prompt or tool schema itself. Model-visible effects are owned by the packages that consume this service.
-
-#### Token effect
-
-Zero direct token effect from this package; consumers add any model-visible tokens.
-
-#### KV Cache effect
-
-Independent of request-prefix construction. This package does not alter the assembled prompt or tool list.
-
-## Known Limitations and Deferred Work
+## Known limitations
 
 - The memory drift snapshot covers only the node label (first line, first 80 chars): a change confined to a later line of the same entry is not detected, because the rendered node label is the comparison anchor.
 - A hand-typed bare `memory:<source>:<index>` id has no snapshot and skips the drift check (legacy path; the rendered ids always carry the snapshot).
 
 **Runtime invariant:** No companion is published. The platform auto-assembles nothing and the family mounts no `<pkg>/invariant` cordis row, so a companion here would never execute (v37 S2.1 / I-3).
+## Notes and history
+
+- `graph edit`/`graph delete` route through the evolution approval seam when it is mounted (soft-probed; the write executes directly when it is absent): for BOTH skill nodes and memory nodes (P2-6, v15; memory staged args mirror the `memory` tool runner's replay shape).
+- Before staging, the seam's `hasRunner` is checked (P2-7, v15): with approval enabled, a session policy that is not `never`, and the write actually staging (foreground stages per approval `stageForeground`; subagent origins always stage), a missing replay runner's row (tool-skill-manage / tool-memory) refuses the write; allow-direct combinations execute unchanged instead of creating a pending record no approver could replay.
+- The command invocation's session rides the approval request (v12 N1), so a `never`-policy session stages nothing instead of deriving every graph write as foreground.

package/lib/index.js

@@ -1,6 +1,6 @@
 import { effectiveSessionPolicy } from "@lmzhen/dsh-evolution-approval";
 import z from "@deepseek-ai/schemastery";
-import { SKILL_NAME_RE, contentHash, evolutionIoAdapter, newSkillLibrary, relatedSkillNames, resolveOrigins } from "@lmzhen/dsh-evolution-core";
+import { SKILL_NAME_RE, contentHash, evolutionIoAdapter, newSkillLibrary, relatedSkillNames, resolveExecOrigins } from "@lmzhen/dsh-evolution-core";
 //#region lib/types/index.js
 /**
 * Learning graph over skills and memory, plus node-level commands
@@ -243,12 +243,16 @@ function apply(ctx, rawConfig = {}) {
 				const usage = usageService;
 				const memory = memoryService;
 				const io = ioService;
-				const session = invocation.agent.session;
+				const invocationAgent = invocation.agent;
+				const session = invocationAgent?.session;
+				const sessionMissing = (need) => session === void 0 ? err(`E-305: this invocation carries no agent — \`${need}\` needs a session-backed call (run it from a session in the GUI or the CLI).`) : void 0;
 				const input = invocation.rawInput.trim();
 				const detail = /^detail\s+(\S+)$/.exec(input);
 				if (detail && detail[1]) return await nodeDetail(detail[1]);
 				const edit = /^edit\s+(\S+)\s+([\s\S]+)$/.exec(input);
 				if (edit && edit[1] && edit[2] !== void 0) {
+					const gated = sessionMissing("/graph edit");
+					if (gated) return gated;
 					if (!edit[1].startsWith("memory:")) {
 						const stageRead = await probeSkillForStaging(edit[1]);
 						if (stageRead !== null) return err(stageRead);
@@ -257,6 +261,8 @@ function apply(ctx, rawConfig = {}) {
 				}
 				const remove = /^delete\s+(\S+)$/.exec(input);
 				if (remove && remove[1]) {
+					const gated = sessionMissing("/graph delete");
+					if (gated) return gated;
 					if (!remove[1].startsWith("memory:")) {
 						const stageRead = await probeSkillForStaging(remove[1]);
 						if (stageRead !== null) return err(stageRead);
@@ -333,7 +339,7 @@ function apply(ctx, rawConfig = {}) {
 					if (parsed === null) return err(`Invalid node id "${id}". Skill names or memory:<source>:<index> expected.`);
 					if (parsed.kind === "skill") {
 						const approval = ctx.get("evolutionApproval");
-						const origins = resolveOrigins(session.header.origin);
+						const origins = resolveExecOrigins(invocationAgent && { agent: invocationAgent });
 						if (approval) {
 							const sessionPolicyEdit = effectiveSessionPolicy(ctx, session);
 							const stagesForeground = approval.stageForeground !== false;
@@ -354,7 +360,7 @@ function apply(ctx, rawConfig = {}) {
 									libraryOrigin: origins.library
 								},
 								origin: origins.approval,
-								...session.id ? { sessionId: session.id } : {},
+								...session?.id ? { sessionId: session.id } : {},
 								session,
 								...sessionPolicy !== void 0 ? { sessionPolicy } : {}
 							});
@@ -368,7 +374,7 @@ function apply(ctx, rawConfig = {}) {
 					if (!check.ok) return err(check.message ?? "Memory index check failed.");
 					const memoryApproval = ctx.get("evolutionApproval");
 					if (memoryApproval) {
-						const originsM = resolveOrigins(session.header.origin);
+						const originsM = resolveExecOrigins(invocationAgent && { agent: invocationAgent });
 						const sessionPolicyM = effectiveSessionPolicy(ctx, session);
 						if (memoryApproval.isEnabled !== false && sessionPolicyM !== "never" && (originsM.approval === "background_review" || memoryApproval.stageForeground !== false) && !memoryApproval.hasRunner("memory")) return err("Graph memory write cannot be staged: no memory replay runner is registered — mount the tool-memory row (evolution-all bundle, or the evolution-preset overlay; the host bundle does not carry it) or disable evolution-approval.");
 						const decision = await memoryApproval.request({
@@ -383,7 +389,7 @@ function apply(ctx, rawConfig = {}) {
 								}]
 							},
 							origin: originsM.approval,
-							...session.id ? { sessionId: session.id } : {},
+							...session?.id ? { sessionId: session.id } : {},
 							session,
 							...sessionPolicyM !== void 0 ? { sessionPolicy: sessionPolicyM } : {}
 						});
@@ -402,7 +408,7 @@ function apply(ctx, rawConfig = {}) {
 					if (parsed.kind === "skill") {
 						const approval = ctx.get("evolutionApproval");
 						if (approval) {
-							const origins = resolveOrigins(session.header.origin);
+							const origins = resolveExecOrigins(invocationAgent && { agent: invocationAgent });
 							const sessionPolicy = effectiveSessionPolicy(ctx, session);
 							if (approval.isEnabled !== false && sessionPolicy !== "never" && (origins.approval === "background_review" || approval.stageForeground !== false) && !approval.hasRunner("skill")) return err("Graph skill delete cannot be staged: no skill replay runner is registered — mount the tool-skill-manage row (evolution-agent preset) or disable evolution-approval.");
 							const decision = await approval.request({
@@ -417,7 +423,7 @@ function apply(ctx, rawConfig = {}) {
 									libraryOrigin: origins.library
 								},
 								origin: origins.approval,
-								...session.id ? { sessionId: session.id } : {},
+								...session?.id ? { sessionId: session.id } : {},
 								session,
 								...sessionPolicy !== void 0 ? { sessionPolicy } : {}
 							});
@@ -431,7 +437,7 @@ function apply(ctx, rawConfig = {}) {
 					if (!check.ok) return err(check.message ?? "Memory index check failed.");
 					const memoryApproval = ctx.get("evolutionApproval");
 					if (memoryApproval) {
-						const origins = resolveOrigins(session.header.origin);
+						const origins = resolveExecOrigins(invocationAgent && { agent: invocationAgent });
 						const sessionPolicy = effectiveSessionPolicy(ctx, session);
 						if (memoryApproval.isEnabled !== false && sessionPolicy !== "never" && (origins.approval === "background_review" || memoryApproval.stageForeground !== false) && !memoryApproval.hasRunner("memory")) return err("Graph memory delete cannot be staged: no memory replay runner is registered — mount the tool-memory row (evolution-all bundle, or the evolution-preset overlay; the host bundle does not carry it) or disable evolution-approval.");
 						const decision = await memoryApproval.request({
@@ -445,7 +451,7 @@ function apply(ctx, rawConfig = {}) {
 								}]
 							},
 							origin: origins.approval,
-							...session.id ? { sessionId: session.id } : {},
+							...session?.id ? { sessionId: session.id } : {},
 							session,
 							...sessionPolicy !== void 0 ? { sessionPolicy } : {}
 						});

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@lmzhen/dsh-evolution-learning-graph",
   "description": "Learning graph over skills and memory (community build)",
-  "version": "0.3.83",
+  "version": "0.4.1",
   "publishConfig": {
     "access": "public"
   },
@@ -27,8 +27,8 @@
   "license": "MIT",
   "dependencies": {
     "@deepseek-ai/schemastery": "^3.18.1",
-    "@lmzhen/dsh-evolution-approval": "^0.3.83",
-    "@lmzhen/dsh-evolution-core": "^0.3.83"
+    "@lmzhen/dsh-evolution-approval": "^0.4.1",
+    "@lmzhen/dsh-evolution-core": "^0.4.1"
   },
   "peerDependencies": {
     "@deepseek-ai/dsh-commands": "^0.1.5-rc.2",