dsh-continual-evolve 0.2.0 → 0.4.0

package/lib/wrapup.js

@@ -1,7 +1,10 @@
-import { BlockAssembler, createUserMessage, ReasoningEffortId } from "@deepseek-ai/dsh-llm";
 import { ARCHIVED_AT_KEY, PROMOTED_AT_KEY, PROMOTED_TO_KEY, SOURCE_SEQS_KEY, SOURCE_SESSION_KEY, SOURCED_FROM_KEY, isArchived } from "./types.js";
 import { extractJsonObject } from "./plan.js";
 import { compactText } from "./render.js";
+import { streamText } from "./llm-text.js";
+import { getUsageCount, loadUsage } from "./usage.js";
+import { recencyScore } from "./inject.js";
+import { DEFAULT_PROMOTION_POLICY, mostSimilarGlobalEntry, projectScopedReason } from "./promotion.js";
 export function candidateKey(kind, id) {
     return `${kind}:${id}`;
 }
@@ -71,7 +74,11 @@ export function globalHintsFor(globalState, kind, entry) {
  * `coveredGlobally` flag so the assessor never wastes a promote on a topic
  * the global store already owns.
  */
-export function listLocalCandidates(state, globalState) {
+/** Staleness threshold: entries with recency below this AND zero usage are stale. */
+const STALE_RECENCY_THRESHOLD = 0.1;
+export function listLocalCandidates(state, globalState, baseDir) {
+    const usage = baseDir ? loadUsage(baseDir) : undefined;
+    const now = Date.now();
     const candidates = [];
     for (const kind of Object.keys(state.entries)) {
         for (const entry of Object.values(state.entries[kind])) {
@@ -81,6 +88,8 @@ export function listLocalCandidates(state, globalState) {
                 continue;
             if (typeof entry.metadata[PROMOTED_TO_KEY] === "string")
                 continue;
+            const injectionCount = usage ? getUsageCount(usage, kind, entry.id) : 0;
+            const stale = injectionCount === 0 && recencyScore(entry, now) < STALE_RECENCY_THRESHOLD;
             candidates.push({
                 kind,
                 id: entry.id,
@@ -91,6 +100,8 @@ export function listLocalCandidates(state, globalState) {
                 metadata: entry.metadata,
                 coveredGlobally: globalCoverageDetected(globalState, kind, entry),
                 globalHints: globalHintsFor(globalState, kind, entry),
+                injectionCount,
+                stale,
             });
         }
     }
@@ -148,9 +159,16 @@ export function parseWrapupAssessment(text, candidates) {
  * Apply-time deterministic guard: re-check every promote verdict against the
  * global store right before it lands. The LLM classification may be stale
  * (a gate ran while assessing) or wrong; this ensures a promote never writes
- * a duplicate global entry. Pure and unit-tested.
+ * a duplicate, project-scoped, or too-thin global entry. Pure and unit-tested.
+ *
+ * Guards (2026-08-22 promotion policy):
+ * - audited candidate list + title coverage (pre-existing),
+ * - project-scoped content markers (absolute paths / session ids) — the
+ *   global store is shared across projects and must stay portable,
+ * - thin content below the policy floor (framing outweighs the fact),
+ * - near-duplicate of an existing global entry by content overlap.
  */
-export function filterPromotable(items, globalState, candidates) {
+export function filterPromotable(items, globalState, candidates, policy = DEFAULT_PROMOTION_POLICY) {
     const byKey = new Map(candidates.map((candidate) => [candidateKey(candidate.kind, candidate.id), candidate]));
     const promotable = [];
     const skipped = [];
@@ -166,6 +184,26 @@ export function filterPromotable(items, globalState, candidates) {
             skipped.push({ key: item.key, reason: "already covered globally" });
             continue;
         }
+        const scoped = projectScopedReason(`${candidate.title}\n${candidate.content}`, policy);
+        if (scoped) {
+            skipped.push({ key: item.key, reason: scoped });
+            continue;
+        }
+        if (candidate.content.length < policy.minPromoteChars) {
+            skipped.push({
+                key: item.key,
+                reason: `too thin to promote (${candidate.content.length} < ${policy.minPromoteChars} chars) — keep local or merge`,
+            });
+            continue;
+        }
+        const similar = mostSimilarGlobalEntry(globalState, candidate.kind, candidate.title, candidate.content, policy);
+        if (similar) {
+            skipped.push({
+                key: item.key,
+                reason: `near-duplicate of global ${candidate.kind}:${similar.id} "${similar.title}" (overlap ${similar.score.toFixed(2)}) — update that entry instead`,
+            });
+            continue;
+        }
         promotable.push(item);
     }
     return { promotable, skipped };
@@ -213,16 +251,27 @@ export function splitArchiveGuards(items, candidates) {
 }
 /**
  * Apply-time guard for a split promotion (archive + promote sub-object):
- * the cleaned title must not duplicate a topic already covered globally. A
- * duplicate split is dropped (the entry still archives plain) rather than
- * half-promoting a redundancy.
+ * the cleaned payload must pass the same promotion policy as a whole
+ * promote — no global coverage duplicate, no project-scoped content, not
+ * too thin, no near-duplicate global entry. A blocked split is dropped (the
+ * entry still archives plain) rather than half-promoting a redundancy.
  */
-export function splitPromoteBlocked(item, globalState, kind) {
+export function splitPromoteBlocked(item, globalState, kind, policy = DEFAULT_PROMOTION_POLICY) {
     if (!item.promote)
         return "no split payload";
     if (globalCoverageDetected(globalState, kind, { id: "", title: item.promote.title })) {
         return "split promotion duplicates a globally covered topic";
     }
+    const scoped = projectScopedReason(`${item.promote.title}\n${item.promote.content}`, policy);
+    if (scoped)
+        return `split promotion is ${scoped}`;
+    if (item.promote.content.length < policy.minPromoteChars) {
+        return `split promotion too thin (${item.promote.content.length} < ${policy.minPromoteChars} chars)`;
+    }
+    const similar = mostSimilarGlobalEntry(globalState, kind, item.promote.title, item.promote.content, policy);
+    if (similar) {
+        return `split promotion near-duplicates global ${kind}:${similar.id} "${similar.title}" (overlap ${similar.score.toFixed(2)})`;
+    }
     return undefined;
 }
 /**
@@ -342,12 +391,17 @@ listed entry exactly once:
   procedure or skill. Future sessions would benefit from seeing it.
 - "archive" — the content is session-specific task progress, one-off noise,
   superseded or obsolete, or already covered by the global store (note
-  "covered globally" in the reason).
+  "covered globally" in the reason), or stale (old + never injected — note
+  "stale (injectionCount=0, recency low)" in the reason).
 - "keep" — still actively useful to this session, or genuinely uncertain.
 
 Rules:
 - When an entry is marked "covered globally" in the listing, prefer "archive"
   or "keep" over "promote" — promoting a duplicate gains nothing.
+- When an entry is marked "stale" (injectionCount=0 and low recency), prefer
+  "archive" — the entry has never been used and is old, so it is unlikely to
+  be needed again. Only "keep" if the content is clearly valuable despite low
+  usage (e.g. a safety policy that rarely triggers but is critical).
 - Do not promote local task state, work-in-progress notes, or content tied to
   one session's ephemeral details.
 - Skills: only "promote" a skill entry that is a genuinely reusable procedure
@@ -387,10 +441,11 @@ export async function assessLocalEntries(ctx, agent, candidates, options = {}) {
         .map((candidate) => {
         const key = candidateKey(candidate.kind, candidate.id);
         const covered = candidate.coveredGlobally ? " (covered globally)" : "";
+        const stale = candidate.stale ? ` (stale: injectionCount=${candidate.injectionCount}, recency low)` : "";
         const hints = candidate.globalHints.length > 0
             ? ` | global≈${candidate.globalHints.map((hint) => hint.id + ":" + hint.title).join(", ")}`
             : "";
-        return `- ${key} [${candidate.path}, v${candidate.version}] "${candidate.title}"${covered}${hints}: ${compactText(candidate.content, 220)}`;
+        return `- ${key} [${candidate.path}, v${candidate.version}] "${candidate.title}"${covered}${stale}${hints}: ${compactText(candidate.content, 220)}`;
     })
         .join("\n");
     const userPrompt = [
@@ -398,42 +453,14 @@ export async function assessLocalEntries(ctx, agent, candidates, options = {}) {
         `<local_entries>\n${candidateText}\n</local_entries>`,
         "Return only JSON. Every item must reference one of the keys above.",
     ].join("\n\n");
-    const assembler = new BlockAssembler();
-    for await (const chunk of ctx.llm.stream({
+    const text = await streamText(ctx, {
         provider: agent.options.provider,
         model: agent.options.model,
         system: WRAPUP_ASSESS_SYSTEM_PROMPT,
-        messages: [
-            createUserMessage({
-                content: [{ type: "text", text: userPrompt }],
-                source: { kind: "plugin", plugin: "dsh-continual-evolve" },
-            }),
-        ],
-        // Force non-reasoning output so the budget lands in the JSON verdicts.
-        reasoningEffort: ReasoningEffortId("off"),
+        prompt: userPrompt,
         maxTokens: options.maxOutputTokens ?? 4096,
-        ...(options.signal ? { signal: options.signal } : {}),
-    })) {
-        assembler.push(chunk);
-    }
-    const finish = assembler.finish;
-    if (finish.kind === "error") {
-        throw new Error(`evolve: wrap-up assessor call failed: ${finish.failure?.message ?? "unknown"}`);
-    }
-    if (finish.kind === "aborted") {
-        throw new Error("evolve: wrap-up assessor call aborted");
-    }
-    if (finish.kind === "max-tokens") {
-        throw new Error("evolve: wrap-up assessor output budget exhausted (max-tokens)");
-    }
-    const text = assembler
-        .blocks()
-        .filter((block) => block.type === "text")
-        .map((block) => block.text)
-        .join("\n");
-    if (text.length === 0) {
-        throw new Error("evolve: wrap-up assessor produced no text");
-    }
+        signal: options.signal,
+    });
     return parseWrapupAssessment(text, candidates);
 }
 //# sourceMappingURL=wrapup.js.map

package/package.json

@@ -1,11 +1,11 @@
 {
   "name": "dsh-continual-evolve",
-  "version": "0.2.0",
+  "version": "0.4.0",
   "description": "Continual self-evolution plugin for DeepSeek Harness: versioned, auditable, rollback-safe harness state (prompt notes, memories, skills, subagent specs) refined from session trajectories.",
   "license": "MIT",
   "repository": {
     "type": "git",
-    "url": "https://github.com/ZK-Andy/dsh-continual-evolve.git"
+    "url": "git+https://github.com/ZK-Andy/dsh-continual-evolve.git"
   },
   "homepage": "https://github.com/ZK-Andy/dsh-continual-evolve",
   "bugs": {
@@ -50,6 +50,7 @@
     "dev": "tsc -p tsconfig.json --watch",
     "typecheck": "tsc -p tsconfig.json --noEmit",
     "test": "vitest run",
+    "test:coverage": "vitest run --coverage",
     "test:watch": "vitest",
     "lint": "oxlint src test",
     "clean": "rm -rf lib"
@@ -62,14 +63,15 @@
   },
   "devDependencies": {
     "@deepseek-ai/cordis": "^4.0.1",
-    "@deepseek-ai/dsh-agent": "0.1.0-rc.6",
-    "@deepseek-ai/dsh-commands": "0.1.0-rc.6",
-    "@deepseek-ai/dsh-home-paths": "0.1.0-rc.6",
-    "@deepseek-ai/dsh-llm": "0.1.0-rc.6",
-    "@deepseek-ai/dsh-system-prompt": "0.1.0-rc.6",
-    "@deepseek-ai/dsh-tools": "0.1.0-rc.6",
+    "@deepseek-ai/dsh-agent": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-commands": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-home-paths": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-llm": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-system-prompt": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-tools": "0.1.1-rc.2",
     "@deepseek-ai/schemastery": "^3.18.1",
     "@types/node": "^22.10.0",
+    "@vitest/coverage-v8": "^3.2.0",
     "oxlint": "^0.16.0",
     "typescript": "^5.9.0",
     "vitest": "^3.2.0"