@hadooppei/hwcode 1.0.11 → 1.0.12

package/.pi/dist/lib/knowledge/extractor.js

@@ -8,14 +8,25 @@ Persist only knowledge that is likely to be useful in future sessions:
 - reusable architecture, testing, debugging, deployment, or operational knowledge.
 Do not persist task summaries, guesses, temporary IDs, credentials, secrets, private data, or facts recoverable by simply reading the repository.
 Do not persist mutable inventory such as current cloud resources, names, availability, status, timestamps, or account snapshots. Preserve the discovery method or selection rule instead. If useful procedure is mixed with volatile facts, remove the volatile facts.
-Use storageHint "rule" only for short, precise, high-value instructions. Use "topic" for multi-step SOPs and detailed experience.
-Use action "revise" only when the delta explicitly corrects prior knowledge; otherwise use "add" or "reinforce".
+Treat failed, unsupported, invalid, or corrected operations as negative evidence. Never present an operation as valid when the supplied evidence says it failed; use only a verified alternative or record an explicit warning. Never invent commands or parameters absent from successful evidence.
+Use storageHint "rule" only for one short imperative instruction of at most 240 characters. A rule must not contain incident narration, example resource names, IDs, timestamps, or evidence details. Use "topic" for multi-step SOPs and detailed experience.
+Relevant existing knowledge may be supplied. If a candidate has the same intent as an existing item, do not add a translated, renamed, or reformatted duplicate. Set targetId to that exact existing ID and action to "reinforce" when the existing content remains correct, or "revise" only when the delta explicitly proves a correction. Use targetId null only for genuinely new knowledge.
 Scope may be "global" only for an explicit user directive intended across projects; otherwise use "project".
-Each summary must be one complete sentence of at most 160 characters. Keep each body under 3000 characters. Escape line breaks and quotes inside JSON strings. Set durability to "stable" only after removing facts likely to change or be cheaply rediscovered.
-Return: {"candidates":[{"key":"stable semantic key","title":"...","summary":"one complete sentence","keywords":["..."],"scope":"global|project","body":"markdown body","evidence":["verified evidence"],"confidence":0.0,"storageHint":"rule|topic","action":"add|reinforce|revise","explicitUserDirective":false,"durability":"stable"}]}
+The prior validation context is only for checking facts and contradictions; do not persist it again unless the new delta independently reinforces it. Prefer accumulated workflow successfulSteps and failedApproaches over assistant narration.
+Each summary must be one complete sentence of at most 160 characters. Keep each topic body under 2200 characters. Escape line breaks and quotes inside JSON strings. Set durability to "stable" only after removing facts likely to change or be cheaply rediscovered. Return at most 3 candidates.
+Return: {"candidates":[{"key":"stable semantic key","targetId":"existing-id or null","title":"...","summary":"one complete sentence","keywords":["..."],"scope":"global|project","body":"markdown body","evidence":["verified evidence"],"confidence":0.0,"storageHint":"rule|topic","action":"add|reinforce|revise","explicitUserDirective":false,"durability":"stable"}]}
 Return {"candidates":[]} when nothing meets the threshold.`;
-export function buildKnowledgeExtractionPrompt(projectRoot, delta) {
-    return `Project root: ${projectRoot}\n\n<conversation_delta>\n${delta}\n</conversation_delta>`;
+export function buildKnowledgeExtractionPrompt(projectRoot, delta, context = "", existing = []) {
+    const related = existing.length > 0 ? JSON.stringify(existing) : "[]";
+    return [
+        `Project root: ${projectRoot}`,
+        "",
+        "<relevant_existing_knowledge>", related, "</relevant_existing_knowledge>",
+        "",
+        "<prior_validation_context>", context, "</prior_validation_context>",
+        "",
+        "<conversation_delta>", delta, "</conversation_delta>",
+    ].join("\n");
 }
 export function parseCandidateEnvelope(text) {
     const trimmed = text.trim();

package/.pi/dist/lib/knowledge/session-scanner.js

@@ -6,6 +6,8 @@ import { KNOWLEDGE_RUNTIME_DEFAULTS } from "../runtime/defaults.js";
 import { knowledgeDeltaDigest } from "./extractor.js";
 import { sanitizeKnowledgeText } from "./sanitize.js";
 import { projectKnowledgeKey } from "./store.js";
+const FAILURE_EVIDENCE_RE = /(?:\b(?:error|failed|failure|invalid|unsupported|mistyp(?:e|ed)|typo)\b|not supported|失败|错误|不支持|无效)/iu;
+const INJECTED_SKILL_RE = /<skill\b[^>]*>[\s\S]*?<\/skill>/giu;
 function hash(value, length = 64) {
     return createHash("sha256").update(value).digest("hex").slice(0, length);
 }
@@ -23,11 +25,11 @@ function textContent(content) {
 }
 function reviewPiece(entry, index) {
     if (entry.type === "compaction" || entry.type === "branch_summary") {
-        return { index, priority: 1, category: "summary", text: `[summary]\n${sanitizeKnowledgeText(entry.summary).slice(0, 6_000)}` };
+        return { index, priority: 2, category: "summary", text: `[summary]\n${sanitizeKnowledgeText(entry.summary).slice(0, 6_000)}` };
     }
     if (entry.type === "custom" && entry.customType.startsWith("hwcode-workflow")) {
         return {
-            index, priority: 0, category: "workflow",
+            index, priority: 1, category: "workflow",
             text: `[workflow_state:${entry.customType}]\n${sanitizeKnowledgeText(JSON.stringify(entry.data ?? {})).slice(0, 6_000)}`,
         };
     }
@@ -35,10 +37,11 @@ function reviewPiece(entry, index) {
         return undefined;
     const message = entry.message;
     const role = typeof message.role === "string" ? message.role : "message";
-    const content = sanitizeKnowledgeText(textContent(message.content));
+    const content = sanitizeKnowledgeText(textContent(message.content).replace(INJECTED_SKILL_RE, "[loaded skill omitted]"));
     if (!content)
         return undefined;
-    const priority = role === "user" ? 0 : role === "toolResult" ? 3 : 2;
+    const priority = role === "user" ? 0 : role === "toolResult" && FAILURE_EVIDENCE_RE.test(content) ? 2
+        : role === "toolResult" ? 4 : 3;
     const cap = role === "user" ? 8_000 : role === "toolResult" ? 1_500 : 4_000;
     const category = role === "user" ? "user" : role === "toolResult" ? "tool" : "assistant";
     return { index, priority, category, text: `[${role}]\n${content.slice(0, cap)}` };
@@ -50,16 +53,49 @@ function buildDelta(entries) {
         user: 14_000, workflow: 6_000, summary: 4_000, assistant: 8_000, tool: 4_000,
     };
     let remaining = KNOWLEDGE_RUNTIME_DEFAULTS.review.maxDeltaChars;
-    for (const piece of pieces.slice().sort((left, right) => left.priority - right.priority || left.index - right.index)) {
+    for (const piece of pieces.slice().sort((left, right) => {
+        if (left.priority !== right.priority)
+            return left.priority - right.priority;
+        if (left.category === right.category && left.category !== "user")
+            return right.index - left.index;
+        return left.index - right.index;
+    })) {
         if (remaining <= 0)
             break;
-        const text = piece.text.slice(0, Math.min(remaining, categoryRemaining[piece.category]));
+        const separatorCharacters = selected.length > 0 ? 2 : 0;
+        if (remaining <= separatorCharacters)
+            break;
+        const text = piece.text.slice(0, Math.min(remaining - separatorCharacters, categoryRemaining[piece.category]));
         if (text)
             selected.push({ ...piece, text });
-        remaining -= text.length;
+        remaining -= text.length + separatorCharacters;
         categoryRemaining[piece.category] -= text.length;
     }
-    return selected.sort((left, right) => left.index - right.index).map((piece) => piece.text).join("\n\n");
+    return selected.sort((left, right) => left.index - right.index).map((piece) => piece.text).join("\n\n")
+        .slice(0, KNOWLEDGE_RUNTIME_DEFAULTS.review.maxDeltaChars);
+}
+function buildValidationContext(entries) {
+    const pieces = entries.map(reviewPiece).filter((piece) => Boolean(piece));
+    const latestWorkflow = pieces.filter((piece) => piece.category === "workflow").at(-1);
+    const evidence = pieces.filter((piece) => ((piece.category === "tool" || piece.category === "assistant" || piece.category === "summary")
+        && FAILURE_EVIDENCE_RE.test(piece.text)));
+    if (latestWorkflow && !evidence.includes(latestWorkflow))
+        evidence.push(latestWorkflow);
+    let remaining = KNOWLEDGE_RUNTIME_DEFAULTS.review.maxPriorContextChars;
+    const selected = [];
+    for (const piece of evidence.slice().sort((left, right) => right.index - left.index)) {
+        if (remaining <= 0)
+            break;
+        const separatorCharacters = selected.length > 0 ? 2 : 0;
+        if (remaining <= separatorCharacters)
+            break;
+        const text = piece.text.slice(0, remaining - separatorCharacters);
+        if (text)
+            selected.push({ ...piece, text });
+        remaining -= text.length + separatorCharacters;
+    }
+    return selected.sort((left, right) => left.index - right.index).map((piece) => piece.text).join("\n\n")
+        .slice(0, KNOWLEDGE_RUNTIME_DEFAULTS.review.maxPriorContextChars);
 }
 function projectRootForSession(header, branch) {
     for (const entry of branch.slice().reverse()) {
@@ -128,6 +164,7 @@ export function readReviewTask(sessionFile, cursor, now = Date.now()) {
     if (pending.length === 0)
         return undefined;
     const delta = buildDelta(pending);
+    const context = cursorIndex >= 0 ? buildValidationContext(branch.slice(0, cursorIndex + 1)) : "";
     const firstEntryId = pending[0].id;
     const lastEntryId = pending[pending.length - 1].id;
     const deltaDigest = knowledgeDeltaDigest(delta || `${firstEntryId}\0${lastEntryId}`);
@@ -138,7 +175,7 @@ export function readReviewTask(sessionFile, cursor, now = Date.now()) {
     return {
         requestId: randomUUID(), reviewKey, sessionId: header.id, sessionKey, sessionFileHash,
         projectRoot, projectKey: projectKnowledgeKey(projectRoot),
-        firstEntryId, lastEntryId, delta, deltaDigest, fileSize: stat.size, fileMtimeMs: stat.mtimeMs,
+        firstEntryId, lastEntryId, context, delta, deltaDigest, fileSize: stat.size, fileMtimeMs: stat.mtimeMs,
     };
 }
 export function findNextReviewTask(sessionsRoot, reviews, now = Date.now()) {

package/.pi/dist/lib/knowledge/store.js

@@ -101,6 +101,57 @@ function normalizeSlug(value) {
 function normalizedFingerprint(candidate) {
     return createHash("sha256").update(`${candidate.scope}\0${candidate.key.trim().toLowerCase()}`).digest("hex");
 }
+function comparisonTerms(value) {
+    return new Set((value.normalize("NFKC").toLowerCase().match(/\p{Script=Han}+|[\p{L}\p{N}]+/gu) ?? [])
+        .filter((term) => term.length >= 2));
+}
+function containment(left, right) {
+    const minimum = Math.min(left.size, right.size);
+    if (minimum === 0)
+        return 0;
+    let shared = 0;
+    for (const term of left)
+        if (right.has(term))
+            shared++;
+    return shared / minimum;
+}
+function semanticDuplicateScore(candidate, entry) {
+    if (candidate.scope !== entry.scope || candidate.storageHint !== entry.track)
+        return 0;
+    const titleScore = containment(comparisonTerms(candidate.title), comparisonTerms(entry.title));
+    const keywordScore = containment(comparisonTerms(candidate.keywords.join(" ")), comparisonTerms(entry.keywords.join(" ")));
+    if (titleScore < 0.8 || keywordScore < 0.5)
+        return 0;
+    return titleScore * 0.7 + keywordScore * 0.3;
+}
+function semanticDuplicateIndex(candidate, catalog) {
+    let bestIndex = -1;
+    let bestScore = 0;
+    for (const [index, entry] of catalog.items.entries()) {
+        const score = semanticDuplicateScore(candidate, entry);
+        if (score > bestScore) {
+            bestIndex = index;
+            bestScore = score;
+        }
+    }
+    return bestIndex;
+}
+function unsupportedOperations(value) {
+    return [...value.matchAll(/\bOperation\s+([A-Za-z][A-Za-z0-9]+)\s+is not supported\b/gu)]
+        .map((match) => match[1].toLowerCase());
+}
+function contradictsVerifiedFailures(body, validationText) {
+    const normalizedBody = body.toLowerCase();
+    for (const operation of unsupportedOperations(validationText)) {
+        const index = normalizedBody.search(new RegExp(`\\b${operation}\\b`, "u"));
+        if (index < 0)
+            continue;
+        const context = normalizedBody.slice(Math.max(0, index - 80), index + operation.length + 80);
+        if (!/(?:not supported|unsupported|do not|don't|avoid|不支持|不要|避免)/u.test(context))
+            return true;
+    }
+    return false;
+}
 function contentHash(candidate) {
     return createHash("sha256").update(`${candidate.title}\0${candidate.summary}\0${candidate.body}`).digest("hex");
 }
@@ -110,7 +161,7 @@ function asStringArray(value, maxItems) {
     return [...new Set(value.filter((item) => typeof item === "string")
             .map((item) => compactKnowledgeText(item, 240)).filter(Boolean))].slice(0, maxItems);
 }
-function normalizeCandidate(value, projectKey) {
+function normalizeCandidate(value, projectKey, validationText) {
     if (!value || typeof value !== "object")
         return undefined;
     const raw = value;
@@ -127,11 +178,16 @@ function normalizeCandidate(value, projectKey) {
     const evidence = asStringArray(raw.evidence, 8);
     if (!body || evidence.length === 0)
         return undefined;
+    if (contradictsVerifiedFailures(body, validationText))
+        return undefined;
     const requestedTrack = raw.storageHint === "rule" ? "rule" : "topic";
     const ruleEligible = body.length <= STORAGE.maxRuleChars && body.split("\n").length <= STORAGE.maxRuleFileLines
+        && body.split(/\n\s*\n/gu).length === 1
         && (explicitUserDirective || raw.confidence >= 0.9);
+    const targetId = typeof raw.targetId === "string" && SAFE_ID_RE.test(raw.targetId) ? raw.targetId : undefined;
     return {
-        key: compactKnowledgeText(raw.key, 160), title: compactKnowledgeText(raw.title, STORAGE.maxTitleChars),
+        key: compactKnowledgeText(raw.key, 160), targetId,
+        title: compactKnowledgeText(raw.title, STORAGE.maxTitleChars),
         summary: compactKnowledgeSummary(raw.summary, STORAGE.maxSummaryChars),
         keywords: asStringArray(raw.keywords, STORAGE.maxKeywordCount).map((keyword) => keyword.toLowerCase()),
         scope, body, evidence, confidence: Math.min(1, raw.confidence),
@@ -269,16 +325,30 @@ export function commitKnowledgeReview(values, task, leaderToken, home = homedir(
         const catalog = structuredClone(current.catalog);
         const result = { saved: 0, updated: 0, pending: 0, skipped: 0 };
         for (const value of values.slice(0, KNOWLEDGE_RUNTIME_DEFAULTS.review.maxCandidates)) {
-            const candidate = normalizeCandidate(value, task.projectKey);
+            const candidate = normalizeCandidate(value, task.projectKey, `${task.context ?? ""}\n${task.delta}`);
             if (!candidate) {
                 result.skipped++;
                 continue;
             }
             const fingerprint = normalizedFingerprint(candidate);
             const hash = contentHash(candidate);
-            const existingIndex = catalog.items.findIndex((item) => item.fingerprint === fingerprint);
+            const targetIndex = candidate.targetId
+                ? catalog.items.findIndex((item) => item.id === candidate.targetId && applicable(item, task.projectKey))
+                : -1;
+            if (candidate.targetId && targetIndex < 0) {
+                result.skipped++;
+                continue;
+            }
+            let existingIndex = targetIndex >= 0
+                ? targetIndex
+                : catalog.items.findIndex((item) => item.fingerprint === fingerprint);
+            let semanticReinforcement = false;
+            if (existingIndex < 0) {
+                existingIndex = semanticDuplicateIndex(candidate, catalog);
+                semanticReinforcement = existingIndex >= 0;
+            }
             const existing = existingIndex >= 0 ? catalog.items[existingIndex] : undefined;
-            if (existing?.contentHash === hash) {
+            if (existing && (existing.contentHash === hash || candidate.action === "reinforce" || semanticReinforcement)) {
                 existing.evidenceCount += candidate.evidence.length;
                 existing.updatedAt = new Date().toISOString();
                 result.updated++;
@@ -301,8 +371,9 @@ export function commitKnowledgeReview(values, task, leaderToken, home = homedir(
             const relativeFile = `${track === "rule" ? "rules" : "topics"}/${id}.md`;
             atomicWrite(join(temporary, relativeFile), renderKnowledgeFile(candidate));
             const entry = {
-                id, fingerprint, contentHash: hash, title: candidate.title, summary: candidate.summary,
-                keywords: candidate.keywords, scope: candidate.scope, track, file: relativeFile,
+                id, fingerprint: existing?.fingerprint ?? fingerprint, contentHash: hash,
+                title: candidate.title, summary: candidate.summary,
+                keywords: candidate.keywords, scope: existing?.scope ?? candidate.scope, track, file: relativeFile,
                 evidenceCount: (existing?.evidenceCount ?? 0) + candidate.evidence.length,
                 createdAt: existing?.createdAt ?? now, updatedAt: now,
             };

package/.pi/dist/lib/runtime/defaults.js

@@ -31,13 +31,17 @@ export const KNOWLEDGE_RUNTIME_DEFAULTS = Object.freeze({
         idleMs: 60_000,
         capabilityPollMs: 5_000,
         modelTimeoutMs: 180_000,
-        maxDeltaChars: 20_000,
-        maxCandidates: 5,
+        maxDeltaChars: 12_000,
+        maxPriorContextChars: 4_000,
+        maxExistingContextItems: 5,
+        maxExistingBodyItems: 2,
+        maxExistingBodyChars: 600,
+        maxCandidates: 3,
         minimumConfidence: 0.72,
     }),
     storage: Object.freeze({
-        maxRuleChars: 600,
-        maxRuleFileLines: 40,
+        maxRuleChars: 320,
+        maxRuleFileLines: 6,
         maxRulesPromptChars: 12_000,
         maxMemoryChars: 25_000,
         maxMemoryLines: 200,

package/.pi/extensions/knowledge.ts

@@ -8,7 +8,7 @@ import { Type } from "@earendil-works/pi-ai";
 import { defineTool, type ExtensionAPI, type ExtensionContext } from "@earendil-works/pi-coding-agent";
 
 import {
-  buildKnowledgeExtractionPrompt, KNOWLEDGE_EXTRACTION_SYSTEM_PROMPT,
+  buildKnowledgeExtractionPrompt, type ExistingKnowledgeContext, KNOWLEDGE_EXTRACTION_SYSTEM_PROMPT,
 } from "../lib/knowledge/extractor.ts";
 import { matchKnowledge } from "../lib/knowledge/matcher.ts";
 import { loadKnowledgeById, loadKnowledgeSnapshot, projectKnowledgeKey } from "../lib/knowledge/store.ts";
@@ -103,13 +103,39 @@ async function runReview(message: Extract<KnowledgeWorkerOutput, { type: "review
     runtime.activeReview = { requestId: message.requestId, leaderToken: message.leaderToken, controller };
     timeout = setTimeout(() => controller.abort(), KNOWLEDGE_RUNTIME_DEFAULTS.review.modelTimeoutMs);
     timeout.unref();
+    const snapshot = loadKnowledgeSnapshot(message.task.projectKey);
+    const related: ExistingKnowledgeContext[] = matchKnowledge(
+      message.task.delta,
+      snapshot.catalog,
+      KNOWLEDGE_RUNTIME_DEFAULTS.review.maxExistingContextItems,
+    ).map((entry, index) => {
+      const reference: ExistingKnowledgeContext = {
+        id: entry.id,
+        title: entry.title,
+        summary: entry.summary,
+        keywords: entry.keywords,
+        track: entry.track,
+      };
+      if (index >= KNOWLEDGE_RUNTIME_DEFAULTS.review.maxExistingBodyItems) return reference;
+      const content = loadKnowledgeById(entry.id, message.task.projectKey)?.content
+        .slice(0, KNOWLEDGE_RUNTIME_DEFAULTS.review.maxExistingBodyChars);
+      return content ? { ...reference, content } : reference;
+    });
     const response = await ctx.modelRegistry.complete(
       ctx.model,
       {
         systemPrompt: KNOWLEDGE_EXTRACTION_SYSTEM_PROMPT,
         messages: [{
           role: "user",
-          content: [{ type: "text", text: buildKnowledgeExtractionPrompt(message.task.projectRoot, message.task.delta) }],
+          content: [{
+            type: "text",
+            text: buildKnowledgeExtractionPrompt(
+              message.task.projectRoot,
+              message.task.delta,
+              message.task.context,
+              related,
+            ),
+          }],
           timestamp: Date.now(),
         }],
       },

package/.pi/lib/knowledge/extractor.ts

@@ -2,6 +2,15 @@ import { createHash } from "node:crypto";
 
 import { KNOWLEDGE_RUNTIME_DEFAULTS } from "../runtime/defaults.ts";
 
+export interface ExistingKnowledgeContext {
+  id: string;
+  title: string;
+  summary: string;
+  keywords: string[];
+  track: "rule" | "topic";
+  content?: string;
+}
+
 export const KNOWLEDGE_EXTRACTION_SYSTEM_PROMPT = `You are HWCode's background knowledge reviewer.
 Review only the supplied conversation delta. Return strict JSON and no markdown.
 Persist only knowledge that is likely to be useful in future sessions:
@@ -10,15 +19,31 @@ Persist only knowledge that is likely to be useful in future sessions:
 - reusable architecture, testing, debugging, deployment, or operational knowledge.
 Do not persist task summaries, guesses, temporary IDs, credentials, secrets, private data, or facts recoverable by simply reading the repository.
 Do not persist mutable inventory such as current cloud resources, names, availability, status, timestamps, or account snapshots. Preserve the discovery method or selection rule instead. If useful procedure is mixed with volatile facts, remove the volatile facts.
-Use storageHint "rule" only for short, precise, high-value instructions. Use "topic" for multi-step SOPs and detailed experience.
-Use action "revise" only when the delta explicitly corrects prior knowledge; otherwise use "add" or "reinforce".
+Treat failed, unsupported, invalid, or corrected operations as negative evidence. Never present an operation as valid when the supplied evidence says it failed; use only a verified alternative or record an explicit warning. Never invent commands or parameters absent from successful evidence.
+Use storageHint "rule" only for one short imperative instruction of at most 240 characters. A rule must not contain incident narration, example resource names, IDs, timestamps, or evidence details. Use "topic" for multi-step SOPs and detailed experience.
+Relevant existing knowledge may be supplied. If a candidate has the same intent as an existing item, do not add a translated, renamed, or reformatted duplicate. Set targetId to that exact existing ID and action to "reinforce" when the existing content remains correct, or "revise" only when the delta explicitly proves a correction. Use targetId null only for genuinely new knowledge.
 Scope may be "global" only for an explicit user directive intended across projects; otherwise use "project".
-Each summary must be one complete sentence of at most 160 characters. Keep each body under 3000 characters. Escape line breaks and quotes inside JSON strings. Set durability to "stable" only after removing facts likely to change or be cheaply rediscovered.
-Return: {"candidates":[{"key":"stable semantic key","title":"...","summary":"one complete sentence","keywords":["..."],"scope":"global|project","body":"markdown body","evidence":["verified evidence"],"confidence":0.0,"storageHint":"rule|topic","action":"add|reinforce|revise","explicitUserDirective":false,"durability":"stable"}]}
+The prior validation context is only for checking facts and contradictions; do not persist it again unless the new delta independently reinforces it. Prefer accumulated workflow successfulSteps and failedApproaches over assistant narration.
+Each summary must be one complete sentence of at most 160 characters. Keep each topic body under 2200 characters. Escape line breaks and quotes inside JSON strings. Set durability to "stable" only after removing facts likely to change or be cheaply rediscovered. Return at most 3 candidates.
+Return: {"candidates":[{"key":"stable semantic key","targetId":"existing-id or null","title":"...","summary":"one complete sentence","keywords":["..."],"scope":"global|project","body":"markdown body","evidence":["verified evidence"],"confidence":0.0,"storageHint":"rule|topic","action":"add|reinforce|revise","explicitUserDirective":false,"durability":"stable"}]}
 Return {"candidates":[]} when nothing meets the threshold.`;
 
-export function buildKnowledgeExtractionPrompt(projectRoot: string, delta: string): string {
-  return `Project root: ${projectRoot}\n\n<conversation_delta>\n${delta}\n</conversation_delta>`;
+export function buildKnowledgeExtractionPrompt(
+  projectRoot: string,
+  delta: string,
+  context = "",
+  existing: ExistingKnowledgeContext[] = [],
+): string {
+  const related = existing.length > 0 ? JSON.stringify(existing) : "[]";
+  return [
+    `Project root: ${projectRoot}`,
+    "",
+    "<relevant_existing_knowledge>", related, "</relevant_existing_knowledge>",
+    "",
+    "<prior_validation_context>", context, "</prior_validation_context>",
+    "",
+    "<conversation_delta>", delta, "</conversation_delta>",
+  ].join("\n");
 }
 
 export function parseCandidateEnvelope(text: string): unknown[] {

package/.pi/lib/knowledge/session-scanner.ts

@@ -19,6 +19,9 @@ interface ReviewPiece {
   text: string;
 }
 
+const FAILURE_EVIDENCE_RE = /(?:\b(?:error|failed|failure|invalid|unsupported|mistyp(?:e|ed)|typo)\b|not supported|失败|错误|不支持|无效)/iu;
+const INJECTED_SKILL_RE = /<skill\b[^>]*>[\s\S]*?<\/skill>/giu;
+
 function hash(value: string, length = 64): string {
   return createHash("sha256").update(value).digest("hex").slice(0, length);
 }
@@ -35,20 +38,21 @@ function textContent(content: unknown): string {
 
 function reviewPiece(entry: SessionEntry, index: number): ReviewPiece | undefined {
   if (entry.type === "compaction" || entry.type === "branch_summary") {
-    return { index, priority: 1, category: "summary", text: `[summary]\n${sanitizeKnowledgeText(entry.summary).slice(0, 6_000)}` };
+    return { index, priority: 2, category: "summary", text: `[summary]\n${sanitizeKnowledgeText(entry.summary).slice(0, 6_000)}` };
   }
   if (entry.type === "custom" && entry.customType.startsWith("hwcode-workflow")) {
     return {
-      index, priority: 0, category: "workflow",
+      index, priority: 1, category: "workflow",
       text: `[workflow_state:${entry.customType}]\n${sanitizeKnowledgeText(JSON.stringify(entry.data ?? {})).slice(0, 6_000)}`,
     };
   }
   if (entry.type !== "message" || !entry.message || typeof entry.message !== "object") return undefined;
   const message = entry.message as unknown as Record<string, unknown>;
   const role = typeof message.role === "string" ? message.role : "message";
-  const content = sanitizeKnowledgeText(textContent(message.content));
+  const content = sanitizeKnowledgeText(textContent(message.content).replace(INJECTED_SKILL_RE, "[loaded skill omitted]"));
   if (!content) return undefined;
-  const priority = role === "user" ? 0 : role === "toolResult" ? 3 : 2;
+  const priority = role === "user" ? 0 : role === "toolResult" && FAILURE_EVIDENCE_RE.test(content) ? 2
+    : role === "toolResult" ? 4 : 3;
   const cap = role === "user" ? 8_000 : role === "toolResult" ? 1_500 : 4_000;
   const category = role === "user" ? "user" : role === "toolResult" ? "tool" : "assistant";
   return { index, priority, category, text: `[${role}]\n${content.slice(0, cap)}` };
@@ -61,14 +65,43 @@ function buildDelta(entries: SessionEntry[]): string {
     user: 14_000, workflow: 6_000, summary: 4_000, assistant: 8_000, tool: 4_000,
   };
   let remaining = KNOWLEDGE_RUNTIME_DEFAULTS.review.maxDeltaChars;
-  for (const piece of pieces.slice().sort((left, right) => left.priority - right.priority || left.index - right.index)) {
+  for (const piece of pieces.slice().sort((left, right) => {
+    if (left.priority !== right.priority) return left.priority - right.priority;
+    if (left.category === right.category && left.category !== "user") return right.index - left.index;
+    return left.index - right.index;
+  })) {
     if (remaining <= 0) break;
-    const text = piece.text.slice(0, Math.min(remaining, categoryRemaining[piece.category]));
+    const separatorCharacters = selected.length > 0 ? 2 : 0;
+    if (remaining <= separatorCharacters) break;
+    const text = piece.text.slice(0, Math.min(remaining - separatorCharacters, categoryRemaining[piece.category]));
     if (text) selected.push({ ...piece, text });
-    remaining -= text.length;
+    remaining -= text.length + separatorCharacters;
     categoryRemaining[piece.category] -= text.length;
   }
-  return selected.sort((left, right) => left.index - right.index).map((piece) => piece.text).join("\n\n");
+  return selected.sort((left, right) => left.index - right.index).map((piece) => piece.text).join("\n\n")
+    .slice(0, KNOWLEDGE_RUNTIME_DEFAULTS.review.maxDeltaChars);
+}
+
+function buildValidationContext(entries: SessionEntry[]): string {
+  const pieces = entries.map(reviewPiece).filter((piece): piece is ReviewPiece => Boolean(piece));
+  const latestWorkflow = pieces.filter((piece) => piece.category === "workflow").at(-1);
+  const evidence = pieces.filter((piece) => (
+    (piece.category === "tool" || piece.category === "assistant" || piece.category === "summary")
+    && FAILURE_EVIDENCE_RE.test(piece.text)
+  ));
+  if (latestWorkflow && !evidence.includes(latestWorkflow)) evidence.push(latestWorkflow);
+  let remaining = KNOWLEDGE_RUNTIME_DEFAULTS.review.maxPriorContextChars;
+  const selected: ReviewPiece[] = [];
+  for (const piece of evidence.slice().sort((left, right) => right.index - left.index)) {
+    if (remaining <= 0) break;
+    const separatorCharacters = selected.length > 0 ? 2 : 0;
+    if (remaining <= separatorCharacters) break;
+    const text = piece.text.slice(0, remaining - separatorCharacters);
+    if (text) selected.push({ ...piece, text });
+    remaining -= text.length + separatorCharacters;
+  }
+  return selected.sort((left, right) => left.index - right.index).map((piece) => piece.text).join("\n\n")
+    .slice(0, KNOWLEDGE_RUNTIME_DEFAULTS.review.maxPriorContextChars);
 }
 
 function projectRootForSession(header: SessionHeader, branch: SessionEntry[]): string {
@@ -120,6 +153,7 @@ export function readReviewTask(
   const pending = branch.slice(cursorIndex + 1);
   if (pending.length === 0) return undefined;
   const delta = buildDelta(pending);
+  const context = cursorIndex >= 0 ? buildValidationContext(branch.slice(0, cursorIndex + 1)) : "";
   const firstEntryId = pending[0].id;
   const lastEntryId = pending[pending.length - 1].id;
   const deltaDigest = knowledgeDeltaDigest(delta || `${firstEntryId}\0${lastEntryId}`);
@@ -130,7 +164,7 @@ export function readReviewTask(
   return {
     requestId: randomUUID(), reviewKey, sessionId: header.id, sessionKey, sessionFileHash,
     projectRoot, projectKey: projectKnowledgeKey(projectRoot),
-    firstEntryId, lastEntryId, delta, deltaDigest, fileSize: stat.size, fileMtimeMs: stat.mtimeMs,
+    firstEntryId, lastEntryId, context, delta, deltaDigest, fileSize: stat.size, fileMtimeMs: stat.mtimeMs,
   };
 }
 

package/.pi/lib/knowledge/store.ts

@@ -117,6 +117,56 @@ function normalizedFingerprint(candidate: KnowledgeCandidate): string {
   return createHash("sha256").update(`${candidate.scope}\0${candidate.key.trim().toLowerCase()}`).digest("hex");
 }
 
+function comparisonTerms(value: string): Set<string> {
+  return new Set((value.normalize("NFKC").toLowerCase().match(/\p{Script=Han}+|[\p{L}\p{N}]+/gu) ?? [])
+    .filter((term) => term.length >= 2));
+}
+
+function containment(left: Set<string>, right: Set<string>): number {
+  const minimum = Math.min(left.size, right.size);
+  if (minimum === 0) return 0;
+  let shared = 0;
+  for (const term of left) if (right.has(term)) shared++;
+  return shared / minimum;
+}
+
+function semanticDuplicateScore(candidate: KnowledgeCandidate, entry: KnowledgeCatalogEntry): number {
+  if (candidate.scope !== entry.scope || candidate.storageHint !== entry.track) return 0;
+  const titleScore = containment(comparisonTerms(candidate.title), comparisonTerms(entry.title));
+  const keywordScore = containment(
+    comparisonTerms(candidate.keywords.join(" ")),
+    comparisonTerms(entry.keywords.join(" ")),
+  );
+  if (titleScore < 0.8 || keywordScore < 0.5) return 0;
+  return titleScore * 0.7 + keywordScore * 0.3;
+}
+
+function semanticDuplicateIndex(candidate: KnowledgeCandidate, catalog: KnowledgeCatalog): number {
+  let bestIndex = -1;
+  let bestScore = 0;
+  for (const [index, entry] of catalog.items.entries()) {
+    const score = semanticDuplicateScore(candidate, entry);
+    if (score > bestScore) { bestIndex = index; bestScore = score; }
+  }
+  return bestIndex;
+}
+
+function unsupportedOperations(value: string): string[] {
+  return [...value.matchAll(/\bOperation\s+([A-Za-z][A-Za-z0-9]+)\s+is not supported\b/gu)]
+    .map((match) => match[1]!.toLowerCase());
+}
+
+function contradictsVerifiedFailures(body: string, validationText: string): boolean {
+  const normalizedBody = body.toLowerCase();
+  for (const operation of unsupportedOperations(validationText)) {
+    const index = normalizedBody.search(new RegExp(`\\b${operation}\\b`, "u"));
+    if (index < 0) continue;
+    const context = normalizedBody.slice(Math.max(0, index - 80), index + operation.length + 80);
+    if (!/(?:not supported|unsupported|do not|don't|avoid|不支持|不要|避免)/u.test(context)) return true;
+  }
+  return false;
+}
+
 function contentHash(candidate: KnowledgeCandidate): string {
   return createHash("sha256").update(`${candidate.title}\0${candidate.summary}\0${candidate.body}`).digest("hex");
 }
@@ -127,7 +177,7 @@ function asStringArray(value: unknown, maxItems: number): string[] {
     .map((item) => compactKnowledgeText(item, 240)).filter(Boolean))].slice(0, maxItems);
 }
 
-function normalizeCandidate(value: unknown, projectKey: string): KnowledgeCandidate | undefined {
+function normalizeCandidate(value: unknown, projectKey: string, validationText: string): KnowledgeCandidate | undefined {
   if (!value || typeof value !== "object") return undefined;
   const raw = value as Record<string, unknown>;
   if (typeof raw.key !== "string" || typeof raw.title !== "string" || typeof raw.summary !== "string"
@@ -139,11 +189,15 @@ function normalizeCandidate(value: unknown, projectKey: string): KnowledgeCandid
   const body = compactKnowledgeText(raw.body, 20_000);
   const evidence = asStringArray(raw.evidence, 8);
   if (!body || evidence.length === 0) return undefined;
+  if (contradictsVerifiedFailures(body, validationText)) return undefined;
   const requestedTrack: KnowledgeTrack = raw.storageHint === "rule" ? "rule" : "topic";
   const ruleEligible = body.length <= STORAGE.maxRuleChars && body.split("\n").length <= STORAGE.maxRuleFileLines
+    && body.split(/\n\s*\n/gu).length === 1
     && (explicitUserDirective || raw.confidence >= 0.9);
+  const targetId = typeof raw.targetId === "string" && SAFE_ID_RE.test(raw.targetId) ? raw.targetId : undefined;
   return {
-    key: compactKnowledgeText(raw.key, 160), title: compactKnowledgeText(raw.title, STORAGE.maxTitleChars),
+    key: compactKnowledgeText(raw.key, 160), targetId,
+    title: compactKnowledgeText(raw.title, STORAGE.maxTitleChars),
     summary: compactKnowledgeSummary(raw.summary, STORAGE.maxSummaryChars),
     keywords: asStringArray(raw.keywords, STORAGE.maxKeywordCount).map((keyword) => keyword.toLowerCase()),
     scope, body, evidence, confidence: Math.min(1, raw.confidence),
@@ -275,13 +329,24 @@ export function commitKnowledgeReview(
     const catalog: KnowledgeCatalog = structuredClone(current.catalog);
     const result: PersistKnowledgeResult = { saved: 0, updated: 0, pending: 0, skipped: 0 };
     for (const value of values.slice(0, KNOWLEDGE_RUNTIME_DEFAULTS.review.maxCandidates)) {
-      const candidate = normalizeCandidate(value, task.projectKey);
+      const candidate = normalizeCandidate(value, task.projectKey, `${task.context ?? ""}\n${task.delta}`);
       if (!candidate) { result.skipped++; continue; }
       const fingerprint = normalizedFingerprint(candidate);
       const hash = contentHash(candidate);
-      const existingIndex = catalog.items.findIndex((item) => item.fingerprint === fingerprint);
+      const targetIndex = candidate.targetId
+        ? catalog.items.findIndex((item) => item.id === candidate.targetId && applicable(item, task.projectKey))
+        : -1;
+      if (candidate.targetId && targetIndex < 0) { result.skipped++; continue; }
+      let existingIndex = targetIndex >= 0
+        ? targetIndex
+        : catalog.items.findIndex((item) => item.fingerprint === fingerprint);
+      let semanticReinforcement = false;
+      if (existingIndex < 0) {
+        existingIndex = semanticDuplicateIndex(candidate, catalog);
+        semanticReinforcement = existingIndex >= 0;
+      }
       const existing = existingIndex >= 0 ? catalog.items[existingIndex] : undefined;
-      if (existing?.contentHash === hash) {
+      if (existing && (existing.contentHash === hash || candidate.action === "reinforce" || semanticReinforcement)) {
         existing.evidenceCount += candidate.evidence.length;
         existing.updatedAt = new Date().toISOString();
         result.updated++;
@@ -304,8 +369,9 @@ export function commitKnowledgeReview(
       const relativeFile = `${track === "rule" ? "rules" : "topics"}/${id}.md`;
       atomicWrite(join(temporary, relativeFile), renderKnowledgeFile(candidate));
       const entry: KnowledgeCatalogEntry = {
-        id, fingerprint, contentHash: hash, title: candidate.title, summary: candidate.summary,
-        keywords: candidate.keywords, scope: candidate.scope, track, file: relativeFile,
+        id, fingerprint: existing?.fingerprint ?? fingerprint, contentHash: hash,
+        title: candidate.title, summary: candidate.summary,
+        keywords: candidate.keywords, scope: existing?.scope ?? candidate.scope, track, file: relativeFile,
         evidenceCount: (existing?.evidenceCount ?? 0) + candidate.evidence.length,
         createdAt: existing?.createdAt ?? now, updatedAt: now,
       };

package/.pi/lib/knowledge/types.ts

@@ -5,6 +5,7 @@ export type KnowledgeDurability = "stable";
 
 export interface KnowledgeCandidate {
   key: string;
+  targetId?: string;
   title: string;
   summary: string;
   keywords: string[];
@@ -86,6 +87,7 @@ export interface KnowledgeReviewTask {
   projectKey: string;
   firstEntryId: string;
   lastEntryId: string;
+  context?: string;
   delta: string;
   deltaDigest: string;
   fileSize: number;

package/.pi/lib/runtime/defaults.ts

@@ -32,13 +32,17 @@ export const KNOWLEDGE_RUNTIME_DEFAULTS = Object.freeze({
     idleMs: 60_000,
     capabilityPollMs: 5_000,
     modelTimeoutMs: 180_000,
-    maxDeltaChars: 20_000,
-    maxCandidates: 5,
+    maxDeltaChars: 12_000,
+    maxPriorContextChars: 4_000,
+    maxExistingContextItems: 5,
+    maxExistingBodyItems: 2,
+    maxExistingBodyChars: 600,
+    maxCandidates: 3,
     minimumConfidence: 0.72,
   }),
   storage: Object.freeze({
-    maxRuleChars: 600,
-    maxRuleFileLines: 40,
+    maxRuleChars: 320,
+    maxRuleFileLines: 6,
     maxRulesPromptChars: 12_000,
     maxMemoryChars: 25_000,
     maxMemoryLines: 200,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hadooppei/hwcode",
-  "version": "1.0.11",
+  "version": "1.0.12",
   "description": "A customizable terminal coding agent with local-model support and HWCode workflows.",
   "type": "module",
   "bin": {