@hadooppei/hwcode 1.0.20 → 1.0.24

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

@@ -81,8 +81,11 @@ function boundedExcerpt(value, limit = 900) {
     if (clean.length <= limit)
         return clean;
     const lines = clean.split("\n");
-    const significant = lines.filter((line) => FAILURE_RE.test(line) || VERIFICATION_RE.test(line));
-    const selected = [...lines.slice(0, 3), ...significant.slice(0, 5), ...lines.slice(-3)];
+    const priority = (line) => REVERSAL_RE.test(line) ? 0
+        : FAILURE_RE.test(line) && VERIFICATION_RE.test(line) ? 1
+            : FAILURE_RE.test(line) ? 2 : 3;
+    const significant = lines.filter((line) => (REVERSAL_RE.test(line) || FAILURE_RE.test(line) || VERIFICATION_RE.test(line))).sort((left, right) => priority(left) - priority(right));
+    const selected = [...lines.slice(0, 2), ...significant.slice(0, 7), ...lines.slice(-2)];
     return [...new Set(selected)].join("\n").slice(0, limit);
 }
 function compactArguments(args) {
@@ -152,10 +155,8 @@ function entryIdOf(entry, index) {
 function evidenceId(entryId, suffix) {
     return `ev-${digest(`${entryId}\0${suffix}`)}`;
 }
-function outcomeFor(message, text) {
-    if (message.isError === true || FAILURE_RE.test(text))
-        return "failure";
-    return "success";
+function outcomeFor(message) {
+    return message.isError === true ? "failure" : "success";
 }
 function userKind(text) {
     return CORRECTION_RE.test(text) ? "user-correction" : "user-directive";
@@ -206,11 +207,11 @@ export function buildKnowledgeEvidence(entries) {
             // lookup result as verification would let knowledge reinforce itself merely by loading it.
             if (call?.name === "hwcode_knowledge_lookup")
                 return;
-            const outcome = outcomeFor(message, text);
+            const outcome = outcomeFor(message);
             const callText = call ? `[call] ${call.name}: ${compactArguments(call.args)}\n` : "";
             evidence.push({
                 id: evidenceId(entryId, toolCallId || "tool"), entryId, timestamp,
-                kind: outcome === "failure" ? "tool-failure" : VERIFICATION_RE.test(text) ? "verification" : "tool-success",
+                kind: outcome === "failure" ? "tool-failure" : "tool-success",
                 subject: evidenceSubject(call, text), operation: operationName(call), outcome,
                 excerpt: boundedExcerpt(`${callText}[result] ${text}`),
             });
@@ -219,7 +220,7 @@ export function buildKnowledgeEvidence(entries) {
         if (role === "assistant" && text && (FAILURE_RE.test(text) || VERIFICATION_RE.test(text) || REVERSAL_RE.test(text))) {
             evidence.push({
                 id: evidenceId(entryId, "assistant"), entryId, timestamp,
-                kind: VERIFICATION_RE.test(text) ? "verification" : "assistant-claim",
+                kind: "assistant-claim",
                 subject: "Assistant diagnosis or conclusion", outcome: "unknown", excerpt: boundedExcerpt(text, 800),
             });
         }

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

@@ -14,9 +14,11 @@ Treat failed, unsupported, invalid, or corrected operations as negative 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.
 Rank knowledge by future time saved, repeated failures, hypothesis reversals, user corrections, verified resolution, reuse, and novelty. Do not favor the final task result merely because it completed the objective. Every supplied episode scoring 45 or higher must have an episodeDecisions entry, even when skipped.
 Relevant existing knowledge may be supplied. Cheap recall only proposes related items; it does not prove duplication. Compare execution environment, symptom, and root cause. Use action "reinforce" only when the exact target already covers every claim and there is no new content. Use "extend" for new compatible claims, "correct" when direct evidence disproves old content, "retire" when the whole target is invalid, and "add" for a distinct environment or root cause. Never turn lexical similarity into reinforcement.
+Prefer the most specific existing topic that already covers a claim over a broad conventions or SOP topic.
 Scope may be "global" only for an explicit user directive intended across projects; otherwise use "project".
 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.
 Every technical claim must have a stable claim key and cite evidenceIds from the supplied evidence ledger. Tool success, tool failure, and verification evidence outrank assistant narration. Do not support a service-wide or universal claim with one operation-specific failure. Evidence for an unrelated parameter error cannot support an API-version claim.
+Keep each candidate atomic: put separable claim sets in separate candidates, omit any claim without direct cited evidence, and prioritize the highest-value supported candidates. One unsupported claim rejects its entire candidate.
 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.
 The body must not repeat the title as a Markdown H1. The storage renderer supplies the H1. Replace any necessary example resource value with an obvious placeholder instead of a real project, namespace, image, network, or resource name.
 Return: {"episodeDecisions":[{"episodeId":"ep-id","decision":"add|reinforce|extend|correct|retire|skip","targetId":"existing-id or omitted","reason":"..."}],"candidates":[{"key":"stable semantic key","identityKey":"platform/runtime/problem/root-cause","targetId":"existing-id or null","episodeIds":["ep-id"],"title":"...","summary":"one complete sentence","keywords":["..."],"scope":"global|project","body":"markdown body or only the new section for extend","claims":[{"key":"stable claim key","text":"one precise claim","evidenceIds":["ev-id"]}],"evidenceIds":["ev-id"],"confidence":0.0,"storageHint":"rule|topic","action":"add|reinforce|extend|correct|retire","explicitUserDirective":false,"durability":"stable"}]}

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

@@ -30,6 +30,7 @@ export const KNOWLEDGE_RUNTIME_DEFAULTS = Object.freeze({
         intervalMs: 60_000,
         idleMs: 60_000,
         capabilityPollMs: 5_000,
+        requestTimeoutMs: 45_000,
         modelTimeoutMs: 180_000,
         maxAttempts: 2,
         maxFailureDetailChars: 4_000,
@@ -39,7 +40,7 @@ export const KNOWLEDGE_RUNTIME_DEFAULTS = Object.freeze({
         maxEvidenceItems: 24,
         maxEvidenceChars: 12_000,
         mandatoryEpisodeScore: 45,
-        maxExistingContextItems: 5,
+        maxExistingContextItems: 8,
         maxExistingBodyItems: 2,
         maxExistingBodyChars: 600,
         maxCandidates: 3,

package/.pi/extensions/knowledge.ts

@@ -1,4 +1,3 @@
-import { randomUUID } from "node:crypto";
 import { existsSync } from "node:fs";
 import { dirname } from "node:path";
 import { fileURLToPath } from "node:url";
@@ -54,7 +53,7 @@ function compactReviewEvidence(task: KnowledgeReviewTask): {
   let characters = 0;
   const add = (item: KnowledgeEvidence | undefined): void => {
     if (!item || seen.has(item.id) || selected.length >= KNOWLEDGE_RUNTIME_DEFAULTS.review.maxEvidenceItems) return;
-    const compact = { ...item, excerpt: item.excerpt.slice(0, 500) };
+    const compact = { ...item, excerpt: item.excerpt.slice(0, 800) };
     const size = JSON.stringify(compact).length;
     if (characters + size > KNOWLEDGE_RUNTIME_DEFAULTS.review.maxEvidenceChars) return;
     selected.push(compact);
@@ -183,13 +182,20 @@ async function runReview(message: Extract<KnowledgeWorkerOutput, { type: "review
       seenRelated.add(entry.id);
     };
     const episodeMatches = (message.task.episodes ?? []).map((episode) => (
-      matchKnowledge(`${episode.title}\n${episode.query}`, applicableCatalog, 2)
+      matchKnowledge(
+        `${episode.title}\n${episode.query}`,
+        applicableCatalog,
+        KNOWLEDGE_RUNTIME_DEFAULTS.review.maxExistingContextItems,
+      )
     ));
-    for (const matches of episodeMatches.slice(0, 3)) addRelated(matches[0]);
+    for (const matches of episodeMatches) addRelated(matches[0]);
     for (const id of message.task.loadedKnowledgeIds ?? []) {
       addRelated(applicableCatalog.items.find((entry) => entry.id === id));
     }
-    for (const matches of episodeMatches) for (const entry of matches.slice(1)) addRelated(entry);
+    const matchDepth = Math.max(0, ...episodeMatches.map((matches) => matches.length));
+    for (let rank = 1; rank < matchDepth; rank++) {
+      for (const matches of episodeMatches) addRelated(matches[rank]);
+    }
     for (const entry of matchKnowledge(
       message.task.recallQuery || message.task.delta,
       applicableCatalog,
@@ -219,7 +225,7 @@ async function runReview(message: Extract<KnowledgeWorkerOutput, { type: "review
       promptContext.evidence,
       promptContext.episodes,
     );
-    review.diagnostic = [
+    const diagnostic = [
       `reviewKey=${message.task.reviewKey}`,
       `model=${ctx.model.provider}/${ctx.model.id}`,
       `promptChars=${prompt.length}`,
@@ -227,7 +233,12 @@ async function runReview(message: Extract<KnowledgeWorkerOutput, { type: "review
       `evidenceItems=${promptContext.evidence.length}`,
       `episodeItems=${promptContext.episodes.length}`,
       `existingItems=${related.length}`,
-    ].join("; ");
+    ];
+    const requestStartedAt = Date.now();
+    const setRequestPhase = (phase: string): void => {
+      review!.diagnostic = [...diagnostic, `phase=${phase}`, `requestElapsedMs=${Date.now() - requestStartedAt}`].join("; ");
+    };
+    setRequestPhase("preparing-request");
     const response = await ctx.modelRegistry.complete(
       ctx.model,
       {
@@ -241,11 +252,32 @@ async function runReview(message: Extract<KnowledgeWorkerOutput, { type: "review
           timestamp: Date.now(),
         }],
       },
-      { signal: controller.signal, reasoningEffort: "low", cacheRetention: "none", sessionId: randomUUID() },
+      {
+        signal: controller.signal,
+        reasoningEffort: "low",
+        cacheRetention: "none",
+        temperature: 0,
+        timeoutMs: KNOWLEDGE_RUNTIME_DEFAULTS.review.requestTimeoutMs,
+        maxRetries: 0,
+        onPayload: () => { setRequestPhase("awaiting-http-response"); },
+        onResponse: (providerResponse: { status: number }) => {
+          setRequestPhase(`streaming-response-http-${providerResponse.status}`);
+        },
+      },
     );
     if (controller.signal.aborted) throw new Error(review.abortReason ?? "knowledge-review-interrupted");
+    if (response.stopReason === "error" || response.stopReason === "aborted") {
+      throw new Error(response.errorMessage ?? `knowledge-review-model-${response.stopReason}`);
+    }
+    const responseDetail = [
+      `stopReason=${response.stopReason}`,
+      `contentTypes=${response.content.map((item) => item.type).join(",") || "none"}`,
+      `outputTokens=${response.usage.output}`,
+    ].join("; ");
+    if (response.stopReason === "length") throw new Error(`knowledge-review-output-truncated; ${responseDetail}`);
     const raw = response.content.filter((item): item is { type: "text"; text: string } => item.type === "text")
       .map((item) => item.text).join("\n");
+    if (!raw.trim()) throw new Error(`knowledge-review-model-returned-no-text; ${responseDetail}`);
     post({ type: "review_result", leaderToken: message.leaderToken, requestId: message.requestId, raw });
   } catch (error) {
     const baseReason = review?.controller.signal.aborted

package/.pi/extensions/model-providers.ts

@@ -28,6 +28,7 @@ type OpenAIModel = Model<"openai-completions">;
 interface RemoteModel {
   id: string;
   name?: string;
+  reasoning?: boolean;
   input?: ModelInput[];
   contextWindow?: number;
 }
@@ -69,7 +70,7 @@ function toPiModel(
     api: "openai-completions",
     provider: provider.id,
     baseUrl,
-    reasoning: override?.reasoning ?? defaults?.reasoning ?? false,
+    reasoning: override?.reasoning ?? defaults?.reasoning ?? remote.reasoning ?? false,
     input: override?.input ?? remote.input ?? defaults?.input ?? ["text"],
     contextWindow,
     maxTokens: Math.min(maxTokens, contextWindow),
@@ -104,7 +105,7 @@ function remoteContextWindow(model: Record<string, unknown>): number | undefined
   return undefined;
 }
 
-function parseRemoteModels(payload: unknown): RemoteModel[] {
+export function parseRemoteModels(payload: unknown): RemoteModel[] {
   const entries = Array.isArray(payload)
     ? payload
     : isRecord(payload) && Array.isArray(payload.data)
@@ -122,6 +123,7 @@ function parseRemoteModels(payload: unknown): RemoteModel[] {
     models.push({
       id,
       name: typeof entry.name === "string" && entry.name.trim() ? entry.name.trim() : undefined,
+      reasoning: typeof entry.reasoning === "boolean" ? entry.reasoning : isRecord(entry.reasoning) || undefined,
       input: remoteInput(entry),
       contextWindow: remoteContextWindow(entry),
     });

package/.pi/lib/knowledge/evidence.ts

@@ -99,8 +99,13 @@ function boundedExcerpt(value: string, limit = 900): string {
   const clean = sanitizeKnowledgeText(value).replace(/\n{3,}/gu, "\n\n").trim();
   if (clean.length <= limit) return clean;
   const lines = clean.split("\n");
-  const significant = lines.filter((line) => FAILURE_RE.test(line) || VERIFICATION_RE.test(line));
-  const selected = [...lines.slice(0, 3), ...significant.slice(0, 5), ...lines.slice(-3)];
+  const priority = (line: string): number => REVERSAL_RE.test(line) ? 0
+    : FAILURE_RE.test(line) && VERIFICATION_RE.test(line) ? 1
+      : FAILURE_RE.test(line) ? 2 : 3;
+  const significant = lines.filter((line) => (
+    REVERSAL_RE.test(line) || FAILURE_RE.test(line) || VERIFICATION_RE.test(line)
+  )).sort((left, right) => priority(left) - priority(right));
+  const selected = [...lines.slice(0, 2), ...significant.slice(0, 7), ...lines.slice(-2)];
   return [...new Set(selected)].join("\n").slice(0, limit);
 }
 
@@ -163,9 +168,8 @@ function evidenceId(entryId: string, suffix: string): string {
   return `ev-${digest(`${entryId}\0${suffix}`)}`;
 }
 
-function outcomeFor(message: Record<string, unknown>, text: string): "success" | "failure" {
-  if (message.isError === true || FAILURE_RE.test(text)) return "failure";
-  return "success";
+function outcomeFor(message: Record<string, unknown>): "success" | "failure" {
+  return message.isError === true ? "failure" : "success";
 }
 
 function userKind(text: string): KnowledgeEvidenceKind {
@@ -215,11 +219,11 @@ export function buildKnowledgeEvidence(entries: SessionEntry[]): KnowledgeEviden
       // Recalled knowledge is comparison context, not fresh execution evidence. Treating a
       // lookup result as verification would let knowledge reinforce itself merely by loading it.
       if (call?.name === "hwcode_knowledge_lookup") return;
-      const outcome = outcomeFor(message, text);
+      const outcome = outcomeFor(message);
       const callText = call ? `[call] ${call.name}: ${compactArguments(call.args)}\n` : "";
       evidence.push({
         id: evidenceId(entryId, toolCallId || "tool"), entryId, timestamp,
-        kind: outcome === "failure" ? "tool-failure" : VERIFICATION_RE.test(text) ? "verification" : "tool-success",
+        kind: outcome === "failure" ? "tool-failure" : "tool-success",
         subject: evidenceSubject(call, text), operation: operationName(call), outcome,
         excerpt: boundedExcerpt(`${callText}[result] ${text}`),
       });
@@ -228,7 +232,7 @@ export function buildKnowledgeEvidence(entries: SessionEntry[]): KnowledgeEviden
     if (role === "assistant" && text && (FAILURE_RE.test(text) || VERIFICATION_RE.test(text) || REVERSAL_RE.test(text))) {
       evidence.push({
         id: evidenceId(entryId, "assistant"), entryId, timestamp,
-        kind: VERIFICATION_RE.test(text) ? "verification" : "assistant-claim",
+        kind: "assistant-claim",
         subject: "Assistant diagnosis or conclusion", outcome: "unknown", excerpt: boundedExcerpt(text, 800),
       });
     }

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

@@ -33,9 +33,11 @@ Treat failed, unsupported, invalid, or corrected operations as negative 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.
 Rank knowledge by future time saved, repeated failures, hypothesis reversals, user corrections, verified resolution, reuse, and novelty. Do not favor the final task result merely because it completed the objective. Every supplied episode scoring 45 or higher must have an episodeDecisions entry, even when skipped.
 Relevant existing knowledge may be supplied. Cheap recall only proposes related items; it does not prove duplication. Compare execution environment, symptom, and root cause. Use action "reinforce" only when the exact target already covers every claim and there is no new content. Use "extend" for new compatible claims, "correct" when direct evidence disproves old content, "retire" when the whole target is invalid, and "add" for a distinct environment or root cause. Never turn lexical similarity into reinforcement.
+Prefer the most specific existing topic that already covers a claim over a broad conventions or SOP topic.
 Scope may be "global" only for an explicit user directive intended across projects; otherwise use "project".
 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.
 Every technical claim must have a stable claim key and cite evidenceIds from the supplied evidence ledger. Tool success, tool failure, and verification evidence outrank assistant narration. Do not support a service-wide or universal claim with one operation-specific failure. Evidence for an unrelated parameter error cannot support an API-version claim.
+Keep each candidate atomic: put separable claim sets in separate candidates, omit any claim without direct cited evidence, and prioritize the highest-value supported candidates. One unsupported claim rejects its entire candidate.
 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.
 The body must not repeat the title as a Markdown H1. The storage renderer supplies the H1. Replace any necessary example resource value with an obvious placeholder instead of a real project, namespace, image, network, or resource name.
 Return: {"episodeDecisions":[{"episodeId":"ep-id","decision":"add|reinforce|extend|correct|retire|skip","targetId":"existing-id or omitted","reason":"..."}],"candidates":[{"key":"stable semantic key","identityKey":"platform/runtime/problem/root-cause","targetId":"existing-id or null","episodeIds":["ep-id"],"title":"...","summary":"one complete sentence","keywords":["..."],"scope":"global|project","body":"markdown body or only the new section for extend","claims":[{"key":"stable claim key","text":"one precise claim","evidenceIds":["ev-id"]}],"evidenceIds":["ev-id"],"confidence":0.0,"storageHint":"rule|topic","action":"add|reinforce|extend|correct|retire","explicitUserDirective":false,"durability":"stable"}]}

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

@@ -31,6 +31,7 @@ export const KNOWLEDGE_RUNTIME_DEFAULTS = Object.freeze({
     intervalMs: 60_000,
     idleMs: 60_000,
     capabilityPollMs: 5_000,
+    requestTimeoutMs: 45_000,
     modelTimeoutMs: 180_000,
     maxAttempts: 2,
     maxFailureDetailChars: 4_000,
@@ -40,7 +41,7 @@ export const KNOWLEDGE_RUNTIME_DEFAULTS = Object.freeze({
     maxEvidenceItems: 24,
     maxEvidenceChars: 12_000,
     mandatoryEpisodeScore: 45,
-    maxExistingContextItems: 5,
+    maxExistingContextItems: 8,
     maxExistingBodyItems: 2,
     maxExistingBodyChars: 600,
     maxCandidates: 3,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hadooppei/hwcode",
-  "version": "1.0.20",
+  "version": "1.0.24",
   "description": "A customizable terminal coding agent with local-model support and HWCode workflows.",
   "type": "module",
   "bin": {