kc-beta 0.5.4 → 0.5.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "kc-beta",
-  "version": "0.5.4",
+  "version": "0.5.6",
   "description": "KC Agent — LLM document verification agent (pure Node.js CLI)",
   "type": "module",
   "bin": {

package/src/agent/context-window.js

@@ -38,8 +38,18 @@ export class ContextWindow {
       return { messages, wasWindowed: false, removedCount: 0 };
     }
 
-    // Split into older and recent
-    const splitPoint = Math.max(0, messages.length - this.recentWindowSize);
+    // Split into older and recent. The recent slice is fed directly to the
+    // LLM, so it must not begin with an orphan "tool" message — those carry a
+    // tool_call_id that references an assistant `tool_calls` entry, and if
+    // that assistant message ended up in the compressed older slice the
+    // provider rejects the request (OpenAI: "tool messages must follow an
+    // assistant with tool_calls"; Anthropic: unpaired tool_use/tool_result).
+    // Walk the split point forward past any leading tool rows so the recent
+    // window always starts on a turn boundary.
+    let splitPoint = Math.max(0, messages.length - this.recentWindowSize);
+    while (splitPoint < messages.length && messages[splitPoint]?.role === "tool") {
+      splitPoint++;
+    }
     const recentMessages = messages.slice(splitPoint);
     const olderMessages = messages.slice(0, splitPoint);
 

package/src/agent/engine.js

@@ -341,7 +341,10 @@ export class AgentEngine {
     if (!this.contextWindow) return;
     const windowed = this.contextWindow.window(this.history.messages, this._phaseSummaries);
     if (windowed.wasWindowed) {
-      this.history.messages = windowed.messages;
+      // `messages` is a getter-only property on ConversationHistory; write the
+      // backing field and persist (same pattern as compact()).
+      this.history._messages = windowed.messages;
+      this.history._save();
       this.eventLog.append("context_windowed", {
         removed: windowed.removedCount,
         trigger: "post_tool_result",

package/src/cli/index.js

@@ -235,6 +235,7 @@ function App({ engine, config }) {
             return true;
           }
           const ok = engine._advancePhase(next, "manual /phase advance");
+          if (ok) setPhase(engine.currentPhase);
           addMessage({
             role: "system",
             content: ok
@@ -246,12 +247,28 @@ function App({ engine, config }) {
         }
 
         // /phase <name> — force-jump. Uses {force:true} to allow backward jumps.
+        // Whitelist against known phases first so an unknown name doesn't
+        // silently corrupt engine state (_advancePhase with {force:true}
+        // would otherwise accept any string and mutate currentPhase).
+        const validPhases = Object.keys(engine.pipelines);
+        if (!validPhases.includes(sub)) {
+          addMessage({
+            role: "system",
+            content: `Unknown phase: ${sub}. Valid: ${validPhases.join(", ")}`,
+          });
+          return true;
+        }
+        if (sub === engine.currentPhase) {
+          addMessage({ role: "system", content: `Already in phase ${sub.toUpperCase()}.` });
+          return true;
+        }
         const ok = engine._advancePhase(sub, "manual /phase <name>", { force: true });
+        if (ok) setPhase(engine.currentPhase);
         addMessage({
           role: "system",
           content: ok
             ? `→ phase set to ${sub.toUpperCase()}.`
-            : `Unknown phase: ${sub}. Valid: bootstrap, extraction, skill_authoring, skill_testing, distillation, production_qc`,
+            : `Failed to set phase to ${sub}.`,
         });
         updateContextStats();
         return true;
@@ -290,6 +307,15 @@ function App({ engine, config }) {
 
       case "/compact": {
         addMessage({ role: "system", content: "Compacting conversation history..." });
+        // Gate the prompt while compact() is in flight. Without this,
+        // InputPrompt stays active (isActive: !streaming) and a concurrent
+        // user submission routes into runTurn → history.addUser(...), which
+        // appends to _messages AFTER compact()'s pre-await snapshot. When
+        // compact resolves it overwrites _messages with [summary, ack,
+        // ...recentMessages] and silently drops the concurrent turn.
+        streamingRef.current = true;
+        setStreaming(true);
+        setSpinnerStatus("Compacting...");
         (async () => {
           try {
             const result = await engineRef.current.compact();
@@ -312,6 +338,14 @@ function App({ engine, config }) {
             updateContextStats();
           } catch (err) {
             addMessage({ role: "system", content: `Compact failed: ${err.message}` });
+          } finally {
+            streamingRef.current = false;
+            setStreaming(false);
+            setSpinnerStatus(null);
+            if (queueRef.current.length > 0) {
+              const next = queueRef.current.shift();
+              runTurn(next);
+            }
           }
         })();
         return true;

package/src/cli/onboard.js

@@ -171,13 +171,19 @@ export async function onboard() {
   }
 
   // --- API Key ---
-  const maskedExisting = existing.api_key ? existing.api_key.slice(0, 6) + "..." + existing.api_key.slice(-4) : "";
+  // Only offer to "keep" an existing key when the user hasn't switched
+  // providers. Otherwise an accidental Enter would silently save the OLD
+  // provider's key against the NEW provider's base URL — silent breakage.
+  const keyIsForSameProvider = existing.provider === provider.id;
+  const maskedExisting = keyIsForSameProvider && existing.api_key
+    ? existing.api_key.slice(0, 6) + "..." + existing.api_key.slice(-4)
+    : "";
   const keyHint = maskedExisting ? t.apiKeyKeep : t.apiKeyRequired;
   const keyPrompt = maskedExisting
     ? `  ${CYAN}${t.apiKey}${RESET} ${DIM}(${maskedExisting})${RESET}`
     : `  ${CYAN}${t.apiKey}${RESET} ${YELLOW}(${t.apiKeyRequired})${RESET}`;
   const apiKey = await ask(rl, keyPrompt, "", keyHint);
-  const finalKey = apiKey || existing.api_key || "";
+  const finalKey = apiKey || (keyIsForSameProvider ? existing.api_key : "") || "";
   if (!finalKey) { console.log(`  ${RED}${t.apiKeyMissing}${RESET}`); rl.close(); process.exit(1); }
   console.log();
 
@@ -245,34 +251,12 @@ export async function onboard() {
   );
   console.log();
 
-  // --- Worker LLM tiers ---
-  console.log(`  ${CYAN}${t.workerTiers}${RESET} ${DIM}(${t.tierHint})${RESET}`);
-  const tiers = {};
-  for (const tier of ["tier1", "tier2", "tier3", "tier4"]) {
-    const def = suggestedTiers?.[tier] || provider.defaultTiers[tier] || existing?.tiers?.[tier] || "";
-    tiers[tier] = await ask(
-      rl,
-      `    ${tier.toUpperCase()}`,
-      def,
-      t.discoveryAccept ? "" : "",
-    );
-  }
-  console.log();
-
-  // --- VLM tiers (vision/OCR) ---
-  console.log(`  ${CYAN}${t.vlmTiers}${RESET} ${DIM}(${t.tierHint})${RESET}`);
-  const vlmTiers = {};
-  for (const tier of ["tier1", "tier2", "tier3"]) {
-    const def = provider.defaultVlm?.[tier] || existing?.vlm_tiers?.[tier] || "";
-    vlmTiers[tier] = await ask(
-      rl,
-      `    ${tier.toUpperCase()}`,
-      def,
-    );
-  }
-  console.log();
-
   // --- Worker LLM provider (optional) ---
+  // Ask worker-provider BEFORE tier prompts so that when worker differs from
+  // conductor (e.g. conductor=xfyun single-model, worker=siliconflow) the
+  // tier-default suggestions come from the WORKER provider's model-tiers.json
+  // entry, not the conductor's. Previous ordering defaulted tiers from the
+  // conductor and produced nonsensical defaults for single-model conductors.
   console.log(`  ${CYAN}${t.workerConfig}${RESET}`);
   const sameProvider = await ask(rl, `  ${t.workerSameProvider}`, "Y", t.yesNo);
   let workerProvider = "";
@@ -280,6 +264,7 @@ export async function onboard() {
   let workerBaseUrl = "";
   let workerAuthType = "";
   let workerApiFormat = "";
+  let tierProviderDef = provider;  // where tier defaults come from
 
   if (sameProvider.toLowerCase() === "n" || sameProvider.toLowerCase() === "no") {
     // Pick a different provider for workers
@@ -294,21 +279,58 @@ export async function onboard() {
     workerAuthType = wp.authType;
     workerApiFormat = wp.apiFormat;
     workerBaseUrl = wp.baseUrl;
+    tierProviderDef = wp;
 
     if (wp.id === "custom") {
       workerBaseUrl = await ask(rl, `  ${t.baseUrl}`, existing.worker_base_url || "");
     }
 
-    // Worker API key
-    const wMasked = existing.worker_api_key ? existing.worker_api_key.slice(0, 6) + "..." + existing.worker_api_key.slice(-4) : "";
+    // Worker API key. Show masked existing key in the prompt (matches the
+    // main-provider prompt style) so the user can confirm what's saved
+    // without guessing. Like the main key, only offer to "keep" the existing
+    // value if the WORKER provider itself hasn't changed — otherwise Enter
+    // would silently carry the previous worker provider's key across.
+    const workerKeyIsForSameProvider = existing.worker_provider === wp.id;
+    const wMasked = workerKeyIsForSameProvider && existing.worker_api_key
+      ? existing.worker_api_key.slice(0, 6) + "..." + existing.worker_api_key.slice(-4)
+      : "";
     const wKeyHint = wMasked ? t.apiKeyKeep : t.apiKeyRequired;
-    workerApiKey = await ask(
+    const wKeyPrompt = wMasked
+      ? `  ${CYAN}${t.apiKey} (Worker)${RESET} ${DIM}(${wMasked})${RESET}`
+      : `  ${CYAN}${t.apiKey} (Worker)${RESET} ${YELLOW}(${t.apiKeyRequired})${RESET}`;
+    workerApiKey = await ask(rl, wKeyPrompt, "", wKeyHint);
+    workerApiKey = workerApiKey || (workerKeyIsForSameProvider ? existing.worker_api_key : "") || "";
+  }
+  console.log();
+
+  // --- Worker LLM tiers (defaults come from tierProviderDef set above) ---
+  // When worker==conductor, these default from the conductor's model-tiers.json
+  // entry. When worker is a separate provider, they default from the WORKER
+  // provider's entry — so e.g. siliconflow's GLM-5.1 tier1 defaults apply.
+  console.log(`  ${CYAN}${t.workerTiers}${RESET} ${DIM}(${t.tierHint})${RESET}`);
+  const tiers = {};
+  const tierSuggested = tierProviderDef.id === provider.id ? suggestedTiers : null;
+  for (const tier of ["tier1", "tier2", "tier3", "tier4"]) {
+    const def = tierSuggested?.[tier] || tierProviderDef.defaultTiers?.[tier] || existing?.tiers?.[tier] || "";
+    tiers[tier] = await ask(
+      rl,
+      `    ${tier.toUpperCase()}`,
+      def,
+      t.discoveryAccept ? "" : "",
+    );
+  }
+  console.log();
+
+  // --- VLM tiers (vision/OCR) — also from worker provider when split ---
+  console.log(`  ${CYAN}${t.vlmTiers}${RESET} ${DIM}(${t.tierHint})${RESET}`);
+  const vlmTiers = {};
+  for (const tier of ["tier1", "tier2", "tier3"]) {
+    const def = tierProviderDef.defaultVlm?.[tier] || existing?.vlm_tiers?.[tier] || "";
+    vlmTiers[tier] = await ask(
       rl,
-      `  ${CYAN}${t.apiKey} (Worker)${RESET}`,
-      "",
-      wKeyHint,
+      `    ${tier.toUpperCase()}`,
+      def,
     );
-    workerApiKey = workerApiKey || existing.worker_api_key || "";
   }
   console.log();
 

package/src/model-tiers.json

@@ -32,9 +32,10 @@
   },
 
   "volcanocloud": {
-    "conductor": "doubao-seed-2-0-pro-260215",
+    "_comment": "Coding plan (api/coding/v3) serves glm-5.1 only. Regular plan (api/v3) serves doubao/deepseek/glm-4-7. Pick conductor per the plan you onboarded.",
+    "conductor": "glm-5.1",
     "llm": {
-      "tier1": "doubao-seed-2-0-pro-260215, deepseek-v3-2-251201",
+      "tier1": "glm-5.1, doubao-seed-2-0-pro-260215, deepseek-v3-2-251201",
       "tier2": "glm-4-7-251222, doubao-1-5-pro-32k-250115",
       "tier3": "doubao-seed-2-0-mini-260215",
       "tier4": "doubao-seed-2-0-lite-260215, doubao-1-5-lite-32k-250115"
@@ -46,6 +47,22 @@
     }
   },
 
+  "xfyun": {
+    "_comment": "iFlytek Astro coding plan exposes only astron-code-latest; no VLM.",
+    "conductor": "astron-code-latest",
+    "llm": {
+      "tier1": "astron-code-latest",
+      "tier2": "",
+      "tier3": "",
+      "tier4": ""
+    },
+    "vlm": {
+      "tier1": "",
+      "tier2": "",
+      "tier3": ""
+    }
+  },
+
   "anthropic": {
     "conductor": "claude-sonnet-4-20250514",
     "llm": {

package/src/providers.js

@@ -77,14 +77,20 @@ const PROVIDERS = [
   {
     id: "volcanocloud",
     name: "VolcanoCloud",
+    // Regular Ark API — serves doubao / deepseek / glm-4-7-251222.
+    // Coding plan uses api/coding/v3 and serves glm-5.1 (aliased to glm-4.7
+    // server-side, thinking model).
     baseUrl: "https://ark.cn-beijing.volces.com/api/v3",
+    codingPlanUrl: "https://ark.cn-beijing.volces.com/api/coding/v3",
     authType: "bearer",
     apiFormat: "openai",
-    modelsEndpoint: null, // VolcanoCloud coding plan — use curated list
+    modelsEndpoint: null, // VolcanoCloud — use curated list
+    supportsCodingPlanKey: true,
     defaultModel: getTierConfig("volcanocloud").conductor || "doubao-seed-2-0-pro-260215",
     defaultTiers: getTierConfig("volcanocloud").llm,
     defaultVlm: getTierConfig("volcanocloud").vlm,
     curatedModels: [
+      { id: "glm-5.1", ownedBy: "zhipu" },
       { id: "doubao-seed-2-0-pro-260215", ownedBy: "bytedance" },
       { id: "deepseek-v3-2-251201", ownedBy: "deepseek" },
       { id: "glm-4-7-251222", ownedBy: "zhipu" },
@@ -98,6 +104,27 @@ const PROVIDERS = [
       zh: "火山云（字节跳动）",
     },
   },
+  {
+    id: "xfyun",
+    name: "XfYun Astro",
+    // iFlytek Astro coding plan — OpenAI-compatible endpoint. Only exposes
+    // a single model (astron-code-latest) today, so no /models discovery and
+    // the tier assignment in model-tiers.json only fills tier1 / conductor.
+    baseUrl: "https://maas-coding-api.cn-huabei-1.xf-yun.com/v2",
+    authType: "bearer",
+    apiFormat: "openai",
+    modelsEndpoint: null,
+    defaultModel: getTierConfig("xfyun").conductor || "astron-code-latest",
+    defaultTiers: getTierConfig("xfyun").llm,
+    defaultVlm: getTierConfig("xfyun").vlm,
+    curatedModels: [
+      { id: "astron-code-latest", ownedBy: "iflytek" },
+    ],
+    labels: {
+      en: "iFlytek XfYun Astro (coding plan, single-model)",
+      zh: "科大讯飞 Astro 编程套餐（单模型）",
+    },
+  },
   {
     id: "anthropic",
     name: "Anthropic",
@@ -238,12 +265,15 @@ const MODEL_RANKING = {
   "qwen3.5-122b": 75,
   "qwen3.5-35b": 65,
   // Zhipu
+  "glm-5.1": 92,
   "glm-5": 90,
   "glm-4.7": 80,
   "glm-4": 75,
   // Others
   "kimi-k2.5": 85,
   "kimi-k2": 80,
+  // iFlytek Astro
+  "astron-code": 90,
   "minimax-m2": 80,
   "deepseek-v3": 85,
   "deepseek-r1": 90,