@kenkaiiii/gg-boss 4.11.1 → 4.11.3

package/dist/{chunk-MFM536F4.js → chunk-WECKKLXP.js}

@@ -63845,16 +63845,17 @@ function toAnthropicThinking(level, maxTokens, model) {
       outputConfig: { effort }
     };
   }
+  const VISIBLE_FLOOR = 1024;
   const effectiveLevel = level === "xhigh" || level === "max" ? "high" : level;
   const budgetMap = {
-    low: Math.max(1024, Math.floor(maxTokens * 0.25)),
-    medium: Math.max(2048, Math.floor(maxTokens * 0.5)),
-    high: Math.max(4096, maxTokens)
+    low: Math.max(1024, Math.floor(maxTokens * 0.2)),
+    medium: Math.max(2048, Math.floor(maxTokens * 0.45)),
+    high: Math.max(4096, Math.floor(maxTokens * 0.8))
   };
-  const budget = budgetMap[effectiveLevel];
+  const budget = Math.max(0, Math.min(budgetMap[effectiveLevel], maxTokens - VISIBLE_FLOOR));
   return {
     thinking: { type: "enabled", budget_tokens: budget },
-    maxTokens: maxTokens + budget
+    maxTokens
   };
 }
 function remapToolCallId(id2, idMap) {
@@ -72278,7 +72279,7 @@ var SubAgentParams = external_exports.object({
   task: external_exports.string().describe("The task to delegate to the sub-agent"),
   agent: external_exports.string().optional().describe("Named agent definition to use (from ~/.gg/agents/ or .gg/agents/)")
 });
-function createSubAgentTool(cwd2, agents, parentProvider, parentModel, getParentCacheKey, planModeRef) {
+function createSubAgentTool(cwd2, agents, getParentProvider, getParentModel, getParentCacheKey, planModeRef) {
   const agentList = agents.map((a) => `- ${a.name}: ${a.description}`).join("\n");
   const agentDesc = agentList ? `
 
@@ -72308,13 +72309,14 @@ ${agentList}` : "\n\nNo named agents configured.";
           };
         }
       }
-      const useProvider = parentProvider;
+      const useProvider = getParentProvider();
+      const useModel = getParentModel();
       const cliArgs = [
         "--json",
         "--provider",
         useProvider,
         "--model",
-        parentModel,
+        useModel,
         "--max-turns",
         String(SUB_AGENT_MAX_TURNS)
       ];
@@ -72341,7 +72343,7 @@ ${agentList}` : "\n\nNo named agents configured.";
       log("INFO", "subagent", "Sub-agent spawn", {
         cacheKey: subCacheKey,
         provider: useProvider,
-        model: parentModel,
+        model: useModel,
         agent: agentDef?.name ?? "(default)"
       });
       let childExited = false;
@@ -74271,7 +74273,7 @@ function createTools(cwd2, opts) {
     tools.push(createWebSearchTool());
   }
   if (opts?.agents && opts.agents.length > 0 && opts.provider && opts.model) {
-    tools.push(createSubAgentTool(cwd2, opts.agents, opts.provider, opts.model, opts.getCacheKey, planModeRef));
+    tools.push(createSubAgentTool(cwd2, opts.agents, () => opts.getProvider?.() ?? opts.provider, () => opts.getModel?.() ?? opts.model, opts.getCacheKey, planModeRef));
   }
   if (opts?.skills && opts.skills.length > 0) {
     tools.push(createSkillTool(opts.skills));
@@ -83716,6 +83718,10 @@ var AgentSession = class {
   customSystemPrompt;
   /** Shared with the tool layer so plan-mode restrictions read live state. */
   planModeRef = { current: false };
+  /** Path of the approved plan currently being implemented, or undefined. When
+   *  set, the system prompt carries the `[DONE:n]` progress contract so the
+   *  model emits step-completion markers the UI's plan-progress widget reads. */
+  approvedPlanPath;
   sessionId = "";
   sessionPath = "";
   lastPersistedIndex = 0;
@@ -83728,10 +83734,25 @@ var AgentSession = class {
     this.model = options.model;
     this.cwd = options.cwd;
     this.baseUrl = options.baseUrl;
-    this.maxTokens = options.maxTokens ?? getModel(options.model)?.maxOutputTokens ?? 16384;
+    this.maxTokens = this.resolveMaxTokens(options.model);
     this.thinkingLevel = options.thinkingLevel;
     this.customSystemPrompt = options.systemPrompt;
   }
+  /**
+   * Derive the output-token cap for a model. Follows the active model's
+   * `maxOutputTokens` so a session booted on a large-output model (e.g. Kimi's
+   * 256K) doesn't carry that cap to a smaller one (e.g. Opus's 128K) after a
+   * model switch — that mismatch surfaces from the provider as
+   * `max_tokens: 262144 > 128000, which is the maximum allowed …`. An explicit
+   * `maxTokens` override is honored but clamped to the model's ceiling.
+   */
+  resolveMaxTokens(modelId) {
+    const modelInfo = getModel(modelId);
+    if (this.opts.maxTokens) {
+      return modelInfo ? Math.min(this.opts.maxTokens, modelInfo.maxOutputTokens) : this.opts.maxTokens;
+    }
+    return modelInfo?.maxOutputTokens ?? 16384;
+  }
   async initialize() {
     setEstimatorModel(this.model);
     const paths = await ensureAppDirs();
@@ -83758,8 +83779,10 @@ var AgentSession = class {
       provider: this.provider,
       model: this.model,
       lspDiagnostics: this.settingsManager.get("lspDiagnostics"),
-      // Lazy — sessionId isn't assigned yet when createTools() runs, so we
-      // must defer reading the cache key until the sub-agent actually fires.
+      // Lazy — sessionId/model/provider can change after createTools() runs, so
+      // sub-agent spawns read the current parent state at execution time.
+      getProvider: () => this.provider,
+      getModel: () => this.model,
       getCacheKey: () => this.getPromptCacheKey(),
       // Plan mode: only wired when the host supplies callbacks. The ref is
       // shared so bash/edit/write enforce read-only restrictions live.
@@ -84140,6 +84163,7 @@ ${fileNotes.join("\n")}`);
       this.provider = provider;
     this.model = model;
     setEstimatorModel(model);
+    this.maxTokens = this.resolveMaxTokens(model);
     this.eventBus.emit("model_change", { provider: this.provider, model: this.model });
     if (provider && provider !== prevProvider) {
       const hasWebSearch = this.tools.some((t) => t.name === "web_search");
@@ -84203,6 +84227,8 @@ ${fileNotes.join("\n")}`);
     });
   }
   async newSession() {
+    this.planModeRef.current = false;
+    this.approvedPlanPath = void 0;
     const basePrompt = this.customSystemPrompt ?? await buildSystemPrompt(this.cwd, this.skills, false, void 0, this.tools.map((tool) => tool.name), void 0, this.provider);
     this.messages = [{ role: "system", content: basePrompt }];
     await this.createNewSession();
@@ -84296,9 +84322,25 @@ ${fileNotes.join("\n")}`);
    */
   async setPlanMode(active) {
     this.planModeRef.current = active;
+    if (active)
+      this.approvedPlanPath = void 0;
+    await this.rebuildSystemPromptInPlace();
+  }
+  /**
+   * Bake an approved plan into the system prompt so the model is told to emit
+   * `[DONE:n]` markers as it completes each step (the contract the UI's
+   * plan-progress widget reads). Pass `undefined` to clear it. No-op when a
+   * custom system prompt is in force (the host owns the prompt then).
+   */
+  async setApprovedPlan(approvedPlanPath) {
+    this.approvedPlanPath = approvedPlanPath;
+    await this.rebuildSystemPromptInPlace();
+  }
+  /** Rebuild messages[0] from current plan-mode + approved-plan state. */
+  async rebuildSystemPromptInPlace() {
     if (this.customSystemPrompt)
       return;
-    const rebuilt = await buildSystemPrompt(this.cwd, this.skills, active, void 0, this.tools.map((tool) => tool.name), void 0, this.provider);
+    const rebuilt = await buildSystemPrompt(this.cwd, this.skills, this.planModeRef.current, this.approvedPlanPath, this.tools.map((tool) => tool.name), void 0, this.provider);
     if (this.messages[0]?.role === "system") {
       this.messages[0] = { role: "system", content: rebuilt };
     } else {
@@ -116472,4 +116514,4 @@ react/cjs/react-jsx-runtime.development.js:
    * LICENSE file in the root directory of this source tree.
    *)
 */
-//# sourceMappingURL=chunk-MFM536F4.js.map
+//# sourceMappingURL=chunk-WECKKLXP.js.map