@caupulican/pi-adaptative 0.80.95 → 0.80.97

package/dist/core/agent-session.js

@@ -85,6 +85,7 @@ import { CURRENT_SESSION_VERSION, getLatestCompactionEntry } from "./session-man
 import { matchesResourceProfilePattern, } from "./settings-manager.js";
 import { createSyntheticSourceInfo } from "./source-info.js";
 import { buildSystemPrompt } from "./system-prompt.js";
+import { buildReflexUserPrompt, parseReflexPlan, REFLEX_INTERPRETER_SYSTEM_PROMPT, } from "./toolkit/reflex-interpreter.js";
 import { executeToolkitScript } from "./toolkit/script-runner.js";
 import { createLocalBashOperations } from "./tools/bash.js";
 import { createDelegateToolDefinition } from "./tools/delegate.js";
@@ -2215,6 +2216,7 @@ export class AgentSession {
         }
         const previousModel = this.agent.state.model;
         const previousThinkingLevel = this.agent.state.thinkingLevel;
+        const previousTurnTools = this.agent.state.tools;
         const previousActiveModelRouterIntent = this._activeModelRouterIntent;
         const previousActiveModelRouterRoute = this._activeModelRouterRoute;
         const previousModelRouterSessionBuffer = this._modelRouterSessionBuffer;
@@ -2247,6 +2249,17 @@ export class AgentSession {
         if (!modelsAreEqual(this.model, routedModel)) {
             this.agent.state.model = routedModel;
             this.agent.state.thinkingLevel = clampThinkingLevel(routedModel, previousThinkingLevel);
+            // G4: capability tool-filtering follows the ROUTED model for the turn. Without this a
+            // cheap/local routed model inherits the session model's full tool surface — schemas it
+            // pays for on every request and may not be able to drive at all.
+            const routedProfile = deriveModelCapabilityProfile({
+                contextWindow: routedModel.contextWindow,
+                mode: this.settingsManager.getModelCapabilitySettings().mode,
+            });
+            if (routedProfile.class !== "full") {
+                const allowed = new Set(filterToolNamesForCapability(previousTurnTools.map((tool) => tool.name), routedProfile));
+                this.agent.state.tools = previousTurnTools.filter((tool) => allowed.has(tool.name));
+            }
         }
         try {
             await this._runAgentPrompt(messages);
@@ -2284,6 +2297,7 @@ export class AgentSession {
             if (modelsAreEqual(this.agent.state.model, routedModel)) {
                 this.agent.state.model = previousModel;
                 this.agent.state.thinkingLevel = previousThinkingLevel;
+                this.agent.state.tools = previousTurnTools;
                 // The registry may have changed mid-turn (command-time registerProvider): re-resolve
                 // the restored model so a provider override is not dropped with the routed model.
                 this._refreshCurrentModelFromRegistry();
@@ -4083,6 +4097,31 @@ export class AgentSession {
             const runToolkitScriptToolDefinition = createRunToolkitScriptToolDefinition({
                 getScripts: () => this.settingsManager.getToolkitScripts(),
                 execute: (script, scriptArgs) => executeToolkitScript({ script, scriptArgs, cwd: this._cwd }),
+                // Reflex brain (fitness-gated local model): resolves ambiguous requests into a
+                // registry pick. Best-effort — absent/unfit brain keeps the shortlist behavior.
+                interpret: async (request, scripts) => {
+                    const model = this._resolveCurationModelIfFit();
+                    if (!model)
+                        return undefined;
+                    const completion = await this.runIsolatedCompletion({
+                        systemPrompt: REFLEX_INTERPRETER_SYSTEM_PROMPT,
+                        messages: [
+                            {
+                                role: "user",
+                                content: [{ type: "text", text: buildReflexUserPrompt(request, scripts) }],
+                                timestamp: Date.now(),
+                            },
+                        ],
+                        model,
+                        thinkingLevel: "off",
+                        maxTokens: 256,
+                        cacheRetention: "short",
+                    });
+                    if (completion.usage.cost.total > 0 || completion.usage.totalTokens > 0) {
+                        this.addSpawnedUsage(completion.usage, { label: "toolkit-brain" });
+                    }
+                    return parseReflexPlan(completion.text);
+                },
             });
             this._baseToolDefinitions.set(runToolkitScriptToolDefinition.name, runToolkitScriptToolDefinition);
         }