@caupulican/pi-adaptative 0.81.13 → 0.81.15

package/dist/core/agent-session.js

@@ -151,6 +151,7 @@ export class AgentSession {
     _repairModeSessionCounts = new Map();
     _textProtocolParseFailures = new Map();
     _textProtocolParseObservedThisTurn = false;
+    _textProtocolValidationOutcomeThisTurn;
     /** Assembles the session's base system prompt from live session state (see
      * system-prompt-builder.ts); owns the paired _baseSystemPromptOptions. */
     _systemPromptBuilder;
@@ -507,6 +508,7 @@ export class AgentSession {
             previousToolArgumentValidation?.(taggedEvent);
             this._analytics.recordToolArgumentValidation(taggedEvent);
             this._recordToolValidationBounce(taggedEvent);
+            this._handleTextToolProtocolValidationOutcome(taggedEvent);
             this._handleModelAdaptationTelemetry(taggedEvent);
         };
         this._treeNavigator = new SessionTreeNavigator({
@@ -837,7 +839,6 @@ export class AgentSession {
         return {
             systemPrompt: `${primer}\n\n${instruction}`,
             messages: [{ role: "user", content: [{ type: "text", text: instruction }], timestamp: Date.now() }],
-            tools: [TEXT_TOOL_PROTOCOL_ECHO_TOOL],
         };
     }
     _messageHasEchoProbe(message, token) {
@@ -850,17 +851,17 @@ export class AgentSession {
             systemPrompt: instruction,
             messages: [{ role: "user", content: [{ type: "text", text: instruction }], timestamp: Date.now() }],
             tools: [TEXT_TOOL_PROTOCOL_ECHO_TOOL],
-        }, { textToolCallProtocol: false, maxRetries: 0 });
+        }, { textToolCallProtocol: false, maxRetries: 0, temperature: 0, maxTokens: 256 });
         return this._messageHasEchoProbe(await stream.result(), token);
     }
     async _runTextProtocolTrial(model, variant, token) {
         const stream = await this._streamForToolProbe(model, this._textProtocolCalibrationContext(variant, token), {
             textToolCallProtocol: false,
             maxRetries: 0,
+            temperature: 0,
+            maxTokens: 256,
         });
         const message = await stream.result();
-        if (this._messageHasEchoProbe(message, token))
-            return true;
         const text = message.content
             .filter((block) => block.type === "text")
             .map((block) => block.text)
@@ -1007,10 +1008,8 @@ export class AgentSession {
     _handleTextToolProtocolParse(event) {
         this._textProtocolParseObservedThisTurn = true;
         const modelKey = `${event.provider}/${event.model}`;
-        if (event.status === "parsed") {
-            this._textProtocolParseFailures.delete(modelKey);
+        if (event.status === "parsed")
             return;
-        }
         const signature = `${event.variant}:${event.reason ?? "failed"}`;
         const previous = this._textProtocolParseFailures.get(modelKey);
         const repeats = previous?.signature === signature ? previous.repeats + 1 : 1;
@@ -1024,7 +1023,40 @@ export class AgentSession {
         }
         this._textProtocolParseFailures.delete(modelKey);
     }
+    _handleTextToolProtocolValidationOutcome(event) {
+        if (event.source !== "text-protocol")
+            return;
+        const protocol = this.agent.textToolCallProtocol;
+        const variant = protocol === true ? "tool-tag" : protocol ? protocol.variant : undefined;
+        if (!variant)
+            return;
+        const status = event.outcome === "bounced" ? "failed" : "parsed";
+        if (this._textProtocolValidationOutcomeThisTurn?.status === "parsed" && status === "failed")
+            return;
+        this._textProtocolValidationOutcomeThisTurn = {
+            provider: event.provider ?? this.agent.state.model.provider,
+            model: event.model ?? this.agent.state.model.id,
+            variant,
+            status,
+            callCount: 1,
+            textLength: 0,
+            ...(status === "failed" && {
+                reason: event.errorKeywords?.includes("unknown_tool") ? "unknown-tool" : "validation-failed",
+            }),
+        };
+    }
     _recordTextToolProtocolParseOutcomeFromLastAssistant() {
+        const validationOutcome = this._textProtocolValidationOutcomeThisTurn;
+        this._textProtocolValidationOutcomeThisTurn = undefined;
+        if (validationOutcome?.status === "parsed") {
+            this._textProtocolParseObservedThisTurn = true;
+            this._textProtocolParseFailures.delete(`${validationOutcome.provider}/${validationOutcome.model}`);
+            return;
+        }
+        if (validationOutcome) {
+            this._handleTextToolProtocolParse(validationOutcome);
+            return;
+        }
         if (this._textProtocolParseObservedThisTurn)
             return;
         const protocol = this.agent.textToolCallProtocol;
@@ -2181,6 +2213,7 @@ export class AgentSession {
         }
         preflightResult?.(true);
         this._textProtocolParseObservedThisTurn = false;
+        this._textProtocolValidationOutcomeThisTurn = undefined;
         await this._modelRouter.runRoutedTurn(messages, routedTurnModel, routedTurnRouteDecision);
         this._recordTextToolProtocolParseOutcomeFromLastAssistant();
         // R4: score whether the agent actually used the recalled context, so the recall gate can adapt.