@caupulican/pi-adaptative 0.81.15 → 0.81.17

package/dist/core/agent-session.js

@@ -1,6 +1,7 @@
-import { readFileSync } from "node:fs";
+import { readFileSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
 import { basename, dirname, join } from "node:path";
-import { classifyFailure, compactToolResultDetailsForRetention, computeRetryDelayMs, createCustomMessage, DEFAULT_RETRY_POLICY, RetryController, sleepAbortable, withStreamIdleWatchdog, } from "@caupulican/pi-agent-core";
+import { classifyFailure, compactToolResultDetailsForRetention, computeRetryDelayMs, createCustomMessage, DEFAULT_RETRY_POLICY, DEFAULT_STREAM_IDLE, RetryController, sleepAbortable, withStreamIdleWatchdog, } from "@caupulican/pi-agent-core";
 import { calculateContextTokens, compact, createDeterministicCompaction, estimateContextTokens, getLatestCompactionEntry, prepareCompaction, runCompactionLoop, shouldCompact, } from "@caupulican/pi-agent-core/node";
 import { cleanupSessionResources, formatToolRepairStandingRule, generateTextToolProtocolPrimer, isContextOverflow, parseTextToolCalls, streamSimple, } from "@caupulican/pi-ai";
 import { Type } from "typebox";
@@ -30,6 +31,7 @@ import { deriveModelCapabilityProfile, filterToolNamesForCapability, } from "./m
 import { formatModelRouterModel, ModelRouterController } from "./model-router-controller.js";
 import { ModelSelectionController } from "./model-selection-controller.js";
 import { ModelAdaptationStore } from "./models/adaptation-store.js";
+import { DEFAULT_ADAPTIVE_STREAM_IDLE_CEILING_MS, estimateContextPromptTokens, resolveAdaptiveStreamIdleOptions, withModelPerfProfile, } from "./models/perf-profile.js";
 import { ProfileFilterController } from "./profile-filter-controller.js";
 import { expandPromptTemplate } from "./prompt-templates.js";
 import { ReflectionController } from "./reflection-controller.js";
@@ -60,12 +62,22 @@ const MODEL_ADAPTATION_REPAIR_THRESHOLD = 3;
 const TEXT_TOOL_PROTOCOL_VERSION = 1;
 const TEXT_TOOL_PROTOCOL_TRIALS_PER_VARIANT = 2;
 const TEXT_TOOL_PROTOCOL_PARSE_FAILURE_THRESHOLD = 3;
-const TEXT_TOOL_PROTOCOL_VARIANTS = ["tool-tag", "tool-call", "fenced-json"];
+const TEXT_TOOL_PROTOCOL_VARIANTS = [
+    "tool-tag",
+    "tool-call",
+    "fenced-json",
+    "function-xml",
+];
 const TEXT_TOOL_PROTOCOL_ECHO_TOOL = {
     name: "echo",
     description: "Echo calibration data",
     parameters: Type.Object({ data: Type.String() }),
 };
+const NATIVE_TOOL_PROBE_READ_TOOL = {
+    name: "read",
+    description: "Read file contents",
+    parameters: Type.Object({ path: Type.String() }),
+};
 /** Test-only override of the stream-idle bounds. Read per-request by the wiring's resolver. */
 let streamIdleOptionsOverride;
 /**
@@ -148,6 +160,8 @@ export class AgentSession {
     _collectWorkspaceSources;
     _localRuntimeController;
     _modelAdaptationStore;
+    _prefixWarmer;
+    _completedPrefixWarms = new Set();
     _repairModeSessionCounts = new Map();
     _textProtocolParseFailures = new Map();
     _textProtocolParseObservedThisTurn = false;
@@ -229,14 +243,35 @@ export class AgentSession {
         // withStreamIdleWatchdog's contract), so no extra drain is added at this wiring site.
         // Wrapping also breaks the `streamFn === streamSimple` identity the auth-injection checks
         // use, so the wrapper carries a rawness marker that _isRawStreamSimple reads.
+        const agentDir = config.agentDir ?? getAgentDir();
+        const modelAdaptationStore = ModelAdaptationStore.forAgentDir(agentDir);
         const baseStreamFn = this.agent.streamFn;
+        const profiledStreamFn = withModelPerfProfile(baseStreamFn, {
+            modelKey: (model) => formatModelRouterModel(model),
+            recordSample: (modelKey, sample) => {
+                modelAdaptationStore.recordPerfSample(modelKey, sample);
+            },
+        });
         // `this.settingsManager` is assigned below; the resolver closes over the config reference
         // because the wrapper must be installed before that assignment runs.
         const stallSettingsSource = config.settingsManager;
-        this.agent.streamFn = tagRawness(withStreamIdleWatchdog(baseStreamFn, () => ({
-            ...stallSettingsSource.getStreamStallSettings(),
-            ...streamIdleOptionsOverride,
-        })), baseStreamFn === streamSimple);
+        this.agent.streamFn = tagRawness(withStreamIdleWatchdog(profiledStreamFn, (model, context) => {
+            const configured = {
+                ...stallSettingsSource.getStreamStallSettings(),
+                ...streamIdleOptionsOverride,
+            };
+            const httpIdleTimeoutMs = stallSettingsSource.getHttpIdleTimeoutMs();
+            const profile = modelAdaptationStore.get(formatModelRouterModel(model)).perf;
+            const adaptive = resolveAdaptiveStreamIdleOptions({
+                base: { ...DEFAULT_STREAM_IDLE, ...configured },
+                profile,
+                promptTokens: estimateContextPromptTokens(context),
+                ceilingMs: httpIdleTimeoutMs === 0
+                    ? DEFAULT_ADAPTIVE_STREAM_IDLE_CEILING_MS
+                    : Math.max(DEFAULT_STREAM_IDLE.quietIdleMs, httpIdleTimeoutMs - 60_000),
+            });
+            return { ...configured, ...adaptive };
+        }), baseStreamFn === streamSimple);
         this.sessionManager = config.sessionManager;
         this.settingsManager = config.settingsManager;
         // Auto-retry rides the reliability kernel: the controller owns the attempt counter and the
@@ -260,8 +295,8 @@ export class AgentSession {
         this._resourceLoader = config.resourceLoader;
         this._customTools = config.customTools ?? [];
         this._cwd = config.cwd;
-        this._agentDir = config.agentDir ?? getAgentDir();
-        this._modelAdaptationStore = ModelAdaptationStore.forAgentDir(this._agentDir);
+        this._agentDir = agentDir;
+        this._modelAdaptationStore = modelAdaptationStore;
         this.agent.onTextToolProtocolParse = (event) => this._handleTextToolProtocolParse(event);
         this._applyToolRepairLayerSettings();
         this._collectWorkspaceSources = config.collectWorkspaceSources ?? collectWorkspaceSources;
@@ -578,11 +613,82 @@ export class AgentSession {
             activeToolNames: this._initialActiveToolNames,
             includeAllExtensionTools: true,
         });
+        this._scheduleLocalPrefixWarm(this.agent.state.model, "session-start");
     }
     /** Model registry for API key resolution and model discovery */
     get modelRegistry() {
         return this._modelRegistry;
     }
+    _scheduleLocalPrefixWarm(model, _reason) {
+        if (!model || !this._isWarmableLocalModel(model))
+            return;
+        const modelKey = formatModelRouterModel(model);
+        if (this._completedPrefixWarms.has(modelKey) || this._prefixWarmer?.modelKey === modelKey)
+            return;
+        this._cancelPrefixWarm();
+        const controller = new AbortController();
+        const timer = setTimeout(() => {
+            const warmer = this._prefixWarmer;
+            if (!warmer || warmer.controller !== controller || controller.signal.aborted)
+                return;
+            warmer.timer = undefined;
+            void this._runLocalPrefixWarm(model, modelKey, controller);
+        }, 0);
+        timer.unref?.();
+        this._prefixWarmer = { modelKey, controller, timer };
+    }
+    _cancelPrefixWarm() {
+        const warmer = this._prefixWarmer;
+        if (!warmer)
+            return;
+        if (warmer.timer)
+            clearTimeout(warmer.timer);
+        warmer.controller.abort(new Error("prefix warmer preempted"));
+        this._prefixWarmer = undefined;
+    }
+    async _runLocalPrefixWarm(model, modelKey, controller) {
+        try {
+            const options = {
+                maxTokens: 1,
+                signal: controller.signal,
+                onPayload: this.agent.onPayload,
+                onResponse: this.agent.onResponse,
+            };
+            if (this._isRawStreamSimple(this.agent.streamFn)) {
+                const auth = await this._getRequiredRequestAuth(model);
+                options.apiKey = auth.apiKey;
+                options.headers = auth.headers;
+            }
+            if (controller.signal.aborted)
+                return;
+            const stream = await this.agent.streamFn(model, {
+                systemPrompt: this._baseSystemPrompt,
+                tools: this.agent.state.tools,
+                messages: [],
+            }, options);
+            await stream.result();
+            if (!controller.signal.aborted)
+                this._completedPrefixWarms.add(modelKey);
+        }
+        catch {
+            // Best-effort cache warm only; a miss must never affect the real turn.
+        }
+        finally {
+            if (this._prefixWarmer?.controller === controller)
+                this._prefixWarmer = undefined;
+        }
+    }
+    _isWarmableLocalModel(model) {
+        if (model.api !== "openai-completions")
+            return false;
+        try {
+            const hostname = new URL(model.baseUrl).hostname.toLowerCase();
+            return hostname === "localhost" || hostname === "127.0.0.1" || hostname === "::1" || hostname === "[::1]";
+        }
+        catch {
+            return false;
+        }
+    }
     /**
      * True when the session's stream fn is the raw `streamSimple` provider entry (directly, or as the
      * base wrapped by the idle watchdog at construction). Callers use this to decide whether request
@@ -841,18 +947,56 @@ export class AgentSession {
             messages: [{ role: "user", content: [{ type: "text", text: instruction }], timestamp: Date.now() }],
         };
     }
-    _messageHasEchoProbe(message, token) {
-        return message.content.some((block) => block.type === "toolCall" && block.name === "echo" && block.arguments.data === token);
+    _messageHasToolCallWithStringArgument(message, toolName, argName, argValue) {
+        return message.content.some((block) => {
+            if (block.type !== "toolCall" || block.name !== toolName)
+                return false;
+            const args = block.arguments;
+            return (typeof args === "object" &&
+                args !== null &&
+                !Array.isArray(args) &&
+                args[argName] === argValue);
+        });
+    }
+    _nativeToolProbeSystemPrompt(instruction) {
+        const base = (this.agent.state.systemPrompt ?? "").trim();
+        return base ? `${base}\n\n${instruction}` : instruction;
+    }
+    async _runNativeReadTaskProbeTrial(model, path) {
+        const instruction = `Native tool-call capability probe: task-scale read. Use provider-native tool calling, not prose. ` +
+            `Call read exactly once with path exactly "${path}".`;
+        const stream = await this._streamForToolProbe(model, {
+            systemPrompt: this._nativeToolProbeSystemPrompt(instruction),
+            messages: [{ role: "user", content: [{ type: "text", text: instruction }], timestamp: Date.now() }],
+            tools: [NATIVE_TOOL_PROBE_READ_TOOL],
+        }, { textToolCallProtocol: false, maxRetries: 0, temperature: 0, maxTokens: 768 });
+        return this._messageHasToolCallWithStringArgument(await stream.result(), "read", "path", path);
     }
-    async _runNativeToolProbeTrial(model, token) {
-        const instruction = `Native tool-call capability probe. Use provider-native tool calling, not prose. ` +
+    async _runNativeEchoToolProbeTrial(model, token) {
+        const instruction = `Native tool-call capability probe: echo-only. Use provider-native tool calling, not prose. ` +
             `Call echo with data exactly "${token}".`;
         const stream = await this._streamForToolProbe(model, {
-            systemPrompt: instruction,
+            systemPrompt: this._nativeToolProbeSystemPrompt(instruction),
             messages: [{ role: "user", content: [{ type: "text", text: instruction }], timestamp: Date.now() }],
             tools: [TEXT_TOOL_PROTOCOL_ECHO_TOOL],
         }, { textToolCallProtocol: false, maxRetries: 0, temperature: 0, maxTokens: 256 });
-        return this._messageHasEchoProbe(await stream.result(), token);
+        return this._messageHasToolCallWithStringArgument(await stream.result(), "echo", "data", token);
+    }
+    async _gradeNativeToolCallingForModel(model, token) {
+        const path = join(tmpdir(), `pi-native-probe-${process.pid}-${Date.now()}.txt`);
+        writeFileSync(path, token, "utf-8");
+        try {
+            const taskPassed = await this._runNativeReadTaskProbeTrial(model, path);
+            if (taskPassed)
+                return "task";
+            const echoPassed = await this._runNativeEchoToolProbeTrial(model, token);
+            if (echoPassed)
+                return "echo-only";
+            return "absent";
+        }
+        finally {
+            rmSync(path, { force: true });
+        }
     }
     async _runTextProtocolTrial(model, variant, token) {
         const stream = await this._streamForToolProbe(model, this._textProtocolCalibrationContext(variant, token), {
@@ -933,12 +1077,17 @@ export class AgentSession {
         return `${model.provider}/${model.id}`;
     }
     _formatToolProbeReport(results) {
-        const lines = ["Tool probe results:", "Model | Verdict | Variant | Diagnostic", "--- | --- | --- | ---"];
+        const lines = [
+            "Tool probe results:",
+            "Model | Verdict | Variant | Native grade | Diagnostic",
+            "--- | --- | --- | --- | ---",
+        ];
         for (const result of results) {
             lines.push([
                 result.model,
                 result.verdict,
                 result.variant ?? "-",
+                result.nativeGrade ?? "-",
                 result.diagnostic ? result.diagnostic.replace(/\s+/g, " ").slice(0, 160) : "-",
             ].join(" | "));
         }
@@ -950,12 +1099,23 @@ export class AgentSession {
     async _probeToolCallingForModel(model) {
         const modelKey = this._modelRef(model);
         const probedAt = new Date().toISOString();
+        let nativeGrade = "absent";
         let diagnostic;
         try {
-            if (await this._runNativeToolProbeTrial(model, "pi-native-probe")) {
-                this._storeToolProbe(modelKey, { version: TEXT_TOOL_PROTOCOL_VERSION, status: "native", probedAt });
-                return { model: modelKey, verdict: "native" };
+            nativeGrade = await this._gradeNativeToolCallingForModel(model, "pi-native-probe");
+            if (nativeGrade === "task") {
+                this._storeToolProbe(modelKey, {
+                    version: TEXT_TOOL_PROTOCOL_VERSION,
+                    status: "native",
+                    probedAt,
+                    nativeGrade,
+                });
+                return { model: modelKey, verdict: "native", nativeGrade };
             }
+            diagnostic =
+                nativeGrade === "echo-only"
+                    ? "Native echo probe passed but task-scale read probe failed."
+                    : "Native task-scale read and echo probes did not produce provider-native tool calls.";
         }
         catch (error) {
             diagnostic = error instanceof Error ? error.message : String(error);
@@ -968,16 +1128,24 @@ export class AgentSession {
                     status: "text-protocol",
                     probedAt: calibrated.calibratedAt,
                     variant: calibrated.variant,
+                    nativeGrade,
+                    diagnostic,
                 });
-                return { model: modelKey, verdict: "text-protocol", variant: calibrated.variant };
+                return { model: modelKey, verdict: "text-protocol", variant: calibrated.variant, nativeGrade, diagnostic };
             }
-            diagnostic ??= `Text protocol variants failed: ${calibrated.variantsTried.join(", ")}`;
+            diagnostic = `${diagnostic ? `${diagnostic} ` : ""}Text protocol variants failed: ${calibrated.variantsTried.join(", ")}`;
         }
         catch (error) {
             diagnostic = error instanceof Error ? error.message : String(error);
         }
-        this._storeToolProbe(modelKey, { version: TEXT_TOOL_PROTOCOL_VERSION, status: "none", probedAt, diagnostic });
-        return { model: modelKey, verdict: "none", diagnostic };
+        this._storeToolProbe(modelKey, {
+            version: TEXT_TOOL_PROTOCOL_VERSION,
+            status: "none",
+            probedAt,
+            nativeGrade,
+            diagnostic,
+        });
+        return { model: modelKey, verdict: "none", nativeGrade, diagnostic };
     }
     async _resolveToolProbeModels(target) {
         const trimmed = target?.trim();
@@ -1687,6 +1855,7 @@ export class AgentSession {
             this.abortCompaction();
             this.abortBranchSummary();
             this.abortBash();
+            this._cancelPrefixWarm();
             this.agent.abort();
             // R8: stop any deployment-registered gateway channels / schedulers.
             void this._gatewayRegistry.stop().catch(() => { });
@@ -1935,6 +2104,9 @@ export class AgentSession {
     getLocalRuntime(baseUrl) {
         return this._localRuntimeController.getLocalRuntime(baseUrl);
     }
+    getTransformersRuntime(modelId, baseUrl) {
+        return this._localRuntimeController.getTransformersRuntime(modelId, baseUrl);
+    }
     /** models.json registers a local model's baseUrl as `<server>/v1` (OpenAI-compat); the runtime's
      * own health/boot endpoints are on the Ollama-native server root. Delegates to
      * {@link LocalRuntimeController}; kept here for `_warnIfManualModelChoiceIsRisky`'s own use. */
@@ -2010,6 +2182,7 @@ export class AgentSession {
     }
     async _promptUnserialized(text, options) {
         this._applyToolRepairLayerSettings();
+        this._cancelPrefixWarm();
         const expandPromptTemplates = options?.expandPromptTemplates ?? true;
         const processSlashCommands = options?.processSlashCommands ?? expandPromptTemplates;
         const preflightResult = options?.preflightResult;
@@ -2521,10 +2694,13 @@ export class AgentSession {
     // Model Management
     // =========================================================================
     async setModel(model, options = {}) {
-        return this._modelSelection.setModel(model, options);
+        await this._modelSelection.setModel(model, options);
+        this._scheduleLocalPrefixWarm(this.agent.state.model, "selection");
     }
     async cycleModel(direction = "forward") {
-        return this._modelSelection.cycleModel(direction);
+        const result = await this._modelSelection.cycleModel(direction);
+        this._scheduleLocalPrefixWarm(result?.model, "selection");
+        return result;
     }
     // =========================================================================
     // Thinking Level Management