open-multi-agent-kit 0.90.4 → 0.90.6

package/dist/core/agent-session.js

@@ -23,19 +23,19 @@ import { formatNoApiKeyFoundMessage, formatNoModelSelectedMessage } from "./auth
 import { parseBangInvocation } from "./bang-skill-invocation.js";
 import { executeBashWithOperations } from "./bash-executor.js";
 import { classifyShellCommand } from "./command-safety.js";
-import { calculateContextTokens, collectEntriesForBranchSummary, compact, estimateContextTokens, estimateProjectedContextTokens, generateBranchSummary, prepareCompaction, shouldCompact, } from "./compaction/index.js";
+import { calculateContextTokens, collectEntriesForBranchSummary, compact, estimateContextTokens, estimateProjectedContextTokens, generateBranchSummary, prepareCompaction, resolveCompactionModel, shouldCompact, } from "./compaction/index.js";
+import { compactionEmitWillRetry } from "./compaction/resume-policy.js";
 import { DEFAULT_THINKING_LEVEL } from "./defaults.js";
 import { exportSessionToHtml } from "./export-html/index.js";
 import { createToolHtmlRenderer } from "./export-html/tool-renderer.js";
 import { buildBlockedBashResult, evaluateCommandGate, isCommandSafetyAssumeYesEnabled, } from "./extensions/builtin/command-safety-gate.js";
 import { ExtensionRunner, wrapRegisteredTools, } from "./extensions/index.js";
 import { emitSessionShutdownEvent } from "./extensions/runner.js";
+import { assertTextChatModelForCompletion } from "./grok-harness.js";
+import { grokPlaybookAppendForProvider } from "./grok-playbook.js";
 import { assertLoadoutAccess, decideLoadoutAccess } from "./loadout-access-policy.js";
 import { expandPromptTemplate } from "./prompt-templates.js";
-import { classifyTask, resolveThinkingLevel } from "./reasoning-router.js";
 import { getBiasStepsForCell, getDefaultRouterBiasSnapshotPath, parseRouterBiasSnapshot, } from "./reasoning-router-bias.js";
-import { classifyTaskV2, DEFAULT_WEIGHTS, resolveThinkingLevelV2ForAuto, } from "./reasoning-router-v2.js";
-import { classifyTaskV3, resolveThinkingLevelV3ForAuto } from "./reasoning-router-v3.js";
 import { classifyTaskV4, resolveThinkingLevelV4WithUncertainty } from "./reasoning-router-v4.js";
 import { appendRouterFeedbackRecord, } from "./router-feedback-collector.js";
 import { detectSandboxBackend } from "./sandbox/backend.js";
@@ -109,18 +109,9 @@ export class AgentSession {
      * auto-resolved levels are never written to the user's persisted settings.
      */
     _thinkingMode = "manual";
-    /**
-     * Reasoning router version used by auto thinking mode. "v1" (default) keeps
-     * the frozen v1 classifyTask/resolveThinkingLevel; "v2" uses the Goal 004
-     * weighted router; "v3" uses the Goal 007 calibrated classifier; "v4" uses
-     * the Goal 009 confidence-bearing classifier (classifyTaskV4 +
-     * resolveThinkingLevelV4WithUncertainty). Default "v1" so behavior is
-     * unchanged until explicitly opted in.
-     */
-    _thinkingRouterVersion = "v1";
     /**
      * N=8 ring buffer of recent auto-turn task classes (newest first), feeding
-     * the v2/v3 classifiers' multi-turn prior feature. Never persisted to settings.
+     * v4's multi-turn prior feature. Never persisted to settings.
      */
     _taskClassHistory = [];
     /**
@@ -837,7 +828,12 @@ export class AgentSession {
         }
         const loaderSystemPrompt = this._resourceLoader.getSystemPrompt();
         const loaderAppendSystemPrompt = this._resourceLoader.getAppendSystemPrompt();
-        const appendSystemPrompt = loaderAppendSystemPrompt.length > 0 ? loaderAppendSystemPrompt.join("\n\n") : undefined;
+        const grokAppend = grokPlaybookAppendForProvider(this.model?.provider);
+        const appendParts = [...loaderAppendSystemPrompt];
+        if (grokAppend) {
+            appendParts.push(grokAppend);
+        }
+        const appendSystemPrompt = appendParts.length > 0 ? appendParts.join("\n\n") : undefined;
         const loadedSkills = this._resourceLoader.getSkills().skills;
         const loadedContextFiles = this._resourceLoader.getAgentsFiles().agentsFiles;
         this._baseSystemPromptOptions = {
@@ -972,6 +968,8 @@ export class AgentSession {
             if (!this.model) {
                 throw new Error(formatNoModelSelectedMessage());
             }
+            // Grok OAuth: refuse Imagine ids on the chat/completions path (tool-only).
+            assertTextChatModelForCompletion(this.model.id, this.model.provider);
             if (!this._modelRegistry.hasConfiguredAuth(this.model)) {
                 const isOAuth = this._modelRegistry.isUsingOAuth(this.model);
                 if (isOAuth) {
@@ -1330,6 +1328,8 @@ export class AgentSession {
         if (!this._modelRegistry.hasConfiguredAuth(model)) {
             throw new Error(`No API key for ${model.provider}/${model.id}`);
         }
+        // Grok OAuth: block selecting Imagine models as the session chat model.
+        assertTextChatModelForCompletion(model.id, model.provider);
         const previousModel = this.model;
         const thinkingLevel = this._getThinkingLevelForModelSwitch();
         this.agent.state.model = model;
@@ -1431,18 +1431,6 @@ export class AgentSession {
     setThinkingMode(mode) {
         this._thinkingMode = mode;
     }
-    /**
-     * Reasoning router version selection for auto thinking mode.
-     * Default "v1". Switching to "v2", "v3", or "v4" activates the
-     * corresponding opt-in router; the multi-turn history buffer is only
-     * consulted there.
-     */
-    get thinkingRouterVersion() {
-        return this._thinkingRouterVersion;
-    }
-    setThinkingRouterVersion(version) {
-        this._thinkingRouterVersion = version;
-    }
     /**
      * In auto thinking mode, resolve and apply this turn's thinking level from the
      * prompt content. Updates agent state, records the change in the session, and
@@ -1455,79 +1443,12 @@ export class AgentSession {
             return;
         if (!this.supportsThinking())
             return;
-        if (this._thinkingRouterVersion === "v4") {
-            this._applyAutoThinkingLevelV4(promptText);
-            return;
-        }
-        if (this._thinkingRouterVersion === "v3") {
-            this._applyAutoThinkingLevelV3(promptText);
-            return;
-        }
-        if (this._thinkingRouterVersion === "v2") {
-            this._applyAutoThinkingLevelV2(promptText);
-            return;
-        }
-        const taskClass = classifyTask({ prompt: promptText });
-        const resolved = resolveThinkingLevel(taskClass, this.getAvailableThinkingLevels());
-        const previousLevel = this.agent.state.thinkingLevel;
-        if (resolved === previousLevel)
-            return;
-        this.agent.state.thinkingLevel = resolved;
-        this.sessionManager.appendThinkingLevelChange(resolved);
-        this._emit({ type: "thinking_level_changed", level: resolved });
-    }
-    /**
-     * v2 auto-mode resolver (Goal 004). Builds the caller-side features the pure
-     * classifier cannot reach on its own (multi-turn ring buffer, context-pressure
-     * bucket), then applies classifyTaskV2 + resolveThinkingLevelV2ForAuto. The
-     * resolved class is pushed to the N=8 history for the NEXT turn's prior.
-     */
-    _applyAutoThinkingLevelV2(promptText) {
-        const availableLevels = this.getAvailableThinkingLevels();
-        const taskClass = classifyTaskV2({
-            prompt: promptText,
-            history: this._taskClassHistory,
-            pressureBucket: this._computePressureBucket(),
-        }, DEFAULT_WEIGHTS);
-        this._taskClassHistory.unshift(taskClass);
-        if (this._taskClassHistory.length > 8)
-            this._taskClassHistory.length = 8;
-        const resolved = resolveThinkingLevelV2ForAuto(taskClass, availableLevels, undefined);
-        const previousLevel = this.agent.state.thinkingLevel;
-        if (resolved === previousLevel)
-            return;
-        this.agent.state.thinkingLevel = resolved;
-        this.sessionManager.appendThinkingLevelChange(resolved);
-        this._emit({ type: "thinking_level_changed", level: resolved });
-    }
-    /**
-     * v3 auto-mode resolver (Goal 007). Reuses the same N=8 recent-class history
-     * and context-pressure bucket as v2, but routes through the v3 classifier and
-     * resolver pair.
-     */
-    _applyAutoThinkingLevelV3(promptText) {
-        const availableLevels = this.getAvailableThinkingLevels();
-        const taskClass = classifyTaskV3({
-            prompt: promptText,
-            history: this._taskClassHistory,
-            pressureBucket: this._computePressureBucket(),
-        });
-        this._taskClassHistory.unshift(taskClass);
-        if (this._taskClassHistory.length > 8)
-            this._taskClassHistory.length = 8;
-        const resolved = resolveThinkingLevelV3ForAuto(taskClass, availableLevels, undefined);
-        const previousLevel = this.agent.state.thinkingLevel;
-        if (resolved === previousLevel)
-            return;
-        this.agent.state.thinkingLevel = resolved;
-        this.sessionManager.appendThinkingLevelChange(resolved);
-        this._emit({ type: "thinking_level_changed", level: resolved });
+        this._applyAutoThinkingLevelV4(promptText);
     }
     /**
-     * v4 auto-mode resolver (Goal 009 Wave 3 Lane V1; wired to the opt-in
-     * learning bias by Goal 010 Lane I). Reuses the same N=8 recent-class
-     * history and context-pressure bucket as v2/v3, routed through the
-     * confidence-bearing v4 classifier and its uncertainty-aware resolver. No
+     * Auto-mode resolver. Reuses the N=8 recent-class history and
+     * context-pressure bucket, routed through the confidence-bearing v4 classifier
+     * and its uncertainty-aware resolver. No
      * `laneType` applies to the main session (always "none"/`undefined`);
      * `hint` is permanently `null` -- the Adaptorch advisory bridge has no
      * transport wired into the session (out of this lane's scope; see
@@ -1626,9 +1547,8 @@ export class AgentSession {
      * Locally derives the same three bounded, privacy-safe feedback-ledger
      * features (`lenBucket`, `hadFence`, `hadDiff`) the v4 classifier computes
      * internally (reasoning-router-v4.ts keeps those helpers file-private, so
-     * they cannot be imported here). Mirrors reasoning-router-v2.ts's
-     * `clampLenBucket` and reasoning-router-v4.ts's `hasCodeFence`/
-     * `hasDiffMarkers` byte-for-byte. Operates on the same trimmed prompt text
+     * they cannot be imported here). Mirrors reasoning-router-v4.ts's
+     * length-bucket, `hasCodeFence`, and `hasDiffMarkers` helpers. Operates on the same trimmed prompt text
      * the classifier scores; never returns raw prompt content or its exact
      * length, only the clamped [0,7] bucket and two booleans.
      */
@@ -1743,7 +1663,8 @@ export class AgentSession {
             if (!this.model) {
                 throw new Error(formatNoModelSelectedMessage());
             }
-            const { apiKey, headers } = await this._getCompactionRequestAuth(this.model);
+            const compactionModel = resolveCompactionModel(this.model, this._modelRegistry.getAvailable());
+            const { apiKey, headers } = await this._getCompactionRequestAuth(compactionModel);
             const pathEntries = this.sessionManager.getBranch();
             const settings = this.settingsManager.getCompactionSettings();
             const preparation = prepareCompaction(pathEntries, settings);
@@ -1786,7 +1707,7 @@ export class AgentSession {
             }
             else {
                 // Generate compaction result
-                const result = await compact(preparation, this.model, apiKey, headers, customInstructions, this._compactionAbortController.signal, this.thinkingLevel, this.agent.streamFn);
+                const result = await compact(preparation, compactionModel, apiKey, headers, customInstructions, this._compactionAbortController.signal, this.thinkingLevel, this.agent.streamFn);
                 summary = result.summary;
                 firstKeptEntryId = result.firstKeptEntryId;
                 tokensBefore = result.tokensBefore;
@@ -1976,17 +1897,21 @@ export class AgentSession {
                 });
                 return false;
             }
+            const compactionModel = resolveCompactionModel(this.model, this._modelRegistry.getAvailable());
             let apiKey;
             let headers;
             if (this.agent.streamFn === streamSimple) {
-                const authResult = await this._modelRegistry.getApiKeyAndHeaders(this.model);
+                const authResult = await this._modelRegistry.getApiKeyAndHeaders(compactionModel);
                 if (!authResult.ok || !authResult.apiKey) {
+                    const providerLabel = compactionModel.provider;
                     this._emit({
                         type: "compaction_end",
                         reason,
                         result: undefined,
                         aborted: false,
                         willRetry: false,
+                        errorMessage: `Auto-compaction could not authenticate for "${providerLabel}" (${compactionModel.id}). ` +
+                            `Check proxy/OAuth health and run '/login ${providerLabel}' if needed.`,
                     });
                     return false;
                 }
@@ -1994,7 +1919,7 @@ export class AgentSession {
                 headers = authResult.headers;
             }
             else {
-                ({ apiKey, headers } = await this._getCompactionRequestAuth(this.model));
+                ({ apiKey, headers } = await this._getCompactionRequestAuth(compactionModel));
             }
             const pathEntries = this.sessionManager.getBranch();
             const preparation = prepareCompaction(pathEntries, settings);
@@ -2046,7 +1971,7 @@ export class AgentSession {
             }
             else {
                 // Generate compaction result
-                const compactResult = await compact(preparation, this.model, apiKey, headers, undefined, this._autoCompactionAbortController.signal, this.thinkingLevel, this.agent.streamFn);
+                const compactResult = await compact(preparation, compactionModel, apiKey, headers, undefined, this._autoCompactionAbortController.signal, this.thinkingLevel, this.agent.streamFn);
                 summary = compactResult.summary;
                 firstKeptEntryId = compactResult.firstKeptEntryId;
                 tokensBefore = compactResult.tokensBefore;
@@ -2081,7 +2006,8 @@ export class AgentSession {
                 tokensBefore,
                 details,
             };
-            this._emit({ type: "compaction_end", reason, result, aborted: false, willRetry });
+            const emitWillRetry = compactionEmitWillRetry(willRetry, this.agent.hasQueuedMessages());
+            this._emit({ type: "compaction_end", reason, result, aborted: false, willRetry: emitWillRetry });
             if (willRetry) {
                 const messages = this.agent.state.messages;
                 const lastMsg = messages[messages.length - 1];
@@ -2270,6 +2196,10 @@ export class AgentSession {
             },
             getThinkingLevel: () => this.thinkingLevel,
             setThinkingLevel: (level) => this.setThinkingLevel(level),
+            // Optional: no in-process MCP client manager yet. Leave unbound so
+            // ExtensionAPI.callMcpTool remains present (load-time capture) but throws
+            // until a session-level handler is provided. Tests / future MCP hub bind here.
+            callMcpTool: undefined,
         }, {
             getModel: () => this.model,
             isIdle: () => !this.isStreaming,