open-multi-agent-kit 0.90.0 → 0.90.4

package/dist/core/agent-session.js

@@ -20,6 +20,7 @@ import { stripFrontmatter } from "../utils/frontmatter.js";
 import { resolvePath } from "../utils/paths.js";
 import { sleep } from "../utils/sleep.js";
 import { formatNoApiKeyFoundMessage, formatNoModelSelectedMessage } from "./auth-guidance.js";
+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";
@@ -31,6 +32,12 @@ import { ExtensionRunner, wrapRegisteredTools, } from "./extensions/index.js";
 import { emitSessionShutdownEvent } from "./extensions/runner.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";
 import { CURRENT_SESSION_VERSION, getLatestCompactionEntry } from "./session-manager.js";
 import { createSyntheticSourceInfo } from "./source-info.js";
@@ -53,6 +60,17 @@ export function parseSkillBlock(text) {
         userMessage: match[4]?.trim() || undefined,
     };
 }
+function mergePromptActiveSkillNames(first, second) {
+    const names = [];
+    const seen = new Set();
+    for (const name of [...first, ...second]) {
+        if (!seen.has(name)) {
+            seen.add(name);
+            names.push(name);
+        }
+    }
+    return names;
+}
 // ============================================================================
 // Constants
 // ============================================================================
@@ -86,6 +104,36 @@ export class AgentSession {
     sessionManager;
     settingsManager;
     _scopedModels;
+    /**
+     * Thinking mode (default "manual"). Persists across turns within the session;
+     * 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.
+     */
+    _taskClassHistory = [];
+    /**
+     * Compiled reasoning-router bias snapshot for the opt-in v4 learning path
+     * (Goal 010 Lane I). Loaded at most once per session and pinned thereafter:
+     * `_reasoningRouterBiasSnapshotLoaded` flips to `true` on the first v4
+     * auto-turn that has learning enabled, and `_reasoningRouterBiasSnapshot`
+     * then stays fixed for the rest of the session (even across later settings
+     * reloads, and even if the on-disk file changes or the load failed/was
+     * invalid, in which case it stays `null`). See `_getReasoningRouterBiasSnapshot`.
+     */
+    _reasoningRouterBiasSnapshot = null;
+    _reasoningRouterBiasSnapshotLoaded = false;
     // Event subscription state
     _unsubscribeAgent;
     _eventListeners = [];
@@ -540,6 +588,10 @@ export class AgentSession {
     get thinkingLevel() {
         return this.agent.state.thinkingLevel;
     }
+    /** Current thinking mode ("manual" = user-selected level, "auto" = per-turn router) */
+    get thinkingMode() {
+        return this._thinkingMode;
+    }
     /** Whether agent is currently streaming a response */
     get isStreaming() {
         return this.agent.state.isStreaming;
@@ -852,12 +904,29 @@ export class AgentSession {
     async prompt(text, options) {
         const expandPromptTemplates = options?.expandPromptTemplates ?? true;
         const preflightResult = options?.preflightResult;
+        let currentText = text;
+        let promptActiveSkillNames = [...(options?.activeSkillNames ?? [])];
+        let promptActiveSkillSource = options?.activeSkillSource;
+        let isBangSkillInvocation = false;
+        if (expandPromptTemplates) {
+            const bangInvocation = parseBangInvocation(text, {
+                hasSkill: (name) => this._resourceLoader.getSkills().skills.some((skill) => skill.name === name),
+            });
+            if (bangInvocation.kind === "skill") {
+                isBangSkillInvocation = true;
+                currentText = bangInvocation.prompt
+                    ? `/skill:${bangInvocation.skillName} ${bangInvocation.prompt}`
+                    : `/skill:${bangInvocation.skillName}`;
+                promptActiveSkillNames = mergePromptActiveSkillNames(promptActiveSkillNames, bangInvocation.activeSkillNames);
+                promptActiveSkillSource = bangInvocation.source;
+            }
+        }
         let messages;
         try {
             // Handle extension commands first (execute immediately, even during streaming)
             // Extension commands manage their own LLM interaction via omk.sendMessage()
-            if (expandPromptTemplates && text.startsWith("/")) {
-                const handled = await this._tryExecuteExtensionCommand(text);
+            if (expandPromptTemplates && !isBangSkillInvocation && currentText.startsWith("/")) {
+                const handled = await this._tryExecuteExtensionCommand(currentText);
                 if (handled) {
                     // Extension command executed, no prompt to send
                     preflightResult?.(true);
@@ -865,7 +934,6 @@ export class AgentSession {
                 }
             }
             // Emit input event for extension interception (before skill/template expansion)
-            let currentText = text;
             let currentImages = options?.images;
             if (this._extensionRunner.hasHandlers("input")) {
                 const inputResult = await this._extensionRunner.emitInput(currentText, currentImages, options?.source ?? "interactive", this.isStreaming ? options?.streamingBehavior : undefined);
@@ -926,6 +994,9 @@ export class AgentSession {
                     this._flushPendingBashMessages();
                 }
             }
+            // Auto thinking mode: resolve this turn's level from the prompt content.
+            // Manual mode never enters the router, so /think <level> always wins.
+            this._applyAutoThinkingLevelForTurn(expandedText);
             // Build messages array (custom message if any, then user message)
             messages = [];
             // Add user message
@@ -943,8 +1014,16 @@ export class AgentSession {
                 messages.push(msg);
             }
             this._pendingNextTurnMessages = [];
+            const turnSystemPromptOptions = promptActiveSkillNames.length > 0
+                ? {
+                    ...this._baseSystemPromptOptions,
+                    activeSkillNames: promptActiveSkillNames,
+                    ...(promptActiveSkillSource ? { activeSkillSource: promptActiveSkillSource } : {}),
+                }
+                : this._baseSystemPromptOptions;
+            const turnSystemPrompt = promptActiveSkillNames.length > 0 ? buildSystemPrompt(turnSystemPromptOptions) : this._baseSystemPrompt;
             // Emit before_agent_start extension event
-            const result = await this._extensionRunner.emitBeforeAgentStart(expandedText, currentImages, this._baseSystemPrompt, this._baseSystemPromptOptions);
+            const result = await this._extensionRunner.emitBeforeAgentStart(expandedText, currentImages, turnSystemPrompt, turnSystemPromptOptions);
             // Add all custom messages from extensions
             if (result?.messages) {
                 for (const msg of result.messages) {
@@ -964,7 +1043,7 @@ export class AgentSession {
             }
             else {
                 // Ensure we're using the base prompt (in case previous turn had modifications)
-                this.agent.state.systemPrompt = this._baseSystemPrompt;
+                this.agent.state.systemPrompt = turnSystemPrompt;
             }
             await this._checkProjectedCompaction(messages);
         }
@@ -1344,6 +1423,249 @@ export class AgentSession {
             });
         }
     }
+    /**
+     * Set the thinking mode. "auto" resolves a level per turn via the reasoning
+     * router; "manual" keeps the explicitly selected level. The mode persists for
+     * the session lifetime and is never written to user settings.
+     */
+    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
+     * notifies observers - but never overwrites the user's persisted default
+     * thinking level in settings. Models without reasoning support bypass the
+     * router entirely (level stays "off").
+     */
+    _applyAutoThinkingLevelForTurn(promptText) {
+        if (this._thinkingMode !== "auto")
+            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 });
+    }
+    /**
+     * 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
+     * `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
+     * adaptorch-bridge.ts). `bias` stays `0` unless BOTH hold: (a) the global,
+     * owner-only `reasoningRouterLearning.enabled` setting is `true` (default
+     * off; a project-scope `.omk/settings.json` value for this key is never
+     * consulted -- see settings-manager.ts), and (b) a compiled
+     * `RouterBiasSnapshot` was found and passed strict validation at the
+     * configured/default path, loaded and cached ("pinned") at most once per
+     * session (see `_getReasoningRouterBiasSnapshot`). When learning is
+     * enabled, exactly one bounded "accepted" feedback record (no raw prompt/
+     * path/diff/session/model/provider/tool/hook content; see
+     * router-feedback-collector.ts's exact ten-key schema) is appended to the
+     * local ledger after every v4 auto-turn resolution, for a future, separate
+     * offline compile step to learn from -- this lane never records an
+     * override/fail/hook-outcome signal, only the neutral "accepted" one. The
+     * resolver's own confidence-band/fallback-reason escalation (see
+     * reasoning-router-v4.ts) still applies on top of the base+lane+bias
+     * target, so a low-confidence or fallback-decided verdict can still only
+     * match or exceed the confident-path level.
+     */
+    _applyAutoThinkingLevelV4(promptText) {
+        const availableLevels = this.getAvailableThinkingLevels();
+        const verdict = classifyTaskV4({
+            prompt: promptText,
+            history: this._taskClassHistory,
+            pressureBucket: this._computePressureBucket(),
+        });
+        this._taskClassHistory.unshift(verdict.taskClass);
+        if (this._taskClassHistory.length > 8)
+            this._taskClassHistory.length = 8;
+        const learningEnabled = this.settingsManager.getReasoningRouterLearningEnabled();
+        const snapshot = learningEnabled ? this._getReasoningRouterBiasSnapshot() : null;
+        const features = this._deriveRouterFeedbackFeatures(promptText);
+        const bias = snapshot === null
+            ? 0
+            : getBiasStepsForCell(snapshot, {
+                predictedClass: verdict.taskClass,
+                laneType: "none",
+                lenBucket: features.lenBucket,
+                hadFence: features.hadFence,
+                hadDiff: features.hadDiff,
+            });
+        const resolved = resolveThinkingLevelV4WithUncertainty(verdict, availableLevels, undefined, bias, null);
+        if (learningEnabled && resolved !== "off") {
+            const record = {
+                routerVersion: "v4",
+                laneType: "none",
+                predictedClass: verdict.taskClass,
+                resolvedLevel: resolved,
+                acceptedLevel: resolved,
+                signal: "s2-accept",
+                outcome: "accepted",
+                lenBucket: features.lenBucket,
+                hadFence: features.hadFence,
+                hadDiff: features.hadDiff,
+            };
+            appendRouterFeedbackRecord(record, {
+                enabled: true,
+                ledgerPath: this.settingsManager.getReasoningRouterLearningFeedbackLedgerPath(),
+            });
+        }
+        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 });
+    }
+    /**
+     * Loads and strictly validates the compiled reasoning-router bias snapshot
+     * for the opt-in v4 learning path (Goal 010 Lane I), at most once per
+     * session ("pinned"): the first call attempts the read and caches whatever
+     * it finds (including a `null` miss/failure); every later call in the same
+     * session reuses that cached result without touching disk again. Returns
+     * `null` when no file exists at the configured/default path, the file
+     * cannot be read, or its contents fail `parseRouterBiasSnapshot`'s schema
+     * validation -- never throws.
+     */
+    _getReasoningRouterBiasSnapshot() {
+        if (this._reasoningRouterBiasSnapshotLoaded)
+            return this._reasoningRouterBiasSnapshot;
+        this._reasoningRouterBiasSnapshotLoaded = true;
+        const path = this.settingsManager.getReasoningRouterLearningBiasSnapshotPath() ?? getDefaultRouterBiasSnapshotPath();
+        try {
+            if (existsSync(path)) {
+                this._reasoningRouterBiasSnapshot = parseRouterBiasSnapshot(readFileSync(path, "utf-8"));
+            }
+        }
+        catch {
+            this._reasoningRouterBiasSnapshot = null;
+        }
+        return this._reasoningRouterBiasSnapshot;
+    }
+    /**
+     * 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
+     * the classifier scores; never returns raw prompt content or its exact
+     * length, only the clamped [0,7] bucket and two booleans.
+     */
+    _deriveRouterFeedbackFeatures(promptText) {
+        const trimmed = promptText.trim();
+        let lenBucket = 0;
+        let remaining = trimmed.length + 1;
+        while (remaining > 1 && lenBucket < 7) {
+            remaining >>= 1;
+            lenBucket++;
+        }
+        const hadFence = trimmed.includes("```");
+        const hadDiff = /^@@[^\n]*@@/m.test(trimmed) ||
+            /^diff --git /m.test(trimmed) ||
+            (/^\+(?!\+)/m.test(trimmed) && /^-(?!-)/m.test(trimmed));
+        return { lenBucket: lenBucket, hadFence, hadDiff };
+    }
+    /**
+     * Context-pressure band 0..3 from the projected token estimate over the
+     * model context window (reuses estimateProjectedContextTokens). 0..<0.5,
+     * 1..<0.75, 2..<0.9, 3..>=0.9. Inert under DEFAULT_WEIGHTS (no pressure
+     * coefficient); computed for future calibrated weight presets.
+     */
+    _computePressureBucket() {
+        const contextWindow = this.model?.contextWindow ?? 0;
+        if (contextWindow <= 0)
+            return 0;
+        const estimate = estimateProjectedContextTokens(this.agent.state.messages, []);
+        const pressure = estimate.tokens / contextWindow;
+        if (pressure >= 0.9)
+            return 3;
+        if (pressure >= 0.75)
+            return 2;
+        if (pressure >= 0.5)
+            return 1;
+        return 0;
+    }
     /**
      * Cycle to next thinking level.
      * @returns New level, or undefined if model doesn't support thinking