@code-yeongyu/senpi 2026.7.31 → 2026.8.1

package/dist/core/agent-session.js

@@ -195,10 +195,14 @@ export class AgentSession {
         this._isAgentRunActive = false;
         this._toolExecutionDepth = 0;
         this._promptStartPending = false;
+        this._nextInputId = 0;
         /** Tracks pending steering messages for UI display. Removed when delivered. */
         this._steeringMessages = [];
         /** Tracks pending follow-up messages for UI display. Removed when delivered. */
         this._followUpMessages = [];
+        /** Recovery-only order across both native queue modes and TUI compaction ownership. */
+        this._queuedInputOrder = [];
+        this._nextQueuedInputOrder = 0;
         /** Messages queued to be included with the next user prompt as context ("asides"). */
         this._pendingNextTurnMessages = [];
         // Queues held while the first post-compaction response is classified. Agent
@@ -234,7 +238,7 @@ export class AgentSession {
         this._userAbortPromise = undefined;
         this._agentAbortSource = undefined;
         this._suppressQueuedContinuationAfterUserAbort = false;
-        /** Set when clearQueue() drains non-empty queues, so abort() can detect the gap even after queues are cleared. */
+        /** Set when clearQueue({ abortWillFollow: true }) drains queues immediately before abort(). */
         this._hadClearedQueuedMessages = false;
         this._extensionEventSignal = undefined;
         // Bash execution state
@@ -252,7 +256,8 @@ export class AgentSession {
         // Base system prompt (without extension appends) - used to apply fresh appends each turn
         this._baseSystemPrompt = "";
         this._currentServiceTier = undefined;
-        this._lastHighReasoningWarningKey = undefined;
+        this._sessionFastMode = false;
+        this._shownHighReasoningWarningKeys = new Set();
         // Track last assistant message for auto-compaction check
         this._lastAssistantMessage = undefined;
         /** Internal handler for agent events - shared by subscribe and reconnect */
@@ -733,7 +738,9 @@ export class AgentSession {
                     this.agent.steer(message);
                 const userMessage = queuedMessages.find((message) => message.role === "user");
                 if (userMessage?.role === "user") {
-                    this._steeringMessages.push(this._extractUserMessageText(userMessage.content));
+                    const text = this._extractUserMessageText(userMessage.content);
+                    this._steeringMessages.push(text);
+                    this._recordQueuedInput(text, "steer");
                     this._emitQueueUpdate();
                 }
                 return;
@@ -941,6 +948,7 @@ export class AgentSession {
                 const steeringIndex = this._steeringMessages.indexOf(messageText);
                 if (steeringIndex !== -1) {
                     this._steeringMessages.splice(steeringIndex, 1);
+                    this._removeQueuedInput(messageText, "steer");
                     this._emitQueueUpdate();
                 }
                 else {
@@ -948,6 +956,7 @@ export class AgentSession {
                     const followUpIndex = this._followUpMessages.indexOf(messageText);
                     if (followUpIndex !== -1) {
                         this._followUpMessages.splice(followUpIndex, 1);
+                        this._removeQueuedInput(messageText, "followUp");
                         this._emitQueueUpdate();
                     }
                 }
@@ -1342,6 +1351,22 @@ export class AgentSession {
     get serviceTier() {
         return this._currentServiceTier;
     }
+    /**
+     * True when the active model is served at the priority ("fast") tier: either the model
+     * itself is configured for it (an `openai` `-fast` catalog variant, a scoped
+     * `provider/id:priority`) or an extension turned fast mode on for this session.
+     *
+     * Display-only: it never feeds request composition, so an extension toggling fast mode
+     * for one provider cannot leak `service_tier` into another provider's payload.
+     * `serviceTier` stays the single request-side source.
+     */
+    isFastModeActive() {
+        return this._sessionFastMode || this._currentServiceTier === "priority";
+    }
+    /** Session-scoped fast-mode indicator; never persisted, reset on session start. */
+    setSessionFastMode(enabled) {
+        this._sessionFastMode = enabled;
+    }
     /**
      * Explicit scoped/favorite tiers win; otherwise fall back to the model's
      * configured serviceTier from models.json/extension compatibility config.
@@ -1683,6 +1708,12 @@ export class AgentSession {
         let messages;
         let titlePrompt;
         let consumedNextTurnMessages;
+        let inputId;
+        const emitInputDisposition = async (disposition) => {
+            if (inputId === undefined)
+                return;
+            await this._extensionRunner.emit({ type: "input_disposition", inputId, disposition });
+        };
         try {
             // Handle extension commands first (execute immediately, even during streaming)
             // Extension commands manage their own LLM interaction via pi.sendMessage()
@@ -1700,9 +1731,11 @@ export class AgentSession {
             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);
+                inputId = `${this.sessionManager.getSessionId()}:${++this._nextInputId}`;
+                const inputResult = await this._extensionRunner.emitInput(currentText, currentImages, options?.source ?? "interactive", this.isStreaming ? options?.streamingBehavior : undefined, inputId);
                 throwIfCancelled();
                 if (inputResult.action === "handled") {
+                    await emitInputDisposition("handled");
                     promptDisposition?.("handled");
                     preflightResult?.(true);
                     return;
@@ -1733,6 +1766,7 @@ export class AgentSession {
                 else {
                     await this._queueSteer(expandedText, currentImages);
                 }
+                await emitInputDisposition("queued");
                 promptDisposition?.("queued");
                 preflightResult?.(true);
                 return;
@@ -1750,6 +1784,7 @@ export class AgentSession {
                 else {
                     await this._queueSteer(expandedText, currentImages);
                 }
+                await emitInputDisposition("queued");
                 promptDisposition?.("queued");
                 preflightResult?.(true);
                 return;
@@ -1773,6 +1808,7 @@ export class AgentSession {
                     else {
                         await this._queueSteer(expandedText, currentImages);
                     }
+                    await emitInputDisposition("queued");
                     promptDisposition?.("queued");
                     preflightResult?.(true);
                     return;
@@ -1793,6 +1829,7 @@ export class AgentSession {
                 else {
                     await this._queueSteer(expandedText, currentImages);
                 }
+                await emitInputDisposition("queued");
                 promptDisposition?.("queued");
                 preflightResult?.(true);
                 return;
@@ -1866,8 +1903,11 @@ export class AgentSession {
             // next-turn, and before_agent_start additions are also provider-visible,
             // so make one final admission decision against the complete request.
             await this._enforceFinalProviderAdmission(messages);
+            throwIfCancelled();
+            await emitInputDisposition("started");
         }
         catch (error) {
+            await emitInputDisposition("rejected");
             if (consumedNextTurnMessages && consumedNextTurnMessages.length > 0) {
                 this._pendingNextTurnMessages = [...consumedNextTurnMessages, ...this._pendingNextTurnMessages];
             }
@@ -1881,7 +1921,6 @@ export class AgentSession {
         if (!messages) {
             return;
         }
-        throwIfCancelled();
         promptDisposition?.("started");
         preflightResult?.(true);
         if (options?.thinkingLevel !== undefined) {
@@ -2001,7 +2040,7 @@ export class AgentSession {
      * @param images Optional image attachments to include with the message
      * @throws Error if text is an extension command
      */
-    async steer(text, images) {
+    async steer(text, images, recovery) {
         // Check for extension commands (cannot be queued)
         if (text.startsWith("/")) {
             this._throwIfExtensionCommand(text);
@@ -2009,7 +2048,7 @@ export class AgentSession {
         // Expand skill commands and prompt templates
         let expandedText = this._expandSkillCommand(text);
         expandedText = expandPromptTemplate(expandedText, [...this.promptTemplates]);
-        await this._queueSteer(expandedText, images);
+        await this._queueSteer(expandedText, images, recovery?.enqueueOrder);
     }
     /**
      * Queue a follow-up message to be processed after the agent finishes.
@@ -2018,7 +2057,7 @@ export class AgentSession {
      * @param images Optional image attachments to include with the message
      * @throws Error if text is an extension command
      */
-    async followUp(text, images) {
+    async followUp(text, images, recovery) {
         // Check for extension commands (cannot be queued)
         if (text.startsWith("/")) {
             this._throwIfExtensionCommand(text);
@@ -2026,7 +2065,7 @@ export class AgentSession {
         // Expand skill commands and prompt templates
         let expandedText = this._expandSkillCommand(text);
         expandedText = expandPromptTemplate(expandedText, [...this.promptTemplates]);
-        await this._queueFollowUp(expandedText, images);
+        await this._queueFollowUp(expandedText, images, recovery?.enqueueOrder);
     }
     _startSessionTitleGeneration(firstPrompt) {
         if (!this._autoTitleSessions || this.sessionManager.getSessionName() || shouldSkipSessionTitle(firstPrompt)) {
@@ -2100,8 +2139,9 @@ export class AgentSession {
     /**
      * Internal: Queue a steering message (already expanded, no extension command check).
      */
-    async _queueSteer(text, images) {
+    async _queueSteer(text, images, enqueueOrder) {
         this._steeringMessages.push(text);
+        this._recordQueuedInput(text, "steer", enqueueOrder);
         this._sessionLogger.debug("queue_enqueue", { mode: "steer", count: this._steeringMessages.length });
         this._emitQueueUpdate();
         const content = [{ type: "text", text }];
@@ -2118,8 +2158,9 @@ export class AgentSession {
     /**
      * Internal: Queue a follow-up message (already expanded, no extension command check).
      */
-    async _queueFollowUp(text, images) {
+    async _queueFollowUp(text, images, enqueueOrder) {
         this._followUpMessages.push(text);
+        this._recordQueuedInput(text, "followUp", enqueueOrder);
         this._sessionLogger.debug("queue_enqueue", { mode: "followUp", count: this._followUpMessages.length });
         this._emitQueueUpdate();
         const content = [{ type: "text", text }];
@@ -2311,26 +2352,46 @@ export class AgentSession {
             finishSessionWork?.();
         }
     }
+    /** Reserve a global order for input temporarily owned outside the native queues. */
+    reserveQueuedInputOrder() {
+        this._nextQueuedInputOrder += 1;
+        return this._nextQueuedInputOrder;
+    }
+    _recordQueuedInput(text, mode, enqueueOrder) {
+        const order = enqueueOrder ?? this.reserveQueuedInputOrder();
+        this._nextQueuedInputOrder = Math.max(this._nextQueuedInputOrder, order);
+        this._queuedInputOrder.push({ text, mode, enqueueOrder: order });
+    }
+    _removeQueuedInput(text, mode) {
+        const index = this._queuedInputOrder.findIndex((message) => message.mode === mode && message.text === text);
+        if (index !== -1)
+            this._queuedInputOrder.splice(index, 1);
+    }
     /**
-     * Clear all queued messages and return them.
-     * Useful for restoring to editor when user aborts.
-     * @returns Object with steering and followUp arrays
+     * Clear all queued messages and return them. The non-enumerable `ordered`
+     * view preserves legacy object equality and native queue semantics.
+     * @param options.abortWillFollow Mark a non-empty drain so an immediately following abort can emit session_abort.
      */
-    clearQueue() {
+    clearQueue(options = { abortWillFollow: false }) {
         const steering = [...this._steeringMessages];
         const followUp = [...this._followUpMessages];
-        if (steering.length > 0 || followUp.length > 0)
+        const ordered = [...this._queuedInputOrder].sort((a, b) => a.enqueueOrder - b.enqueueOrder);
+        if (options.abortWillFollow && (steering.length > 0 || followUp.length > 0)) {
             this._hadClearedQueuedMessages = true;
+        }
         // Clear every queue synchronously. Deferred post-compaction messages are
         // already represented in visible bookkeeping, so they must not be returned
         // a second time or later resurrected into Agent's native queues.
         this._steeringMessages = [];
         this._followUpMessages = [];
+        this._queuedInputOrder = [];
         this._postCompactionDeferredSteeringMessages = [];
         this._postCompactionDeferredFollowUpMessages = [];
         this.agent.clearAllQueues();
         this._emitQueueUpdate();
-        return { steering, followUp };
+        const cleared = { steering, followUp };
+        Object.defineProperty(cleared, "ordered", { value: ordered, enumerable: false });
+        return cleared;
     }
     /** Number of pending messages (includes both steering and follow-up) */
     get pendingMessageCount() {
@@ -2487,9 +2548,6 @@ export class AgentSession {
         if (opts.ephemeralThinkingLevel !== undefined) {
             this._applyEphemeralThinkingLevel(thinkingLevel);
         }
-        else if (opts.persistDefault) {
-            this.setThinkingLevel(thinkingLevel);
-        }
         else {
             this.setSessionThinkingLevel(thinkingLevel);
         }
@@ -2543,10 +2601,10 @@ export class AgentSession {
         this.settingsManager.setDefaultModelAndProvider(next.model.provider, next.model.id);
         this._currentServiceTier = this._resolveServiceTier(next.model, next.serviceTier);
         // Apply thinking level.
-        // - Explicit favorite model thinking level overrides current session level
-        // - Undefined favorite model thinking level inherits the current session preference
-        // setThinkingLevel clamps to model capabilities.
-        this.setThinkingLevel(thinkingLevel);
+        // - Explicit favorite model thinking level overrides the effective session level
+        // - Undefined favorite model thinking level restores the remembered user preference
+        // setSessionThinkingLevel clamps without replacing that preference.
+        this.setSessionThinkingLevel(thinkingLevel);
         const systemPromptChange = await this._emitModelSelect(next.model, currentModel, "cycle");
         const cycleResult = { model: next.model, thinkingLevel: this.thinkingLevel, isScoped: true };
         if (systemPromptChange) {
@@ -2599,15 +2657,12 @@ export class AgentSession {
     _emitHighReasoningWarningIfNeeded() {
         const model = this.model;
         const level = this.thinkingLevel;
-        if (!model || !shouldWarnHighReasoning(model, level)) {
-            this._lastHighReasoningWarningKey = undefined;
+        if (!model || !shouldWarnHighReasoning(model, level))
             return;
-        }
-        const key = `${model.provider}/${model.id}:${level}`;
-        if (this._lastHighReasoningWarningKey === key) {
+        const key = `${model.provider}/${model.id}`;
+        if (this._shownHighReasoningWarningKeys.has(key))
             return;
-        }
-        this._lastHighReasoningWarningKey = key;
+        this._shownHighReasoningWarningKeys.add(key);
         this._emit({
             type: "high_reasoning_warning",
             modelId: model.id,
@@ -2660,10 +2715,7 @@ export class AgentSession {
         if (explicitLevel !== undefined) {
             return explicitLevel;
         }
-        if (!this.supportsThinking()) {
-            return this.settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL;
-        }
-        return this.thinkingLevel;
+        return this.settingsManager.getDefaultThinkingLevel() ?? this.thinkingLevel ?? DEFAULT_THINKING_LEVEL;
     }
     _clampThinkingLevel(level, availableLevels) {
         const available = new Set(availableLevels);
@@ -3287,6 +3339,8 @@ export class AgentSession {
             if (compacted)
                 return true;
         }
+        if (this._isCompactionOnCooldown())
+            return false;
         throw new RequiredCompactionError();
     }
     /**
@@ -3325,6 +3379,8 @@ export class AgentSession {
             throw new RequiredCompactionError();
         }
         const compacted = await this._runPrePromptCompaction(lastAssistantMessage, false, "pre_prompt");
+        if (!compacted && !isOversized() && this._isCompactionOnCooldown())
+            return;
         if (!compacted || isOversized()) {
             throw new RequiredCompactionError();
         }
@@ -3470,6 +3526,10 @@ export class AgentSession {
         }
         return false;
     }
+    _isCompactionOnCooldown() {
+        const state = this._compactionLifecycle.state;
+        return state.status === "failed" && state.rejectionCause === "circuit-breaker";
+    }
     async _runPrePromptCompaction(lastAssistantMessage, skipAbortedCheck, reason = "pre_prompt", willRetry = false, allowSummaryOnly = false) {
         const controller = new AbortController();
         this._claimCompactionController(controller, "compaction");
@@ -3530,8 +3590,11 @@ export class AgentSession {
         if (!shouldCompact(contextTokens, model.contextWindow, settings))
             return;
         const compacted = await this._runPrePromptCompaction(this._findLastAssistantMessage(), true, "pre_prompt");
-        if (!compacted)
+        if (!compacted) {
+            if (this._isCompactionOnCooldown())
+                return;
             throw new RequiredCompactionError();
+        }
         this._scheduledContinuationRecompacted = true;
     }
     async _continueAgentAfterCurrentRun(options = {}) {
@@ -3909,6 +3972,7 @@ export class AgentSession {
                 return true;
             },
             setSessionThinkingLevel: (level) => this.setSessionThinkingLevel(level),
+            setSessionFastMode: (enabled) => this.setSessionFastMode(enabled),
         }, {
             getModel: () => this.model,
             getServiceTier: () => this.serviceTier,
@@ -4398,6 +4462,15 @@ export class AgentSession {
             // Refusals are only retried through a new chain candidate. They never use
             // same-model retries or the transient over-budget fallback escape hatch.
             if (this._retryAttempt + 1 > settings.maxRetries) {
+                if (this._retryAttempt > 0) {
+                    this._emit({
+                        type: "auto_retry_end",
+                        success: false,
+                        attempt: this._retryAttempt,
+                        finalError: message.errorMessage,
+                    });
+                }
+                this._retryAttempt = 0;
                 this._resolveRetry();
                 return "not-handled";
             }
@@ -4407,6 +4480,15 @@ export class AgentSession {
                 if (exhaustedChainKey) {
                     this._emit({ type: "retry_fallback_exhausted", chainKey: exhaustedChainKey, lastError: errorMessage });
                 }
+                if (this._retryAttempt > 0) {
+                    this._emit({
+                        type: "auto_retry_end",
+                        success: false,
+                        attempt: this._retryAttempt,
+                        finalError: message.errorMessage,
+                    });
+                }
+                this._retryAttempt = 0;
                 this._resolveRetry();
                 return "not-handled";
             }
@@ -4541,7 +4623,8 @@ export class AgentSession {
         if (compactionSettings.enabled &&
             model &&
             shouldCompact(contextTokens, model.contextWindow, compactionSettings)) {
-            if (!(await this._runPrePromptCompaction(message, true, "threshold", true, true))) {
+            const preRetryCompaction = await this._runPrePromptCompaction(message, true, "threshold", true, true);
+            if (!preRetryCompaction && !this._isCompactionOnCooldown()) {
                 const attempt = this._retryAttempt;
                 this._retryAttempt = 0;
                 this._emit({