@earendil-works/pi-coding-agent 0.84.2 → 0.84.4

package/dist/core/agent-session.js

@@ -12,26 +12,26 @@
  *
  * Modes use this class and add their own I/O layer on top.
  */
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { readFileSync } from "node:fs";
 import { basename, dirname } from "node:path";
 import { contentText } from "@earendil-works/pi-ai";
 import { clampThinkingLevel, cleanupSessionResources, getSupportedThinkingLevels, isContextOverflow, isRecoverableLength, isRetryableAssistantError, modelsAreEqual, resetApiProviders, streamSimple, } from "@earendil-works/pi-ai/compat";
 import { getThemeByName, theme } from "../modes/interactive/theme/theme.js";
 import { stripFrontmatter } from "../utils/frontmatter.js";
-import { resolvePath } from "../utils/paths.js";
 import { sleep } from "../utils/sleep.js";
 import { normalizeToolResultImages } from "../utils/tool-result-images.js";
 import { formatNoApiKeyFoundMessage, formatNoModelSelectedMessage } from "./auth-guidance.js";
 import { executeBashWithOperations } from "./bash-executor.js";
 import { calculateContextTokens, collectEntriesForBranchSummary, compact, estimateContextTokens, estimateTokens, generateBranchSummary, prepareCompaction, shouldCompact, } from "./compaction/index.js";
-import { DEFAULT_THINKING_LEVEL } from "./defaults.js";
+import { DEFAULT_THINKING_LEVEL, THINKING_LEVEL_OPTIONS } from "./defaults.js";
 import { exportSessionToHtml } from "./export-html/index.js";
 import { createToolHtmlRenderer } from "./export-html/tool-renderer.js";
 import { ExtensionRunner, wrapRegisteredTools, } from "./extensions/index.js";
 import { emitSessionShutdownEvent } from "./extensions/runner.js";
 import { ModelRegistry } from "./model-registry.js";
 import { expandPromptTemplate } from "./prompt-templates.js";
-import { CURRENT_SESSION_VERSION, getLatestCompactionEntry } from "./session-manager.js";
+import { exportSessionToJsonl } from "./session-export.js";
+import { getLatestCompactionEntry } from "./session-manager.js";
 import { createSyntheticSourceInfo } from "./source-info.js";
 import { buildSystemPrompt } from "./system-prompt.js";
 import { createLocalBashOperations } from "./tools/bash.js";
@@ -71,8 +71,6 @@ function estimateMessagesTokens(messages) {
 // ============================================================================
 // Constants
 // ============================================================================
-/** Standard thinking levels */
-const THINKING_LEVELS = ["off", "minimal", "low", "medium", "high"];
 // ============================================================================
 // AgentSession Class
 // ============================================================================
@@ -93,6 +91,8 @@ export class AgentSession {
     _followUpMessages = [];
     /** Messages queued to be included with the next user prompt as context ("asides"). */
     _pendingNextTurnMessages = [];
+    /** Context-only custom messages queued during a run, flushed once the current turn's tool results are in. */
+    _pendingCustomMessages = [];
     // Compaction state
     _compactionAbortController = undefined;
     _autoCompactionAbortController = undefined;
@@ -271,18 +271,33 @@ export class AgentSession {
             };
         };
     }
+    async _compactBeforeNextAssistantResponse(context) {
+        const model = this.model;
+        const settings = this.settingsManager.getCompactionSettings();
+        if (!model ||
+            model.contextWindow <= 0 ||
+            !shouldCompact(estimateContextTokens(context.messages).tokens, model.contextWindow, settings)) {
+            return context;
+        }
+        await this._runAutoCompaction("threshold", false);
+        return {
+            ...context,
+            messages: this.agent.state.messages.slice(),
+        };
+    }
     _installAgentNextTurnRefresh() {
         const previousPrepareNextTurnWithContext = this.agent.prepareNextTurnWithContext ??
             (this.agent.prepareNextTurn
                 ? async (_turn, signal) => await this.agent.prepareNextTurn?.(signal)
                 : undefined);
         this.agent.prepareNextTurnWithContext = async (turn, signal) => {
-            const previousSnapshot = await previousPrepareNextTurnWithContext?.(turn, signal);
-            const previousContext = previousSnapshot?.context ?? turn.context;
+            const context = await this._compactBeforeNextAssistantResponse(turn.context);
+            const previousSnapshot = await previousPrepareNextTurnWithContext?.({ ...turn, context }, signal);
+            const nextContext = previousSnapshot?.context ?? context;
             return {
                 ...previousSnapshot,
                 context: {
-                    ...previousContext,
+                    ...nextContext,
                     systemPrompt: this._systemPromptOverride ?? this._baseSystemPrompt,
                     tools: this.agent.state.tools.slice(),
                 },
@@ -307,6 +322,11 @@ export class AgentSession {
             followUp: [...this._followUpMessages],
         });
     }
+    async _emitSessionCompactFailed(event) {
+        if (this._extensionRunner.hasHandlers("session_compact_failed")) {
+            await this._extensionRunner.emit({ type: "session_compact_failed", ...event });
+        }
+    }
     _getIdleWaitPromise() {
         if (!this._idleWaitPromise) {
             this._idleWaitPromise = new Promise((resolve) => {
@@ -397,6 +417,14 @@ export class AgentSession {
                 }
             }
         }
+        // A turn ends after its assistant message and every tool result has been appended,
+        // so this is the first point in the run where a context-only custom message can be
+        // inserted without landing between a tool call and its result. Flushing after the
+        // extension and listener dispatch above also picks up messages that turn_end
+        // handlers queued.
+        if (event.type === "turn_end") {
+            this._flushPendingCustomMessages();
+        }
     };
     _willRetryAfterAgentEnd(event) {
         const settings = this.settingsManager.getRetrySettings();
@@ -752,6 +780,7 @@ export class AgentSession {
         finally {
             this._systemPromptOverride = undefined;
             this._flushPendingBashMessages();
+            this._flushPendingCustomMessages();
             await this._emitAgentSettled();
         }
     }
@@ -841,8 +870,9 @@ export class AgentSession {
                 preflightResult?.(true);
                 return;
             }
-            // Flush any pending bash messages before the new prompt
+            // Flush any pending bash and custom messages before the new prompt
             this._flushPendingBashMessages();
+            this._flushPendingCustomMessages();
             // Validate model
             if (!this.model) {
                 throw new Error(formatNoModelSelectedMessage());
@@ -1056,8 +1086,9 @@ export class AgentSession {
     /**
      * Send a custom message to the session. Creates a CustomMessageEntry.
      *
-     * Handles three cases:
+     * Handles four cases:
      * - Streaming: queues message, processed when loop pulls from queue
+     * - Streaming + triggerTurn false: appended to state/session once the current turn ends
      * - Not streaming + triggerTurn: appends to state/session, starts new turn
      * - Not streaming + no trigger: appends to state/session, no turn
      *
@@ -1089,11 +1120,34 @@ export class AgentSession {
         else if (options?.triggerTurn) {
             await this._runAgentPrompt(appMessage);
         }
+        else if (this.isStreaming) {
+            // Appending now would put the message between an assistant tool call and its
+            // result, which providers that validate message order reject on replay. Defer
+            // to the end of the turn. Nothing is emitted yet: message events must not
+            // describe messages the session tree does not contain.
+            this._pendingCustomMessages.push(appMessage);
+        }
         else {
-            this.agent.state.messages.push(appMessage);
-            this.sessionManager.appendCustomMessageEntry(message.customType, message.content, message.display, message.details);
-            this._emit({ type: "message_start", message: appMessage });
-            this._emit({ type: "message_end", message: appMessage });
+            this._appendCustomMessage(appMessage);
+        }
+    }
+    _appendCustomMessage(appMessage) {
+        this.agent.state.messages.push(appMessage);
+        this.sessionManager.appendCustomMessageEntry(appMessage.customType, appMessage.content, appMessage.display, appMessage.details);
+        this._emit({ type: "message_start", message: appMessage });
+        this._emit({ type: "message_end", message: appMessage });
+    }
+    /**
+     * Append custom messages queued while the agent was running.
+     * Called once the current turn's tool results are in agent state and session history.
+     */
+    _flushPendingCustomMessages() {
+        if (this._pendingCustomMessages.length === 0)
+            return;
+        const pending = this._pendingCustomMessages;
+        this._pendingCustomMessages = [];
+        for (const appMessage of pending) {
+            this._appendCustomMessage(appMessage);
         }
     }
     /**
@@ -1191,35 +1245,55 @@ export class AgentSession {
     }
     /**
      * Set model directly.
-     * Validates that auth is configured, saves to session and settings.
+     * Validates that auth is configured and saves to the session transcript.
+     * Persists to global defaults only when options.persist is true.
      * @throws Error if no auth is configured for the model
      */
-    async setModel(model) {
+    async setModel(model, options = {}) {
         if (!(await this._modelRuntime.checkAuth(model.provider))) {
             throw new Error(`No API key for ${model.provider}/${model.id}`);
         }
         const previousModel = this.model;
-        const thinkingLevel = this._getThinkingLevelForModelSwitch();
+        const thinkingLevel = this._getThinkingLevelForModelSwitch(model);
         this.agent.state.model = model;
         this.sessionManager.appendModelChange(model.provider, model.id);
-        this.settingsManager.setDefaultModelAndProvider(model.provider, model.id);
-        // Re-clamp thinking level for new model's capabilities
+        if (options.persist) {
+            this.settingsManager.setDefaultModelAndProvider(model.provider, model.id);
+            this._addPersistedDefaultToNonEmptyScope(model);
+        }
+        // Apply thinking level for the new model.
+        // Per-model thinking level overrides take priority over the global default.
+        // Model persistence does not implicitly rewrite the global thinking default.
         this.setThinkingLevel(thinkingLevel);
         await this._emitModelSelect(model, previousModel, "set");
     }
+    _addPersistedDefaultToNonEmptyScope(model) {
+        if (this._scopedModels.length === 0)
+            return;
+        if (this._scopedModels.some((scoped) => modelsAreEqual(scoped.model, model)))
+            return;
+        this._scopedModels = [...this._scopedModels, { model }];
+        const enabledModels = this.settingsManager.getEnabledModels();
+        if (!enabledModels?.length)
+            return;
+        const modelReference = `${model.provider}/${model.id}`;
+        if (enabledModels.some((pattern) => pattern.toLowerCase() === modelReference.toLowerCase()))
+            return;
+        this.settingsManager.setEnabledModels([...enabledModels, modelReference]);
+    }
     /**
      * Cycle to next/previous model.
      * Uses scoped models (from --models flag) if available, otherwise all available models.
      * @param direction - "forward" (default) or "backward"
      * @returns The new model info, or undefined if only one model available
      */
-    async cycleModel(direction = "forward") {
+    async cycleModel(direction = "forward", options = {}) {
         if (this._scopedModels.length > 0) {
-            return this._cycleScopedModel(direction);
+            return this._cycleScopedModel(direction, options);
         }
-        return this._cycleAvailableModel(direction);
+        return this._cycleAvailableModel(direction, options);
     }
-    async _cycleScopedModel(direction) {
+    async _cycleScopedModel(direction, options) {
         const availableIds = new Set(this._modelRuntime.getAvailableSnapshot().map((model) => `${model.provider}\0${model.id}`));
         const scopedModels = this._scopedModels.filter((scoped) => availableIds.has(`${scoped.model.provider}\0${scoped.model.id}`));
         if (scopedModels.length <= 1)
@@ -1231,20 +1305,24 @@ export class AgentSession {
         const len = scopedModels.length;
         const nextIndex = direction === "forward" ? (currentIndex + 1) % len : (currentIndex - 1 + len) % len;
         const next = scopedModels[nextIndex];
-        const thinkingLevel = this._getThinkingLevelForModelSwitch(next.thinkingLevel);
+        const thinkingLevel = this._getThinkingLevelForModelSwitch(next.model, next.thinkingLevel);
         // Apply model
         this.agent.state.model = next.model;
         this.sessionManager.appendModelChange(next.model.provider, next.model.id);
-        this.settingsManager.setDefaultModelAndProvider(next.model.provider, next.model.id);
-        // Apply thinking level.
-        // - Explicit scoped model thinking level overrides current session level
-        // - Undefined scoped model thinking level inherits the current session preference
+        if (options.persist) {
+            this.settingsManager.setDefaultModelAndProvider(next.model.provider, next.model.id);
+            this._addPersistedDefaultToNonEmptyScope(next.model);
+        }
+        // Apply thinking level for the new model.
+        // - Explicit scoped model thinking level overrides defaults
+        // - Per-model thinking level overrides take priority over the global default
         // setThinkingLevel clamps to model capabilities.
+        // Model persistence does not implicitly rewrite the global thinking default.
         this.setThinkingLevel(thinkingLevel);
         await this._emitModelSelect(next.model, currentModel, "cycle");
         return { model: next.model, thinkingLevel: this.thinkingLevel, isScoped: true };
     }
-    async _cycleAvailableModel(direction) {
+    async _cycleAvailableModel(direction, options) {
         const availableModels = this._modelRuntime.getAvailableSnapshot();
         if (availableModels.length <= 1)
             return undefined;
@@ -1255,11 +1333,15 @@ export class AgentSession {
         const len = availableModels.length;
         const nextIndex = direction === "forward" ? (currentIndex + 1) % len : (currentIndex - 1 + len) % len;
         const nextModel = availableModels[nextIndex];
-        const thinkingLevel = this._getThinkingLevelForModelSwitch();
+        const thinkingLevel = this._getThinkingLevelForModelSwitch(nextModel);
         this.agent.state.model = nextModel;
         this.sessionManager.appendModelChange(nextModel.provider, nextModel.id);
-        this.settingsManager.setDefaultModelAndProvider(nextModel.provider, nextModel.id);
-        // Re-clamp thinking level for new model's capabilities
+        if (options.persist) {
+            this.settingsManager.setDefaultModelAndProvider(nextModel.provider, nextModel.id);
+            this._addPersistedDefaultToNonEmptyScope(nextModel);
+        }
+        // Apply thinking level for the new model.
+        // Model persistence does not implicitly rewrite the global thinking default.
         this.setThinkingLevel(thinkingLevel);
         await this._emitModelSelect(nextModel, currentModel, "cycle");
         return { model: nextModel, thinkingLevel: this.thinkingLevel, isScoped: false };
@@ -1270,20 +1352,21 @@ export class AgentSession {
     /**
      * Set thinking level.
      * Clamps to model capabilities based on available thinking levels.
-     * Saves to session and settings only if the level actually changes.
+     * Saves the clamped level to the session transcript only if the level actually changes.
+     * Persists the requested level to global defaults only when options.persist is true.
      */
-    setThinkingLevel(level) {
+    setThinkingLevel(level, options = {}) {
         const availableLevels = this.getAvailableThinkingLevels();
         const effectiveLevel = availableLevels.includes(level) ? level : this._clampThinkingLevel(level, availableLevels);
         // Only persist if actually changing
         const previousLevel = this.agent.state.thinkingLevel;
         const isChanging = effectiveLevel !== previousLevel;
         this.agent.state.thinkingLevel = effectiveLevel;
+        if (options.persist) {
+            this.settingsManager.setDefaultThinkingLevel(level);
+        }
         if (isChanging) {
             this.sessionManager.appendThinkingLevelChange(effectiveLevel);
-            if (this.supportsThinking() || effectiveLevel !== "off") {
-                this.settingsManager.setDefaultThinkingLevel(effectiveLevel);
-            }
             this._emit({ type: "thinking_level_changed", level: effectiveLevel });
             void this._extensionRunner.emit({
                 type: "thinking_level_select",
@@ -1296,14 +1379,14 @@ export class AgentSession {
      * Cycle to next thinking level.
      * @returns New level, or undefined if model doesn't support thinking
      */
-    cycleThinkingLevel() {
+    cycleThinkingLevel(options = {}) {
         if (!this.supportsThinking())
             return undefined;
         const levels = this.getAvailableThinkingLevels();
         const currentIndex = levels.indexOf(this.thinkingLevel);
         const nextIndex = (currentIndex + 1) % levels.length;
         const nextLevel = levels[nextIndex];
-        this.setThinkingLevel(nextLevel);
+        this.setThinkingLevel(nextLevel, options);
         return nextLevel;
     }
     /**
@@ -1312,7 +1395,7 @@ export class AgentSession {
      */
     getAvailableThinkingLevels() {
         if (!this.model)
-            return THINKING_LEVELS;
+            return [...THINKING_LEVEL_OPTIONS];
         return getSupportedThinkingLevels(this.model);
     }
     /**
@@ -1321,14 +1404,18 @@ export class AgentSession {
     supportsThinking() {
         return !!this.model?.reasoning;
     }
-    _getThinkingLevelForModelSwitch(explicitLevel) {
+    _getThinkingLevelForModelSwitch(targetModel, explicitLevel) {
         if (explicitLevel !== undefined) {
             return explicitLevel;
         }
-        if (!this.supportsThinking()) {
-            return this.settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL;
+        // Per-model default takes priority when switching to a model that has one
+        if (targetModel) {
+            const perModel = this.settingsManager.getModelThinkingLevel(targetModel.provider, targetModel.id);
+            if (perModel !== undefined) {
+                return perModel;
+            }
         }
-        return this.thinkingLevel;
+        return this.settingsManager.getDefaultThinkingLevel() ?? this.thinkingLevel ?? DEFAULT_THINKING_LEVEL;
     }
     _clampThinkingLevel(level, _availableLevels) {
         return this.model ? clampThinkingLevel(this.model, level) : "off";
@@ -1359,15 +1446,30 @@ export class AgentSession {
     // =========================================================================
     // Compaction
     // =========================================================================
+    /** Generate Pi's built-in compaction summary for manual and automatic compaction. */
+    async _runDefaultCompaction(preparation, requestModel, apiKey, headers, customInstructions, signal, env, reason) {
+        return compact(preparation, requestModel, apiKey, headers, customInstructions, signal, this.thinkingLevel, this.agent.streamFunction, env, this.settingsManager.getRetrySettings(), this._summarizationRetryCallbacks({ source: "compaction", reason }), undefined);
+    }
     /**
      * Manually compact the session context.
-     * Aborts current agent operation first.
+     *
+     * This is the manual entry point used by `/compact`, RPC, and extensions. It is
+     * separate from automatic threshold/overflow compaction, which enters through
+     * `_checkCompaction()` and `_runAutoCompaction()`. After preparation and the
+     * `session_before_compact` hook, both paths call the lower-level `compact()`
+     * function imported from `./compaction/index.ts`, unless the hook cancels or
+     * supplies a custom result.
+     *
+     * Aborts the current agent operation first. Manual compaction never retries or
+     * continues the interrupted agent turn.
+     *
      * @param customInstructions Optional instructions for the compaction summary
      */
     async compact(customInstructions) {
         await this.abort();
         this._compactionAbortController = new AbortController();
         this._emit({ type: "compaction_start", reason: "manual" });
+        let fromExtension = false;
         try {
             if (!this.model) {
                 throw new Error(formatNoModelSelectedMessage());
@@ -1385,7 +1487,6 @@ export class AgentSession {
                 throw new Error("Nothing to compact (session too small)");
             }
             let extensionCompaction;
-            let fromExtension = false;
             if (this._extensionRunner.hasHandlers("session_before_compact")) {
                 const result = (await this._extensionRunner.emit({
                     type: "session_before_compact",
@@ -1418,8 +1519,8 @@ export class AgentSession {
                 details = extensionCompaction.details;
             }
             else {
-                // Generate compaction result
-                const result = await compact(preparation, requestModel, apiKey, headers, customInstructions, this._compactionAbortController.signal, this.thinkingLevel, this.agent.streamFunction, env, this.settingsManager.getRetrySettings(), this._summarizationRetryCallbacks({ source: "compaction", reason: "manual" }));
+                // Shared default summary generator, also used by automatic compaction.
+                const result = await this._runDefaultCompaction(preparation, requestModel, apiKey, headers, customInstructions, this._compactionAbortController.signal, env, "manual");
                 summary = result.summary;
                 firstKeptEntryId = result.firstKeptEntryId;
                 tokensBefore = result.tokensBefore;
@@ -1467,6 +1568,7 @@ export class AgentSession {
         catch (error) {
             const message = error instanceof Error ? error.message : String(error);
             const aborted = message === "Compaction cancelled" || (error instanceof Error && error.name === "AbortError");
+            const errorMessage = aborted ? undefined : `Compaction failed: ${message}`;
             this._compactionAbortController = undefined;
             this._emit({
                 type: "compaction_end",
@@ -1474,7 +1576,14 @@ export class AgentSession {
                 result: undefined,
                 aborted,
                 willRetry: false,
-                errorMessage: aborted ? undefined : `Compaction failed: ${message}`,
+                errorMessage,
+            });
+            await this._emitSessionCompactFailed({
+                reason: "manual",
+                errorMessage,
+                aborted,
+                willRetry: false,
+                fromExtension,
             });
             throw error;
         }
@@ -1496,16 +1605,25 @@ export class AgentSession {
         this._branchSummaryAbortController?.abort();
     }
     /**
-     * Check if compaction is needed and run it.
-     * Called after agent_end and before prompt submission.
+     * Dispatch automatic compaction after `agent_end` or before prompt submission.
+     * Manual compaction does not call this method; it enters through `compact()`.
+     *
+     * Automatic cases:
+     * 1. Overflow with retry: a context-overflow error or recoverable length stop;
+     *    remove the failed assistant message, compact, and retry the turn once.
+     * 2. Overflow without retry: a successful response exceeded the configured
+     *    context window; compact but preserve the completed response.
+     * 3. Threshold without retry: valid or estimated context usage crossed the
+     *    configured threshold; compact without retrying the completed response.
      *
-     * Two cases:
-     * 1. Recoverable failure: LLM returned context overflow or stopped below its desired output limit;
-     *    remove the assistant message, compact, and auto-retry once
-     * 2. Threshold: Context over threshold, compact, NO auto-retry (user continues manually)
+     * Each case calls `_runAutoCompaction()`. After preparation and the
+     * `session_before_compact` hook, that method calls the lower-level `compact()`
+     * function imported from `./compaction/index.ts`, unless the hook cancels or
+     * supplies a custom result.
      *
      * @param assistantMessage The assistant message to check
      * @param skipAbortedCheck If false, include aborted messages (for pre-prompt check). Default: true
+     * @returns Whether the post-run loop should call `agent.continue()` for overflow recovery or queued messages
      */
     async _checkCompaction(assistantMessage, skipAbortedCheck = true) {
         const settings = this.settingsManager.getCompactionSettings();
@@ -1528,38 +1646,49 @@ export class AgentSession {
         if (assistantIsFromBeforeCompaction) {
             return false;
         }
-        // Case 1: Recoverable failure. Explicit/silent context overflow still uses context metadata.
+        // Automatic cases 1 and 2: context overflow.
         // A length stop is recoverable when output ended below the model's original desired limit,
         // independent of the configured context size or any context-clamped provider request limit.
-        // A successful response over the configured window should compact but must not retry: the
-        // assistant answer already completed and agent.continue() cannot continue from an assistant.
+        const contextOverflow = sameModel && isContextOverflow(assistantMessage, contextWindow);
         const recoverableLength = sameModel && isRecoverableLength(assistantMessage, this.model?.maxTokens ?? 0);
-        if (sameModel && (isContextOverflow(assistantMessage, contextWindow) || recoverableLength)) {
+        if (contextOverflow || recoverableLength) {
             const willRetry = assistantMessage.stopReason !== "stop";
+            // Case 2: the response completed successfully. Compact, but do not retry because
+            // agent.continue() cannot continue from a completed assistant response.
             if (!willRetry) {
                 return await this._runAutoCompaction("overflow", false);
             }
             if (this._overflowRecoveryAttempted) {
+                const errorMessage = contextOverflow
+                    ? "Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model."
+                    : "Truncated response recovery failed after one compact-and-retry attempt.";
                 this._emit({
                     type: "compaction_end",
                     reason: "overflow",
                     result: undefined,
                     aborted: false,
                     willRetry: false,
-                    errorMessage: "Context overflow recovery failed after one compact-and-retry attempt. Try reducing context or switching to a larger-context model.",
+                    errorMessage,
+                });
+                await this._emitSessionCompactFailed({
+                    reason: "overflow",
+                    errorMessage,
+                    aborted: false,
+                    willRetry: false,
+                    fromExtension: false,
                 });
                 return false;
             }
+            // Case 1: remove the failed or truncated message from agent state, compact, and
+            // retry once. The message remains in session history but is excluded from retry context.
             this._overflowRecoveryAttempted = true;
-            // Remove the failed or truncated message from agent state. It remains in session history,
-            // but must not be included in the compact-and-retry context.
             const messages = this.agent.state.messages;
             if (messages.length > 0 && messages[messages.length - 1].role === "assistant") {
                 this.agent.state.messages = messages.slice(0, -1);
             }
             return await this._runAutoCompaction("overflow", willRetry);
         }
-        // Case 2: Threshold - context is getting large
+        // Case 3: threshold compaction without retry.
         // For error messages or all-zero usage messages, estimate from the last valid response.
         // This ensures sessions that hit persistent API errors (e.g. 529) or malformed zero-usage
         // responses can still compact and do not reset context accounting.
@@ -1568,16 +1697,18 @@ export class AgentSession {
         if (assistantMessage.stopReason === "error" || directContextTokens === 0) {
             const messages = this.agent.state.messages;
             const estimate = estimateContextTokens(messages);
-            if (estimate.lastUsageIndex === null)
-                return false; // No usage data at all
-            // Verify the usage source is post-compaction. Kept pre-compaction messages
-            // have stale usage reflecting the old (larger) context and would falsely
-            // trigger compaction right after one just finished.
-            const usageMsg = messages[estimate.lastUsageIndex];
-            if (compactionEntry &&
-                usageMsg.role === "assistant" &&
-                usageMsg.timestamp <= new Date(compactionEntry.timestamp).getTime()) {
-                return false;
+            // Without provider usage, estimate.tokens is the pure message-size estimate.
+            // Only usage-backed estimates need the stale pre-compaction check.
+            if (estimate.lastUsageIndex !== null) {
+                // Verify the usage source is post-compaction. Kept pre-compaction messages
+                // have stale usage reflecting the old (larger) context and would falsely
+                // trigger compaction right after one just finished.
+                const usageMsg = messages[estimate.lastUsageIndex];
+                if (compactionEntry &&
+                    usageMsg.role === "assistant" &&
+                    usageMsg.timestamp <= new Date(compactionEntry.timestamp).getTime()) {
+                    return false;
+                }
             }
             contextTokens = estimate.tokens;
         }
@@ -1590,11 +1721,19 @@ export class AgentSession {
         return false;
     }
     /**
-     * Internal: Run auto-compaction with events.
+     * Execute threshold or overflow compaction. Manual compaction uses
+     * `AgentSession.compact()` instead. Both paths call the lower-level `compact()`
+     * function imported from `./compaction/index.ts` after preparation and extension
+     * interception.
+     *
+     * @param reason Automatic trigger selected by `_checkCompaction()`
+     * @param willRetry Whether to continue the interrupted turn after overflow compaction
+     * @returns Whether the post-run loop should call `agent.continue()`
      */
     async _runAutoCompaction(reason, willRetry) {
         const settings = this.settingsManager.getCompactionSettings();
         let started = false;
+        let fromExtension = false;
         try {
             if (!this.model) {
                 return false;
@@ -1609,7 +1748,6 @@ export class AgentSession {
             this._autoCompactionAbortController = new AbortController();
             started = true;
             let extensionCompaction;
-            let fromExtension = false;
             if (this._extensionRunner.hasHandlers("session_before_compact")) {
                 const extensionResult = (await this._extensionRunner.emit({
                     type: "session_before_compact",
@@ -1628,6 +1766,12 @@ export class AgentSession {
                         aborted: true,
                         willRetry: false,
                     });
+                    await this._emitSessionCompactFailed({
+                        reason,
+                        aborted: true,
+                        willRetry: false,
+                        fromExtension: false,
+                    });
                     return false;
                 }
                 if (extensionResult?.compaction) {
@@ -1649,8 +1793,8 @@ export class AgentSession {
                 details = extensionCompaction.details;
             }
             else {
-                // Generate compaction result
-                const compactResult = await compact(preparation, requestModel, apiKey, headers, undefined, this._autoCompactionAbortController.signal, this.thinkingLevel, this.agent.streamFunction, env, this.settingsManager.getRetrySettings(), this._summarizationRetryCallbacks({ source: "compaction", reason }));
+                // Shared default summary generator, also used by manual compaction.
+                const compactResult = await this._runDefaultCompaction(preparation, requestModel, apiKey, headers, undefined, this._autoCompactionAbortController.signal, env, reason);
                 summary = compactResult.summary;
                 firstKeptEntryId = compactResult.firstKeptEntryId;
                 tokensBefore = compactResult.tokensBefore;
@@ -1665,6 +1809,12 @@ export class AgentSession {
                     aborted: true,
                     willRetry: false,
                 });
+                await this._emitSessionCompactFailed({
+                    reason,
+                    aborted: true,
+                    willRetry: false,
+                    fromExtension,
+                });
                 return false;
             }
             this.sessionManager.appendCompaction(summary, firstKeptEntryId, tokensBefore, details, fromExtension, usage);
@@ -1711,15 +1861,23 @@ export class AgentSession {
         catch (error) {
             const errorMessage = error instanceof Error ? error.message : "compaction failed";
             if (started) {
+                const formattedErrorMessage = reason === "overflow"
+                    ? `Context overflow recovery failed: ${errorMessage}`
+                    : `Auto-compaction failed: ${errorMessage}`;
                 this._emit({
                     type: "compaction_end",
                     reason,
                     result: undefined,
                     aborted: false,
                     willRetry: false,
-                    errorMessage: reason === "overflow"
-                        ? `Context overflow recovery failed: ${errorMessage}`
-                        : `Auto-compaction failed: ${errorMessage}`,
+                    errorMessage: formattedErrorMessage,
+                });
+                await this._emitSessionCompactFailed({
+                    reason,
+                    errorMessage: formattedErrorMessage,
+                    aborted: false,
+                    willRetry: false,
+                    fromExtension,
                 });
             }
             return false;
@@ -2607,29 +2765,7 @@ export class AgentSession {
      * @returns The resolved output file path.
      */
     exportToJsonl(outputPath) {
-        const filePath = resolvePath(outputPath ?? `session-${new Date().toISOString().replace(/[:.]/g, "-")}.jsonl`, process.cwd());
-        const dir = dirname(filePath);
-        if (!existsSync(dir)) {
-            mkdirSync(dir, { recursive: true });
-        }
-        const header = {
-            type: "session",
-            version: CURRENT_SESSION_VERSION,
-            id: this.sessionManager.getSessionId(),
-            timestamp: new Date().toISOString(),
-            cwd: this.sessionManager.getCwd(),
-        };
-        const branchEntries = this.sessionManager.getBranch();
-        const lines = [JSON.stringify(header)];
-        // Re-chain parentIds to form a linear sequence
-        let prevId = null;
-        for (const entry of branchEntries) {
-            const linear = { ...entry, parentId: prevId };
-            lines.push(JSON.stringify(linear));
-            prevId = entry.id;
-        }
-        writeFileSync(filePath, `${lines.join("\n")}\n`);
-        return filePath;
+        return exportSessionToJsonl(this.sessionManager, outputPath);
     }
     // =========================================================================
     // Utilities