@vellumai/assistant 0.8.6 → 0.8.7

package/src/agent/loop.ts

@@ -7,23 +7,34 @@ import {
   getCalibrationProviderKey,
 } from "../context/token-estimator.js";
 import { calculateMaxToolResultChars } from "../context/tool-result-truncation.js";
+import type { ContextWindowResult } from "../context/window-manager.js";
 import type { ToolActivityMetadata } from "../daemon/message-types/web-activity.js";
-import { defaultEmptyResponseTerminal } from "../plugins/defaults/empty-response.js";
-import { defaultToolErrorTerminal } from "../plugins/defaults/tool-error.js";
-import { defaultToolResultTruncateTerminal } from "../plugins/defaults/tool-result-truncate.js";
+import { defaultCompactionTerminal } from "../plugins/defaults/compaction/terminal.js";
+import { defaultEmptyResponseTerminal } from "../plugins/defaults/empty-response/terminal.js";
+import { defaultTokenEstimateTerminal } from "../plugins/defaults/token-estimate/terminal.js";
+import { defaultToolErrorTerminal } from "../plugins/defaults/tool-error/terminal.js";
+import { defaultToolResultTruncateTerminal } from "../plugins/defaults/tool-result-truncate/terminal.js";
+import type {
+  ToolResultTruncateArgs,
+  ToolResultTruncateResult,
+} from "../plugins/defaults/tool-result-truncate/types.js";
 import { DEFAULT_TIMEOUTS, runPipeline } from "../plugins/pipeline.js";
 import { getMiddlewaresFor } from "../plugins/registry.js";
 import type {
+  CompactionArgs,
+  CompactionCircuitEvent,
+  CompactionResult,
   EmptyResponseArgs,
   EmptyResponseDecision,
+  EstimateArgs,
+  EstimateResult,
   LLMCallArgs,
   LLMCallResult,
   ToolErrorArgs,
   ToolErrorDecision,
-  ToolResultTruncateArgs,
-  ToolResultTruncateResult,
   TurnContext,
 } from "../plugins/types.js";
+import { PluginTimeoutError } from "../plugins/types.js";
 import { normalizeThinkingConfigForWire } from "../providers/thinking-config.js";
 import type {
   ContentBlock,
@@ -40,9 +51,17 @@ import {
 import { AssistantError, ErrorCode, ProviderError } from "../util/errors.js";
 import { getLogger } from "../util/logger.js";
 import { isRetryableNetworkError } from "../util/retry.js";
+import { CompactionCircuit } from "./compaction-circuit.js";
 
 const log = getLogger("agent-loop");
 
+/** Fraction of the preflight budget at which a checkpoint triggers mid-loop compaction. */
+const MID_LOOP_YIELD_THRESHOLD_RATIO = 0.85;
+
+/** In-context message count above which the budget gate raises the safety-margin floor. */
+const LONG_HISTORY_MESSAGE_THRESHOLD = 50;
+const LONG_HISTORY_SAFETY_MARGIN_FLOOR = 0.15;
+
 export interface AgentLoopConfig {
   maxTokens: number;
   maxInputTokens?: number; // context window size for tool result truncation
@@ -66,7 +85,28 @@ export interface CheckpointInfo {
   history: Message[]; // current history snapshot for token estimation
 }
 
-export type CheckpointDecision = "continue" | "yield";
+/**
+ * Why a checkpoint paused the loop. Surfaced back to the caller via
+ * {@link AgentLoopRunResult.exitReason} so the orchestrator reacts to
+ * the loop's own signal (hand off to a queued message vs. compact and
+ * re-enter) instead of the checkpoint callback mutating orchestrator state.
+ */
+export type ExitReason = "handoff" | "budget";
+
+export type CheckpointDecision = "continue" | ExitReason;
+
+/**
+ * Result of {@link AgentLoop.run}.
+ *
+ * `exitReason` carries the reason the loop paused at a checkpoint so the
+ * orchestrator reads the loop's own signal instead of inferring it from
+ * callback side-effects. It is `null` whenever the loop reached a terminal
+ * stop (completion, error, abort, or a tool-requested yield-to-user).
+ */
+export interface AgentLoopRunResult {
+  history: Message[];
+  exitReason: ExitReason | null;
+}
 
 /**
  * Why an agent turn reached a terminal state.
@@ -246,6 +286,24 @@ export type AgentEvent =
        */
       estimatedInputTokens?: number;
     }
+  | {
+      /**
+       * Emitted when the loop begins compacting the running history because
+       * the mid-loop budget gate tripped. The daemon's event dispatcher
+       * translates it into a "compacting context" activity state so clients
+       * surface that the turn paused to summarize context.
+       */
+      type: "context_compacting";
+    }
+  /**
+   * Circuit-breaker transitions emitted when auto-compaction is paused
+   * (`compaction_circuit_open`, after three consecutive summary-LLM
+   * failures) or resumed (`compaction_circuit_closed`). These are already
+   * in wire-contract shape; the daemon's event dispatcher forwards them to
+   * the client unchanged so the "auto-compaction paused" banner shows and
+   * dismisses.
+   */
+  | CompactionCircuitEvent
   | {
       /**
        * Emitted when an agent turn reaches a terminal state. Checkpoint
@@ -378,6 +436,86 @@ export interface ResolvedSystemPrompt {
   model?: string;
 }
 
+/**
+ * Orchestrator-supplied hooks the loop invokes when the mid-loop budget gate
+ * trips and inline compaction runs. The loop owns the trigger, the
+ * `compaction` pipeline call, the result interpretation (circuit-breaker
+ * bookkeeping + the exhaustion decision), and the inline continue; these hooks
+ * bridge the durable / injection state the loop is intentionally blind to.
+ * Durable persistence ({@link applyResult}) and re-injection
+ * ({@link reinject}) remain orchestrator-supplied for now and are expected to
+ * move into the loop in a future change.
+ */
+export interface MidLoopCompaction {
+  /** Strip runtime injections, commit stripped messages, and resolve pipeline options. */
+  prepare: (history: Message[]) => {
+    rawHistory: Message[];
+    options: CompactionArgs["options"];
+  };
+  /** Commit a successful compaction result to durable state. */
+  applyResult: (
+    result: ContextWindowResult,
+    rawHistory: Message[],
+  ) => Promise<void>;
+  /** Re-apply runtime injections and return the history to continue from. */
+  reinject: () => Promise<Message[]>;
+}
+
+export interface AgentLoopRunOptions {
+  signal?: AbortSignal;
+  requestId?: string;
+  onCheckpoint?: (
+    checkpoint: CheckpointInfo,
+  ) => CheckpointDecision | Promise<CheckpointDecision>;
+  callSite?: LLMCallSite;
+  /**
+   * Per-turn context supplied by the orchestrator. Every pipeline
+   * invocation inside the loop clones from this value (overwriting only
+   * `turnIndex`/`requestId`) so middleware sees the real conversation
+   * identity, trust class, and `contextWindowManager` rather than the
+   * `"agent-loop"` sentinel used when the loop is instantiated standalone
+   * in unit tests.
+   */
+  turnContext?: TurnContext;
+  /**
+   * Ad-hoc inference-profile override applied to every LLM call the loop
+   * issues. When set, each `SendMessageOptions.config` carries
+   * `overrideProfile = <name>` so the provider's resolver layers
+   * `llm.profiles[<name>]` between the workspace `activeProfile` and any
+   * call-site named profile. Missing profile names silently fall through.
+   */
+  overrideProfile?: string;
+  resolveOverrideProfile?: () => string | undefined;
+  /**
+   * Resolves the orchestrator's effective context window for this turn: the
+   * provider max-input-token ceiling (read by tool-result truncation) plus the
+   * `overflowRecovery` config that drives the mid-loop budget gate. Resolved
+   * fresh per checkpoint so a mid-turn profile change is reflected. Absent →
+   * truncation falls back to `this.config.maxInputTokens` and the budget gate
+   * is skipped (agent wakes pass `overflowRecovery.enabled = false`).
+   */
+  resolveContextWindow?: () => {
+    maxInputTokens: number;
+    overflowRecovery: { enabled: boolean; safetyMarginRatio: number };
+  };
+  /**
+   * Hooks for inline mid-loop compaction. When supplied and the budget gate
+   * trips, the loop compacts in place and continues instead of yielding
+   * `exitReason = "budget"`. Callers without a compaction path (agent wakes,
+   * convergence/auto-compress reruns) omit this and keep yielding for budget.
+   */
+  compaction?: MidLoopCompaction;
+  /**
+   * When true, the latest user message carries a volatile per-turn block
+   * (e.g. a memory-v3 `<memory>` injection) that varies across otherwise
+   * identical turns. Forwarded to each `SendMessageOptions.config` so the
+   * provider anchors its long-TTL cache breakpoint on the most recent STABLE
+   * user message instead of the volatile latest one, keeping the cached
+   * prefix reusable. Default unset → existing behavior.
+   */
+  mutableLatestUserMessage?: boolean;
+}
+
 /**
  * Callback shape the loop uses to execute a tool invocation.
  *
@@ -425,6 +563,20 @@ export type LoopToolExecutor = (
   activityMetadata?: ToolActivityMetadata;
 }>;
 
+export interface AgentLoopConstructorOptions {
+  config?: Partial<AgentLoopConfig>;
+  tools?: ToolDefinition[];
+  toolExecutor?: LoopToolExecutor;
+  resolveTools?: (history: Message[]) => ToolDefinition[];
+  resolveSystemPrompt?: (history: Message[]) => ResolvedSystemPrompt;
+  /**
+   * Conversation this loop drives. Used to scope the loop-held compaction
+   * circuit breaker; defaults to an empty key for test loops that never
+   * exercise compaction.
+   */
+  conversationId?: string;
+}
+
 export class AgentLoop {
   private provider: Provider;
   private systemPrompt: string;
@@ -436,15 +588,28 @@ export class AgentLoop {
     | null;
   private toolExecutor: LoopToolExecutor | null;
 
+  /**
+   * Loop-held compaction circuit breaker. The loop has a 1:1 lifetime with its
+   * conversation, so it is the source of truth for the cross-turn failure
+   * counter and cooldown deadline. Non-loop callers (the orchestrator's
+   * compaction paths, `Conversation.forceCompact`, and the dev-only playground
+   * routes) reach it via `agentLoop.compactionCircuit`.
+   */
+  readonly compactionCircuit: CompactionCircuit;
+
   constructor(
     provider: Provider,
     systemPrompt: string,
-    config?: Partial<AgentLoopConfig>,
-    tools?: ToolDefinition[],
-    toolExecutor?: LoopToolExecutor,
-    resolveTools?: (history: Message[]) => ToolDefinition[],
-    resolveSystemPrompt?: (history: Message[]) => ResolvedSystemPrompt,
+    options?: AgentLoopConstructorOptions,
   ) {
+    const {
+      config,
+      tools,
+      toolExecutor,
+      resolveTools,
+      resolveSystemPrompt,
+      conversationId,
+    } = options ?? {};
     this.provider = provider;
     this.systemPrompt = systemPrompt;
     this.config = { ...DEFAULT_CONFIG, ...config };
@@ -452,6 +617,7 @@ export class AgentLoop {
     this.resolveTools = resolveTools ?? null;
     this.resolveSystemPrompt = resolveSystemPrompt ?? null;
     this.toolExecutor = toolExecutor ?? null;
+    this.compactionCircuit = new CompactionCircuit(conversationId ?? "");
   }
 
   /**
@@ -481,44 +647,148 @@ export class AgentLoop {
     return estimateToolsTokens(this.getResolvedTools(history));
   }
 
+  /**
+   * Estimate total prompt tokens for `history` via the `tokenEstimate`
+   * pipeline. Args are shallow-frozen so a mutating middleware cannot strip
+   * context from the loop's live `history`.
+   */
+  private estimateTokens(
+    history: Message[],
+    turnContext: TurnContext,
+  ): Promise<EstimateResult> {
+    return runPipeline<EstimateArgs, EstimateResult>(
+      "tokenEstimate",
+      getMiddlewaresFor("tokenEstimate"),
+      defaultTokenEstimateTerminal,
+      {
+        history: Object.freeze([...history]) as Message[],
+        systemPrompt: this.systemPrompt,
+        tools: Object.freeze([
+          ...this.getResolvedTools(history),
+        ]) as ToolDefinition[],
+        providerName: getCalibrationProviderKey(this.provider),
+      },
+      turnContext,
+      DEFAULT_TIMEOUTS.tokenEstimate,
+    );
+  }
+
+  /**
+   * Record a compaction outcome against the loop's circuit breaker. Three
+   * consecutive failures trip a cooldown that suspends auto-compaction; a
+   * success resets the counter. Any open/closed transition is emitted on the
+   * loop's own event channel via `onEvent`.
+   *
+   * Bookkeeping is best-effort — a failure here must not turn a recoverable
+   * compaction outcome into a user-visible turn failure.
+   */
+  private async recordCompactionOutcome(
+    turnContext: TurnContext,
+    summaryFailed: boolean,
+    onEvent: (event: AgentEvent) => void | Promise<void>,
+  ): Promise<void> {
+    try {
+      await this.compactionCircuit.recordOutcome(
+        {
+          currentRequestId: turnContext.requestId,
+          currentTurnTrustContext: turnContext.trust,
+          turnCount: turnContext.turnIndex,
+        },
+        summaryFailed,
+        onEvent,
+      );
+    } catch (recordError) {
+      log.error(
+        { err: recordError, requestId: turnContext.requestId },
+        "Recording a compaction outcome against the circuit breaker failed; suppressing to keep the agent loop alive",
+      );
+    }
+  }
+
+  /**
+   * Compact the running history in place when the mid-loop budget gate trips.
+   *
+   * Runs the `compaction` pipeline natively (like {@link estimateTokens}) on
+   * the stripped history, then re-applies injections via the supplied hooks.
+   * Returns the history to continue from, or `null` when the compactor timed
+   * out or exhausted its retry budget so the caller yields
+   * `exitReason = "budget"` and the orchestrator escalates.
+   */
+  private async compact(
+    history: Message[],
+    turnContext: TurnContext,
+    compaction: MidLoopCompaction,
+    signal: AbortSignal | undefined,
+    onEvent: (event: AgentEvent) => void | Promise<void>,
+  ): Promise<Message[] | null> {
+    await onEvent({ type: "context_compacting" });
+    const { rawHistory, options } = compaction.prepare(history);
+    let result: CompactionResult;
+    try {
+      result = await runPipeline<CompactionArgs, CompactionResult>(
+        "compaction",
+        getMiddlewaresFor("compaction"),
+        (args) => defaultCompactionTerminal(args, turnContext),
+        { messages: rawHistory, signal, options },
+        turnContext,
+        DEFAULT_TIMEOUTS.compaction,
+      );
+    } catch (error) {
+      if (error instanceof PluginTimeoutError) {
+        // A timeout counts as a compaction failure against the circuit breaker.
+        await this.recordCompactionOutcome(turnContext, true, onEvent);
+        return null;
+      }
+      throw error;
+    }
+    // `CompactionResult` is intentionally `unknown` at the plugin boundary so
+    // plugin consumers don't import the window manager; the loop ran the
+    // pipeline, so it interprets the concrete result here.
+    const compactResult = result as ContextWindowResult;
+    // `force: true` bypasses the cooldown/threshold gates, but early returns
+    // for "no eligible messages" / "insufficient messages" still leave
+    // `summaryFailed` undefined. Only record an outcome when the summary LLM
+    // actually ran.
+    if (compactResult.summaryFailed !== undefined) {
+      await this.recordCompactionOutcome(
+        turnContext,
+        compactResult.summaryFailed,
+        onEvent,
+      );
+    }
+    if (compactResult.compacted) {
+      await compaction.applyResult(compactResult, rawHistory);
+    }
+    if (compactResult.exhausted ?? false) {
+      return null;
+    }
+    return compaction.reinject();
+  }
+
   async run(
     messages: Message[],
     onEvent: (event: AgentEvent) => void | Promise<void>,
-    signal?: AbortSignal,
-    requestId?: string,
-    onCheckpoint?: (
-      checkpoint: CheckpointInfo,
-    ) => CheckpointDecision | Promise<CheckpointDecision>,
-    callSite?: LLMCallSite,
-    /**
-     * Optional per-turn context supplied by the orchestrator. Every pipeline
-     * invocation inside the loop clones from this value (overwriting only
-     * `turnIndex`/`requestId`) so middleware sees the real conversation
-     * identity, trust class, and `contextWindowManager` rather than the
-     * `"agent-loop"` sentinel used when the loop is instantiated standalone
-     * in unit tests.
-     */
-    turnContext?: TurnContext,
-    /**
-     * Optional ad-hoc inference-profile override applied to every LLM call
-     * the loop issues. When set, each `SendMessageOptions.config` carries
-     * `overrideProfile = <name>` so the provider's resolver layers
-     * `llm.profiles[<name>]` between the workspace `activeProfile` and any
-     * call-site named profile. Missing profile names silently fall through.
-     * Used by per-conversation pinned profiles to override the workspace
-     * default for the lifetime of an agent loop run.
-     */
-    overrideProfile?: string,
-    effectiveMaxInputTokens?: number,
-    resolveOverrideProfile?: () => string | undefined,
-    resolveEffectiveMaxInputTokens?: () => number | undefined,
-  ): Promise<Message[]> {
-    const history = [...messages];
-    const initialHistoryLength = messages.length;
+    options?: AgentLoopRunOptions,
+  ): Promise<AgentLoopRunResult> {
+    const {
+      signal,
+      requestId,
+      onCheckpoint,
+      callSite,
+      turnContext,
+      overrideProfile,
+      resolveOverrideProfile,
+      resolveContextWindow,
+      compaction,
+      mutableLatestUserMessage,
+    } = options ?? {};
+    let history = [...messages];
+    let producedVisibleTextThisRun = false;
     let toolUseTurns = 0;
     let consecutiveErrorTurns = 0;
     let emptyResponseRetries = 0;
     let lastLlmCallTime = 0;
+    let exitReason: ExitReason | null = null;
     const rlog = requestId ? log.child({ requestId }) : log;
 
     // Per-run substitution map for sensitive output placeholders.
@@ -617,6 +887,15 @@ export class AgentLoop {
           providerConfig.cacheTtl = this.config.cacheTtl;
         }
 
+        // Cache-anchor signal for volatile latest-user-message turns (e.g.
+        // memory-v3 injects its `<memory>` block into the latest user
+        // message). Not part of the call-site schema, so it is always sourced
+        // from the per-run option regardless of `callSite`. Only set when true
+        // so the wire/config stays byte-identical when off.
+        if (mutableLatestUserMessage) {
+          providerConfig.mutableLatestUserMessage = true;
+        }
+
         // Per-call LLM call-site identifier. Surfaces on the per-call
         // `config.callSite` so `RetryProvider.normalizeSendMessageOptions`
         // can route through `resolveCallSiteConfig` against
@@ -692,7 +971,7 @@ export class AgentLoop {
         // Also strip old AX tree snapshots to keep TTFT from growing
         // linearly with step count in computer-use sessions.
         const providerHistory = compactAxTreeHistory(
-          stripOldImageBlocks(history),
+          stripOldMediaBlocks(history),
         );
 
         // Wrap the provider call in the `llmCall` pipeline so middleware
@@ -711,9 +990,9 @@ export class AgentLoop {
         const llmCallArgs: LLMCallArgs = {
           provider: this.provider,
           messages: providerHistory,
-          tools: currentTools.length > 0 ? currentTools : undefined,
-          systemPrompt: turnSystemPrompt,
           options: {
+            tools: currentTools.length > 0 ? currentTools : undefined,
+            systemPrompt: turnSystemPrompt,
             config: providerConfig,
             onEvent: (event) => {
               if (event.type === "text_delta") {
@@ -811,13 +1090,7 @@ export class AgentLoop {
           response = await runPipeline<LLMCallArgs, LLMCallResult>(
             "llmCall",
             getMiddlewaresFor("llmCall"),
-            (args) =>
-              args.provider.sendMessage(
-                args.messages,
-                args.tools,
-                args.systemPrompt,
-                args.options,
-              ),
+            (args) => args.provider.sendMessage(args.messages, args.options),
             llmCallArgs,
             turnCtx,
             DEFAULT_TIMEOUTS.llmCall,
@@ -839,8 +1112,8 @@ export class AgentLoop {
             const rawRequest = {
               provider: this.provider.name,
               messages: llmCallArgs.messages,
-              tools: llmCallArgs.tools,
-              systemPrompt: llmCallArgs.systemPrompt,
+              tools: llmCallArgs.options?.tools,
+              systemPrompt: llmCallArgs.options?.systemPrompt,
               config: llmCallArgs.options?.config,
             };
             onEvent({
@@ -972,25 +1245,20 @@ export class AgentLoop {
         // The actual decision (nudge vs. accept vs. error) is delegated to
         // the `emptyResponse` plugin pipeline. The pipeline returns a
         // decision; the loop carries out the side-effect (pushing the nudge
-        // or surfacing the error). See `plugins/defaults/empty-response.ts`
+        // or surfacing the error). See `plugins/defaults/empty-response/register.ts`
         // for the default decision logic.
         const hasVisibleText = response.content.some(
           (block) => block.type === "text" && block.text.trim().length > 0,
         );
-        const priorAssistantHadVisibleText = (() => {
-          for (let i = history.length - 1; i >= initialHistoryLength; i--) {
-            const msg = history[i];
-            if (msg.role !== "assistant") continue;
-            const hasText = msg.content.some(
-              (block) =>
-                block.type === "text" &&
-                typeof (block as { text?: unknown }).text === "string" &&
-                (block as { text: string }).text.trim().length > 0,
-            );
-            if (hasText) return true;
-          }
-          return false;
-        })();
+        // Track whether the model produced visible text earlier in this
+        // run() invocation. Run-scoped rather than derived from `history` so
+        // it survives inline compaction rewriting the message array: an empty
+        // completion after a compaction must not be nudged into re-sending
+        // text the user already saw.
+        const priorAssistantHadVisibleText = producedVisibleTextThisRun;
+        if (hasVisibleText) {
+          producedVisibleTextThisRun = true;
+        }
 
         const emptyResponseArgs: EmptyResponseArgs = {
           responseContent: response.content,
@@ -999,6 +1267,7 @@ export class AgentLoop {
           emptyResponseRetries,
           maxEmptyResponseRetries: MAX_EMPTY_RESPONSE_RETRIES,
           priorAssistantHadVisibleText,
+          stopReason: response.stopReason,
         };
         const emptyResponseCtx = resolveLoopTurnContext(
           turnContext,
@@ -1201,8 +1470,7 @@ export class AgentLoop {
         // truncation strategy (e.g. a summariser) while the default
         // middleware preserves the historical tail-drop behaviour.
         const contextWindowTokens =
-          resolveEffectiveMaxInputTokens?.() ??
-          effectiveMaxInputTokens ??
+          resolveContextWindow?.().maxInputTokens ??
           this.config.maxInputTokens ??
           180_000;
         const maxChars = calculateMaxToolResultChars(contextWindowTokens);
@@ -1352,8 +1620,8 @@ export class AgentLoop {
         history.push({ role: "user", content: resultBlocks });
 
         // Invoke checkpoint callback after tool results are in history.
-        // The callback may be async — the mid-loop budget check delegates
-        // to the `tokenEstimate` plugin pipeline, which is asynchronous.
+        // Handoff is offered first so a queued message takes precedence over
+        // the mid-loop budget yield below.
         if (onCheckpoint) {
           const decision = await onCheckpoint({
             turnIndex: toolUseTurns - 1, // 0-based (toolUseTurns was already incremented)
@@ -1361,7 +1629,60 @@ export class AgentLoop {
             hasToolUse: true,
             history,
           });
-          if (decision === "yield") {
+          if (decision !== "continue") {
+            exitReason = decision;
+            break;
+          }
+        }
+
+        // Mid-loop budget gate: when overflow recovery is enabled, estimate
+        // the running context size as it approaches the preflight budget.
+        // With a `compaction` hook the loop compacts in place and continues;
+        // without one it yields (`exitReason = "budget"`) so the orchestrator
+        // can recover before the next provider call risks a hard
+        // context-too-large rejection. Keyed off the loop's own
+        // `history.length` (the messages actually in context this turn,
+        // including tool iterations) rather than the durable conversation
+        // count.
+        const contextWindow = resolveContextWindow?.();
+        if (contextWindow?.overflowRecovery.enabled) {
+          const { maxInputTokens, overflowRecovery } = contextWindow;
+          const safetyMargin =
+            history.length > LONG_HISTORY_MESSAGE_THRESHOLD
+              ? Math.max(
+                  overflowRecovery.safetyMarginRatio,
+                  LONG_HISTORY_SAFETY_MARGIN_FLOOR,
+                )
+              : overflowRecovery.safetyMarginRatio;
+          const preflightBudget = Math.floor(
+            maxInputTokens * (1 - safetyMargin),
+          );
+          const midLoopThreshold =
+            preflightBudget * MID_LOOP_YIELD_THRESHOLD_RATIO;
+          const estimated = await this.estimateTokens(history, turnCtx);
+          if (estimated > midLoopThreshold) {
+            if (compaction) {
+              rlog.info(
+                { phase: "mid-loop", estimated, threshold: midLoopThreshold },
+                "Token estimate approaching budget — compacting in place",
+              );
+              const compacted = await this.compact(
+                history,
+                turnCtx,
+                compaction,
+                signal,
+                onEvent,
+              );
+              if (compacted) {
+                history = compacted;
+                continue;
+              }
+            }
+            rlog.warn(
+              { phase: "mid-loop", estimated, threshold: midLoopThreshold },
+              "Token estimate approaching budget — yielding for compaction",
+            );
+            exitReason = "budget";
             break;
           }
         }
@@ -1412,7 +1733,7 @@ export class AgentLoop {
       "Agent loop exited",
     );
 
-    return history;
+    return { history, exitReason };
   }
 }
 
@@ -1513,7 +1834,7 @@ export function compactAxTreeHistory(messages: Message[]): Message[] {
  * turn. Using the last user message unconditionally would leave the most
  * recent tool screenshots unprotected from stripping.
  */
-function stripOldImageBlocks(history: Message[]): Message[] {
+function stripOldMediaBlocks(history: Message[]): Message[] {
   // Find the last user message that contains tool_result blocks.
   let lastToolResultUserIdx = -1;
   for (let i = history.length - 1; i >= 0; i--) {
@@ -1530,29 +1851,32 @@ function stripOldImageBlocks(history: Message[]): Message[] {
     // Keep the most recent tool-result user message intact (current turn)
     if (idx === lastToolResultUserIdx || msg.role !== "user") return msg;
 
-    // Check if any tool_result blocks have image contentBlocks
-    const hasImages = msg.content.some(
+    // Check if any tool_result blocks carry embedded media (image or audio).
+    const isMedia = (cb: ContentBlock) =>
+      cb.type === "image" || cb.type === "file";
+    const hasMedia = msg.content.some(
       (b) =>
         b.type === "tool_result" &&
-        (b as ToolResultContent).contentBlocks?.some(
-          (cb) => cb.type === "image",
-        ),
+        (b as ToolResultContent).contentBlocks?.some(isMedia),
     );
-    if (!hasImages) return msg;
+    if (!hasMedia) return msg;
 
-    // Strip images from tool_result blocks, replacing with text marker
+    // Strip media from tool_result blocks, replacing with a text marker. The
+    // model already saw/heard the media in the turn it was captured; resending
+    // the bytes every turn (a 12 MB audio clip isn't optimized like images)
+    // bloats the request until compaction.
     return {
       ...msg,
       content: msg.content.map((b) => {
         if (b.type !== "tool_result") return b;
         const tr = b as ToolResultContent;
-        if (!tr.contentBlocks?.some((cb) => cb.type === "image")) return b;
+        if (!tr.contentBlocks?.some(isMedia)) return b;
         return {
           ...tr,
           contentBlocks: undefined,
           content:
             (tr.content || "") +
-            "\n[Screenshot was captured and shown previously — image data removed to save context.]",
+            "\n[Media (image/audio) was captured and shown previously — binary data removed to save context.]",
         };
       }),
     };