@vellumai/assistant 0.8.7 → 0.8.8-dev.202606052332.17fc8ea

package/src/daemon/conversation-agent-loop.ts

@@ -7,8 +7,6 @@
  * runAgentLoop method here via the AgentLoopConversationContext interface.
  */
 
-import { join } from "node:path";
-
 import { v4 as uuid } from "uuid";
 
 import { optimizeImageForTransport } from "../agent/image-optimize.js";
@@ -46,10 +44,13 @@ import {
 } from "../context/post-turn-tool-result-truncation.js";
 import {
   estimatePromptTokens,
+  estimatePromptTokensWithTools,
   getCalibrationProviderKey,
 } from "../context/token-estimator.js";
-import type { ContextWindowManager } from "../context/window-manager.js";
-import { getDocumentsForConversation } from "../documents/document-store.js";
+import type {
+  ContextWindowCompactOptions,
+  ContextWindowManager,
+} from "../context/window-manager.js";
 import type { ToolProfiler } from "../events/tool-profiling-listener.js";
 import { writeRelationshipState } from "../home/relationship-state-writer.js";
 import {
@@ -57,9 +58,9 @@ import {
   setSentryConversationContext,
 } from "../instrument.js";
 import { commitAppTurnChanges } from "../memory/app-git-service.js";
-import { getApp, listAppFiles, resolveAppDir } from "../memory/app-store.js";
 import { enqueueAutoAnalysisOnCompaction } from "../memory/auto-analysis-enqueue.js";
 import {
+  addMessage,
   deleteMessageById,
   getConversation,
   getConversationOriginChannel,
@@ -68,77 +69,40 @@ import {
   getLastUserTimestampBefore,
   getMessageById,
   provenanceFromTrustContext,
-  setConversationHistoryStrippedAt,
-  setLastNotifiedInferenceProfile,
   updateConversationContextWindow,
   updateConversationSlackContextWatermark,
+  updateMessageMetadata,
 } from "../memory/conversation-crud.js";
 import { getResolvedConversationDirPath } from "../memory/conversation-directories.js";
 import { syncMessageToDisk } from "../memory/conversation-disk-view.js";
-import {
-  isReplaceableTitle,
-  queueRegenerateConversationTitle,
-} from "../memory/conversation-title-service.js";
+import { isReplaceableTitle } from "../memory/conversation-title-service.js";
 import { isBackgroundConversationType } from "../memory/conversation-types.js";
 import type { ConversationGraphMemory } from "../memory/graph/conversation-graph-memory.js";
 import {
   backfillMessageIdOnLogs,
   recordSyntheticAgentErrorMessageLog,
 } from "../memory/llm-request-log-store.js";
-import { recordMemoryRecallLog } from "../memory/memory-recall-log-store.js";
 import { enqueueMemoryRetrospectiveOnCompaction } from "../memory/memory-retrospective-enqueue.js";
-import { PKB_WORKSPACE_SCOPE } from "../memory/pkb/types.js";
-import type { QdrantSparseVector } from "../memory/qdrant-client.js";
-import {
-  readMemoryV2StaticContent,
-  shouldExposePersonalMemory,
-} from "../memory/v2/static-context.js";
 import type { PermissionPrompter } from "../permissions/prompter.js";
 import { HOOKS } from "../plugin-api/constants.js";
 import type { UserPromptSubmitContext } from "../plugin-api/types.js";
-import { defaultCompactionTerminal } from "../plugins/defaults/compaction/terminal.js";
+import { defaultCompact } from "../plugins/defaults/compaction/compact.js";
 import { deepRepairHistory } from "../plugins/defaults/history-repair/terminal.js";
-import {
-  asDefaultGraphPayload,
-  type DefaultMemoryRetrievalDeps,
-  type GraphMemoryPayload,
-  runDefaultMemoryRetrieval,
-} from "../plugins/defaults/memory-retrieval/register.js";
-import { defaultPersistenceTerminal } from "../plugins/defaults/persistence/terminal.js";
-import { defaultTitleGenerateTerminal } from "../plugins/defaults/title-generate/terminal.js";
-import { defaultTokenEstimateTerminal } from "../plugins/defaults/token-estimate/terminal.js";
-import { DEFAULT_TIMEOUTS, runHook, runPipeline } from "../plugins/pipeline.js";
-import { getMiddlewaresFor } from "../plugins/registry.js";
-import type {
-  CompactionArgs,
-  CompactionResult,
-  EstimateArgs,
-  EstimateResult,
-  MemoryArgs,
-  MemoryResult,
-  OverflowReduceArgs,
-  OverflowReduceResult,
-  PersistAddResult,
-  PersistArgs,
-  PersistResult,
-  TurnContext as PluginTurnContext,
-} from "../plugins/types.js";
-import { PluginExecutionError, PluginTimeoutError } from "../plugins/types.js";
-import type {
-  ContentBlock,
-  Message,
-  ToolDefinition,
-} from "../providers/types.js";
+import postCompactReinject from "../plugins/defaults/memory-retrieval/hooks/post-compact.js";
+import userPromptSubmitMemoryRetrieval, {
+  type MemoryRetrievalHookContext,
+} from "../plugins/defaults/memory-retrieval/hooks/user-prompt-submit-temp.js";
+import { runHook } from "../plugins/pipeline.js";
+import type { TurnContext as PluginTurnContext } from "../plugins/types.js";
+import type { ContentBlock, Message } from "../providers/types.js";
 import type { Provider } from "../providers/types.js";
 import { resolveActorTrust } from "../runtime/actor-trust-resolver.js";
 import { broadcastMessage } from "../runtime/assistant-event-hub.js";
 import { DAEMON_INTERNAL_ASSISTANT_ID } from "../runtime/assistant-scope.js";
 import { publishConversationMessagesChanged } from "../runtime/sync/resource-sync-events.js";
-import { redactSecrets } from "../security/secret-scanner.js";
 import { getSubagentManager } from "../subagent/index.js";
 import type { UsageActor } from "../usage/actors.js";
 import { getLogger } from "../util/logger.js";
-import { getWorkspaceDir } from "../util/platform.js";
 import { timeAgo } from "../util/time.js";
 import { truncate } from "../util/truncate.js";
 import { getWorkspaceGitService } from "../workspace/git-service.js";
@@ -147,7 +111,6 @@ import {
   type AssistantAttachmentDraft,
   cleanAssistantContent,
 } from "./assistant-attachments.js";
-import { cleanupBootstrapAfterTurnThreshold } from "./bootstrap-turn-cleanup.js";
 import { resolveOverflowAction } from "./context-overflow-policy.js";
 import {
   createInitialReducerState,
@@ -158,6 +121,8 @@ import {
   createEventHandlerState,
   dispatchAgentEvent,
   type EventHandlerDeps,
+  finalizePendingToolResultRow,
+  markHistoryStrippedBestEffort,
 } from "./conversation-agent-loop-handlers.js";
 import {
   approveHostAttachmentRead,
@@ -173,7 +138,6 @@ import { raceWithTimeout } from "./conversation-media-retry.js";
 import type { MessageQueue } from "./conversation-queue-manager.js";
 import type { QueueDrainReason } from "./conversation-queue-manager.js";
 import type {
-  ActiveSurfaceContext,
   ChannelCapabilities,
   InboundActorContext,
   InjectionMode,
@@ -182,8 +146,6 @@ import {
   applyRuntimeInjections,
   buildSubagentStatusBlock,
   buildUnifiedTurnContextBlock,
-  findLastInjectedNowContent,
-  getPkbAutoInjectList,
   getSlackCompactionWatermarkForPrefix,
   inboundActorContextFromTrust,
   inboundActorContextFromTrustContext,
@@ -194,7 +156,6 @@ import {
 } from "./conversation-runtime-assembly.js";
 import type { SkillProjectionCache } from "./conversation-skill-tools.js";
 import { markSurfaceCompleted } from "./conversation-surfaces.js";
-import { resolveTrustClass } from "./conversation-tool-setup.js";
 import { recordUsage } from "./conversation-usage.js";
 import {
   formatTurnTimestamp,
@@ -203,45 +164,27 @@ import {
 import { getDiskPressureStatus } from "./disk-pressure-guard.js";
 import { classifyDiskPressureTurnPolicy } from "./disk-pressure-policy.js";
 import type {
-  DynamicPageSurfaceData,
   ServerMessage,
   SurfaceData,
   SurfaceType,
   UsageStats,
 } from "./message-protocol.js";
-import type { MemoryRecalled } from "./message-types/memory.js";
 import type { ConfirmationStateChanged } from "./message-types/messages.js";
-import { conversationMetadataSyncTag } from "./message-types/sync.js";
+import {
+  type OverflowReduceArgs,
+  runOverflowReductionLoop,
+} from "./overflow-reduction-loop.js";
 import { parseActualTokensFromError } from "./parse-actual-tokens-from-error.js";
+import {
+  persistUnsendableImageDowngrades,
+  UNSENDABLE_IMAGE_NOTE,
+} from "./persist-unsendable-image.js";
 import type { TraceEmitter } from "./trace-emitter.js";
-import type { TrustContext } from "./trust-context.js";
+import { resolveTrustClass, type TrustContext } from "./trust-context.js";
 import { stripHistoricalWebSearchResults } from "./web-search-history.js";
 
 const log = getLogger("conversation-agent-loop");
 
-/**
- * Best-effort persistence of the history-stripped marker after an
- * injection-strip event (compaction / overflow recovery). The marker is a
- * durability hint, not turn-critical state — a transient SQLite write failure
- * (SQLITE_BUSY, disk-full, read-only FS) must not abort the turn. Logs a
- * warning and continues on failure, preserving the long-standing non-fatal
- * contract for this metadata write.
- */
-function markHistoryStrippedBestEffort(
-  conversationId: string,
-  strippedAt: number,
-  logger: ReturnType<typeof getLogger>,
-): void {
-  try {
-    setConversationHistoryStrippedAt(conversationId, strippedAt);
-  } catch (err) {
-    logger.warn(
-      { err },
-      "Failed to persist history-stripped marker after compaction strip (non-fatal)",
-    );
-  }
-}
-
 const DISK_PRESSURE_ERROR_CODE = "DISK_SPACE_CRITICAL" as const;
 const DISK_PRESSURE_ERROR_CATEGORY = "disk_pressure";
 
@@ -270,12 +213,12 @@ function formatDiskPressureBlockedMessage(): string {
 // ── Plugin pipeline helpers ──────────────────────────────────────────
 //
 // Canonical {@link PluginTurnContext} builder threaded into every
-// `runPipeline` call inside `runAgentLoopImpl`. The orchestrator composes
+// `runHook` call inside `runAgentLoopImpl`. The orchestrator composes
 // the context on demand at each call site from ambient state rather than
 // carrying a persistent `TurnContext` instance across the turn.
 
 /**
- * Synthetic fallback trust context used when the orchestrator fires a pipeline
+ * Synthetic fallback trust context used when the orchestrator fires a hook
  * before the per-turn trust snapshot has been captured (e.g. invocations that
  * bypass `processMessage` / `drainQueue`). We bias to `unknown` rather than
  * `guardian` so a missing snapshot cannot accidentally grant elevated trust
@@ -287,14 +230,14 @@ const FALLBACK_TURN_TRUST: TrustContext = {
 };
 
 /**
- * Build the {@link TurnContext} passed to {@link runPipeline}.
+ * Build the {@link TurnContext} passed to {@link runHook}.
  *
- * Canonical source of truth for every pipeline call site inside the agent
- * loop. Every `runPipeline` invocation in `runAgentLoopImpl` (and in the
+ * Canonical source of truth for every hook call site inside the agent
+ * loop. Every `runHook` invocation in `runAgentLoopImpl` (and in the
  * handlers that share its ambient state) must route through this helper
  * rather than constructing a `TurnContext` literal inline — this keeps
  * `turnIndex`, trust resolution, and the `contextWindowManager` attachment
- * consistent across pipeline slots, which in turn keeps structured logs
+ * consistent across hooks, which in turn keeps structured logs
  * filtered by `conversationId`/`turnIndex` coherent across slots.
  *
  * Behavior:
@@ -306,9 +249,9 @@ const FALLBACK_TURN_TRUST: TrustContext = {
  *   level context, then {@link FALLBACK_TURN_TRUST}. The cascade matches
  *   the one inside the orchestrator's inline injection assembly so
  *   middleware reads the same trust class the runtime sees.
- * - `contextWindowManager` is attached unconditionally. Pipelines that
- *   don't need it can ignore it; the default compaction plugin reads it
- *   via the typed optional field on `TurnContext`.
+ * - `contextWindowManager` is attached unconditionally. Hooks that
+ *   don't need it can ignore it; it remains available via the typed
+ *   optional field on `TurnContext`.
  */
 function buildPluginTurnContext(
   ctx: AgentLoopConversationContext,
@@ -322,9 +265,23 @@ function buildPluginTurnContext(
     turnIndex: ctx.turnCount,
     trust,
     contextWindowManager: ctx.contextWindowManager,
+    callSite: ctx.currentCallSite,
   };
 }
 
+/**
+ * Trust class of the actor whose turn is in progress, for the compactor's
+ * image manifest filter. Prefers the turn-start snapshot
+ * ({@link AgentLoopConversationContext.currentTurnTrustContext}) over the live
+ * trust context so compaction running in a later tool iteration can't pick up
+ * a concurrent request's actor.
+ */
+function resolveTurnActorTrustClass(
+  ctx: AgentLoopConversationContext,
+): TrustContext["trustClass"] | undefined {
+  return (ctx.currentTurnTrustContext ?? ctx.trustContext)?.trustClass;
+}
+
 // ── Context Interface ────────────────────────────────────────────────
 
 /**
@@ -352,9 +309,18 @@ export interface AssistantSurface {
 export interface AgentLoopConversationContext {
   readonly conversationId: string;
   messages: Message[];
-  processing: boolean;
+  isProcessing(): boolean;
+  setProcessing(value: boolean): void;
   abortController: AbortController | null;
   currentRequestId?: string;
+  /**
+   * The {@link LLMCallSite} of the in-flight turn, set at turn start from
+   * `options?.callSite ?? "mainAgent"`. Read by {@link buildPluginTurnContext}
+   * so pipeline/injector plugins can tell the main reply apart from
+   * background agent-loop work (compaction, subagents, …) on this same
+   * conversation. Per-turn mutable, mirroring {@link currentRequestId}.
+   */
+  currentCallSite?: LLMCallSite;
 
   readonly agentLoop: AgentLoop;
   readonly provider: Provider;
@@ -397,8 +363,6 @@ export interface AgentLoopConversationContext {
   currentTurnSurfaces: AssistantSurface[];
 
   workingDir: string;
-  workspaceTopLevelContext: string | null;
-  workspaceTopLevelDirty: boolean;
   channelCapabilities?: ChannelCapabilities;
   /** Per-turn snapshot of trustContext, frozen at message-processing start. */
   currentTurnTrustContext?: TrustContext;
@@ -424,8 +388,6 @@ export interface AgentLoopConversationContext {
   /** Task-run scope for the current turn. Cleared at turn end so queued/drained turns don't inherit it. */
   taskRunId?: string;
   assistantId?: string;
-  voiceCallControlPrompt?: string;
-  transportHints?: string[];
   clientTimezone?: string;
 
   readonly coreToolNames: Set<string>;
@@ -500,7 +462,6 @@ export interface AgentLoopConversationContext {
   getWorkspaceGitService?: (workspaceDir: string) => GitServiceInitializer;
   commitTurnChanges?: typeof commitTurnChanges;
 
-  refreshWorkspaceTopLevelContextIfNeeded(): void;
   markWorkspaceTopLevelDirty(): void;
   getQueueDepth(): number;
   hasQueuedMessages(): boolean;
@@ -561,6 +522,13 @@ export async function runAgentLoopImpl(
   });
   let yieldedForHandoff = false;
   let yieldedForBudget = false;
+  // Whether the most recent agent-loop run produced at least one new assistant
+  // message — the loop's own forward-progress signal, used by the ordering
+  // retry gate and the overflow convergence fold.
+  let lastRunAppendedNewMessages = false;
+  // The messages the most recent agent-loop run appended on top of its base —
+  // the loop's own new-output boundary, persisted as this turn's new messages.
+  let lastRunNewMessages: Message[] = [];
   let pendingCheckpointYield: "budget" | "handoff" | null = null;
   // Captured when the auto_compress_latest_turn rerun yields at the mid-loop
   // budget checkpoint. SSE emission happens immediately at the detection site;
@@ -579,6 +547,9 @@ export async function runAgentLoopImpl(
   // `resolveCallSiteConfig`, picking up any user overrides under
   // `llm.callSites.mainAgent` (falling back to `llm.default` when absent).
   const turnCallSite: LLMCallSite = options?.callSite ?? "mainAgent";
+  // Expose the turn's call site to plugin pipeline/injector contexts (read by
+  // buildPluginTurnContext) so plugins can scope behaviour to the main reply.
+  ctx.currentCallSite = turnCallSite;
 
   // Read the conversation row once for both the override-profile derivation
   // below and the title-replaceability check at turn start. Later reads in
@@ -792,10 +763,6 @@ export async function runAgentLoopImpl(
         : null,
     },
   );
-  const diskPressureContext =
-    diskPressureDecision.action === "allow-cleanup-mode"
-      ? { cleanupModeActive: true }
-      : null;
   ctx.diskPressureCleanupModeActive =
     diskPressureDecision.action === "allow-cleanup-mode";
 
@@ -898,55 +865,6 @@ export async function runAgentLoopImpl(
       }
     }
 
-    // Generate title early — the user message alone is sufficient context.
-    // Firing before the main LLM call removes the delay of waiting for the
-    // full assistant response. The second-pass regeneration at turn 3 will
-    // refine the title with more context.
-    // No abort signal — title generation should complete even if the user
-    // cancels the response, since the user message is already persisted.
-    // Deferred via setTimeout so the main agent loop LLM call enqueues
-    // first, avoiding rate-limit slot contention on strict configs.
-    if (isReplaceableTitle(turnStartConversation?.title ?? null)) {
-      // TurnContext routed through the canonical builder so the pipeline's
-      // log record reports the same `conversationId`/`turnIndex` shape as
-      // every other slot in this turn. Title generation does not depend on
-      // the context-window manager attached by the builder, but sharing the
-      // builder keeps the invariant enforced in one place.
-      const titlePipelineCtx = buildPluginTurnContext(ctx, reqId);
-      const titleArgs = {
-        conversationId: ctx.conversationId,
-        provider: ctx.provider,
-        userMessage: options?.titleText ?? content,
-        onTitleUpdated: (title: string) => {
-          onEvent({
-            type: "conversation_title_updated",
-            conversationId: ctx.conversationId,
-            title,
-          });
-          onEvent({
-            type: "sync_changed",
-            tags: [conversationMetadataSyncTag(ctx.conversationId)],
-          });
-        },
-      };
-      setTimeout(() => {
-        runPipeline(
-          "titleGenerate",
-          getMiddlewaresFor("titleGenerate"),
-          defaultTitleGenerateTerminal,
-          titleArgs,
-          titlePipelineCtx,
-          DEFAULT_TIMEOUTS.titleGenerate,
-        ).catch((err) => {
-          // Fire-and-forget — keep previous non-propagating semantics.
-          // queueGenerateConversationTitle already swallows internal
-          // errors; this catch covers pipeline-layer errors (timeouts,
-          // middleware throws) without surfacing them to the agent loop.
-          rlog.warn({ err }, "titleGenerate pipeline failed (non-fatal)");
-        });
-      }, 0);
-    }
-
     const isFirstMessage = ctx.messages.length === 1;
     // Promote a pending post-compaction re-inject signal (e.g. from `/compact`)
     // into `compactedThisTurn` so NOW.md / PKB / v2 static blocks land on this
@@ -954,7 +872,6 @@ export async function runAgentLoopImpl(
     // so this fires exactly once per `/compact` event.
     const consumedPostCompactReinject = ctx.pendingPostCompactReinject;
     ctx.pendingPostCompactReinject = false;
-    let shouldInjectWorkspace = isFirstMessage || consumedPostCompactReinject;
     let compactedThisTurn = consumedPostCompactReinject;
     let slackCompactedThisTurn = false;
     const isSlackConversation = ctx.channelCapabilities?.channel === "slack";
@@ -1092,70 +1009,32 @@ export async function runAgentLoopImpl(
     // Skip auto-compaction while the circuit breaker is open. Force paths
     // and user-initiated /compact bypass this check.
     const autoCompactAllowed =
-      !(await ctx.agentLoop.compactionCircuit.isOpen(ctx));
+      !(await ctx.agentLoop.compactionCircuit.isOpen());
     if (compactCheck.needed && autoCompactAllowed) {
       ctx.emitActivityState("thinking", "context_compacting", {
         requestId: reqId,
       });
     }
-    const compactionOptions = {
-      lastCompactedAt: ctx.contextCompactedAt ?? undefined,
-      precomputedEstimate: compactCheck.estimatedTokens,
-      conversationOriginChannel:
-        getConversationOriginChannel(ctx.conversationId) ?? undefined,
-      overrideProfile: resolveCurrentOverrideProfile() ?? null,
-      actorTrustClass: ctx.trustContext?.trustClass,
-    };
     let compacted: Awaited<
       ReturnType<typeof ctx.contextWindowManager.maybeCompact>
     > | null = null;
     if (autoCompactAllowed) {
-      try {
-        compacted = (await runPipeline<CompactionArgs, CompactionResult>(
-          "compaction",
-          getMiddlewaresFor("compaction"),
-          (args) =>
-            defaultCompactionTerminal(args, buildPluginTurnContext(ctx, reqId)),
-          {
-            messages: messagesForStartOfTurnCompaction,
-            signal: abortController.signal,
-            options: compactionOptions,
-          },
-          buildPluginTurnContext(ctx, reqId),
-          DEFAULT_TIMEOUTS.compaction,
-        )) as Awaited<ReturnType<typeof ctx.contextWindowManager.maybeCompact>>;
-      } catch (err) {
-        if (err instanceof PluginTimeoutError) {
-          // Pipeline exceeded its budget. Record the failure so the circuit
-          // breaker tracks consecutive timeouts (it trips after three),
-          // then degrade gracefully by skipping compaction this turn —
-          // the turn proceeds with the un-compacted history rather than
-          // hard-failing. The inner summary call has been aborted by the
-          // runner's signal-linking, so updateSummary's local fallback
-          // also ran before this catch block is reached.
-          rlog.warn(
-            { err, phase: "start-of-turn-compaction" },
-            "Compaction pipeline timed out — skipping compaction this turn",
-          );
-          await ctx.agentLoop.compactionCircuit.recordOutcome(
-            ctx,
-            true,
-            onEvent,
-          );
-          compacted = null;
-        } else {
-          throw err;
-        }
-      }
+      compacted = await defaultCompact({
+        manager: ctx.contextWindowManager,
+        messages: messagesForStartOfTurnCompaction,
+        signal: abortController.signal,
+        precomputedEstimate: compactCheck.estimatedTokens,
+        overrideProfile: resolveCurrentOverrideProfile() ?? null,
+        actorTrustClass: resolveTurnActorTrustClass(ctx),
+      });
     }
     // Only track circuit-breaker state when a summary LLM call actually ran.
     // `summaryFailed` is `undefined` on early returns (compaction disabled,
-    // below threshold, cooldown active, no eligible messages, truncation-only
+    // below threshold, no eligible messages, truncation-only
     // path) — treating those as "successful" compactions would silently reset
     // the 3-strike counter and break the invariant.
     if (compacted && compacted.summaryFailed !== undefined) {
       await ctx.agentLoop.compactionCircuit.recordOutcome(
-        ctx,
         compacted.summaryFailed,
         onEvent,
       );
@@ -1165,7 +1044,6 @@ export async function runAgentLoopImpl(
         compacted,
         messagesForStartOfTurnCompaction,
       );
-      shouldInjectWorkspace = true;
       if (compacted.compactedPersistedMessages > 0) {
         compactedThisTurn = true;
       }
@@ -1203,213 +1081,10 @@ export async function runAgentLoopImpl(
       }
     };
 
-    let runMessages = ctx.messages;
-
-    // Memory retrieval pipeline — fetches PKB, NOW.md, and memory-graph
-    // outputs through a single `memoryRetrieval` pipeline. Plugins may
-    // replace the terminal behavior by registering a middleware that
-    // short-circuits with its own `MemoryResult`; the default terminal
-    // below runs `runDefaultMemoryRetrieval` which reproduces the prior
-    // in-lined behavior (PKB/NOW reads + gated graph call).
-    const isTrustedActor = resolveTrustClass(ctx.trustContext) === "guardian";
-    // Canonical builder — pulls trust from per-turn snapshot, then
-    // conversation-level, then the synthetic fallback. Memory retrieval
-    // does not need the context-window handle the builder attaches, but
-    // keeping every call site on one helper is load-bearing for log
-    // coherence across pipeline slots.
-    const memoryPluginTurnCtx = buildPluginTurnContext(ctx, reqId);
-    const memoryArgs: MemoryArgs = {
-      conversationId: ctx.conversationId,
-      trustContext: ctx.trustContext,
-      turnIndex: ctx.turnCount,
-      // Pass the abort signal via `args` (not `deps`) so the pipeline
-      // runner's `linkAbortSignal` can swap it for a signal linked to the
-      // pipeline's internal controller — on a plugin-set timeout or
-      // external cancel, the linked signal aborts and `prepareMemory`
-      // stops mutating graph state / emitting events after the pipeline
-      // has already errored.
-      signal: abortController.signal,
-    };
-    const memoryDeps: DefaultMemoryRetrievalDeps = {
-      messages: ctx.messages,
-      graphMemory: ctx.graphMemory,
-      config: getConfig(),
-      onEvent,
-      isTrustedActor,
-    };
-    const memoryResult: MemoryResult = await runPipeline(
-      "memoryRetrieval",
-      getMiddlewaresFor("memoryRetrieval"),
-      (args) => runDefaultMemoryRetrieval(args, memoryDeps),
-      memoryArgs,
-      memoryPluginTurnCtx,
-      DEFAULT_TIMEOUTS.memoryRetrieval,
-    );
-
-    // Consume the memory-graph block when the default retriever emitted
-    // one. Custom plugins that substitute their own blocks without the
-    // default discriminator are expected to handle their own side effects
-    // (event emission, metric persistence) inside their middleware; this
-    // block short-circuits to the original no-op behavior in that case.
-    const defaultGraphPayload: GraphMemoryPayload | null =
-      asDefaultGraphPayload(memoryResult.memoryGraphBlocks);
-    let pkbQueryVector: number[] | undefined;
-    let pkbSparseVector: QdrantSparseVector | undefined;
-    if (defaultGraphPayload) {
-      const graphResult = defaultGraphPayload.result;
-      runMessages = graphResult.runMessages;
-      // Select dense+sparse as a matched pair so RRF fusion combines two
-      // signals aligned to the same query text:
-      //   1. Context-load with a user query: user-query dense + user-query
-      //      sparse — the cleanest pairing.
-      //   2. Otherwise (context-load without a user query, or per-turn):
-      //      whatever `queryVector` / `sparseVector` the retriever produced,
-      //      which are themselves co-aligned (both summary-derived in
-      //      context-load, both user-last-message-derived in per-turn).
-      // Never pair a user-query dense with a summary-aligned sparse.
-      if (graphResult.userQueryVector) {
-        pkbQueryVector = graphResult.userQueryVector;
-        pkbSparseVector = graphResult.userQuerySparseVector;
-      } else {
-        pkbQueryVector = graphResult.queryVector;
-        pkbSparseVector = graphResult.sparseVector;
-      }
-
-      // Persist the injected block text in message metadata so it survives
-      // conversation reloads (eviction, restart, fork). loadFromDb re-injects
-      // from metadata. Routed through the `persistence` pipeline so plugins
-      // can observe or override metadata updates alongside add/delete.
-      if (graphResult.injectedBlockText) {
-        try {
-          await runPipeline<PersistArgs, PersistResult>(
-            "persistence",
-            getMiddlewaresFor("persistence"),
-            defaultPersistenceTerminal,
-            {
-              op: "update",
-              messageId: userMessageId,
-              updates: {
-                memoryInjectedBlock: graphResult.injectedBlockText,
-              },
-            },
-            buildPluginTurnContext(ctx, reqId),
-            DEFAULT_TIMEOUTS.persistence,
-          );
-        } catch (err) {
-          rlog.warn(
-            { err },
-            "Failed to persist memory injection to metadata (non-fatal)",
-          );
-        }
-      }
-
-      const m = graphResult.metrics;
-
-      try {
-        recordMemoryRecallLog({
-          conversationId: ctx.conversationId,
-          enabled: true,
-          degraded: false,
-          provider: m?.embeddingProvider ?? undefined,
-          model: m?.embeddingModel ?? undefined,
-          semanticHits: m?.semanticHits ?? 0,
-          mergedCount: m?.mergedCount ?? 0,
-          selectedCount: m?.selectedCount ?? 0,
-          tier1Count: m?.tier1Count ?? 0,
-          tier2Count: m?.tier2Count ?? 0,
-          hybridSearchLatencyMs: m?.hybridSearchLatencyMs ?? 0,
-          sparseVectorUsed: m?.sparseVectorUsed ?? false,
-          injectedTokens: graphResult.injectedTokens,
-          latencyMs: graphResult.latencyMs,
-          topCandidatesJson: (m?.topCandidates ?? []).map((c) => ({
-            key: c.nodeId,
-            type: c.type,
-            kind: "graph",
-            finalScore: c.score,
-            semantic: c.semanticSimilarity,
-            recency: c.recencyBoost,
-          })),
-          injectedText: graphResult.injectedBlockText ?? undefined,
-          reason: `graph:${graphResult.mode}`,
-          queryContext: m?.queryContext ?? undefined,
-        });
-      } catch (err) {
-        log.warn({ err }, "Failed to persist memory recall log (non-fatal)");
-      }
-
-      if (m) {
-        const memoryRecalledEvent: MemoryRecalled = {
-          type: "memory_recalled",
-          provider: m.embeddingProvider ?? "unknown",
-          model: m.embeddingModel ?? "unknown",
-          semanticHits: m.semanticHits,
-          mergedCount: m.mergedCount,
-          selectedCount: m.selectedCount,
-          tier1Count: m.tier1Count,
-          tier2Count: m.tier2Count,
-          hybridSearchLatencyMs: m.hybridSearchLatencyMs,
-          sparseVectorUsed: m.sparseVectorUsed,
-          injectedTokens: graphResult.injectedTokens,
-          latencyMs: graphResult.latencyMs,
-          topCandidates: m.topCandidates.map((c) => ({
-            key: c.nodeId,
-            type: c.type,
-            kind: "graph",
-            finalScore: c.score,
-            semantic: c.semanticSimilarity,
-            recency: c.recencyBoost,
-          })),
-        };
-        onEvent(memoryRecalledEvent);
-      }
-    }
-
-    // Build active surface context
-    let activeSurface: ActiveSurfaceContext | null = null;
-    if (ctx.currentActiveSurfaceId) {
-      const stored = ctx.surfaceState.get(ctx.currentActiveSurfaceId);
-      if (stored && stored.surfaceType === "dynamic_page") {
-        const data = stored.data as DynamicPageSurfaceData;
-        activeSurface = {
-          surfaceId: ctx.currentActiveSurfaceId,
-          html: data.html,
-          currentPage: ctx.currentPage,
-        };
-        if (data.appId) {
-          const app = getApp(data.appId);
-          if (app) {
-            activeSurface.appId = app.id;
-            activeSurface.appName = app.name;
-            activeSurface.appDirName = resolveAppDir(app.id).dirName;
-            activeSurface.appSchemaJson = app.schemaJson;
-            activeSurface.appFiles = listAppFiles(app.id);
-            if (app.pages && Object.keys(app.pages).length > 0) {
-              activeSurface.appPages = app.pages;
-            }
-          }
-        }
-      }
-    }
-
-    // Query active documents for this conversation so the injector chain
-    // can surface them to the assistant (prevents duplicate document_create
-    // calls when existing documents should be targeted with document_update).
-    const conversationDocs = getDocumentsForConversation(ctx.conversationId);
-    const activeDocuments =
-      conversationDocs.length > 0
-        ? conversationDocs.map((d) => ({
-            surfaceId: d.surfaceId,
-            title: d.title,
-            wordCount: d.wordCount,
-            updatedAt: d.updatedAt,
-          }))
-        : null;
-
-    ctx.refreshWorkspaceTopLevelContextIfNeeded();
-
-    // Compute fresh turn timestamp for date grounding.
-    // Absolute "now" is always anchored to assistant host clock, while local
-    // date semantics prefer configured user timezone, then device timezones.
+    // Resolve the turn's timezone cascade up front. It depends only on config
+    // and the inbound request — never on retrieval output — so it can be
+    // settled before context assembly. Local date semantics prefer the
+    // configured user timezone, then device timezones, then the host clock.
     const hostTimeZone = Intl.DateTimeFormat().resolvedOptions().timeZone;
     const timezoneContext = resolveTurnTimezoneContext({
       configuredUserTimeZone: config.ui.userTimezone ?? null,
@@ -1417,9 +1092,6 @@ export async function runAgentLoopImpl(
       detectedTimezone: config.ui.detectedTimezone ?? null,
       hostTimeZone,
     });
-    const timestamp = formatTurnTimestamp({
-      timeZone: timezoneContext.effectiveTimezone,
-    });
 
     // Resolve the inbound actor context for the unified <turn_context> block.
     // When the conversation carries enough identity info, use the unified
@@ -1443,8 +1115,10 @@ export async function runAgentLoopImpl(
       }
     }
 
-    // Build unified turn context block that replaces the separate temporal,
-    // channel, interface, and actor context blocks.
+    // Resolve the channel/interface labels and the guardian flag for this
+    // turn. These derive only from the captured turn context and the resolved
+    // actor trust class — never from retrieval — so they settle before context
+    // assembly.
     const interfaceName =
       capturedTurnInterfaceContext.userMessageInterface ?? undefined;
     const channelName =
@@ -1489,9 +1163,54 @@ export async function runAgentLoopImpl(
       });
       const label = profileEntry?.label ?? effectiveProfileKey;
       modelProfileStr = resolved.model ? `${label} (${resolved.model})` : label;
-      setLastNotifiedInferenceProfile(ctx.conversationId, effectiveProfileKey);
+      // Record the notification for persistence on delivery rather than here:
+      // the model only "learns" the profile once it receives this turn
+      // context, signalled by the first `message_complete`. Persisting inline
+      // would mark the profile notified even if the turn is cancelled or fails
+      // before the model ever sees the notice.
+      state.pendingNotifiedInferenceProfile = effectiveProfileKey;
     }
 
+    // Memory retrieval — fetches PKB, NOW.md, and memory-graph outputs and
+    // persists the retrieval's own side effects (injected-block metadata,
+    // recall log, `memory_recalled` event). Runs at the early "prompt
+    // submitted, before context assembly" moment because its outputs feed the
+    // injection and overflow-reduction transforms below. It is shaped as the
+    // `user-prompt-submit-temp` hook handler but invoked directly for now: it
+    // must run early, while the canonical late `user-prompt-submit` hook
+    // (history repair, title) runs after those transforms, so the two cannot
+    // share a fire site until compaction is cleared from the gap between them.
+    const isTrustedActor = resolveTrustClass(ctx.trustContext) === "guardian";
+    const memoryCtx: MemoryRetrievalHookContext = {
+      graphMemory: ctx.graphMemory,
+      config: getConfig(),
+      onEvent,
+      isTrustedActor,
+      conversationId: ctx.conversationId,
+      userMessageId,
+      logger: rlog,
+      // An external cancel aborts `prepareMemory` instead of letting it run
+      // to completion after the turn has already been torn down.
+      signal: abortController.signal,
+      latestMessages: ctx.messages,
+    };
+    await userPromptSubmitMemoryRetrieval(memoryCtx);
+
+    // The retriever owns its side effects (injected-block metadata, recall
+    // log, `memory_recalled` event) and records the dense/sparse PKB query
+    // pair on the graph handle for the PKB-reminder injector to read back; the
+    // loop only reuses the injected message list downstream.
+    let runMessages = memoryCtx.latestMessages;
+
+    // Capture wall-clock "now" at its point of use, after the blocking memory
+    // retrieval, so the injected `<turn_context>` timestamp reflects current
+    // time rather than the moment the turn began.
+    const timestamp = formatTurnTimestamp({
+      timeZone: timezoneContext.effectiveTimezone,
+    });
+
+    // Build unified turn context block that replaces the separate temporal,
+    // channel, interface, and actor context blocks.
     const baseTurnContext = {
       timestamp,
       interfaceName,
@@ -1513,64 +1232,6 @@ export async function runAgentLoopImpl(
 
     // The `remember` tool handles scratchpad-style memory writes directly to the graph.
 
-    // Personal-memory trust gate: PKB, NOW.md, and v2 static blocks all
-    // hold private user content. Block exposure to non-guardian actors
-    // arriving over a remote channel; internal/local flows pass through.
-    // See `shouldExposePersonalMemory` for the threat model.
-    const personalMemoryAllowed = shouldExposePersonalMemory({
-      sourceChannel: ctx.trustContext?.sourceChannel,
-      isTrustedActor,
-    });
-
-    // Inject NOW.md and PKB content only on the first turn (or after
-    // compaction re-strips them).  Old injections persist in history and
-    // are never stripped on normal turns — this preserves the cached prefix.
-    // PKB/NOW content is sourced from the `memoryRetrieval` pipeline above
-    // so plugins can override either source without touching the agent loop.
-    // NOW.md injection can be disabled via `memory.retrieval.scratchpadInjection.enabled`.
-    const scratchpadInjectionEnabled =
-      getConfig().memory.retrieval.scratchpadInjection.enabled;
-    const currentNowContent =
-      personalMemoryAllowed && scratchpadInjectionEnabled
-        ? memoryResult.nowContent
-        : null;
-    const shouldInjectNowAndPkb = isFirstMessage || compactedThisTurn;
-    const nowScratchpad = shouldInjectNowAndPkb ? currentNowContent : null;
-
-    const currentPkbContent = personalMemoryAllowed
-      ? memoryResult.pkbContent
-      : null;
-    const pkbContext = shouldInjectNowAndPkb ? currentPkbContent : null;
-    const pkbActive = currentPkbContent !== null;
-
-    // V2 static memory block (essentials/threads/recent/buffer).
-    // `currentMemoryV2Static` is the trust-gated content reused by every
-    // re-injection path — it stays non-null on non-full-mode turns so
-    // that mid-turn reducer compaction (which strips the prior `<info>`
-    // block) can restore the freshest content. `memoryV2Static` is the
-    // first-turn / post-compaction cadence-gated value for initial
-    // injection only. `readMemoryV2StaticContent` self-gates on the v2
-    // flag + config and returns null when v2 is off.
-    const currentMemoryV2Static = personalMemoryAllowed
-      ? readMemoryV2StaticContent()
-      : null;
-    const memoryV2Static = shouldInjectNowAndPkb ? currentMemoryV2Static : null;
-
-    // PKB relevance-hint inputs. Resolved once per turn and reused across
-    // re-injections so post-compaction rebuilds pick up fresh hints against
-    // the updated conversation history.
-    const pkbRoot = pkbActive ? join(getWorkspaceDir(), "pkb") : undefined;
-    const pkbAutoInjectList = pkbRoot
-      ? getPkbAutoInjectList(pkbRoot)
-      : undefined;
-    // Pass `ctx` directly — `PkbContextConversation` is structural and
-    // `getInContextPkbPaths` re-reads `conversation.messages` on each call,
-    // so post-compaction re-injects see the updated history.
-    const pkbConversation = pkbActive ? ctx : undefined;
-    // PKB points live under a single workspace sentinel scope.
-    // See `PKB_WORKSPACE_SCOPE` for why.
-    const pkbScopeId = pkbActive ? PKB_WORKSPACE_SCOPE : undefined;
-
     // Subagent status injection — gives the parent LLM visibility into active/completed children.
     // Skipped when this conversation IS a subagent (no nesting) or has no children.
     const subagentStatusBlock = ctx.isSubagent
@@ -1625,21 +1286,14 @@ export async function runAgentLoopImpl(
         )
       : null;
 
-    // Guards the chronological-transcript override on re-injection after
-    // the reducer compacts `ctx.messages`. The captured transcript is the
-    // full persisted history; blindly replaying it on every re-inject would
-    // overwrite the reducer's compacted messages and undo compaction. Flip
-    // to `true` after any compaction so subsequent re-injections fall back
-    // to the reduced `ctx.messages`.
-    let reducerCompacted = compactedThisTurn;
-
-    // memory-v3-live: route the turn's `<memory>` block to the v3 injector.
-    // When on, runtime assembly suppresses v2's `<memory>` injection (only
-    // when the v3 injector actually produced a block — otherwise v2 stays as a
-    // fallback) and the provider anchors its long-TTL cache breakpoint on the
-    // most recent STABLE user message, since the latest user message now
-    // carries the volatile per-turn memory block. Flag off → bit-for-bit
-    // identical to today's v2 path.
+    state.reducerCompacted = compactedThisTurn;
+
+    // memory-v3-live: when on, the provider anchors its long-TTL cache
+    // breakpoint on the most recent STABLE user message, since the latest user
+    // message now carries the volatile per-turn `<memory>` block the v3
+    // injector emits. The matching v2-suppression strip is owned by
+    // `applyRuntimeInjections`, which reads the same flag itself. Flag off →
+    // bit-for-bit identical to today's v2 path.
     const memoryV3Live = isAssistantFeatureFlagEnabled(
       "memory-v3-live",
       getConfig(),
@@ -1647,29 +1301,7 @@ export async function runAgentLoopImpl(
 
     // Shared injection options — reused whenever we need to re-inject after reduction.
     const injectionOpts = {
-      suppressV2MemoryForV3: memoryV3Live,
-      diskPressureContext,
-      activeSurface,
-      activeDocuments,
-      workspaceTopLevelContext: shouldInjectWorkspace
-        ? ctx.workspaceTopLevelContext
-        : null,
-      channelCapabilities: ctx.channelCapabilities ?? null,
-      channelCommandContext: ctx.commandIntent ?? null,
       unifiedTurnContext: unifiedTurnContextStr,
-      pkbContext,
-      pkbActive,
-      pkbQueryVector,
-      pkbSparseVector,
-      pkbScopeId,
-      pkbConversation,
-      pkbAutoInjectList,
-      pkbRoot,
-      pkbWorkingDir: pkbActive ? ctx.workingDir : undefined,
-      memoryV2Static,
-      nowScratchpad,
-      voiceCallControlPrompt: ctx.voiceCallControlPrompt ?? null,
-      transportHints: ctx.transportHints ?? null,
       isNonInteractive: !isInteractiveResolved,
       isBackgroundConversation: isBackgroundConversationType(
         turnStartConversation?.conversationType,
@@ -1689,7 +1321,7 @@ export async function runAgentLoopImpl(
 
     const injection = await applyRuntimeInjections(runMessages, {
       ...injectionOpts,
-      slackChronologicalMessages: reducerCompacted
+      slackChronologicalMessages: state.reducerCompacted
         ? null
         : injectionOpts.slackChronologicalMessages,
       mode: currentInjectionMode,
@@ -1735,18 +1367,7 @@ export async function runAgentLoopImpl(
           metadataUpdates.memoryV2StaticBlock =
             injection.blocks.memoryV2StaticBlock;
         }
-        await runPipeline<PersistArgs, PersistResult>(
-          "persistence",
-          getMiddlewaresFor("persistence"),
-          defaultPersistenceTerminal,
-          {
-            op: "update",
-            messageId: userMessageId,
-            updates: metadataUpdates,
-          },
-          buildPluginTurnContext(ctx, reqId),
-          DEFAULT_TIMEOUTS.persistence,
-        );
+        updateMessageMetadata(userMessageId, metadataUpdates);
       } catch (err) {
         rlog.warn({ err }, "Failed to persist injection metadata (non-fatal)");
       }
@@ -1762,51 +1383,18 @@ export async function runAgentLoopImpl(
     let reducerState: ReducerState | undefined;
 
     const toolTokenBudget = ctx.agentLoop.getToolTokenBudget(runMessages);
-    // Canonical calibration key — passed to the `tokenEstimate` pipeline for
-    // every preflight/mid-loop estimate, the overflow reducer config, and the
-    // convergence-path `estimatePromptTokens` call. Matches the key recorded
-    // by `handleUsage` for wrapper providers (OpenRouter routing to
-    // Anthropic → key is `"anthropic"`).
+    // Canonical calibration key — used by the preflight estimate, the
+    // overflow reducer config, and the convergence-path `estimatePromptTokens`
+    // call. Matches the key recorded by `handleUsage` for wrapper providers
+    // (OpenRouter routing to Anthropic → key is `"anthropic"`).
     const estimationProviderName = getCalibrationProviderKey(ctx.provider);
 
-    // Shared `TurnContext` for every `tokenEstimate` pipeline invocation in
-    // this turn. The pipeline is the extension point for plugins that want
-    // to substitute an alternate estimator (e.g. provider-native tokenization)
-    // without touching orchestrator code.
-    //
-    // Routed through the canonical builder — `turnIndex` is `ctx.turnCount`,
-    // trust cascades through per-turn/conversation-level/fallback, and the
-    // context-window handle rides along so any middleware that wants to
-    // reuse the manager (e.g. to compute compaction-aware estimates) can.
-    const pipelineTurnCtx = buildPluginTurnContext(ctx, reqId);
-
-    const runTokenEstimatePipeline = (
-      history: Message[],
-    ): Promise<EstimateResult> =>
-      runPipeline<EstimateArgs, EstimateResult>(
-        "tokenEstimate",
-        getMiddlewaresFor("tokenEstimate"),
-        defaultTokenEstimateTerminal,
-        {
-          // Shallow-frozen copies so a misbehaving middleware that mutates
-          // `args.history` or `args.tools` in place (e.g. trims the array
-          // before calling next) can't silently strip prompt context from
-          // the orchestrator's live `runMessages` / resolved-tools arrays.
-          // TypeScript `readonly` on `EstimateArgs` does not prevent
-          // `push`/`splice` at runtime; the frozen wrapper throws in strict
-          // mode and isolates any mutation attempts from the call-site state.
-          history: Object.freeze([...history]) as Message[],
-          systemPrompt: ctx.systemPrompt,
-          tools: Object.freeze([
-            ...ctx.agentLoop.getResolvedTools(history),
-          ]) as ToolDefinition[],
-          providerName: estimationProviderName,
-        },
-        pipelineTurnCtx,
-        DEFAULT_TIMEOUTS.tokenEstimate,
-      );
-
-    const preflightTokens = await runTokenEstimatePipeline(runMessages);
+    const preflightTokens = estimatePromptTokensWithTools(
+      runMessages,
+      ctx.systemPrompt,
+      ctx.agentLoop.getResolvedTools(runMessages),
+      estimationProviderName,
+    );
 
     if (overflowRecovery.enabled && preflightTokens > preflightBudget) {
       rlog.warn(
@@ -1818,16 +1406,12 @@ export async function runAgentLoopImpl(
         "Preflight budget exceeded — running overflow reducer before provider call",
       );
 
-      // Overflow reduction runs through the plugin pipeline. The default
-      // middleware (`default-overflow-reduce`, registered at bootstrap)
-      // contains the historical tier loop — forced compaction → tool-result
-      // truncation → media stubbing → injection downgrade — plus the
-      // re-inject/re-estimate convergence check. The callbacks below are
-      // the orchestrator-specific side effects that the plugin coordinates
-      // per iteration (activity emission, compaction application, runtime
-      // injection reassembly, token re-estimation). Registered plugins that
-      // wrap the `overflowReduce` slot see each iteration through their own
-      // middleware `next` callback.
+      // `runOverflowReductionLoop` drives the tier loop — forced compaction →
+      // tool-result truncation → media stubbing → injection downgrade — plus
+      // the re-inject/re-estimate convergence check. The callbacks below are
+      // the orchestrator-specific side effects it coordinates per iteration
+      // (activity emission, compaction application, runtime injection
+      // reassembly, token re-estimation).
       const messagesForPreflightOverflowReduction =
         slackChronologicalContext?.messages ?? ctx.messages;
       const overflowArgs: OverflowReduceArgs = {
@@ -1841,72 +1425,18 @@ export async function runAgentLoopImpl(
         maxAttempts: resolveCurrentContextBudget().overflowRecovery.maxAttempts,
         abortSignal: abortController.signal,
         compactFn: async (msgs, signal, opts) => {
-          // Route the reducer's forced-compaction tier through the
-          // `compaction` pipeline so registered plugins observe these
-          // invocations. Without this, custom compaction middleware only
-          // sees the three orchestrator-owned call sites and misses the
-          // reducer-initiated forced compactions entirely.
-          //
-          // Pipeline timeouts must be caught locally — a `PluginTimeoutError`
-          // bubbling out of here would abort the overflow-reducer tier loop
-          // entirely, skipping fallback tiers (tool-result truncation, media
-          // stubbing, injection downgrade) and bypassing circuit-breaker
-          // bookkeeping. On timeout, record the failure and return a
-          // `compacted: false` result so the reducer falls through to the
-          // next tier.
-          try {
-            return (await runPipeline<CompactionArgs, CompactionResult>(
-              "compaction",
-              getMiddlewaresFor("compaction"),
-              (args) =>
-                defaultCompactionTerminal(
-                  args,
-                  buildPluginTurnContext(ctx, reqId),
-                ),
-              {
-                messages: msgs,
-                signal,
-                options: {
-                  ...(opts ?? {}),
-                  overrideProfile: resolveCurrentOverrideProfile() ?? null,
-                  actorTrustClass: ctx.trustContext?.trustClass,
-                },
-              },
-              buildPluginTurnContext(ctx, reqId),
-              DEFAULT_TIMEOUTS.compaction,
-            )) as Awaited<
-              ReturnType<typeof ctx.contextWindowManager.maybeCompact>
-            >;
-          } catch (err) {
-            if (err instanceof PluginTimeoutError) {
-              rlog.warn(
-                { err, phase: "overflow-reducer-forced-compaction" },
-                "Compaction pipeline timed out — falling through to next reducer tier",
-              );
-              await ctx.agentLoop.compactionCircuit.recordOutcome(
-                ctx,
-                true,
-                onEvent,
-              );
-              return {
-                messages: msgs,
-                compacted: false,
-                previousEstimatedInputTokens: 0,
-                estimatedInputTokens: 0,
-                maxInputTokens: 0,
-                thresholdTokens: 0,
-                compactedMessages: 0,
-                compactedPersistedMessages: 0,
-                summaryCalls: 0,
-                summaryInputTokens: 0,
-                summaryOutputTokens: 0,
-                summaryModel: "",
-                summaryText: "",
-                reason: "compaction pipeline timed out",
-              };
-            }
-            throw err;
-          }
+          // Delegate the reducer's forced-compaction tier to the default
+          // compaction plugin, overlaying the turn's resolved inference
+          // profile and actor trust class onto the reducer-supplied options.
+          const reducerOptions = (opts ?? {}) as ContextWindowCompactOptions;
+          return defaultCompact({
+            manager: ctx.contextWindowManager,
+            messages: msgs,
+            signal,
+            ...reducerOptions,
+            overrideProfile: resolveCurrentOverrideProfile() ?? null,
+            actorTrustClass: resolveTurnActorTrustClass(ctx),
+          });
         },
         emitActivityState: () => {
           ctx.emitActivityState("thinking", "context_compacting", {
@@ -1925,14 +1455,12 @@ export async function runAgentLoopImpl(
           // breaker.
           if (result.summaryFailed !== undefined) {
             await ctx.agentLoop.compactionCircuit.recordOutcome(
-              ctx,
               result.summaryFailed,
               onEvent,
             );
           }
           if (result.compacted) {
             await applySuccessfulCompaction(result, compactedBasis);
-            shouldInjectWorkspace = true;
           }
         },
         reinjectForMode: async (
@@ -1943,27 +1471,25 @@ export async function runAgentLoopImpl(
         ) => {
           // Mirror the pre-PR-23 behavior: `ctx.messages` must track the
           // reducer's latest output before re-injection runs, because other
-          // sites consulted through `injectionOpts` (`workspaceTopLevelContext`,
-          // slack history, etc.) depend on it and `applyCompactionResult`
-          // only updates `ctx.messages` on a compaction tier. Assigning here
+          // sites consulted through `injectionOpts` (slack history, etc.) and
+          // the injectors' own message-presence scans depend on it, and
+          // `applyCompactionResult` only updates `ctx.messages` on a
+          // compaction tier. Assigning here
           // keeps non-compaction tiers (tool-result truncation, media
           // stubbing, injection downgrade) observable to downstream
           // injection assembly on the same turn.
           ctx.messages = reducedMessages;
 
-          // When THIS iteration compacted, it stripped existing NOW.md /
-          // PKB blocks — so we re-inject current content. A later iteration
-          // that only truncates or downgrades must NOT re-force PKB/NOW,
+          // When THIS iteration compacted, it stripped the existing
+          // memory-static block — so we re-inject current content. A later
+          // iteration that only truncates or downgrades must NOT re-force it,
           // or each round would grow the token count.
           // Gate: only the iteration that actually compacted re-injects.
+          // (The `<knowledge_base>`, NOW.md, and v2 static `<info>` blocks
+          // self-gate inside their injectors on whether they are already
+          // present in `reducedMessages`.)
           const injection = await applyRuntimeInjections(reducedMessages, {
             ...injectionOpts,
-            ...(stepCompacted && { pkbContext: currentPkbContent }),
-            ...(stepCompacted && { memoryV2Static: currentMemoryV2Static }),
-            ...(stepCompacted && { nowScratchpad: currentNowContent }),
-            workspaceTopLevelContext: shouldInjectWorkspace
-              ? ctx.workspaceTopLevelContext
-              : null,
             // Once ANY iteration has compacted `ctx.messages`, the captured
             // `slackChronologicalMessages` snapshot (built from the full
             // persisted transcript) would overwrite the compacted history
@@ -1989,41 +1515,17 @@ export async function runAgentLoopImpl(
           }),
       };
 
-      const overflowResult = await runPipeline<
-        OverflowReduceArgs,
-        OverflowReduceResult
-      >(
-        "overflowReduce",
-        getMiddlewaresFor("overflowReduce"),
-        // Terminal — only reached when every registered middleware calls
-        // `next` and delegates past the innermost layer. The default plugin
-        // is a terminal itself (it doesn't call `next`), so in practice
-        // this fallback fires only when the default has been explicitly
-        // deregistered (tests) and no user plugin replaces it. Strict-fail
-        // semantics: throw so the missing terminal surfaces as a visible
-        // error instead of silently returning the history untouched.
-        async () => {
-          throw new PluginExecutionError(
-            "overflowReduce pipeline has no terminal handler — every reducer middleware called next() without providing a replacement",
-            "overflowReduce",
-          );
-        },
-        overflowArgs,
-        buildPluginTurnContext(ctx, reqId),
-        DEFAULT_TIMEOUTS.overflowReduce,
-      );
+      const overflowResult = await runOverflowReductionLoop(overflowArgs);
 
       ctx.messages = overflowResult.messages;
       runMessages = overflowResult.runMessages;
       currentInjectionMode = overflowResult.injectionMode;
       reducerState = overflowResult.reducerState;
       if (overflowResult.reducerCompacted) {
-        reducerCompacted = true;
+        state.reducerCompacted = true;
       }
     }
 
-    let preRepairMessages = runMessages;
-
     // Replace historical web_search_tool_result blocks with text summaries.
     // The opaque `encrypted_content` tokens Anthropic attaches to each result
     // expire / are route-scoped; replaying a stale token is rejected with
@@ -2046,13 +1548,12 @@ export async function runAgentLoopImpl(
     // context with a fresh array; `runHook` forwards whichever the chain
     // settles on. Order is plugin registration order.
     //
-    // Fires BEFORE `preRunHistoryLength` is captured so the boundary
-    // between pre-existing and hook-emitted messages — consumed by the
-    // ordering-error retry gate, the post-run reconcile loop, and the
-    // new-message extraction for persistence — reflects exactly what
-    // `agentLoop.run` receives.
+    // Fires BEFORE the agent loop runs so the hook-emitted messages are part
+    // of the loop's input; the loop then reports its own appended output via
+    // `AgentLoopRunResult.newMessages`, which is what persistence consumes.
     const userPromptCtx: UserPromptSubmitContext = {
       conversationId: ctx.conversationId,
+      prompt: options?.titleText ?? content,
       originalMessages: ctx.messages,
       latestMessages: runMessages,
       logger: rlog,
@@ -2063,8 +1564,6 @@ export async function runAgentLoopImpl(
     );
     runMessages = finalUserPromptCtx.latestMessages;
 
-    let preRunHistoryLength = runMessages.length;
-
     const shouldGenerateTitle = isReplaceableTitle(
       getConversation(ctx.conversationId)?.title ?? null,
     );
@@ -2078,6 +1577,7 @@ export async function runAgentLoopImpl(
       rlog,
       turnChannelContext: capturedTurnChannelContext,
       turnInterfaceContext: capturedTurnInterfaceContext,
+      applyCompaction: applySuccessfulCompaction,
     };
     const eventHandler = (event: AgentEvent): Promise<void> =>
       dispatchAgentEvent(state, deps, event);
@@ -2097,82 +1597,39 @@ export async function runAgentLoopImpl(
     rlog.info({ callSite: turnCallSite }, "Starting agent loop run");
 
     // Thread the orchestrator's canonical per-turn context into the agent
-    // loop so its internal pipeline invocations (llmCall, emptyResponse,
-    // toolError, toolResultTruncate, toolExecute) see the real
-    // conversation identity / trust / contextWindowManager instead of the
-    // synthesized `"agent-loop"` placeholder. The loop clones this value
+    // loop so its internal pipeline invocations (e.g. compaction) see the
+    // real conversation identity / trust / contextWindowManager instead of
+    // the synthesized `"agent-loop"` placeholder. The loop clones this value
     // and overwrites `turnIndex` with its own tool-use iteration counter.
     const loopTurnCtx = buildPluginTurnContext(ctx, reqId);
 
-    // Hooks for the loop-owned mid-loop compaction. The agent loop owns the
+    // Hook for the loop-owned mid-loop compaction. The agent loop owns the
     // trigger (its budget gate), the `compaction` pipeline call, the result
     // interpretation (circuit-breaker bookkeeping + the exhaustion decision),
-    // and the inline continue; these callbacks bridge the durable / injection
-    // state the loop is intentionally blind to. Durable persistence and
-    // re-injection stay orchestrator-supplied for now.
+    // and the inline continue; this callback bridges the injection state the
+    // loop is intentionally blind to. Durable persistence is signalled via
+    // events; re-injection stays orchestrator-supplied for now.
     const midLoopCompaction: MidLoopCompaction = {
-      prepare: (history) => {
-        // Strip injected context so the compactor summarizes the raw
-        // persistent messages, and commit the stripped set to durable state.
-        const rawHistory = stripInjectionsForCompaction(history);
-        ctx.messages = rawHistory;
-        markHistoryStrippedBestEffort(ctx.conversationId, Date.now(), rlog);
-        return {
-          rawHistory,
-          options: {
-            lastCompactedAt: ctx.contextCompactedAt ?? undefined,
-            force: true,
-            targetInputTokensOverride:
-              resolveCurrentContextBudget().preflightBudget,
-            conversationOriginChannel:
-              getConversationOriginChannel(ctx.conversationId) ?? undefined,
-            overrideProfile: resolveCurrentOverrideProfile() ?? null,
-            actorTrustClass: ctx.trustContext?.trustClass,
-          },
-        };
-      },
-      applyResult: async (result, rawHistory) => {
-        await applySuccessfulCompaction(result, rawHistory);
-        reducerCompacted = true;
-        shouldInjectWorkspace = true;
-      },
-      reinject: async () => {
+      postCompactionHook: async ({ history, turnContext }) => {
         // stripInjectionsForCompaction() unconditionally removed the existing
-        // NOW.md block, so re-inject the current content regardless of whether
-        // compaction actually ran.
-        const injection = await applyRuntimeInjections(ctx.messages, {
+        // memory-static block, so re-inject the current content regardless of
+        // whether compaction actually ran. The `<knowledge_base>`, NOW.md, and
+        // v2 static `<info>` blocks self-gate inside their injectors on block
+        // presence.
+        const injection = await postCompactReinject({
           ...injectionOpts,
-          pkbContext: currentPkbContent,
-          memoryV2Static: currentMemoryV2Static,
-          nowScratchpad: currentNowContent,
-          workspaceTopLevelContext: shouldInjectWorkspace
-            ? ctx.workspaceTopLevelContext
-            : null,
           // Suppress the chronological-transcript snapshot once the reducer
           // has collapsed `ctx.messages`; the captured snapshot reflects the
           // full persisted transcript and would overwrite compaction.
-          slackChronologicalMessages: reducerCompacted
+          slackChronologicalMessages: state.reducerCompacted
             ? null
             : injectionOpts.slackChronologicalMessages,
           mode: currentInjectionMode,
-          turnContext: buildPluginTurnContext(ctx, reqId),
+          turnContext,
+          history,
+          logger: rlog,
         });
-        runMessages = injection.messages;
-        if (isTrustedActor && currentInjectionMode !== "minimal") {
-          ctx.graphMemory.retrackCachedNodes();
-        }
-        const midLoopCompactStrip =
-          stripHistoricalWebSearchResults(runMessages);
-        if (midLoopCompactStrip.stats.blocksStripped > 0) {
-          rlog.info(
-            { phase: "mid-loop-compact", ...midLoopCompactStrip.stats },
-            "Converted historical web_search_tool_result blocks to text summaries",
-          );
-          runMessages = midLoopCompactStrip.messages;
-        }
-        preRepairMessages = runMessages;
-        preRunHistoryLength = runMessages.length;
-        return runMessages;
+        return injection.messages;
       },
     };
 
@@ -2188,10 +1645,8 @@ export async function runAgentLoopImpl(
       msgs: Message[],
       compaction?: MidLoopCompaction,
     ): Promise<Message[]> => {
-      const { history, exitReason } = await ctx.agentLoop.run(
-        msgs,
-        eventHandler,
-        {
+      const { history, exitReason, appendedNewMessages, newMessages } =
+        await ctx.agentLoop.run(msgs, eventHandler, {
           signal: abortController.signal,
           requestId: reqId,
           onCheckpoint,
@@ -2205,8 +1660,9 @@ export async function runAgentLoopImpl(
           // `<memory>` block, so anchor the provider's long-TTL cache breakpoint
           // on the most recent stable message instead.
           mutableLatestUserMessage: memoryV3Live,
-        },
-      );
+        });
+      lastRunAppendedNewMessages = appendedNewMessages;
+      lastRunNewMessages = newMessages;
       if (exitReason === "handoff") {
         yieldedForHandoff = true;
         pendingCheckpointYield = "handoff";
@@ -2244,10 +1700,7 @@ export async function runAgentLoopImpl(
     }
 
     // One-shot ordering error retry
-    if (
-      state.orderingErrorDetected &&
-      updatedHistory.length === preRunHistoryLength
-    ) {
+    if (state.orderingErrorDetected && !lastRunAppendedNewMessages) {
       rlog.warn(
         { phase: "retry" },
         "Provider ordering error detected, attempting one-shot deep-repair retry",
@@ -2261,12 +1714,10 @@ export async function runAgentLoopImpl(
       // `user-prompt-submit` hook (the default history-repair plugin runs
       // `repairHistory` there); widening that surface to deep-repair is
       // intentionally deferred until there's a concrete plugin-level use case.
-      const retryRepair = deepRepairHistory(runMessages);
+      const retryRepair = deepRepairHistory(updatedHistory);
       runMessages = retryRepair.messages;
       const retryStrip = stripHistoricalWebSearchResults(runMessages);
       runMessages = retryStrip.messages;
-      preRepairMessages = runMessages;
-      preRunHistoryLength = runMessages.length;
       state.orderingErrorDetected = false;
       state.deferredOrderingError = null;
 
@@ -2319,15 +1770,29 @@ export async function runAgentLoopImpl(
             }
             // Can't resize — replace with a text annotation so the model
             // can explain the situation rather than silently dropping context
-            return [
-              {
-                type: "text" as const,
-                text: "(An image was attached but could not be sent — its dimensions exceed the provider limit and automatic resize was not available. Please resize the image and try again.)",
-              },
-            ];
+            return [{ type: "text" as const, text: UNSENDABLE_IMAGE_NOTE }];
           }),
         };
       });
+      // The transform above only mutates ctx.messages for the current retry.
+      // Persist the downgrade for images that can never be sent so the rejected
+      // upload doesn't rehydrate from the DB and resurface on later turns. This
+      // is cleanup for future turns, so a persistence failure must never abort
+      // the retry that is about to run — log it and continue.
+      try {
+        const rewritten = persistUnsendableImageDowngrades(ctx.conversationId);
+        if (rewritten > 0) {
+          rlog.info(
+            { phase: "image-recovery", rewritten },
+            "Persisted unsendable-image downgrades so they cannot resurface",
+          );
+        }
+      } catch (err) {
+        rlog.warn(
+          { phase: "image-recovery", err },
+          "Failed to persist unsendable-image downgrade; continuing with in-memory recovery",
+        );
+      }
       runMessages = ctx.messages;
       updatedHistory = await runAgentLoop(runMessages);
       if (state.imageTooLargeDetected) {
@@ -2356,19 +1821,9 @@ export async function runAgentLoopImpl(
     // limit), incorporate those new messages into ctx.messages so the
     // convergence loop operates on the full (larger) history.
     if (state.contextTooLargeDetected) {
-      // Detect whether ctx.messages currently lacks NOW.md so we know if
-      // it needs to be re-injected.  Mid-loop compaction (line ~1067) may
-      // have already stripped injections before escalating here, so we
-      // check actual message state rather than tracking mutation sites.
-      let convergenceStripped =
-        findLastInjectedNowContent(ctx.messages) === null;
-
-      if (updatedHistory.length > preRunHistoryLength) {
+      if (lastRunAppendedNewMessages) {
         ctx.messages = stripInjectionsForCompaction(updatedHistory);
-        markHistoryStrippedBestEffort(ctx.conversationId, Date.now(), rlog);
-        convergenceStripped = true;
-        preRepairMessages = updatedHistory;
-        preRunHistoryLength = updatedHistory.length;
+        markHistoryStrippedBestEffort(ctx.conversationId);
       }
       if (!reducerState) {
         reducerState = createInitialReducerState();
@@ -2450,14 +1905,12 @@ export async function runAgentLoopImpl(
             );
             if (emergencyResult.summaryFailed !== undefined) {
               await ctx.agentLoop.compactionCircuit.recordOutcome(
-                ctx,
                 emergencyResult.summaryFailed,
                 onEvent,
               );
             }
             if (emergencyResult.compacted) {
               await applySuccessfulCompaction(emergencyResult, ctx.messages);
-              shouldInjectWorkspace = true;
             }
             // Clear the overflow flag and re-run the agent loop with
             // the compacted context.
@@ -2508,7 +1961,7 @@ export async function runAgentLoopImpl(
             ctx.contextWindowManager.maybeCompact(msgs, signal!, {
               ...(opts ?? {}),
               overrideProfile: resolveCurrentOverrideProfile() ?? null,
-              actorTrustClass: ctx.trustContext?.trustClass,
+              actorTrustClass: resolveTurnActorTrustClass(ctx),
             }),
           abortController.signal,
         );
@@ -2526,7 +1979,6 @@ export async function runAgentLoopImpl(
           step.compactionResult.summaryFailed !== undefined
         ) {
           await ctx.agentLoop.compactionCircuit.recordOutcome(
-            ctx,
             step.compactionResult.summaryFailed,
             onEvent,
           );
@@ -2537,22 +1989,17 @@ export async function runAgentLoopImpl(
             step.compactionResult,
             convergenceCompactionBasis,
           );
-          shouldInjectWorkspace = true;
-          reducerCompacted = true;
+          state.reducerCompacted = true;
         }
 
-        // Only re-inject NOW.md when ctx.messages was actually stripped;
-        // otherwise the existing NOW.md block is still present and
-        // re-injecting would duplicate it.
+        // Only re-inject the memory-static block when ctx.messages was
+        // actually stripped; otherwise the existing block is still present and
+        // re-injecting would duplicate it. (The `<knowledge_base>` and NOW.md
+        // blocks self-gate inside their injectors on whether they are already
+        // present in `ctx.messages`.)
         const injection = await applyRuntimeInjections(ctx.messages, {
           ...injectionOpts,
-          pkbContext: currentPkbContent,
-          memoryV2Static: convergenceStripped ? currentMemoryV2Static : null,
-          nowScratchpad: convergenceStripped ? currentNowContent : null,
-          workspaceTopLevelContext: shouldInjectWorkspace
-            ? ctx.workspaceTopLevelContext
-            : null,
-          slackChronologicalMessages: reducerCompacted
+          slackChronologicalMessages: state.reducerCompacted
             ? null
             : injectionOpts.slackChronologicalMessages,
           mode: currentInjectionMode,
@@ -2570,8 +2017,6 @@ export async function runAgentLoopImpl(
           );
           runMessages = convergenceStrip.messages;
         }
-        preRepairMessages = runMessages;
-        preRunHistoryLength = runMessages.length;
         state.contextTooLargeDetected = false;
         yieldedForBudget = false;
 
@@ -2594,12 +2039,9 @@ export async function runAgentLoopImpl(
           // Fold rerun progress into ctx.messages so the next reducer
           // tier operates on up-to-date history instead of stale
           // pre-rerun messages.
-          if (updatedHistory.length > preRunHistoryLength) {
+          if (lastRunAppendedNewMessages) {
             ctx.messages = stripInjectionsForCompaction(updatedHistory);
-            markHistoryStrippedBestEffort(ctx.conversationId, Date.now(), rlog);
-            convergenceStripped = true;
-            preRepairMessages = updatedHistory;
-            preRunHistoryLength = updatedHistory.length;
+            markHistoryStrippedBestEffort(ctx.conversationId);
           }
         }
       }
@@ -2619,86 +2061,35 @@ export async function runAgentLoopImpl(
           ctx.emitActivityState("thinking", "context_compacting", {
             requestId: reqId,
           });
-          let emergencyCompact: Awaited<
-            ReturnType<typeof ctx.contextWindowManager.maybeCompact>
-          > | null = null;
-          try {
-            emergencyCompact = (await runPipeline<
-              CompactionArgs,
-              CompactionResult
-            >(
-              "compaction",
-              getMiddlewaresFor("compaction"),
-              (args) =>
-                defaultCompactionTerminal(
-                  args,
-                  buildPluginTurnContext(ctx, reqId),
-                ),
-              {
-                messages: ctx.messages,
-                signal: abortController.signal,
-                options: {
-                  lastCompactedAt: ctx.contextCompactedAt ?? undefined,
-                  force: true,
-                  minKeepRecentUserTurns: 0,
-                  targetInputTokensOverride: correctedTarget,
-                  overrideProfile: resolveCurrentOverrideProfile() ?? null,
-                },
-              },
-              buildPluginTurnContext(ctx, reqId),
-              DEFAULT_TIMEOUTS.compaction,
-            )) as Awaited<
-              ReturnType<typeof ctx.contextWindowManager.maybeCompact>
-            >;
-          } catch (err) {
-            if (err instanceof PluginTimeoutError) {
-              // Emergency compaction timed out. Record the circuit-breaker
-              // failure and fall through to the graceful-error path below
-              // (the unsuccessful-compaction fallback) rather than hard-
-              // failing the turn.
-              rlog.warn(
-                { err, phase: "emergency-compaction" },
-                "Emergency compaction pipeline timed out — continuing with overflow fallback",
-              );
-              await ctx.agentLoop.compactionCircuit.recordOutcome(
-                ctx,
-                true,
-                onEvent,
-              );
-              emergencyCompact = null;
-            } else {
-              throw err;
-            }
-          }
+          const emergencyCompact = await defaultCompact({
+            manager: ctx.contextWindowManager,
+            messages: ctx.messages,
+            signal: abortController.signal,
+            force: true,
+            minKeepRecentUserTurns: 0,
+            overrideProfile: resolveCurrentOverrideProfile() ?? null,
+          });
           // Only track when the summary LLM actually ran; `force: true`
-          // bypasses the cooldown but not the early-return paths.
-          if (
-            emergencyCompact &&
-            emergencyCompact.summaryFailed !== undefined
-          ) {
+          // bypasses the auto-threshold gate but not the early-return paths.
+          if (emergencyCompact.summaryFailed !== undefined) {
             await ctx.agentLoop.compactionCircuit.recordOutcome(
-              ctx,
               emergencyCompact.summaryFailed,
               onEvent,
             );
           }
-          if (emergencyCompact?.compacted) {
+          if (emergencyCompact.compacted) {
             await applySuccessfulCompaction(emergencyCompact, ctx.messages);
-            reducerCompacted = true;
-            shouldInjectWorkspace = true;
+            state.reducerCompacted = true;
           }
 
-          // Only re-inject NOW.md when ctx.messages was actually stripped;
-          // otherwise the existing block is still present.
+          // Only re-inject the memory-static block when ctx.messages was
+          // actually stripped; otherwise the existing block is still present.
+          // (The `<knowledge_base>`, NOW.md, and v2 static `<info>` blocks
+          // self-gate inside their injectors on whether they are already
+          // present in `ctx.messages`.)
           const injection = await applyRuntimeInjections(ctx.messages, {
             ...injectionOpts,
-            pkbContext: currentPkbContent,
-            memoryV2Static: convergenceStripped ? currentMemoryV2Static : null,
-            nowScratchpad: convergenceStripped ? currentNowContent : null,
-            workspaceTopLevelContext: shouldInjectWorkspace
-              ? ctx.workspaceTopLevelContext
-              : null,
-            slackChronologicalMessages: reducerCompacted
+            slackChronologicalMessages: state.reducerCompacted
               ? null
               : injectionOpts.slackChronologicalMessages,
             mode: currentInjectionMode,
@@ -2716,8 +2107,6 @@ export async function runAgentLoopImpl(
             );
             runMessages = fallbackStrip.messages;
           }
-          preRepairMessages = runMessages;
-          preRunHistoryLength = runMessages.length;
           state.contextTooLargeDetected = false;
 
           updatedHistory = await runAgentLoop(runMessages);
@@ -2771,44 +2160,11 @@ export async function runAgentLoopImpl(
       onEvent(buildConversationErrorMessage(ctx.conversationId, classified));
     }
 
-    // Reconcile synthesized cancellation tool_results
-    for (let i = preRunHistoryLength; i < updatedHistory.length; i++) {
-      const msg = updatedHistory[i];
-      if (msg.role === "user") {
-        for (const block of msg.content) {
-          if (
-            block.type === "tool_result" &&
-            !state.pendingToolResults.has(block.tool_use_id) &&
-            !state.persistedToolUseIds.has(block.tool_use_id)
-          ) {
-            state.pendingToolResults.set(block.tool_use_id, {
-              content: block.content,
-              isError: block.is_error ?? false,
-            });
-          }
-        }
-      }
-    }
-
-    // Flush remaining tool results
+    // Flush remaining tool results. On a normal turn these drain at the next
+    // `message_complete`; an aborted or yielded loop exits with them still
+    // buffered, so finalize the (possibly already on-arrival-reserved) grouped
+    // row here rather than writing a duplicate.
     if (state.pendingToolResults.size > 0) {
-      const toolResultBlocks = Array.from(
-        state.pendingToolResults.entries(),
-      ).map(([toolUseId, result]) => ({
-        type: "tool_result",
-        tool_use_id: toolUseId,
-        content: redactSecrets(result.content),
-        is_error: result.isError,
-        ...(result.contentBlocks
-          ? {
-              contentBlocks: result.contentBlocks.map((block) =>
-                block.type === "text"
-                  ? { ...block, text: redactSecrets(block.text) }
-                  : block,
-              ),
-            }
-          : {}),
-      }));
       const toolResultMetadata = {
         ...provenanceFromTrustContext(ctx.trustContext),
         userMessageChannel: capturedTurnChannelContext.userMessageChannel,
@@ -2818,21 +2174,12 @@ export async function runAgentLoopImpl(
         assistantMessageInterface:
           capturedTurnInterfaceContext.assistantMessageInterface,
       };
-      await runPipeline<PersistArgs, PersistResult>(
-        "persistence",
-        getMiddlewaresFor("persistence"),
-        defaultPersistenceTerminal,
-        {
-          op: "add",
-          conversationId: ctx.conversationId,
-          role: "user",
-          content: JSON.stringify(toolResultBlocks),
-          metadata: toolResultMetadata,
-        },
-        buildPluginTurnContext(ctx, reqId),
-        DEFAULT_TIMEOUTS.persistence,
+      await finalizePendingToolResultRow(
+        state,
+        ctx.conversationId,
+        toolResultMetadata,
+        rlog,
       );
-      state.pendingToolResults.clear();
     }
 
     // Persist the budget_yield_unrecovered notice now that any pending
@@ -2856,24 +2203,13 @@ export async function runAgentLoopImpl(
       };
       let yieldNoticePersistedId: string | null = null;
       try {
-        const yieldPersistResult = (await runPipeline<
-          PersistArgs,
-          PersistResult
-        >(
-          "persistence",
-          getMiddlewaresFor("persistence"),
-          defaultPersistenceTerminal,
-          {
-            op: "add",
-            conversationId: ctx.conversationId,
-            role: "assistant",
-            content: JSON.stringify(yieldNoticeMessage.content),
-            metadata: yieldNoticeMetadata,
-          },
-          buildPluginTurnContext(ctx, reqId),
-          DEFAULT_TIMEOUTS.persistence,
-        )) as PersistAddResult;
-        yieldNoticePersistedId = yieldPersistResult.message.id;
+        const yieldRow = await addMessage(
+          ctx.conversationId,
+          "assistant",
+          JSON.stringify(yieldNoticeMessage.content),
+          { metadata: yieldNoticeMetadata },
+        );
+        yieldNoticePersistedId = yieldRow.id;
       } catch (err) {
         // Non-fatal — a DB hiccup must not escalate a budget-yield exit into
         // a turn-level throw. The live SSE event was already emitted, so the
@@ -2929,7 +2265,7 @@ export async function runAgentLoopImpl(
     }
 
     // Reconstruct history
-    const newMessages = updatedHistory.slice(preRunHistoryLength).map((msg) => {
+    const newMessages = lastRunNewMessages.map((msg) => {
       if (msg.role !== "assistant") return msg;
       const { cleanedContent } = cleanAssistantContent(msg.content);
       const cleanedBlocks = cleanedContent as ContentBlock[];
@@ -2960,10 +2296,6 @@ export async function runAgentLoopImpl(
         state.assistantRowAwaitingFinalization &&
         state.lastAssistantMessageId
       ) {
-        // Direct `deleteMessageById` (not via the `persistence` pipeline):
-        // see the same rationale on the matching cleanup in
-        // `handleLlmCallStarted` — an unfinalized reservation has no
-        // observable history for plugins.
         try {
           deleteMessageById(state.lastAssistantMessageId);
         } catch (err) {
@@ -2985,20 +2317,12 @@ export async function runAgentLoopImpl(
       const errorAssistantMessage = createAssistantMessage(
         state.providerErrorUserMessage,
       );
-      const errorPersistResult = (await runPipeline<PersistArgs, PersistResult>(
-        "persistence",
-        getMiddlewaresFor("persistence"),
-        defaultPersistenceTerminal,
-        {
-          op: "add",
-          conversationId: ctx.conversationId,
-          role: "assistant",
-          content: JSON.stringify(errorAssistantMessage.content),
-          metadata: errChannelMeta,
-        },
-        buildPluginTurnContext(ctx, reqId),
-        DEFAULT_TIMEOUTS.persistence,
-      )) as PersistAddResult;
+      const errorRow = await addMessage(
+        ctx.conversationId,
+        "assistant",
+        JSON.stringify(errorAssistantMessage.content),
+        { metadata: errChannelMeta },
+      );
       persistedErrorAssistantMessage = true;
       // Repoint `lastAssistantMessageId` at the synthetic error row so the
       // post-loop sync, attachment resolution, and `message_complete`/
@@ -3007,7 +2331,7 @@ export async function runAgentLoopImpl(
       // above. Mark finalization complete so the next LLM call in this run
       // (or a downstream handler) doesn't try to clean up an id that
       // already corresponds to a finalized row.
-      state.lastAssistantMessageId = errorPersistResult.message.id;
+      state.lastAssistantMessageId = errorRow.id;
       state.assistantRowAwaitingFinalization = false;
       newMessages.push(errorAssistantMessage);
       // Pipe the just-assigned message id into any orphaned LLM request log
@@ -3021,10 +2345,7 @@ export async function runAgentLoopImpl(
       // other conversations cannot collide. Non-fatal — a DB hiccup must
       // not escalate a provider rejection into a turn-level throw.
       try {
-        backfillMessageIdOnLogs(
-          ctx.conversationId,
-          errorPersistResult.message.id,
-        );
+        backfillMessageIdOnLogs(ctx.conversationId, errorRow.id);
       } catch (err) {
         rlog.warn(
           { err },
@@ -3037,7 +2358,16 @@ export async function runAgentLoopImpl(
       // would create a duplicate plain-text bubble below the alert card.
     }
 
-    let restoredHistory = [...preRepairMessages, ...newMessages];
+    // Base persisted into `ctx.messages` is the loop's own returned history
+    // (minus the tail it appended this run), with the cleaned `newMessages`
+    // re-appended on top. Sourcing the base from the loop keeps it in lockstep
+    // with any in-loop compaction without the orchestrator maintaining a
+    // parallel snapshot across re-entry sites.
+    const loopBase = updatedHistory.slice(
+      0,
+      updatedHistory.length - lastRunNewMessages.length,
+    );
+    let restoredHistory = [...loopBase, ...newMessages];
 
     // Post-turn tool result truncation: save large results to disk and
     // replace in-context content with a prefix/suffix stub + file pointer.
@@ -3229,30 +2559,6 @@ export async function runAgentLoopImpl(
         publishLoopMessagesChanged();
       }
     }
-
-    // Second title pass: after 3 completed turns, re-generate the title
-    // using the last 3 messages for better context. Only fires when the
-    // current title was auto-generated (isAutoTitle = 1) and the user
-    // has not opted out via `conversations.skipAutoRetitling`.
-    if (ctx.turnCount === 2 && !getConfig().conversations.skipAutoRetitling) {
-      // turnCount is 0-indexed, incremented in finally; 2 = about to become 3rd turn
-      queueRegenerateConversationTitle({
-        conversationId: ctx.conversationId,
-        provider: ctx.provider,
-        onTitleUpdated: (title) => {
-          onEvent({
-            type: "conversation_title_updated",
-            conversationId: ctx.conversationId,
-            title,
-          });
-          onEvent({
-            type: "sync_changed",
-            tags: [conversationMetadataSyncTag(ctx.conversationId)],
-          });
-        },
-        signal: abortController.signal,
-      });
-    }
   } catch (err) {
     const errorCtx = {
       phase: "agent_loop" as const,
@@ -3312,8 +2618,6 @@ export async function runAgentLoopImpl(
     }
   } finally {
     if (turnStarted) {
-      cleanupBootstrapAfterTurnThreshold(ctx.conversationId);
-
       ctx.turnCount++;
       const config = getConfig();
       const maxWait = config.workspaceGit?.turnCommitMaxWaitMs ?? 4000;
@@ -3351,7 +2655,7 @@ export async function runAgentLoopImpl(
     ctx.profiler.emitSummary(ctx.traceEmitter, reqId);
 
     ctx.abortController = null;
-    ctx.processing = false;
+    ctx.setProcessing(false);
     ctx.onConfirmationOutcome = undefined;
     ctx.surfaceActionRequestIds.delete(ctx.currentRequestId ?? "");
     ctx.approvedViaPromptThisTurn = false;
@@ -3498,7 +2802,7 @@ export async function applyCompactionResult(
     result.summaryText,
     ctx.contextCompactedMessageCount,
   );
-  markHistoryStrippedBestEffort(ctx.conversationId, compactedAt, log);
+  markHistoryStrippedBestEffort(ctx.conversationId);
   if (options.slackContextCompactionWatermarkTs) {
     updateConversationSlackContextWatermark(
       ctx.conversationId,