Merge pull request #851 from corbitsdev/cl-7539-stop-compaction-spacer-text-from-becoming-a-model-echo-loop

Stop compaction spacer text from becoming a model echo loop

Author: TheGreatAxios

CHANGELOG.md

@@ -30,6 +30,10 @@ parallel copies under `docs/` or `scripts/notes/`. At cut time: rename
   impersonating Reject; Escape still denies.
 - After context compaction, ChatGPT Codex requests keep the operating prompt as
   instructions.
+- Compaction spacers stamp a reserved harness producer id and a visible
+  sentinel. Spacer-only replies are incomplete and stay on open-task and
+  workflow rails. Frozen-prefix matching ignores model-emitted copies of the
+  marker.
 
 ## [0.3.18] - 2026-09-08
 

docs/ARCHITECTURE.md

@@ -162,7 +162,7 @@ The agent maintains an optional **`manage_tasks`** list (create/update via the h
 
 #### Context compaction (the compaction governor)
 
-When a cycle's input tokens cross a threshold, the director compacts the inference-facing history (the full run is always retained in the context store). The threshold is **model-aware** — roughly 60% of the active model's real context window — so small-window models compact early enough to avoid provider context-overflow while large-window models do not compact prematurely. The compacted prefix is **append-only across passes**: the existing compacted user turn stays byte-identical; new folds become later summary turns with an assistant spacer between them so the prompt head can remain in the provider KV cache. The governor covers three cases:
+When a cycle's input tokens cross a threshold, the director compacts the inference-facing history (the full run is always retained in the context store). The threshold is **model-aware** — roughly 60% of the active model's real context window — so small-window models compact early enough to avoid provider context-overflow while large-window models do not compact prematurely. The compacted prefix is **append-only across passes**: the existing compacted user turn stays byte-identical; new folds become later summary turns with a harness-inserted assistant spacer (identified by reserved `model: "harness"`, plus a visible sentinel; persisted `[compaction]` tokens without a producer still freeze) between them so the prompt head can remain in the provider KV cache. Model-emitted copies of the spacer are not frozen. Spacer-only model replies are incomplete: ChatDirector nudges, then falls through loop-protection, workflow-idle, and open-task rails rather than empty-settling with work still open. The governor covers three cases:
 
 - **Threshold at a tool pause** — Once over threshold, the follow-up `infer` after a tool batch is swapped for a `compact` cycle, and inference resumes via a host continuation message. After a compact that remains over the high watermark, the governor uses **growth hysteresis** (wait for usage to grow by ~10% of the window) instead of re-arming on every cycle; dropping under 60% is not required.
 - **Idle (end-of-turn)** — An interactive turn can end with a reply and then sit idle with no tool batch to intercept; the governor requests a continuation at that pause and compacts when it arrives. An operator message that races the continuation still compacts first, then re-enters inference to answer it.

src/agent/compaction.test.ts

@@ -8,7 +8,12 @@ import type {
 } from "@intx/types/runtime";
 import { createCompactionGovernor } from "./compaction.js";
 import { compactionResumeDeltaFor, compactionThresholdFor } from "../provider/context-window.js";
-import { COMPACTOR_KEEP_RECENT_TURNS, compactorNoOpFloor } from "../session/compactor.js";
+import {
+  COMPACTOR_KEEP_RECENT_TURNS,
+  COMPACT_SPACER_TEXT,
+  LEGACY_COMPACT_SPACER_TEXT,
+  compactorNoOpFloor,
+} from "../session/compactor.js";
 
 const capabilities = {
   infer: (options?: unknown) => ({ type: "infer", ...(options !== undefined ? { options } : {}) }),
@@ -37,10 +42,30 @@ function turnsOfLength(count: number, textLength: number): ConversationTurn[] {
   })) as unknown as ConversationTurn[];
 }
 
-function inferenceDone(input: number): Extract<ReactorInboundEvent, { type: "inference.done" }> {
+function inferenceDone(
+  input: number,
+  text = "",
+): Extract<ReactorInboundEvent, { type: "inference.done" }> {
   return {
     type: "inference.done",
-    turn: { role: "assistant", content: [] },
+    turn: {
+      role: "assistant",
+      content: text.length > 0 ? [{ type: "text", text }] : [],
+    },
+    usage: usage(input),
+    source: { sourceId: "s", provider: "p", model: "m" },
+  } as unknown as Extract<ReactorInboundEvent, { type: "inference.done" }>;
+}
+
+function inferenceDoneWithTools(
+  input: number,
+): Extract<ReactorInboundEvent, { type: "inference.done" }> {
+  return {
+    type: "inference.done",
+    turn: {
+      role: "assistant",
+      content: [{ type: "tool_call", id: "c1", name: "read_file", arguments: { path: "a.ts" } }],
+    },
     usage: usage(input),
     source: { sourceId: "s", provider: "p", model: "m" },
   } as unknown as Extract<ReactorInboundEvent, { type: "inference.done" }>;
@@ -443,4 +468,52 @@ describe("compaction governor", () => {
     expect(actions).not.toBeNull();
     expect(actions?.some((a) => a.type === "compact")).toBe(true);
   });
+
+  test("consecutive threshold and idle compacts are bounded until occupancy", () => {
+    const governor = createCompactionGovernor(() => {});
+    const echo = LEGACY_COMPACT_SPACER_TEXT;
+    governor.noteInferenceDone(inferenceDone(overThreshold, echo), tenTurns);
+    expect(governor.interceptActions(toolDone(), inferAction, capabilities)).not.toBeNull();
+
+    governor.noteInferenceDone(inferenceDone(overThreshold, echo), tenTurns);
+    governor.noteInferenceDone(inferenceDone(overThreshold + resumeDelta, echo), tenTurns);
+    expect(governor.interceptActions(toolDone(), inferAction, capabilities)).not.toBeNull();
+
+    governor.noteInferenceDone(inferenceDone(overThreshold + resumeDelta, echo), tenTurns);
+    governor.noteInferenceDone(inferenceDone(overThreshold + 2 * resumeDelta, echo), tenTurns);
+    expect(governor.interceptActions(toolDone(), inferAction, capabilities)).toBeNull();
+    governor.noteIdleTurn(inferenceDone(overThreshold + 2 * resumeDelta, echo), [
+      { type: "reply", content: "done" },
+    ]);
+    expect(governor.interceptIdleContinuation(emptyMessage(), capabilities)).toBeNull();
+
+    governor.noteInferenceDone(
+      inferenceDone(overThreshold + 3 * resumeDelta, "real work"),
+      tenTurns,
+    );
+    expect(governor.interceptActions(toolDone(), inferAction, capabilities)).toBeNull();
+
+    governor.noteInferenceDone(inferenceDoneWithTools(overThreshold + 4 * resumeDelta), tenTurns);
+    expect(governor.interceptActions(toolDone(), inferAction, capabilities)).not.toBeNull();
+  });
+
+  test("spacer-echo terminal does not arm idle compact", () => {
+    let continuations = 0;
+    const governor = createCompactionGovernor(() => continuations++);
+    governor.noteInferenceDone(inferenceDone(overThreshold, LEGACY_COMPACT_SPACER_TEXT), tenTurns);
+    governor.noteIdleTurn(inferenceDone(overThreshold, LEGACY_COMPACT_SPACER_TEXT), [
+      { type: "reply", content: LEGACY_COMPACT_SPACER_TEXT },
+    ]);
+    expect(continuations).toBe(0);
+    governor.noteIdleTurn(inferenceDone(overThreshold, COMPACT_SPACER_TEXT), [
+      { type: "reply", content: COMPACT_SPACER_TEXT },
+    ]);
+    expect(continuations).toBe(0);
+    expect(governor.interceptIdleContinuation(emptyMessage(), capabilities)).toBeNull();
+
+    governor.noteIdleTurn(inferenceDone(overThreshold, "done"), [
+      { type: "reply", content: "done" },
+    ]);
+    expect(continuations).toBe(1);
+  });
 });

src/agent/compaction.ts

@@ -10,7 +10,12 @@ import {
   compactionThresholdFor,
   contextTokensFromUsage,
 } from "../provider/context-window.js";
-import { COMPACTOR_KEEP_RECENT_TURNS, compactorNoOpFloor } from "../session/compactor.js";
+import {
+  COMPACTOR_KEEP_RECENT_TURNS,
+  assistantTextIsCompactSpacerEcho,
+  compactorNoOpFloor,
+  isCompactSpacerEchoTurn,
+} from "../session/compactor.js";
 import { createContextEstimate, estimateOverheadTokens } from "./context-estimate.js";
 import { onTurnBoundary } from "./reactor-events.js";
 
@@ -22,6 +27,12 @@ const COMPACTOR_NAME = "pruning-compactor";
 // would spend a reactor cycle that shrinks nothing.
 const MIN_TURNS_TO_COMPACT = compactorNoOpFloor(COMPACTOR_KEEP_RECENT_TURNS);
 const MAX_OVERFLOW_RECOVERIES = 2;
+// Last-ditch bound on compact→infer→compact when the post-compact infer never
+// occupies the loop. Reset on tool-call occupancy or when a post-compact
+// measurement lands at or under the high watermark (that infer is not itself
+// a compact). Do not reset merely because assistant text ≠ spacer. Overflow
+// recoveries (above) reset on any successful inference.done instead.
+const MAX_CONSECUTIVE_THRESHOLD_COMPACTS = 2;
 
 // A compact action runs in its own reactor cycle, after which the reactor
 // idles until the next inbound event. Worker loops (sub-agents, the coding
@@ -45,6 +56,7 @@ export function createCompactionGovernor(
   // operator question to answer). Distinct from postCompactInfer.
   let postCompactMeter = false;
   let overflowRecoveries = 0;
+  let consecutiveThresholdCompacts = 0;
   // Set whenever the arming decision fell back to the local estimate because
   // the provider omitted usage or reported zero, so callers rendering a meter
   // can flag the number as approximate instead of implying provider-grade
@@ -90,11 +102,31 @@ export function createCompactionGovernor(
     awaitingPostCompactMeasurement = true;
   }
 
+  function atThresholdCompactCap(): boolean {
+    return consecutiveThresholdCompacts >= MAX_CONSECUTIVE_THRESHOLD_COMPACTS;
+  }
+
+  function issueThresholdCompact(): void {
+    consecutiveThresholdCompacts++;
+    noteCompactIssued();
+  }
+
+  function isSpacerEchoTerminal(event: ReactorInboundEvent, actions: ReactorAction[]): boolean {
+    // Fail-closed only. ChatDirector owns spacer-echo completeness (nudge, then
+    // loop-protection / workflow / open-task rails). This just refuses to treat
+    // that incomplete wait or reply as an idle-compact pause.
+    if (event.type === "inference.done" && isCompactSpacerEchoTurn(event.turn)) return true;
+    return actions.some((a) => a.type === "reply" && assistantTextIsCompactSpacerEcho(a.content));
+  }
+
   function noteInferenceDone(
     event: Extract<ReactorInboundEvent, { type: "inference.done" }>,
     turns: readonly ConversationTurn[],
   ): void {
     overflowRecoveries = 0;
+    if (event.turn.content.some((block) => block.type === "tool_call")) {
+      consecutiveThresholdCompacts = 0;
+    }
     if (requestContinuation === undefined) return;
     syncFromTurns(turns);
     lastModel = event.source?.model;
@@ -109,6 +141,7 @@ export function createCompactionGovernor(
     }
     if (contextTokens <= compactionThresholdFor(lastModel)) {
       tokensAtLastCompact = undefined;
+      consecutiveThresholdCompacts = 0;
     }
     // Assign, don't OR: an under-threshold follow-up must disarm a sticky
     // pending left from an earlier over-threshold turn (e.g. after the
@@ -138,9 +171,10 @@ export function createCompactionGovernor(
     if (event.type !== "tool.done") return null;
     if (!pending && !(usingEstimate && isOverThreshold(estimate.tokens))) return null;
     if (!actions.some((a) => a.type === "infer")) return null;
+    if (atThresholdCompactCap()) return null;
     pending = false;
     postCompactInfer = true;
-    noteCompactIssued();
+    issueThresholdCompact();
     requestContinuation?.();
     return [
       ...actions.filter((a) => a.type !== "infer"),
@@ -154,7 +188,9 @@ export function createCompactionGovernor(
   // compact when it (or the operator's next message) arrives.
   function noteIdleTurn(event: ReactorInboundEvent, actions: ReactorAction[]): void {
     if (!pending || idlePending || requestContinuation === undefined) return;
+    if (atThresholdCompactCap()) return;
     if (!onTurnBoundary(event)) return;
+    if (isSpacerEchoTerminal(event, actions)) return;
     const terminal =
       actions.some((a) => a.type === "reply" || a.type === "wait") &&
       !actions.some((a) => a.type === "infer" || a.type === "execute_tools");
@@ -168,6 +204,10 @@ export function createCompactionGovernor(
     capabilities: ReactorCapabilities,
   ): ReactorAction[] | null {
     if (!idlePending || event.type !== "message.received") return null;
+    if (atThresholdCompactCap()) {
+      idlePending = false;
+      return null;
+    }
     idlePending = false;
     pending = false;
     const content = typeof event.message.content === "string" ? event.message.content : "";
@@ -180,7 +220,7 @@ export function createCompactionGovernor(
     } else {
       postCompactMeter = true;
     }
-    noteCompactIssued();
+    issueThresholdCompact();
     requestContinuation?.();
     return [capabilities.compact(COMPACTOR_NAME, "context-threshold")];
   }

src/agent/director.ts

@@ -10,7 +10,11 @@ import type {
   ConversationTurn,
   RetryPolicy,
 } from "@intx/types/runtime";
-import { type SessionMetadata, type TaskBoundary } from "../session/compactor.js";
+import {
+  type SessionMetadata,
+  type TaskBoundary,
+  isCompactSpacerEchoTurn,
+} from "../session/compactor.js";
 import type { WorkflowCoordinator } from "../workflows/coordinator.js";
 import { createCompactionGovernor, type CompactionGovernor } from "./compaction.js";
 import { onTurnBoundary } from "./reactor-events.js";
@@ -72,11 +76,13 @@ function inferWithNudge(
 //
 // Assumes a terminal bare wait always means the turn is over. That holds for
 // every current wait path: DefaultDirector in conversational mode (the only
-// mode ChatDirector uses) yields one only on an empty model turn, and its halt
-// path already carries a reply; the compaction, workflow, and open-task
-// rewrites either keep those terminals or replace them with an infer.
-// A future wait that pauses mid-turn while expecting more work must not be
-// settled here.
+// mode ChatDirector uses) yields one only on an empty model turn; exhausted
+// spacer-echo incompleteness uses the same wait so loop-protection, workflow,
+// and open-task rails can rewrite it to infer first. This helper only settles
+// a leftover wait into an empty reply. The halt path already carries a reply;
+// compaction, workflow, and open-task rewrites either keep those terminals or
+// replace them with an infer. A future wait that pauses mid-turn while
+// expecting more work must not be settled here.
 function ensureCycleSettlesWithReply(
   actions: ReactorAction | ReactorAction[],
   capabilities: ReactorCapabilities,
@@ -98,6 +104,9 @@ function ensureCycleSettlesWithReply(
 const MAX_OPEN_TASK_NUDGES = 3;
 const MAX_DECLINED_OPEN_TASK_NUDGES = 2;
 const MAX_INFERENCE_RECOVERIES = 2;
+const MAX_SPACER_ECHO_NUDGES = 2;
+
+const SPACER_ECHO_NUDGE = "Continue the task. Do not repeat internal markers.";
 
 const IDLE_OPEN_TASK_NUDGE =
   "\n\nYou are ending your turn while tasks are still open (todo/doing). " +
@@ -394,6 +403,7 @@ class ChatDirectorImpl extends DefaultDirector {
   private idleTerminationNudges = 0;
   private declinedTerminationNudges = 0;
   private inferenceRecoveries = 0;
+  private spacerEchoNudges = 0;
   private lastInferenceTurnHadContent = false;
   private operatorJustResponded = false;
   private tasks: Task[] = [];
@@ -633,6 +643,7 @@ class ChatDirectorImpl extends DefaultDirector {
       this.idleTerminationNudges = 0;
       this.declinedTerminationNudges = 0;
       this.inferenceRecoveries = 0;
+      this.spacerEchoNudges = 0;
       this.toolOnlyStreak = 0;
       this.toolOnlyNudgeFired = false;
       this.pendingToolOnlyNudge = false;
@@ -682,9 +693,10 @@ class ChatDirectorImpl extends DefaultDirector {
     if (onTurnBoundary(event)) {
       this.turnCount++;
       const hasToolCalls = event.turn.content.some((b) => b.type === "tool_call");
-      const hasText = event.turn.content.some(
-        (b) => b.type === "text" && typeof b.text === "string" && b.text.length > 0,
-      );
+      const hasText =
+        event.turn.content.some(
+          (b) => b.type === "text" && typeof b.text === "string" && b.text.length > 0,
+        ) && !isCompactSpacerEchoTurn(event.turn);
       this.lastInferenceTurnHadContent = hasToolCalls || hasText;
 
       // toolOnlyStreak is narration-sensitive: any turn with text clears it
@@ -710,7 +722,12 @@ class ChatDirectorImpl extends DefaultDirector {
         if (hasToolCalls) {
           this.workflowIdleTurns = 0;
         } else {
-          this.workflowIdleTurns++;
+          // Echo-nudge cycles are incompleteness, not a contentful idle beat.
+          // Count them only after the echo budget is spent so the step-nudge
+          // rail still has its three turns before the stuck reply.
+          const spacerEchoStillNudging =
+            isCompactSpacerEchoTurn(event.turn) && this.spacerEchoNudges < MAX_SPACER_ECHO_NUDGES;
+          if (!spacerEchoStillNudging) this.workflowIdleTurns++;
         }
       }
       for (const block of event.turn.content) {
@@ -795,8 +812,20 @@ class ChatDirectorImpl extends DefaultDirector {
       this.compaction.noteInferenceDone(event, turns);
     }
 
+    if (event.type === "inference.done" && isCompactSpacerEchoTurn(event.turn)) {
+      if (this.spacerEchoNudges < MAX_SPACER_ECHO_NUDGES) {
+        this.spacerEchoNudges++;
+        return inferWithNudge(capabilities, SPACER_ECHO_NUDGE);
+      }
+    }
+
     const base = await super.decide(event, state, capabilities);
-    const baseActions = Array.isArray(base) ? base : [base];
+    let baseActions = Array.isArray(base) ? base : [base];
+    const spacerEchoExhausted =
+      event.type === "inference.done" && isCompactSpacerEchoTurn(event.turn);
+    if (spacerEchoExhausted) {
+      baseActions = baseActions.map((a) => (a.type === "reply" ? capabilities.wait() : a));
+    }
 
     this.compaction.noteIdleTurn(event, baseActions);
     const compacted = this.compaction.interceptActions(event, baseActions, capabilities);
@@ -816,12 +845,11 @@ class ChatDirectorImpl extends DefaultDirector {
 
     const coordinator = this.workflowCoordinator;
     if (coordinator?.isActive() && !coordinator.currentStepIsGate()) {
-      const actions = Array.isArray(base) ? base : [base];
-      const hasTerminal = actions.some((a) => a.type === "wait" || a.type === "reply");
-      if (hasTerminal && this.lastInferenceTurnHadContent) {
+      const hasTerminal = baseActions.some((a) => a.type === "wait" || a.type === "reply");
+      if (hasTerminal && (this.lastInferenceTurnHadContent || spacerEchoExhausted)) {
         if (this.operatorJustResponded) {
           this.operatorJustResponded = false;
-          return base;
+          return baseActions;
         }
         if (this.workflowIdleTurns >= 3) {
           if (hasActiveTasks(this.tasks)) this.logTerminationWithOpenTasks("workflow-idle-stall");
@@ -840,7 +868,7 @@ class ChatDirectorImpl extends DefaultDirector {
           `\n\nYou have not yet completed this workflow step. ` +
           `If this step is complete, ${stepClause}. ` +
           `Otherwise continue working with tools.`;
-        const passThrough = actions.filter(
+        const passThrough = baseActions.filter(
           (a): a is Exclude<ReactorAction, { type: "wait" } | { type: "reply" }> =>
             a.type !== "wait" && a.type !== "reply",
         );
@@ -872,7 +900,7 @@ class ChatDirectorImpl extends DefaultDirector {
       }
     }
 
-    return base;
+    return baseActions;
   }
 }