fix(core): committed compaction attempts absorb the turn's pending input (#87)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

packages/core/src/engine/context-engine.ts

@@ -105,6 +105,20 @@ export interface CompactionSettings {
 interface CompactionResult {
   status: StopReason;
   summary?: OmniMessage;
+  /**
+   * Whether at least one summarize attempt was **committed** by AgentHub (only a `completed`
+   * attempt commits — timeout/malformed end an incomplete stream and failed/auth/aborted throw
+   * or cut off before a clean end). The carry rule at every caller is a two-case binary on
+   * this flag (issue #85): committed → the input the caller folded in (mid-Task tool outputs,
+   * or the carry-over a manual `compact()` folds in) now lives in the old LLM object's history
+   * and must never be resent — strict providers reject the duplicates as stale tool_results;
+   * not committed → the folded input is untouched and is resent exactly as before. When
+   * nothing was folded in (idle/boundary compaction), the committed branch is vacuous —
+   * dropping zero outputs, clearing an empty carry — so no separate "was anything absorbed"
+   * signal is needed. Zero committed attempts also implies zero synthesized repairs (repairs
+   * only answer a committed rejection's tool calls).
+   */
+  committed: boolean;
 }
 
 /**
@@ -556,6 +570,9 @@ export class ContextEngine {
       // applies mid-Task — when runTurn returns, all of this turn's
       // tool results are ready and paired with their tool_call.
       const midTask = turn.toolOutputs.length > 0;
+      // Outputs this turn still owes the model: dropped when a committed compaction attempt
+      // consumes them into history (the two-case carry rule below, issue #85).
+      let turnOutputs = turn.toolOutputs;
       const compactionReason = this.compactionTrigger();
       if (compactionReason) {
         const mode = this.deps.compaction!.mode;
@@ -574,11 +591,20 @@ export class ContextEngine {
             signal,
           );
           if (result.status === "aborted") {
-            // User interrupted compaction: keep the original context; if mid-Task, hold the
+            // User interrupted compaction: keep the original context. The carry rule is the
-            // tool outputs as carry-over per case A. Abort is the one path that discards the
+            // same two-case binary as everywhere (issue #85): a committed attempt consumed
-            // steering queue (run's finally — control goes back to the user).
+            // this turn's outputs into history — only the repair stash summarizeContext left
+            // in pendingCarryOver still needs resending; otherwise the outputs are untouched
+            // and are appended behind the stash as case-A carry-over. Abort is the one path
+            // that discards the steering queue (run's finally — control goes back to the
+            // user).
             if (midTask) {
-              this.pendingCarryOver = this.buildCarryOver(attemptInput, turn);
+              if (!result.committed) {
+                this.pendingCarryOver = [
+                  ...this.pendingCarryOver,
+                  ...this.buildCarryOver(attemptInput, turn),
+                ];
+              }
               yield* this.emitAbort("aborted during compaction");
             }
             return;
@@ -605,7 +631,13 @@ export class ContextEngine {
             continue;
           }
           // failed: keep the original context and Trace index; the current Task continues and
-          // retries on the next trigger (no fallback to discard).
+          // retries on the next trigger (no fallback to discard). The carry rule (issue #85):
+          if (result.committed) {
+            // A committed attempt consumed this turn's outputs into history — drop them from
+            // the continuation; only the repair stash remains pending.
+            turnOutputs = [];
+          }
+          // else: nothing committed — the outputs are untouched and resent below as always.
         }
       }
 
@@ -615,9 +647,18 @@ export class ContextEngine {
       // ending the Task — subject to the max-turns guard at the top of the loop).
       const steering = yield* this.deliverSteering();
       // No tool_call this turn and no steering left -> the Task ends (the final reply has
-      // already been streamed out).
+      // already been streamed out). A compaction stash, if any, rides the next run.
       if (!midTask && steering.length === 0) return;
-      nextInput = [...turn.toolOutputs, ...steering];
+      // Anything a failed compaction stashed mid-run (synthesized repair outputs from
+      // rejected attempts) rides the very next request, ahead of the turn outputs so
+      // tool_results stay contiguous and first.
+      const stashed = this.pendingCarryOver;
+      this.pendingCarryOver = [];
+      nextInput = [...stashed, ...turnOutputs, ...steering];
+      // Mid-task, but a committed compaction consumed the outputs and nothing else remains
+      // to send: the run ends here — the failure was surfaced via compaction_end(failed),
+      // the context is intact, and the next prompt continues from the committed state.
+      if (nextInput.length === 0) return;
     }
   }
 
@@ -670,15 +711,27 @@ export class ContextEngine {
       yield* this.discardContext("manual");
       return;
     }
-    const result = yield* this.summarizeContext(
+    // The carry seam is a clean binary on whether the compaction committed anything to
-      "manual",
+    // AgentHub (PR #87 review):
-      this.pendingCarryOver.map(downgradeCarriedGoalInput),
+    //   - nothing committed (every attempt timeout/malformed/failed/auth/aborted): the fold
-      opts?.signal,
+    //     never reached the model context — restore the prior carry-over **verbatim**. Zero
-    );
+    //     committed attempts also means zero synthesized repairs, so there is no stash to
-    if (result.status === "completed") {
+    //     interleave with (pinned by tests);
+    //   - something committed: the carry-over is **consumed** — it lives in the committed
+    //     history now and must never be resent; only the repair stash (unanswered tool_call
+    //     pairing left by a final rejection, already in pendingCarryOver) remains pending.
+    // Dead-goal rounds are downgraded on the drained snapshot (goal mode's consumer-site
+    // rule): a no-commit restore keeps the downgraded copies — the downgrade is idempotent
+    // and every consumer applies it anyway, while non-goal messages keep their identity.
+    const folded = this.pendingCarryOver.map(downgradeCarriedGoalInput);
     this.pendingCarryOver = [];
+    const result = yield* this.summarizeContext("manual", folded, opts?.signal);
+    if (result.status === "completed") {
       this.pendingSummary = result.summary!;
+    } else if (!result.committed) {
+      this.pendingCarryOver = folded;
     }
+    // committed but not completed: the carry-over is deliberately not restored.
   }
 
   /**
@@ -1078,7 +1131,10 @@ export class ContextEngine {
    * then the compaction fails. timeout/malformed reconnect via the existing retry mechanism
    * under the compaction-specific cap (`compactionMaxReconnects`, tighter than the turn loop's
    * ladder), collapsing to failed once retries are exhausted; on failure/abort, the original
-   * context and Trace index are kept — it does not fall back to discard.
+   * context and Trace index are kept — it does not fall back to discard. The first **committed**
+   * attempt absorbs `pendingToolOutputs` into the old context's history (issue #85): later
+   * resends carry only the repairs and the Prompt, and the result's `committed` flag tells
+   * the caller the folded input must not be resent even though the compaction did not complete.
    * Docs: /docs/agent-loop § "Compaction".
    */
   private async *summarizeContext(
@@ -1095,10 +1151,18 @@ export class ContextEngine {
     // executed and aren't recorded again, while carry-over's not-yet-written synthetic content
     // (flatten text, backfilled placeholders) and the compaction Prompt are written now.
     const prompt = userText(settings.prompt);
-    const baseInput = [...pendingToolOutputs, prompt];
+    // The resend base: shrinks to the Prompt alone once an attempt commits — the folded turn
-    let input = baseInput;
+    // input then lives in the old LLM object's history, and resending it would make strict
+    // providers reject the request over duplicate/stale tool_results (issue #85).
+    let base = [...pendingToolOutputs, prompt];
+    let input = base;
     await this.write(prompt);
 
+    // Whether any attempt was committed by AgentHub (only `completed` commits: timeout and
+    // malformed end an incomplete stream, and failed/auth/aborted throw or cut off before a
+    // clean end — none of those reach the stateful commit). Returned as `committed`: the
+    // callers' two-case carry rule branches on it.
+    let committed = false;
     // Synthesized outputs answering the latest rejected attempt's tool calls, not yet carried
     // by a committed request: prepended to the retry input, and stashed as carry-over should
     // the compaction be abandoned first (see stashRepairs).
@@ -1114,13 +1178,16 @@ export class ContextEngine {
       if (signal?.aborted) {
         this.stashRepairs(pendingRepairs);
         yield* this.emitCompactionEnd(reason, "summarize", "aborted");
-        return { status: "aborted" };
+        return { status: "aborted", committed };
       }
       const attempt = await this.runCompactionRequest(input, signal, reconnects);
       if (attempt.status === "completed") {
         // The attempt was committed by AgentHub, so whatever its input carried — including
         // repairs synthesized for a previous rejection — is now in history and must not be
-        // resent.
+        // resent. The first commit absorbs the folded turn input: the base shrinks to the
+        // Prompt alone.
+        committed = true;
+        base = [prompt];
         pendingRepairs = [];
         // A completed response counts as a compaction success only when it is a **usable
         // summary**: the extracted text is non-empty and the response called no tool. The
@@ -1143,7 +1210,7 @@ export class ContextEngine {
           const summary = userText(buildContextSummaryText(summaryText));
           yield* this.emitCompactionEnd(reason, "summarize", "completed");
           await this.startNewContext();
-          return { status: "completed", summary };
+          return { status: "completed", summary, committed };
         }
         // Rejected. Tool calls were never dispatched, yet the assistant turn holding them IS
         // committed on the live LLM object — leaving them unanswered would get every
@@ -1163,13 +1230,14 @@ export class ContextEngine {
           }),
         );
         for (const repair of pendingRepairs) await this.write(repair);
-        // Rebuild from baseInput rather than appending: everything the rejected attempt's
+        // Rebuild from the (shrunken) base rather than appending: everything the rejected
-        // input carried is committed, so only the fresh repairs and the Prompt go out again.
+        // attempt's input carried is committed, so only the fresh repairs and the Prompt go
-        input = pendingRepairs.length > 0 ? [...pendingRepairs, ...baseInput] : baseInput;
+        // out again.
+        input = pendingRepairs.length > 0 ? [...pendingRepairs, ...base] : base;
         if (rejections >= MAX_SUMMARY_REJECTIONS) {
           this.stashRepairs(pendingRepairs);
           yield* this.emitCompactionEnd(reason, "summarize", "failed");
-          return { status: "failed" };
+          return { status: "failed", committed };
         }
         // A rejection is model behavior, not a transport failure: the request pipeline is
         // healthy, so the repaired input is resent immediately — no backoff and no
@@ -1181,7 +1249,7 @@ export class ContextEngine {
       if (attempt.status === "aborted") {
         this.stashRepairs(pendingRepairs);
         yield* this.emitCompactionEnd(reason, "summarize", "aborted");
-        return { status: "aborted" };
+        return { status: "aborted", committed };
       }
       if (attempt.status === "failed" || attempt.status === "auth") {
         // `auth` folds into `failed` here: the compaction event pair keeps its
@@ -1190,7 +1258,7 @@ export class ContextEngine {
         // request will surface it; the compaction request_end is Trace-only).
         this.stashRepairs(pendingRepairs);
         yield* this.emitCompactionEnd(reason, "summarize", "failed");
-        return { status: "failed" };
+        return { status: "failed", committed };
       }
       // timeout / malformed: retried via reconnect — transport-level, never committed by
       // AgentHub (case B), so the input (any pending repairs included) is resent unchanged.
@@ -1200,14 +1268,14 @@ export class ContextEngine {
       if (reconnects >= this.compactionMaxReconnects) {
         this.stashRepairs(pendingRepairs);
         yield* this.emitCompactionEnd(reason, "summarize", "failed");
-        return { status: "failed" };
+        return { status: "failed", committed };
       }
       reconnects += 1;
       const ok = await this.backoff(reconnects, signal);
       if (!ok) {
         this.stashRepairs(pendingRepairs);
         yield* this.emitCompactionEnd(reason, "summarize", "aborted");
-        return { status: "aborted" };
+        return { status: "aborted", committed };
       }
     }
   }

packages/core/test/compaction.test.ts

@@ -819,6 +819,209 @@ describe("context compaction", () => {
     expect(closureRepairs).toEqual([]);
   });
 
+  it("mid-task: a committed rejection absorbs the turn's tool outputs — retries never resend them, and an empty-handed abandonment ends the run", async () => {
+    // Issue #85, strict-provider scenario: the first committed attempt puts the turn's tool
+    // outputs (folded into the compaction input) into the old LLM object's history. From then
+    // on the retries must carry only the repairs and the Prompt — resending the outputs would
+    // be rejected as stale tool_results — and after the compaction is abandoned the
+    // continuation must not resend them either. Here the final rejection is empty (no repairs
+    // left to deliver) and no steering is queued, so the run ends at the failure: the context
+    // is intact and the next prompt continues from the committed state.
+    const llm1 = new ScriptedLLM(
+      [
+        // Turn 1: a tool call, over the threshold -> mid-task compaction.
+        {
+          messages: [toolCall({ name: "t", arguments: "{}", toolCallId: "ct" }), usage(150, 150)],
+        },
+        // Attempt 1: committed but rejected (tool call) -> absorbs [output(ct), prompt].
+        {
+          messages: [
+            assistantText("[summary]nope[/summary]"),
+            toolCall({ name: "t", arguments: "{}", toolCallId: "c1" }),
+          ],
+        },
+        // Attempts 2-5: committed but empty -> rejected; the base has shrunk to the Prompt.
+        { messages: [thinkingMessage("blank 2")] },
+        { messages: [thinkingMessage("blank 3")] },
+        { messages: [thinkingMessage("blank 4")] },
+        { messages: [thinkingMessage("blank 5")] },
+        // Next run after the abandonment: served by the same instance, plain prompt.
+        { messages: [assistantText("continuing"), usage(60, 900)] },
+      ],
+      "llm1",
+    );
+    let created = 0;
+    const trace = new Writer({ tracesDir: traces, sessionId: "sess_absorb" });
+    const engine = new ContextEngine({
+      llm: llm1,
+      environment: fakeEnvironment,
+      trace,
+      sessionMeta: metaMessage,
+      compaction: settings(),
+      createLLM: () => {
+        created += 1;
+        return new ScriptedLLM([], "llm2");
+      },
+      reconnectBackoffMs: 1,
+    });
+    const oldPath = trace.currentPath();
+
+    const out1 = await collect(engine.run([userText("task one")], { approve: allowAll }));
+
+    expect(compactionEvents(out1)[1]).toMatchObject({ type: "compaction_end", status: "failed" });
+    // The run ends at the empty-handed abandonment without an abort: the compaction failure
+    // is the whole story.
+    expect(payloadTypes(out1)).not.toContain("abort");
+    expect(llm1.calls).toHaveLength(6);
+    // Attempt 1 folds the turn's tool output in; attempt 2 carries the repair + Prompt but
+    // NOT the absorbed output; attempts 3-5 are Prompt-only.
+    expect(payloadTypes(llm1.calls[1]!)).toEqual(["tool_call_output", "text"]);
+    expect((llm1.calls[1]![0]!.payload as { tool_call_id?: string }).tool_call_id).toBe("ct");
+    expect(payloadTypes(llm1.calls[2]!)).toEqual(["tool_call_output", "text"]);
+    expect((llm1.calls[2]![0]!.payload as { tool_call_id?: string }).tool_call_id).toBe("c1");
+    expect(textOf(llm1.calls[2]![1]!)).toBe("COMPACT NOW");
+    for (let attempt = 3; attempt <= 5; attempt += 1) {
+      expect(llm1.calls[attempt]!.map(textOf)).toEqual(["COMPACT NOW"]);
+    }
+    // Original context kept: no LLM swap, no Trace rotation.
+    expect(created).toBe(0);
+    expect(trace.currentPath()).toBe(oldPath);
+
+    // The next run continues from the committed state: the absorbed outputs are not resent
+    // and nothing was stashed (the final rejection was empty).
+    await collect(engine.run([userText("task two")], { approve: allowAll }));
+    expect(llm1.calls[6]!.map(textOf)).toEqual(["task two"]);
+  });
+
+  it("mid-task: abandonment with a trailing tool-calling rejection continues the task with the repair alone", async () => {
+    // Any committed attempt absorbs — here the FIRST (empty) rejection commits the turn
+    // outputs, and the FINAL rejection leaves unanswered tool calls. The continuation after
+    // the failure delivers exactly the stashed repairs (never the absorbed outputs), keeping
+    // the live object's history well-formed while the task keeps running.
+    const llm1 = new ScriptedLLM(
+      [
+        {
+          messages: [toolCall({ name: "t", arguments: "{}", toolCallId: "ct" }), usage(150, 150)],
+        },
+        // Attempts 1-4: empty rejections (attempt 1 commits and absorbs the outputs).
+        { messages: [thinkingMessage("blank 1")] },
+        { messages: [thinkingMessage("blank 2")] },
+        { messages: [thinkingMessage("blank 3")] },
+        { messages: [thinkingMessage("blank 4")] },
+        // Attempt 5: rejected with a tool call -> its repair is stashed at the abandonment.
+        { messages: [toolCall({ name: "t", arguments: "{}", toolCallId: "c5" })] },
+        // Continuation: the repair alone; the model wraps the task up (under the threshold).
+        { messages: [assistantText("recovered"), usage(60, 900)] },
+      ],
+      "llm1",
+    );
+    const engine = new ContextEngine({
+      llm: llm1,
+      environment: fakeEnvironment,
+      compaction: settings(),
+      createLLM: () => new ScriptedLLM([], "llm2"),
+      reconnectBackoffMs: 1,
+    });
+
+    const out = await collect(engine.run([userText("go")], { approve: allowAll }));
+    expect(compactionEvents(out)[1]).toMatchObject({ type: "compaction_end", status: "failed" });
+    expect(llm1.calls).toHaveLength(7);
+    // Attempt 1 folds the outputs; attempts 2-5 are Prompt-only (absorbed by the first commit).
+    expect(payloadTypes(llm1.calls[1]!)).toEqual(["tool_call_output", "text"]);
+    for (let attempt = 2; attempt <= 5; attempt += 1) {
+      expect(llm1.calls[attempt]!.map(textOf)).toEqual(["COMPACT NOW"]);
+    }
+    // The continuation input is the repair answering c5 — nothing else: no absorbed outputs,
+    // no prompt.
+    expect(payloadTypes(llm1.calls[6]!)).toEqual(["tool_call_output"]);
+    const cont = llm1.calls[6]![0]!.payload as { tool_call_id: string; output: string };
+    expect(cont.tool_call_id).toBe("c5");
+    expect(cont.output).toBe(
+      "[tool error] the compaction request expects a summary, not tool calls",
+    );
+    // The task finished on the continuation; the stash is spent for later runs.
+    expect(out.map((m) => (m.payload as { text?: string }).text)).toContain("recovered");
+    await collect(engine.run([userText("again")], { approve: allowAll }));
+    expect(llm1.calls[7]!.map(textOf)).toEqual(["again"]);
+  });
+
+  it("mid-task: an all-transport abandonment absorbs nothing — the outputs are resent as before", async () => {
+    // Counter-case: timeout/malformed attempts never commit, so the turn's tool outputs were
+    // never absorbed and the post-failure continuation must resend them exactly as before
+    // (their tool_use pairing is still unanswered on the live object).
+    const llm1 = new ScriptedLLM(
+      [
+        {
+          messages: [toolCall({ name: "t", arguments: "{}", toolCallId: "ct" }), usage(150, 150)],
+        },
+        { messages: [], outcome: { status: "timeout" } },
+        { messages: [], outcome: { status: "timeout" } },
+        // Continuation: the resent tool output; the model wraps up under the threshold.
+        { messages: [assistantText("done on old context"), usage(60, 400)] },
+      ],
+      "llm1",
+    );
+    const engine = new ContextEngine({
+      llm: llm1,
+      environment: fakeEnvironment,
+      compaction: settings(),
+      createLLM: () => new ScriptedLLM([], "llm2"),
+      compactionMaxReconnects: 1,
+      reconnectBackoffMs: 1,
+    });
+
+    const out = await collect(engine.run([userText("go")], { approve: allowAll }));
+    expect(compactionEvents(out)[1]).toMatchObject({ type: "compaction_end", status: "failed" });
+    expect(llm1.calls).toHaveLength(4);
+    expect(payloadTypes(llm1.calls[3]!)).toEqual(["tool_call_output"]);
+    expect((llm1.calls[3]![0]!.payload as { tool_call_id?: string }).tool_call_id).toBe("ct");
+  });
+
+  it("mid-task: an abort after a committed rejection merges the stash — repairs ride, absorbed outputs do not", async () => {
+    // The abort-in-window case (issue #85): the stash must merge, not be overwritten. After a
+    // committed rejection the turn outputs are absorbed, so the case-A carry-over is skipped
+    // entirely and only the unanswered repair rides the next run.
+    const llm1 = new ScriptedLLM(
+      [
+        {
+          messages: [toolCall({ name: "t", arguments: "{}", toolCallId: "ct" }), usage(150, 150)],
+        },
+        // Attempt 1: committed rejection with a tool call -> absorbs outputs, repair pending.
+        { messages: [toolCall({ name: "t", arguments: "{}", toolCallId: "c1" })] },
+        // Attempt 2 (carrying the repair): the user aborts mid-request -> uncommitted, the
+        // repair is still unanswered and gets stashed.
+        { messages: [], outcome: { status: "aborted" } },
+        // Next run: the stash leads, then the new prompt.
+        { messages: [assistantText("back"), usage(60, 500)] },
+      ],
+      "llm1",
+    );
+    const engine = new ContextEngine({
+      llm: llm1,
+      environment: fakeEnvironment,
+      compaction: settings(),
+      createLLM: () => new ScriptedLLM([], "llm2"),
+      reconnectBackoffMs: 1,
+    });
+
+    const out = await collect(engine.run([userText("go")], { approve: allowAll }));
+    const events = compactionEvents(out);
+    expect(events[1]).toMatchObject({ type: "compaction_end", status: "aborted" });
+    expect(payloadTypes(out)).toContain("abort");
+    expect(llm1.calls).toHaveLength(3);
+    // Attempt 2 carried the repair + Prompt (not the absorbed outputs).
+    expect(payloadTypes(llm1.calls[2]!)).toEqual(["tool_call_output", "text"]);
+    expect((llm1.calls[2]![0]!.payload as { tool_call_id?: string }).tool_call_id).toBe("c1");
+
+    // The next run leads with the still-unanswered repair; the absorbed turn outputs are
+    // NOT re-held as carry-over (no duplicate ct output).
+    await collect(engine.run([userText("next")], { approve: allowAll }));
+    const nextRun = llm1.calls[3]!;
+    expect(payloadTypes(nextRun)).toEqual(["tool_call_output", "text"]);
+    expect((nextRun[0]!.payload as { tool_call_id?: string }).tool_call_id).toBe("c1");
+    expect(textOf(nextRun[1]!)).toBe("next");
+  });
+
   it("the compaction request carries exactly the same tools as ordinary requests (prompt-cache pin)", async () => {
     // Owner constraint (#84): compaction runs exactly when the context is at its largest, and
     // the provider's prompt cache only holds if the request prefix — the tool list included —
@@ -1048,6 +1251,104 @@ describe("context compaction", () => {
     ]);
   });
 
+  it("manual compaction that commits nothing restores the prior carry-over verbatim", async () => {
+    // The carry seam's binary (PR #87 review): every attempt was a transport failure, so
+    // nothing reached AgentHub — the folded carry-over comes back exactly as it was (the very
+    // same message objects, not routed through any absorption logic), and with zero committed
+    // attempts there are no repairs to interleave with.
+    const llm1 = new ScriptedLLM(
+      [
+        { messages: [assistantText("hi"), usage(10, 10)] },
+        // Manual compaction attempts: transport failures only — nothing committed.
+        { messages: [], outcome: { status: "timeout" } },
+        { messages: [], outcome: { status: "timeout" } },
+        // Run 3: the restored carry-over leads the input, exactly as before the compact().
+        { messages: [assistantText("resumed"), usage(20, 40)] },
+      ],
+      "llm1",
+    );
+    const engine = new ContextEngine({
+      llm: llm1,
+      environment: fakeEnvironment,
+      compaction: settings(),
+      createLLM: () => new ScriptedLLM([], "llm2"),
+      compactionMaxReconnects: 1,
+      reconnectBackoffMs: 1,
+    });
+    await collect(engine.run([userText("start")], { approve: allowAll }));
+    // An aborted-before-issue run leaves its input as carry-over.
+    const aborted = new AbortController();
+    aborted.abort();
+    const pendingMsg = userText("pending question");
+    await collect(engine.run([pendingMsg], { signal: aborted.signal, approve: allowAll }));
+    expect(llm1.calls).toHaveLength(1); // the aborted run never reached the LLM
+
+    const out = await collect(engine.compact());
+    expect(compactionEvents(out)[1]).toMatchObject({ type: "compaction_end", status: "failed" });
+    // The compaction folded the carry-over into its (uncommitted) attempts...
+    expect(llm1.calls).toHaveLength(3);
+    expect(llm1.calls[1]!.map(textOf)).toEqual(["pending question", "COMPACT NOW"]);
+    expect(llm1.calls[2]!.map(textOf)).toEqual(["pending question", "COMPACT NOW"]);
+
+    // ...and restored it verbatim: the next run resends the very same message object first.
+    await collect(engine.run([userText("run3")], { approve: allowAll }));
+    expect(llm1.calls[3]![0]).toBe(pendingMsg);
+    expect(llm1.calls[3]!.map(textOf)).toEqual(["pending question", "run3"]);
+  });
+
+  it("manual compaction with a committed attempt consumes the carry-over; only repairs remain pending", async () => {
+    // The other side of the binary: the first committed attempt put the folded carry-over
+    // into AgentHub history, so it is disposed of at the seam — never restored — and the next
+    // input reflects the committed state plus the repair stash left by the final rejection.
+    const llm1 = new ScriptedLLM(
+      [
+        { messages: [assistantText("hi"), usage(10, 10)] },
+        // Attempt 1: committed but empty -> absorbs the folded carry-over into history.
+        { messages: [thinkingMessage("blank 1")] },
+        // Attempts 2-4: still empty; the base has shrunk to the Prompt.
+        { messages: [thinkingMessage("blank 2")] },
+        { messages: [thinkingMessage("blank 3")] },
+        { messages: [thinkingMessage("blank 4")] },
+        // Attempt 5: rejected with a tool call -> its repair is stashed at the abandonment.
+        { messages: [toolCall({ name: "t", arguments: "{}", toolCallId: "c9" })] },
+        // Run 3: the repair leads; the consumed carry-over is NOT resent.
+        { messages: [assistantText("resumed"), usage(20, 40)] },
+      ],
+      "llm1",
+    );
+    const engine = new ContextEngine({
+      llm: llm1,
+      environment: fakeEnvironment,
+      compaction: settings(),
+      createLLM: () => new ScriptedLLM([], "llm2"),
+      reconnectBackoffMs: 1,
+    });
+    await collect(engine.run([userText("start")], { approve: allowAll }));
+    const aborted = new AbortController();
+    aborted.abort();
+    await collect(
+      engine.run([userText("pending question")], { signal: aborted.signal, approve: allowAll }),
+    );
+
+    const out = await collect(engine.compact());
+    expect(compactionEvents(out)[1]).toMatchObject({ type: "compaction_end", status: "failed" });
+    expect(llm1.calls).toHaveLength(6);
+    // Attempt 1 folded the carry-over in (and committed it); later attempts resend the Prompt
+    // alone — the absorbed carry-over never again.
+    expect(llm1.calls[1]!.map(textOf)).toEqual(["pending question", "COMPACT NOW"]);
+    for (let attempt = 2; attempt <= 5; attempt += 1) {
+      expect(llm1.calls[attempt]!.map(textOf)).toEqual(["COMPACT NOW"]);
+    }
+
+    // The carry-over is consumed at the seam — the next run leads with the stashed repair
+    // alone, continuing from committed history.
+    await collect(engine.run([userText("run3")], { approve: allowAll }));
+    const run3 = llm1.calls[6]!;
+    expect(payloadTypes(run3)).toEqual(["tool_call_output", "text"]);
+    expect((run3[0]!.payload as { tool_call_id?: string }).tool_call_id).toBe("c9");
+    expect(textOf(run3[1]!)).toBe("run3");
+  });
+
   it("no compaction capability (createLLM missing) means thresholds never fire and compact() is a no-op", async () => {
     const llm1 = new ScriptedLLM(
       [{ messages: [assistantText("big"), usage(999999, 999999)] }],

packages/core/test/replay.test.ts

@@ -566,4 +566,98 @@ describe("resumeTrace regressions (PR #39 review)", () => {
     expect(result.sessionTurns).toBe(1);
     expect(result.carryOver).toEqual([]);
   });
+
+  it("absorbed mid-task compaction (committed rejections) replays with nothing left to resend", () => {
+    // Resume symmetry for issue #85: a mid-task compaction whose first attempt committed has
+    // absorbed the turn's tool output into history; the repair answering the rejected
+    // attempt's tool call rode the second committed attempt. After a process exit at the
+    // abandonment, replay must reconstruct exactly the live engine's state: everything in
+    // history, nothing in carry-over — the next request continues from the committed state
+    // without resending any tool_result.
+    const result = resumeTrace([
+      meta(),
+      userText("go"),
+      requestBegin(),
+      toolCall({ name: "exec_command", arguments: "{}", toolCallId: "ct" }),
+      requestEnd("completed"),
+      tokenUsage(usage(150), usage(150)),
+      toolCallOutput({ output: "tool ran", toolCallId: "ct" }),
+      compactionBegin({ reason: "context", mode: "summarize", context: 150, turns: 1 }),
+      userText("COMPACT NOW"),
+      requestBegin(),
+      assistantText("[summary]nope[/summary]"),
+      toolCall({ name: "t", arguments: "{}", toolCallId: "c1" }),
+      requestEnd("completed"),
+      tokenUsage(usage(160), usage(310)),
+      toolCallOutput({
+        output: "[tool error] the compaction request expects a summary, not tool calls",
+        toolCallId: "c1",
+        stopReason: "failed",
+      }),
+      requestBegin(),
+      thinkingMessage("still nothing"),
+      requestEnd("completed"),
+      tokenUsage(usage(170), usage(480)),
+      compactionEnd({ reason: "context", mode: "summarize", status: "failed" }),
+    ]);
+    expect(result.contextClosed).toBe(false);
+    expect(result.pendingSummary).toBeUndefined();
+    // Both committed compaction rounds enter history: the turn output was absorbed by the
+    // first, the repair by the second.
+    expect(result.history.map((m) => (m.payload as { type?: string }).type)).toEqual([
+      "text",
+      "tool_call",
+      "tool_call_output",
+      "text",
+      "text",
+      "tool_call",
+      "tool_call_output",
+      "thinking",
+    ]);
+    // Nothing left to resend — matching the live engine, which also holds no carry-over here.
+    expect(result.carryOver).toEqual([]);
+    expect(result.sessionTurns).toBe(1);
+  });
+
+  it("absorbed mid-task compaction abandoned with an unanswered repair replays it as carry-over", () => {
+    // The abort-in-window shape: the rejected attempt committed (absorbing the turn output),
+    // its repair was written, and the process exited before any request carried the repair.
+    // Replay's eligibility filter keeps exactly the repair (paired with the committed,
+    // unanswered c1) as carry-over — the same stash the live engine holds — and does not
+    // reclaim the absorbed turn output.
+    const result = resumeTrace([
+      meta(),
+      userText("go"),
+      requestBegin(),
+      toolCall({ name: "exec_command", arguments: "{}", toolCallId: "ct" }),
+      requestEnd("completed"),
+      tokenUsage(usage(150), usage(150)),
+      toolCallOutput({ output: "tool ran", toolCallId: "ct" }),
+      compactionBegin({ reason: "context", mode: "summarize", context: 150, turns: 1 }),
+      userText("COMPACT NOW"),
+      requestBegin(),
+      toolCall({ name: "t", arguments: "{}", toolCallId: "c1" }),
+      requestEnd("completed"),
+      tokenUsage(usage(160), usage(310)),
+      toolCallOutput({
+        output: "[tool error] the compaction request expects a summary, not tool calls",
+        toolCallId: "c1",
+        stopReason: "failed",
+      }),
+      compactionEnd({ reason: "context", mode: "summarize", status: "aborted" }),
+    ]);
+    expect(result.contextClosed).toBe(false);
+    // The absorbed turn output sits in history (attempt 1's committed input), not carry-over.
+    expect(result.history.map((m) => (m.payload as { type?: string }).type)).toEqual([
+      "text",
+      "tool_call",
+      "tool_call_output",
+      "text",
+      "tool_call",
+    ]);
+    expect(
+      result.carryOver.map((m) => (m.payload as { tool_call_id?: string }).tool_call_id),
+    ).toEqual(["c1"]);
+    expect((result.carryOver[0]!.payload as { output?: string }).output).toContain("[tool error]");
+  });
 });

packages/docs/content/agent-loop.en.md

@@ -116,7 +116,7 @@ Three triggers (`compaction_begin.reason`):
 
 Two modes: `summarize` (default) appends the compaction Prompt to the old context, extracts the `[summary]`, wraps it as a `[context_summary]` user text and continues in a **fresh model context**; `discard` simply drops the old context. System markers are written as `[tag]…[/tag]`; the earlier angle-bracket form (`<summary>`, `<context_summary>`, …) is still recognized when reading old Traces and old persisted compaction prompts. Compaction rotates the [Trace file](/sessions-and-traces) (`_002`, `_003`, …) — one Trace file always equals one complete model context. `compactability()` probes feasibility before `session.compact()` (`ok | unsupported | empty | just_compacted`).
 
-The compaction request keeps the session's toolset **unchanged** — the request prefix (tool list included) stays byte-identical to ordinary turns, so the provider's prompt cache remains valid at the moment the context is largest. Compaction still succeeds only with a valid summary: a response that calls a tool or whose extracted summary is empty is rejected — any tool calls are answered with synthesized failed outputs (keeping `tool_use`/`tool_result` pairing intact) and the repaired request is resent immediately (a rejection is model behavior, not a transport failure, so no backoff applies), up to 5 rejected attempts; then the compaction ends `failed`, keeping the original context and Trace file until the next trigger. Transport `timeout`/`malformed` attempts follow the compaction-specific reconnect cap and backoff ladder described under "Automatic reconnect" above.
+The compaction request keeps the session's toolset **unchanged** — the request prefix (tool list included) stays byte-identical to ordinary turns, so the provider's prompt cache remains valid at the moment the context is largest. Compaction still succeeds only with a valid summary: a response that calls a tool or whose extracted summary is empty is rejected — any tool calls are answered with synthesized failed outputs (keeping `tool_use`/`tool_result` pairing intact) and the repaired request is resent immediately (a rejection is model behavior, not a transport failure, so no backoff applies), up to 5 rejected attempts; then the compaction ends `failed`, keeping the original context and Trace file until the next trigger. Transport `timeout`/`malformed` attempts follow the compaction-specific reconnect cap and backoff ladder described under "Automatic reconnect" above. The first **committed** attempt — adopted or rejected — also absorbs whatever turn input was folded into the compaction request (mid-task tool results, or the carry-over a manual `/compact` folds in) into the old context's history: retries resend only the repairs and the Prompt, and the continuation after an abandoned compaction never resends the absorbed input.
 
 ## Concurrency model
 

packages/docs/content/agent-loop.zh.md

@@ -113,7 +113,7 @@ interface CompactionSettings {
 
 两种模式：`summarize`(默认)向旧上下文追加压缩 Prompt，提取 `[summary]` 后包装为 `[context_summary]` 用户文本，在**全新的模型上下文**中继续；`discard` 直接丢弃旧上下文。系统标记统一写作 `[tag]…[/tag]`；读取旧 Trace 与旧压缩 Prompt 时仍识别早期的尖括号形式（`<summary>`、`<context_summary>` 等）。压缩时 [Trace 文件随之轮转](/sessions-and-traces)(`_002`、`_003`……)，一个 Trace 文件恒等于一个完整模型上下文。`session.compact()` 前可用 `compactability()` 探询可行性(`ok | unsupported | empty | just_compacted`)。
 
-压缩请求**保持会话工具集不变**——请求前缀（含工具列表）与普通轮次逐字节一致，确保上下文最大的时刻提供商的提示词缓存依然有效。只有得到有效摘要，压缩才算成功：若响应中出现工具调用、或提取出的摘要为空，则判为无效并重试——工具调用会先以合成的失败输出逐一应答（保持 `tool_use`/`tool_result` 配对完整），修复后的请求**立即重发**（无效摘要是模型行为而非传输故障，不做退避），最多允许 5 次无效尝试，之后压缩以 `failed` 结束，保留原上下文与 Trace 文件，等待下次触发。传输层 `timeout`/`malformed` 仍走上文「自动重连」一节所述的压缩专用重连上限与退避阶梯。
+压缩请求**保持会话工具集不变**——请求前缀（含工具列表）与普通轮次逐字节一致，确保上下文最大的时刻提供商的提示词缓存依然有效。只有得到有效摘要，压缩才算成功：若响应中出现工具调用、或提取出的摘要为空，则判为无效并重试——工具调用会先以合成的失败输出逐一应答（保持 `tool_use`/`tool_result` 配对完整），修复后的请求**立即重发**（无效摘要是模型行为而非传输故障，不做退避），最多允许 5 次无效尝试，之后压缩以 `failed` 结束，保留原上下文与 Trace 文件，等待下次触发。传输层 `timeout`/`malformed` 仍走上文「自动重连」一节所述的压缩专用重连上限与退避阶梯。首个**已提交**的尝试（无论被采纳还是被判无效）同时会把折叠进压缩请求的本轮输入（任务中途的工具结果、手动 `/compact` 折叠的补发内容）吸收进旧上下文：重试只重发修复输出与压缩 Prompt，压缩放弃后的续跑也不再重发已吸收的输入。
 
 ## 并发模型