Rewind "Summarize up to this point": turns completed during the in-flight summarization are silently dropped, and /context then reports stale pre-compact usage

Summary

In a long session (~856k/1M tokens) with background Task agents running, invoking Rewind → "Summarize up to this point" produced two coupled failures:

1. Silent turn loss (race condition): the summarization ran asynchronously for ~4 minutes. During that window, two background agents completed; their task-notification turns were delivered and the assistant replied to each (all visible in the terminal). When the summarization landed, those exchanges vanished from both the terminal transcript and the model context — no warning, no error. The turns had durable side effects (git commits made by tool calls in the replies), so the model's context now omits actions it actually performed.
2. Stale /context after a successful compact: compaction actually succeeded, but /context run 14s and again ~110s after the compact boundary still reported the pre-compact numbers (856.2k / 1M, 86%; "Messages: 839.7k"), which made it look like no compaction had occurred at all.

Environment

• Claude Code v2.1.202 (CLI, entrypoint: cli)
• macOS guest VM (Tart virtualization), darwin 25.5.0
• Model: claude-fable-5 with 1M context window
• Session state at repro: ~856k tokens (86%), 3 background agents launched via the Task tool still running

Repro narrative

1. Long-running session reaches ~856k/1M tokens; 3 background Task agents in flight.
2. Two agents complete; their task-notification turns arrive and the assistant replies to each (user sees both exchanges).

(Note: per the transcript, the notifications actually arrived ~2.5 min after the user had already started step 3 — the point is they were fully rendered and answered while the summarize was still running.)

1. User opens Rewind, selects a checkpoint from ~2h earlier, and chooses "Summarize up to this point" (with custom summary instructions).
2. Summarization runs in the background for ~4m11s; meanwhile the main loop keeps dequeuing and answering task-notifications.
3. Summarization completes → terminal redraws without the notification exchanges; no summary message is visible to the user; /context still shows 856.2k (86%).
4. User concludes "messages disappeared and compaction failed".

Forensic findings (session JSONL, ~/.claude/projects/<project>/<session-id>.jsonl)

Timeline reconstructed from entry timestamps (UTC):

| time | event |
|---|---|
| 18:36:42 | checkpoint node a43e959b (system turn_duration) — the rewind point the user later selected |
| 20:49:10 | last assistant turn before the operation completes; tail node 02de9921 |
| ~20:50:13 | user triggers "Summarize up to this point" (= boundary timestamp minus durationMs) |
| 20:52:51 | agent #1 task-notification enqueued (queue-operation) and processed; assistant replies — parented to 02de9921 |
| 20:53:38–20:54:17 | agent #2 notification processed; assistant replies — same branch |
| 20:54:24.774 | system entry subtype: "compact_boundary" written, followed by a user entry with isCompactSummary: true (~11k chars). |
| 20:54:38 | /context → 856.2k / 1M (86%), Messages 839.7k |
| 20:56:14 | /context again → still 856.2k (86%), Messages 839.7k |
| 20:56:57 | next real assistant API call usage: input_tokens: 7822, cache_creation: 5660, cache_read: 108702 ≈ 122k actual context; subsequent turns ~125k |

Key fields of the compact_boundary entry:

{
  "subtype": "compact_boundary",
  "parentUuid": null,
  "logicalParentUuid": "a43e959b-…",          // the 18:36 checkpoint the user selected
  "compactMetadata": {
    "trigger": "manual",
    "preTokens": 858844,
    "postTokens": 30147,
    "messagesSummarized": 817,
    "durationMs": 251396,                      // ≈ 4m11s in flight
    "cumulativeDroppedTokens": 828697,
    "preservedSegment": {
      "headUuid": "51153d6d-…",               // first message after the checkpoint (18:43)
      "anchorUuid": "a3b6245b-…",             // the isCompactSummary message
      "tailUuid": "02de9921-…"                // last turn at operation START (20:49:10)
    }
  }
}

What this shows:

• The preserved segment was snapshotted when the operation started (tailUuid = the 20:49:10 node). The turns generated during the 4-minute in-flight window (20:52:51 → 20:54:17) chain off that same node in the old branch. When the boundary + summary + preserved chain became canonical, those turns were orphaned — still physically present in the JSONL as a dead branch, but absent from the active conversation and the redrawn terminal. Nothing was deleted from disk; nothing was surfaced to the user either.
• Compaction itself worked: 858,844 → ~122k actual tokens on the next API call. The summary message exists in the transcript (isCompactSummary: true, parented to the boundary; the boundary's logicalParentUuid correctly points at the user-selected checkpoint).
• /context did not reflect it: two invocations at +14s and +110s after the boundary still reported the pre-compact estimate (856.2k total / Messages 839.7k) — while the very same UI had already redrawn the transcript in its post-compact shape. The stale readout is what convinced the user that the whole operation had silently failed.

Expected behavior

1. Turns that complete while a rewind-summarize is in flight must not be silently discarded. Either (a) hold the notification queue while the summarization runs, (b) re-anchor/append turns that landed mid-flight onto the compacted chain, or (c) at minimum warn: "N turns completed during summarization and were dropped (recoverable via rewind)".
2. /context immediately after a compact should report the post-compact context (or explicitly say it is an estimate pending the next request).
3. Nice-to-have: the Rewind menu could warn when background Task agents are running, since their completion notifications are exactly what races with the summarize.

Actual behavior

• Two completed-agent notification turns + assistant replies silently vanished from terminal and context (orphaned branch in JSONL); their durable side effects (git commits) remain, unknown to the model.
• /context reported pre-compact usage (856.2k/86%) at least twice after a compact that had actually reduced the context to ~122k, so the user had no way to tell the compaction had succeeded.

🤖 Generated with Claude Code