Repository navigation
Compaction failed: received empty response from model (3x retry, manual /compact on Opus 4.6) #2861
Description
Describe the bug
Running /compact manually in a short session (< 30 turns) on Claude Opus 4.6 produces three consecutive failures:
X Compaction Failed: Error: Compaction failed: received empty response from model
X Compaction Failed: Error: Compaction failed: received empty response from model
X Compaction Failed: Error: Compaction failed: received empty response from model
The session remained usable — after one additional turn (wasting a premium request), a subsequent /compact succeeded normally.
Affected version
1.0.34
Steps to reproduce the behavior
- Start a session using Claude Opus 4.6.
- Work through a task (commit pushed, environment reloaded mid-session — MCP servers, extensions, hooks, agents, skills all reloaded at the same time).
- Run
/compactmanually. - Observe three consecutive
Compaction failed: received empty response from modelerrors.
Expected behavior
/compact should succeed, or at minimum degrade gracefully (e.g., retry with a different summarization strategy rather than repeating the same failing call three times).
Additional context
- Model: Claude Opus 4.6
- Session length: Short (< 30 turns)
- Trigger: Manual
/compactcommand - Environment: Windows, VS Code Insiders connected, 2 custom instructions, 1 extension, 1 hook, 6 agents, 14 skills
- MCP servers: 0 connected at the time of failure (had just reloaded)
- Recovery: Session continued normally after one additional turn, then
/compactworked — but this wasted a premium request on the recovery turn.
The environment reload (MCP Servers reloaded: 0 servers connected, mermaid-render extension loaded, Environment loaded: ...) occurred immediately before the compaction failures, which may be relevant — the compaction model may have received an unusual message history containing reload notifications.
Related issues
- Compaction killed the session #2500 — Compaction killed the session (different symptom but same area)
- CAPIError: 400 Unrecognized request argument 'reasoning_effort' when compacting after switching models #2553 — Compaction fails after model switching (different root cause)
- Toggle off for auto compaction!!!! #2333 — Request to toggle off auto-compaction (workaround request)
Activity
jay-tau commentedon Aug 30, 2026
I can reproduce this on a substantially newer version, with evidence that the model response is not actually empty.
Environment
1.0.82/compactTwo consecutive reproductions
Both
/responsescalls completed normally over HTTP withfinish_reason: stop,responses_message_status: completed, and a non-emptyfinal_answer. The returned compaction text was structurally valid (<overview>,<next_steps>, and<checkpoint_title>sections).The streamed event sequence contained several reasoning-only assistant messages with empty
content, followed by one final assistant message with the non-empty summary. Despite that, the nativecompactionRunModelCallpath returnedCompaction failed: received empty response from model, so the summary was not applied.The session remained usable and its original transcript was preserved. This looks like stream aggregation selecting an empty reasoning message instead of the later non-empty final-answer message, rather than the model returning no response.
MattPD commentedon Aug 31, 2026
I reproduced this behavior in two Copilot CLI releases: 1.0.81 and 1.0.82. In the 1.0.82 run, three streamed assistant messages contained reasoning content but no text before a complete 22,877-character final answer. The CLI still logged
received empty response from model343 ms after that answer. My next/compactsucceeded after I switched from GPT-5.6 Sol to Claude Opus 5, then failed after I switched back. I have also seen/compactfail with Claude Opus 5 and succeed after a switch to GPT-5.6 Sol. The pattern suggests that the CLI may be combining streamed messages incorrectly. Changing models and retrying has worked in either direction. No particular model has been consistently successful.DavidTeju commentedon Sep 2, 2026
I reproduced this btw
megasilk53 commentedon Sep 3, 2026
Reproduces readily with GPT-5.6 Luna · 1.1M context (63%)
GitHub Copilot CLI 1.0.82 (Linux x86_64)
ja552588 commentedon Sep 3, 2026
PureWeen commentedon Sep 8, 2026
Additional reproduction with GPT-6 Astra long context, observed through GitHub Copilot App with Copilot CLI 1.0.84-1.
Compaction failed: received empty response from model.This was not accompanied by a 413 or serialized-request-size error, so it appears distinct from the 5 MB request-limit failure mode. The contrasting failed and successful Astra attempts suggest the issue is intermittent or input/stream-shape dependent rather than a simple advertised-context-limit overflow.
I did not capture the raw
/responsesstream for the failed attempt, so I cannot independently confirm whether a valid final answer was present and later discarded. However, the exact CLI error and the model-switch recovery behavior align with the stream-aggregation evidence already reported in this thread.copilot-cli-bot commentedon Sep 29, 2026
This appears to be resolved as of GitHub Copilot CLI v1.0.89. Please update to that release or a newer stable version. We're closing this as completed. If it still happens, please comment with your version and reproduction steps.