Compaction can return an empty summary while leaving the over-cap history unchanged

[yxlyx]

Summary

Auto-compaction failed in the session Using Codex image generation (worktree /Users/limyuxi/my-website) at approximately 218k context tokens:

[compacting ~218185 tokens…]
[compaction failed: empty summary, history unchanged]

The session then remained usable only with the original, oversized history; the next visible event was a tool call. No non-empty compacted summary replaced the history.

Evidence

• Provider/model: codex / gpt-5.6-luna
• Session file: .graff/sessions/using-codex-image-generation.session.json
• Session title: Using Codex image generation
• Session state after the failure still recorded context_tokens=212546 and 730 messages.
• Trace evidence from .graff/traces/bcaa6e126cd944f6.jsonl shows compaction-sized requests at context 64930 followed immediately by context values 218185 and 218862, rather than a durable reduction of the active history. The trace run ID is bcaa6e126cd944f6 and its session ID is c6e9e2f02f3bc8b9bdfc98891d43764c.

This is distinct from a transport WriteFailed: the observed user-facing failure specifically says the summary was empty and that history was unchanged.

Expected behavior

When compaction is triggered, the harness should either:

1. atomically replace the old history with a validated, non-empty summary and materially lower the context count; or
2. keep the old history but clearly enter a recoverable fallback (emergency trim/retry) before sending another model request.

An empty summary must never be accepted as a completed compaction result, and a failed compaction should not leave the session silently operating at the same over-cap context.

Impact

Long-running sessions can repeatedly hit the compaction threshold, fail to reduce context, and eventually become unable to send further turns or require manual session recovery. The failure also makes it unclear whether the user's conversation state was safely summarized.

Related prior compaction reports: #88 and #163. This report covers the newer empty-summary/history-unchanged failure mode.

[justrach]

Shipped in v0.0.239. Two changes: (1) a compact_summary_failures streak counts consecutive complete-but-unusable summaries (empty or truncated) — unlike the transport counter it survives a complete response, and two of them while over compact@ unlock the same bounded emergencyTrim escape as the WriteFailed streak, so an over-cap history can no longer be re-shipped forever; (2) compaction summary requests run at low reasoning effort on the Responses wire — a high-effort reasoner completing with only reasoning items and zero output text was the empty-summary trigger observed here. The mid-turn re-attempt debounce from the analysis was deliberately left out: escalation at two failures already bounds the loop to one extra request; can be a follow-up if that one request matters.