Compaction summarizes the current turn and then appends the same turn again

[leseb]

Clawpatch finding

• Severity: medium
• Category: bug
• Confidence: high
• Triage: confirmed-bug
• Feature: feat_custom_apis_openai_responses_retrieval_proxy
• Finding ID: fnd_sig-feat-custom-apis-openai-resp_b177a277d5
• Run ID: 20260824T070943-dd5f81
• Run report: .clawpatch/reports/20260824T070943-dd5f81.md
• Prior-run overlap: No exact prior-run overlap found; related issue Compaction restores stale file references after resolution, bypassing the protected rewrite #661 covers a different compaction state-rewrite hazard.

Evidence

• apis/src/openai/responses/compact/mod.rs:291-322 (should_compact) — let conversation_text = build_conversation_text(&state.messages);
• apis/src/openai/responses/compact/mod.rs:505-515 (replace_messages) — new_messages.extend(state.input.iter().cloned());
• apis/src/openai/responses/compact/tests.rs:239-249 (conversation_text_array_content) — let text = build_conversation_text(&messages);

Description

After rehydration, state.messages contains prior history followed by the current request's normalized state.input. should_compact sends all of state.messages to the summarizer, so the generated summary already includes the current user turn. replace_messages then constructs the outbound context as the summary followed by state.input, duplicating the current turn semantically. This can cause the model to answer or weigh the latest request twice and wastes the context compaction was intended to reclaim.

Reproduction

Construct state with one historical exchange and current input "CURRENT UNIQUE TURN", set history_rehydrated and a zero compaction threshold, and capture the summarization request. Its user message contains CURRENT UNIQUE TURN. After applying the returned summary, inspect state.messages: the summary represents that turn and the original current-input item is also present.

Recommendation

Use the full context for threshold accounting if desired, but build the summarization text only from the history prefix state.messages[..state.messages.len() - state.input.len()]. Preserve state.input exactly once after the compaction item.

Minimum fix scope

Split history from the current-input tail when constructing conversation_text, retaining existing full-context token accounting where appropriate, and add one focused compaction lifecycle test.

Human review gate

This issue requires human review before any work can begin. Do not self-assign and Do not submit a PR before the issue has been validated.

[leseb]

This issue requires human review before any work can begin. Do not self-assign and Do not submit a PR before the issue has been validated.

Current Clawpatch run update:

• Finding ID: fnd_sig-feat-custom-apis-openai-resp_b177a277d5
• Run ID: 20260824T070943-dd5f81
• Severity/category/confidence: medium / bug / high
• Prior-run overlap: No exact prior-run overlap found; related issue Compaction restores stale file references after resolution, bypassing the protected rewrite #661 covers a different compaction state-rewrite hazard.
• Evidence: apis/src/openai/responses/compact/mod.rs:291-322; apis/src/openai/responses/compact/mod.rs:505-515; apis/src/openai/responses/compact/tests.rs:239-249

Reproduction:
Construct state with one historical exchange and current input "CURRENT UNIQUE TURN", set history_rehydrated and a zero compaction threshold, and capture the summarization request. Its user message contains CURRENT UNIQUE TURN. After applying the returned summary, inspect state.messages: the summary represents that turn and the original current-input item is also present.

[crstrn13]

Issue is dealt with in #726

[leseb]

Current Codex/xhigh review re-confirms current-turn double summarization. Finding fnd_sig-feat-custom-apis-openai-resp_cf03a964ef (high / data-loss / high confidence) cites apis/src/openai/responses/compact/mod.rs:328,558-563; PR #854 addresses the stale rewritten-tail half, but the summarizer still includes the current turn before that turn is appended again.