[Bug]: Automatic context compression truncates older turns without a usable summary

Summary

Automatic context compression can discard older conversation turns without producing a usable summary, leaving only roughly the latest six messages.

Steps to reproduce

1. Use Cherry Studio v2.0.9 on Windows 10 x64 with DeepSeek V4 Pro.
2. Enable automatic context compression.
3. Continue a conversation until compression is triggered.
4. Ask the model to recall constraints from early turns and inspect the post-compression context behavior.

Actual result

The reporter observed only approximately six recent messages remaining and no effective summary of earlier turns. Earlier conversation context was lost.

Expected result

Compression produces a usable summary that preserves important instructions, decisions, and task state. It must not silently behave as simple history truncation.

Impact

Long-running conversations can lose user constraints and prior decisions without warning. Using a very large context window is the reported workaround.

Environment

• Cherry Studio: v2.0.9
• OS: Windows 10 x64
• Model: DeepSeek V4 Pro
• Context window and compression threshold: To be confirmed

Investigation notes

• #17727 tracks compression not triggering in a Work Agent; this report covers a different failure mode after compression triggers.
• #18271 and #18977 concern transparency/control and do not establish that this truncation behavior is fixed.

Acceptance criteria

• Triggered compression produces and retains a summary before old turns are removed.
• Important system/user constraints survive compression and can be recalled afterward.
• Failure to summarize does not silently delete history; the user receives an actionable error or the operation safely retries.
• Regression coverage verifies content preservation across more than one compression cycle.

Feedback sources

• Cherry Studio in-app feedback form, August 28, 2026 (Beijing time), original archive retained as protected evidence.