Auto-compaction silently resets long-running conversations: English-only checkpoints, zero-tail reset on overflow, and no user control — request for a “continuous conversation” mode

[TuringCorp-net]

The real-world usage that breaks

A common — for me, daily — pattern is one project ↔ one very long conversation. We keep chatting for hundreds of rounds: multi-stage plans, debugging sessions in the middle of step 2 of a fix, live architecture discussions. At the moment the thread matters most — often while a long autonomous task is still running in the background — the conversation "blows up":

1. Trigger A — silent checkpoint compaction. When pressure crosses ~80% of the context window, auto-compaction (on by default) rewrites the older history into a single checkpoint summary. The summarizer prompt is hard-coded English, so a Chinese conversation gets condensed into an English digest, and the model's replies suddenly flip to English, missing details discussed only minutes earlier. No notice, no preview, no confirmation.
2. Trigger B — hard overflow, full reset. When a large tool result or a background-subagent report lands and the very next request exceeds the API limit, overflow recovery force-compacts with a zero retained tail and retries inside the same send. The conversation is effectively reset to a few-thousand-token digest. The user must re-explain everything from scratch — often a thread they were minutes from finishing is lost forever.

Opening a new conversation mid-task is not an option (the live thread is the work), and a manual /compact produces the same amnesia. The net effect is the worst possible experience: the assistant turns into a stranger at the worst possible moment.

Code references (0.1.1-rc.2)

• dsh-compaction-basic: auto-compaction registers on agent/pre-step (pressure trigger) and agent/request-error (context-window-exceeded trigger).
• Defaults: auto: config.auto ?? true, thresholdRatio = 0.8, retainRatio = 0.16, summary budget maxTokens = 8192; the overflow path calls the range selector with retainTokens = 0.
• The summarizer instruction is hard-coded English ("You are now acting as a compaction engine for this AI coding assistant…"), independent of the conversation's language.

Feature request

1. Make auto-compaction opt-in or at least user-visible: surface a notice/preview and let the user decide (run at idle, never silently mid-turn).
2. Preserve the conversation's language in the checkpoint.
3. Prefer incremental rolling memory over one-shot full rewrites: keep the recent working set (last N user rounds) verbatim, condense the middle structurally, and fold only the ancient part into compact summaries — bounded steady state, never a reset.
4. Consider an official "continuous conversation" mode so users never need to abandon a live thread.

Reference implementation

We hit this badly in daily use and open-sourced a plugin implementing point 3: MosaicMemoryCompress — https://github.com/TuringCorp-net/mosaic-memory-compress (also listed on https://awesome-dsh-plugin.com). Recent ~30 rounds stay verbatim, middle rounds are dehydrated incrementally, ancient rounds fold into a compact summary pair, the original language is preserved, the steady-state message count is constant (same at round 60 or round 15,000), and the conversation never resets. Happy to adapt or contribute if that direction is useful.

Related threads sharing the "long conversations are not first-class" theme: #211 (UI freeze at long context), #370 (long sessions fail to open), #238 (TokenMeter O(n²) on long sessions).