Session wedged over the 1M window: repeated 400 prompt-too-long despite auto-compaction (v1.58.0)

Summary

A long, tool-heavy session is stuck: every request fails with 400 This model's maximum context length is 1048576 tokens ... requested 1057726 (993726 in the messages, 64000 in the completion). The transcript only exceeds the window once the fixed 64k completion budget is added. Each continue retry appends the error text to the transcript, so the request grows ~+2.2k tokens per attempt and can never fit. Auto-compaction has already run twice and then reports "Conversation is already compact.", so the documented "prompt too long" recovery never clears the session.

Expected Behavior

Per docs (Context & Compaction): a too-long request is classified, the window is compacted on the spot, and the request is retried once - only a second consecutive failure surfaces. Auto-compaction Tier 3 should also fire at >= 90% of the window.

Actual Behavior

The 400 surfaces on every attempt (3 in a row). Compaction meta: count=2, tokensSaved=646529 total (last pass only 29011). After that, continue reports "Conversation is already compact." and the next request still overflows. The session is unusable without /rewind or /clear.

Steps to reproduce the issue

1. Run a long, tool-heavy session on a 1M-window model with the transcript near the window.
2. Let auto-compaction run (or force it), then keep the session growing.
3. Type continue - every request fails with the 400 above.

Command Code Version

1.58.0

Operating System

Windows

Terminal/IDE

No response

Shell

No response

Session file (optional)

No response

Fix prompt (optional)

If a compaction pass reclaims too little and the next request still overflows, escalate instead of backing off: force a tool-call trim (keep last 10) regardless of compact mode, or retry once with a reduced completion budget so the request can fit under the window.

Additional context

• session id: 272ee4d1-930f-41ec-8eea-5e3d715fc450 (fork of 80401b9d-fbd9-4b6c-9f2e-6a2a44813ef0)
• model: deepseek/deepseek-v4.1-flash; project slug: c-windows-system32
• three consecutive failures:
  1. requested 1053393 (989393 messages + 64000 completion) - trace f1357f1faeee53e3709f485ef4468454
  2. requested 1055558 (991558 + 64000) - trace 028ff3b7779a4557810d3de8c158986f
  3. requested 1057726 (993726 + 64000) - trace 17e437ff33af0695c0169e0bcd17147d
• lastCompactedAt: 2026-09-19T22:02:39.097Z
• session JSONL (9.2 MB) can be provided on request

Không có tệp nguồn hoặc bài kiểm thử nào được nêu tên. Hãy tái hiện ba lỗi 400 liên tiếp bằng các bước dài phiên được cung cấp, sau đó truy vết đường đi của quá trình tự động nén gọn và thử lại; được xem là hoàn tất khi một yêu cầu quá lớn được nén gọn hoặc được giảm theo cách khác, và phiên không còn bị mắc kẹt sau lần thử lại.