V2: compaction fires on the first request and repeats every turn when compaction.buffer >= model context limit #56813

Description

Summary

With compaction.buffer set greater than or equal to the model's context limit, V2 starts automatic compaction on the very first request and re-triggers it on every turn or two. This emits meaningless structured summaries at session start and sometimes surfaces an internal Session.StepFailedError: Nothing to compact yet at server start.

Plugins

No response

OpenCode version

• opencode v2.0.32 (Linux x64). The same config is also read by V1 1.19.0, but the looping behavior was observed on V2.

Steps to reproduce

• Reproduced on opencode/mimo-v2.6-flash-free; the defect is model-agnostic (it depends only on the arithmetic below).

Reproduction

opencode.json:

{
  "compaction": {
    "auto": true,
    "buffer": 600000,
    "keep": { "tokens": 15000 }
  }
}

1. Start a V2 session (any directory, e.g. one with an AGENTS.md).
2. Send a single first message.
3. Watch the message stream: a compaction message is created immediately after the first user message (reason: "auto"), summarizing an empty history, and it repeats every 1-2 turns.

Startup log:

ERROR "Failed to drain Session" cause="Session.StepFailedError: Nothing to compact yet
    at SessionRunner.runStep
    at SessionRunner.drain
    at ServerProcess.start"

Expected

• An oversized buffer should be clamped to < context limit, or rejected with a clear warning at config-load time.
• Compaction should not run when there is no older history to compact.
• The internal StepFailedError should not surface as a user-facing failure at server start.

Actual

The trigger condition is estimated_tokens > context_limit - max(requested_output_tokens, buffer) (the docs describe buffer's default as 10% of the context limit). When buffer >= context_limit, the right-hand side is zero or negative, so the condition is always true and compaction fires on every step, including the first.

Because buffer is a single global value evaluated against each model's own limit, a value chosen for a large-context model (e.g. 400000 for a 1M-token model) breaks every smaller model in the same config.

Additional context / feature request

buffer and keep.tokens are configurable only in the top-level compaction block; per-model settings.compaction accepts only type: native|summary. That forces one compaction policy across models with very different context windows. Allowing a per-model (or per-provider) buffer, and/or an absolute "compact at N tokens" threshold, would let users choose safe compaction points for a 200k model and a 1M model without a single global value that is unsafe for one of them.

Screenshot and/or share link

No response

Operating System

Artix Linux Kernel 7.1.9-artix1-2

Terminal

Wezterm