V2: compaction fires on first request / loops when compaction.buffer >= model context limit #56783

Activity

1. reynold-hu commented on Oct 18, 2026

   

   reynold-hu

   on Oct 18, 2026

   Contributor

   More actions

   Opened #56786 with a fix: the ceiling now keeps at most 90% of the window free, automatic compaction waits for an older assistant reply before it summarizes, and an automatic no-op is reported as skipped instead of a failed compaction. The new test fails on the parent commit (both cases complete) and passes with the fix; session-runner's existing buffer cases still pass.
2. opencode-agent commented on Oct 18, 2026

   

   opencode-agentBot

   on Oct 18, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report. I was able to reproduce this on v2.0.32, and it still happens on the latest dev (da9d436).

   Steps:

   1. Use a model with a 128k context window.
   2. Set "compaction": { "auto": true, "buffer": 600000, "keep": { "tokens": 15000 } }.
   3. Send hello, then a second message.

   An automatic compaction starts straight after each user message, including the very first one, and neither message gets a normal reply. The recording and screenshot below show it. The script below reproduces it with a simulated model, so the compaction summaries fail on a template check. What matters is that compaction runs at all on a near-empty session. I didn't see the exact Nothing to compact yet log line in this run, but compaction fired on every turn as you described.

   As you worked out, the automatic trigger uses window - buffer with no lower limit. A buffer at or above the model's window makes compaction due on every turn. Until this is addressed, keeping buffer below the smallest context window you use avoids it.

   The per-model buffer part of your request is tracked in #43703, and #56443 is an open PR for it.

   recording.mp4

   

   Reproduction script (drive.ts)

   // Repro for anomalyco/opencode#56783: compaction.buffer >= model context limit
   // makes auto compaction fire on every turn, starting with the very first one.
   // Simulated model has limit.context = 128000; buffer = 600000.
   import { Effect } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       tui: { recording: true },
       project: { git: true, files: { "AGENTS.md": "# Project\n" } },
       config: {
         compaction: { auto: true, buffer: 600000, keep: { tokens: 15000 } },
       } as any,
     },
     ({ ui, llm }) =>
       Effect.gen(function* () {
         // Enough responses for the reply and any (unexpected) compaction summaries.
         for (let i = 0; i < 6; i++) yield* llm.queue(Llm.text(`Reply number ${i + 1}.`))
         yield* ui.submit("hello")
         yield* Effect.sleep(6000)
         yield* ui.screenshot("after-first-message")
         yield* ui.submit("second message")
         yield* Effect.sleep(6000)
         yield* ui.screenshot("after-second-message")
       }),
   )

   Run it with opencode-drive: opencode-drive run ./drive.ts. It drives opencode from your PATH; set OC_DEV to the path of an OpenCode checkout to run that from source instead.
3. opencode-agent

   added

   reproduced

   on Oct 18, 2026
4. rekram1-node

   closed this as completedin #58313on Oct 23, 2026