[Bug] Auto-compaction fires at ~447k on a 1M-context model after model switch and fails with compaction.interrupted #57392

Summary

Auto-compaction fired at ~447k conversation tokens on a model configured with a 1,000,000-token context, about 1 second after sending a message, and then failed with compaction.interrupted. The session had just been switched from opencode/beam-free to opencode/deepseek-v4.1-flash-free (same provider). The UI showed ~447.6K (43%).

This may be related to the catalog-vs-gateway limit mismatch in #50574 / #56296, but the model-switch timing and the full-estimate behavior look like a separate bug.

Environment

• OpenCode: 2.1.0 (opencode --version on the CLI shim reports 2.0.32, the running server logs version=2.1.0)
• OS: Windows 11
• Provider/model: opencode/deepseek-v4.1-flash-free
• Config (~/.config/opencode/opencode.jsonc):

{
  "compaction": { "auto": true, "keep": { "tokens": 24000 }, "buffer": 16000 },
  "providers": {
    "opencode": {
      "models": {
        "deepseek-v4.1-flash-free": {
          "limit": { "context": 1000000, "output": 128000 }
        }
      }
    }
  }
}

The running server's /api/model reports limit: { context: 1000000, output: 128000 } for this model.

Reproduction

1. Have a long session (~447k tokens according to the previous model's usage) using opencode/beam-free.
2. Switch the session model to opencode/deepseek-v4.1-flash-free (same provider).
3. Send a short message.

Expected

Either:

• the request is accepted and the session continues, or
• if the provider rejects it as too long, a single overflow-recovery compaction runs and completes.

Actual

• A compaction message is created ~1.05s after the user message.
• Its status is failed, reason auto, error compaction.interrupted ("Compaction was interrupted").
• The session goes idle with outcome interrupted.
• No assistant response is produced.

Evidence

Session message rows (read-only from opencode.db):

seq 143529  type=assistant  model=opencode/beam-free
  tokens: input=280 cache.read=446592 output=732 reasoning=6  (usage ≈ 447,610)

seq 143538  type=model-switched  beam-free -> deepseek-v4.1-flash-free
seq 143541  type=system  (3,541 chars)
seq 143551  type=system  (8,980 chars)
seq 143561  type=system  (14,561 chars)
seq 143571  type=system  (10,560 chars)
seq 143572  type=user    (46 chars)
seq 143573  type=compaction  status=failed reason=auto
  error: {"type":"compaction.interrupted","message":"Compaction was interrupted"}
seq 143575  type=idle  outcome=interrupted

The session had ~165 tools loaded (godot 71, unityMCP 48, open_browser_use 19, blender 9, open_computer_use 9, opencode 8, blindfold 1).

Additional context (from follow-up)

• I did not press Esc or cancel the turn; the TUI was idle when the compaction started.
• Server log around that moment (~/.local/share/opencode/log/opencode.log) has no ERROR/WARN for compaction, only:

timestamp=2026-10-20T11:56:09.579Z level=INFO run=73cd3b04 message=session.revert.stage sessionID=ses_ec9ed0effffezdoXCLpbtIqGZK ...
timestamp=2026-10-20T11:56:09.736Z level=INFO run=73cd3b04 message=session.revert.commit sessionID=ses_ec9ed0effffezdoXCLpbtIqGZK
timestamp=2026-10-20T11:56:10.867Z  (compaction event created, status=failed, reason=auto, error=compaction.interrupted)

• Without the 1M override (default 200k), auto-compaction would trigger around 184k (200000 - 16000), i.e. ~2.4x earlier. I overrode to 1M because the free model plainly serves more than 200k: this session had already reached ~447k under beam-free and ~861k under space-bunny-free without the provider rejecting.

• The free deepseek-v4.1-flash-free is not actually capped at ~200k. The 200k value comes from models.opencode.ai, but real usage exceeded 447k and earlier sessions reached 861k on space-bunny-free (1M catalog) with no 400. The catalog under-reports real capacity (tracked in [FEATURE]:increase session limit of deepseek V4.1 flash the free model #56034 / compaction: 1M-context sessions never auto-compact before the provider returns HTTP 400, and the 400 is not treated as context overflow #50574). The override to 1M is a workaround; the early compaction here is a side effect.

• Could maintainers confirm the real input limit for opencode/deepseek-v4.1-flash-free? If it is genuinely ~450-500k, updating the catalog to that value (and letting the overflow classifier surface the provider body) would let users avoid both the premature 200k compactions and the surprise compaction at 447k.

Hypothesis

In SessionCompaction.compact, the auto preflight check is:

if (reason === "auto" && !F(context, nm(context.model.limit, buffer))) return { status: "skipped" }

nm(limit, buffer) uses limit.input || limit.context minus buffer, i.e. 984,000 here. The measured usage is ~447k, so the preflight threshold should not fire. Two possibilities:

1. The request was rejected by the gateway as too large (real free-tier input limit lower than the advertised 1M), and the overflow-recovery path triggered the compaction. The failed event is recorded with reason: "auto" regardless.
2. After the model switch, the usage estimator falls back to a full request estimate (system + all 165 tool schemas + all messages), which can exceed the threshold even though the provider-reported conversation usage is ~447k.

Either way, the user-visible result is an early compaction at 43% followed by compaction.interrupted, rather than a completed compaction or a usable session.

Related

• compaction: 1M-context sessions never auto-compact before the provider returns HTTP 400, and the 400 is not treated as context overflow #50574 — 1M-context sessions and compaction ceiling / catalog-vs-provider limit
• opencode-go: long sessions stuck at provider input limit; auto-compaction never triggers #56296 — opencode-go long sessions stuck at provider input limit; auto-compaction never triggers
• [FEATURE]:increase session limit of deepseek V4.1 flash the free model #56034 — Zen's ~200k context for the free DeepSeek V4.1 Flash
• [FEATURE]: per-model compaction threshold (reserved/buffer) #43703 — per-model compaction threshold (reserved/buffer)
• keep.tokens is not honoured: the compaction split walk-back is unbounded and skips synthetic messages, so agent-driven sessions carry 234K past a 15K setting #43250 — keep.tokens not honoured