compaction.auto: false is ignored when provider throws a context overflow error

[timwkosatec]

Description

Setting compaction: { "auto": false } prevents proactive compaction (the token threshold check), but compaction still runs when the provider returns a hard context overflow error. The auto: false flag is completely ignored in this path.

---

The catch block for ContextOverflowError in processor.ts bypasses isOverflow() entirely and schedules compaction unconditionally. Once scheduled, the compaction task marker is written into the session message stream and SessionCompaction.process() runs on every subsequent loop iteration regardless of config.
The practical consequence: when a session hits the context limit, stopping and switching to a model with a higher token limit has no effect. The session remains stuck in the compaction task. The only escape is starting a new session entirely or manually changing the session json and removing the error entirely.

---

Reverting the last message is not an option as the agent would lose track of file edits already made in the session

Plugins

None

OpenCode version

1.2.24

Steps to reproduce

1. Set compaction: { "auto": false } in opencode.json
2. Run a session until the provider returns a context overflow error (this triggers the auto compaction task)
   (optional)
3. Stop the session and switch to a model with a higher context limit
4. Resume - the session is stuck in the compaction task

Screenshot and/or share link

No response

Operating System

Windows 11 Enterprise

Terminal

Windows Terminal

[MaheshBhushan]

this looks like a clear config-consistency bug in the overflow path.

suggested fix:

1. in the ContextOverflowError catch path, apply the same compaction.auto gate used by proactive overflow checks (isOverflow() path)
2. if compaction.auto === false, do not enqueue a compaction task marker; surface a recoverable error instead
3. on resume / model-switch, clear any stale queued compaction marker when auto compaction is disabled (or when current model context no longer overflows)

right now the session can get trapped in a compaction loop, and switching to a higher-context model doesn’t recover because the queued task keeps re-running.

expected behavior: auto:false should consistently mean no automatic compaction, regardless of whether overflow is predicted proactively or returned by provider error.

[timwkosatec]

Found the PR where the auto flag was overlooked:

• Commit: be20f86
• Issue: Request entity too large error #6574 (linked from PR)
• PR: fix: recover from 413 Request Entity Too Large via auto-compaction #14707

packages/opencode/src/session/processor.ts, lines 356-364 in the diff:
// Before (fell through to error → stop):
if (MessageV2.ContextOverflowError.isInstance(error)) {
  // TODO: Handle context overflow error
}
// After (forces compaction, skips error path):
if (MessageV2.ContextOverflowError.isInstance(error)) {
  needsCompaction = true          // line 360
  Bus.publish(Session.Event.Error, {
    sessionID: input.sessionID,
    error,
  })
} else {
The else on line 365 is the bug — overflow no longer reaches input.assistantMessage.error = error (line 379), so the session never stops.

[timwkosatec]

Proposed Design:

When auto: false: treat ContextOverflowError as a normal error → set assistantMessage.error → return "stop". No compaction marker is written. Session stops cleanly. User can resume the same session with a bigger model.

---

When auto: true (default): keep current behavior — needsCompaction = true → compacts and recovers. PR #14707's fix stays intact for the majority of users.