Summary

Fixes #15533.

Auto-compaction triggers after an assistant turn whose token usage overflows the model's context window. Before this change, SessionCompaction.process() unconditionally injected a synthetic Continue if you have next steps... user message whenever auto: true and no replay was needed.

When the assistant had already ended its turn naturally (finish ≠ "tool-calls" / "unknown" and no pending tool calls), that synthetic Continue pushed the model into yet another turn — which overflowed again, triggered another compaction, injected another Continue, and so on. Sessions got stuck in the Build → Compaction → Build → Compaction → … loop reported repeatedly in #15533, including reproductions with Claude Opus via Copilot, OpenAI GPT 5.5, local llama.cpp, mlx, and others.

What this PR changes

packages/opencode/src/session/compaction.ts:

1. Before deciding whether to inject the synthetic Continue, walk input.messages backwards and find the most recent non-summary assistant message that has finish set — i.e. the message whose overflow caused this auto-compaction.
2. Classify it as "ended naturally" iff:
   • finish is not "tool-calls" or "unknown", and
   • it has no non-providerExecuted tool parts pending.
     This matches the finish-reason set the prompt loop itself uses at prompt.ts:1267-1275 to decide it can break out of the run loop.
3. If the assistant ended naturally:
   • Skip the synthetic Continue injection.
   • Return "stop" from process() so runLoop in prompt.ts exits cleanly instead of re-entering with a synthetic user message that is now newer than lastAssistant.

What this PR does not change

• Manual compaction (auto: false) — never injected Continue, untouched.
• The overflow-with-replay path (input.overflow === true with a recoverable prior user message) — still replays the original user turn.
• Mid-tool-call compactions (finish === "tool-calls" or pending tool calls) — still inject Continue so the model can finish its in-flight work.
• The experimental.compaction.autocontinue plugin hook — still consulted exactly as before for the (now narrower) cases where we do want to autocontinue.

Test plan

• bun test test/session/compaction.test.ts — 52/52 pass.
• Existing "adds synthetic continue prompt when auto is enabled" test still passes (assistant with no prior finish ⇒ assistantEndedNaturally stays false ⇒ Continue still injected).
• New regression test added: skips synthetic continue and stops when prior assistant ended naturally — sets up user → assistant(finish=end_turn) → compaction-marker user, calls process({ auto: true }), asserts (a) no compaction_continue-marked synthetic user is produced, and (b) process returns "stop".
• bun run typecheck clean (full monorepo, 15/15 packages).

Why "stop" instead of "continue" at the end

If we only skipped the Continue injection but still returned "continue", the prompt loop in prompt.ts:1294-1304 would continue back to the top. There it would find the just-created compaction user message as lastUser, with lastUser.id > lastAssistant.id, so the lastUser.id < lastAssistant.id natural break condition (prompt.ts:1271) would be false, and the loop would call the model again with the compacted summary in context — re-triggering #15533 by a slightly different path. Returning "stop" ends the loop the moment we know the assistant had nothing more to do. The user can pick up the conversation with their next prompt; the compacted summary is preserved.

🤖 Generated with Claude Code