session.summarize with a non-session model makes the synthetic post-compaction continue (and the following turns) run on that model

What happens

session.summarize accepts a model for the compaction request. That model is stored on the synthetic compaction trigger message, and OpenCode's post-compaction autocontinue then reuses it for the synthetic continue turn:

packages/opencode/src/session/compaction.ts (v1.18.31, lines ~497–548):

const continueMsg = yield* session.updateMessage({
  id: MessageID.ascending(),
  role: "user",
  sessionID: input.sessionID,
  time: { created: Date.now() },
  agent: userMessage.agent,
  model: userMessage.model,   // <-- the summarize request's model
})
const text = /* ... */ "Continue if you have next steps, or stop and ask for clarification if you are unsure how to proceed."

and SessionCompaction.create (same file, ~559–582) stores the summarize request's model on the trigger message. So when a caller summarizes with a model different from the session's working model, every turn after the auto-continue runs on the summarize model instead of the model the user/session is actually using — and the session only returns to the working model when the next real user message arrives.

This is not reachable through the built-in auto-compact path (which passes lastUser.model), but it is reachable through the public session.summarize API/SDK, which is exactly what orchestration plugins use to summarize on a cheaper/faster model while the session keeps running on its selected model.

Reproduction

1. Session on model A (providerID/modelID of your choice), a few turns of context.
2. POST /session/{id}/summarize (or SDK client.session.summarize) with { providerID, modelID } of model B ≠ A, auto: true.
3. Observe the recorded messages:

Message rows from a live session (opencode.db) showing exactly this (B = zai-coding-plan/glm-5.3-flash, A = opencode-go/deepseek-v4.1-flash), 2026-09-16:

17:23:18 user      zai-coding-plan/glm-5.3-flash   [part: compaction, auto:true]
17:24:20 assistant zai-coding-plan/glm-5.3-flash   agent=compaction summary=true
17:31:38 user      zai-coding-plan/glm-5.3-flash   "Continue if you have next steps, ..."
17:31:38 assistant zai-coding-plan/glm-5.3-flash   agent=<working agent>   <- wrong model
17:33:25 assistant zai-coding-plan/glm-5.3-flash   agent=<working agent>
17:36:11 user      opencode-go/deepseek-v4.1-flash (next real prompt — everything after is back on A)

Expected

The synthetic continue turn (and the turns it triggers) should run on the session's working model — i.e. the model of the last non-compaction user/assistant message — regardless of which model was used to produce the summary. Summaries can legitimately run on a different model; continuations should not silently switch the session's model.

Suggested fix

Resolve the continue message's model from the last non-compaction user message instead of the trigger message, e.g.:

const continueModel = lastNonCompactionUserMessage?.model ?? userMessage.model

or pass the working model explicitly through SessionCompaction.create/processCompaction separately from the summarize model.

Note on the plugin workaround (why the hook is not enough)

The experimental.compaction.autocontinue plugin hook can only enable/disable the synthetic continue — the payload contains the model but there is no way to correct it. Downstream plugins therefore have to disable the synthetic continue and re-issue the continue prompt themselves via session.promptAsync with the working model. That workaround has been implemented for the oh-my-openagent plugin in code-yeongyu/oh-my-openagent#8397; a core-side fix would make the workaround unnecessary and would also cover every non-plugin caller of session.summarize.

Environment: opencode-ai 1.18.31 (Windows, x64); source references are from tag v1.18.31.