fix(compaction): keep the post-compaction autocontinue on the session working model - #8397

Summary

• Keeps the post-compaction autocontinue on the session's working model when the agent's compaction.model differs from it.
• OpenCode's synthetic continue turn inherits the summarize request's model (session/compaction.ts: model: userMessage.model), so with a per-agent compaction.model configured the compaction model silently ran the continuation turns. The existing guard only matches agent === "compaction" and can never catch a model-only mismatch.
• On mismatch the handler now disables OpenCode's synthetic continue and re-issues the same continue prompt through session.promptAsync with the working model, the working agent, and the compaction_continue metadata marker.
• New helper module plugin/compaction-autocontinue-reroute.ts + three regression tests.

Problem

With agents.<agent>.compaction.model configured (preemptive compaction enabled), a live session was observed continuing on the compaction model after auto-compaction. Message rows from that session's opencode.db:

Root cause chain

1. Preemptive compaction calls client.session.summarize({ providerID, modelID: <compaction.model>, auto: true }) (hooks/preemptive-compaction-trigger.ts).
2. OpenCode stores the compaction trigger user message with that model (SessionCompaction.create).
3. OpenCode's autocompaction creates the synthetic continue message with model: userMessage.model — i.e. the summarize model — and gates it only on the experimental.compaction.autocontinue plugin hook (packages/opencode/src/session/compaction.ts, around lines 497–548 in v1.18.31).
4. createCompactionAutocontinueHandler disables the continue only when isCompactionAgent(input.agent). When the trigger's agent is the working agent (normal case) and only the model differs, the guard cannot match, so the continue runs on the compaction model.

Fix

plugin/compaction-autocontinue-reroute.ts

• Resolves the trigger model (newest message carrying a compaction part) and the working model (nearest non-compaction message, via the existing findNearestMessageWithFieldsFromMessages, which already skips compaction messages).
• createCompactionAutocontinueHandler now, after the existing checkpoint restores:
  • sets output.enabled = false to drop OpenCode's poisoned synthetic continue, then
  • re-issues the identical continue prompt with session.promptAsync({ body: { agent, model: <working>, parts: [{ type: "text", text: <same continue text>, metadata: { compaction_continue: true } }] } }).

Behavior matrix:

Wiring: testing/create-plugin-module.ts passes client and directory into the handler options. Both are optional — without a client the handler keeps the previous behavior, so existing unit tests and non-client call sites are unaffected.

Notes / limitations

• The SDK session.promptAsync body has no variant field, so the re-issued continue runs at the model's default effort; model selection itself is exact.
• Scope is limited to the post-compaction autocontinue; summaries still run on the configured compaction.model.
• Applies to every agent with a compaction.model, not just one.

Verification

• bun test src/index.compacting.test.ts src/index.compaction-model-agnostic.static.test.ts → 16 pass / 0 fail, 52 expect() calls (3 new cases: reroute on mismatch; unchanged on equal models; unchanged on lookup failure)
• bunx tsgo --noEmit -p packages/omo-opencode/tsconfig.json → passed
• bun run build → passed
• Live reproduction recorded before the fix (table above); the guard logic is additionally covered by unit tests.

Only the intended files are included (4 files, +308/−1); unrelated local worktree edits were not staged.

Summary by cubic

Fixes post-compaction autocontinue running on the configured compaction.model instead of the session's working model.

OpenCode's synthetic continue inherits the summarize request's model, so with a per-agent compaction.model configured, the compaction model silently ran continuation turns after auto-compaction. The handler now compares the compaction trigger's model with the working model, resolved from the newest real assistant turn while skipping compaction summaries and internal continue prompts so the reroute doesn't cascade onto itself. On mismatch it disables the synthetic continue and reissues the same prompt on the working model and agent through the shared prompt gate (dispatchInternalPrompt, queued when busy); a sent or queued dispatch counts as reissued, while equal models, lookup failures, and rejected dispatches keep the previous flow.

• New plugin/compaction-autocontinue-reroute.ts module plus client/directory wiring in testing/create-plugin-module.ts.
• Three regression tests cover rerouting, equal models, and lookup failure.
• Changelog entry added.

^{Written for commit 7de402d. Summary will update on new commits.}

Rebased onto v5.0.0-beta.72 (2026-09-18)

The fix has been ported forward from beta.62 to the current release so it can merge cleanly:

• 8 local fix commits cherry-picked onto v5.0.0-beta.72; the only conflict was changes.md (changelog) and the base version was kept there.
• The generated-artifacts commit (lockfile + extension bundles) was dropped and regenerated instead (bun install + bun run build) — replaying it collides with extension files this release removed.
• packages/omo-opencode/src/index.compacting.test.ts: 14 pass / 0 fail; typecheck clean on omo-opencode, omo-senpi, omo-codex.
• Verified this fix is not present upstream at v5.0.0-beta.72 (compaction-autocontinue-reroute absent from the tree), so it is still required.

Reproduction this addresses: after auto-compaction, the synthetic continuation ran on the compaction model instead of the session working model (session/compaction.ts stamps the trigger message with the per-agent compaction.model, and the continuation inherits userMessage.model). The added compaction-autocontinue-reroute module disables core's synthetic continue and re-issues the prompt via session.promptAsync with the working model.