Distillation

Date: Sep 16, 2026

• 🟡 (16:13) /tmp/opencode/session-recovery-v7-check/packages/core/test/session-runner.test.ts ownership-fencing test "never dispatches a provider after ownership changes" pauses model resolution with modelResolveHook = Deferred.succeed(resolving, undefined).pipe(Effect.andThen(Deferred.await(resumeResolution))), changes ownership in SessionExecutionTable, and verifies the stale runner does not dispatch the provider.
• 🟡 (16:13) Ownership-fencing test "never publishes streamed output after ownership changes" configures streamed text "stale output" via fragmentFixture("text", "text-fenced", ["stale output"]).completeEvents, changes ownership in SessionExecutionTable, and verifies the stale output is not durably published.
• 🟡 (16:13) Ownership-fencing test "never starts a local tool after ownership changes" clears executions, changes ownership before tool execution, and verifies no execution occurs and no event with SessionEvent.Tool.Called.type is published.
• 🟡 (16:13) Ownership-fencing test "never fails interrupted tools after ownership changes" uses assistant tool call ID "call-interrupted-fence" and tool "echo"; after ownership changes, the projected tool content remains { type: "tool", id: "call-interrupted-fence", state: { status: "running" } } rather than being failed by the stale runner.
• 🟡 (16:13) The globally attached application-tool coverage in session-runner.test.ts defines structured output as Schema.Struct({ answer: Schema.String }).
• 🟡 (16:13) session-runner.test.ts includes compaction test "retains only complete serialized messages during compaction"; its setup records a recent prompt with session.prompt({ sessionID, prompt: Prompt.make({ text: recent }), resume: false }).
• 🟡 (16:13) Compaction recovery coverage verifies compacted history contains no message whose type === "synthetic".
• 🟡 (16:13) Context-overflow recovery coverage verifies the first user text in requests[2] contains the exact summary wrapper "<summary>\n## Objective\n- Recover overflow\n</summary>".
• 🟡 (16:13) Raw context-overflow recovery uses summary text "## Objective\n- Recover raw overflow" through fragmentFixture("text", "text-summary", ["## Objective\n- Recover raw overflow"]).completeEvents.
• 🟡 (16:13) Overflow-recovery interruption coverage awaits the runner fiber with Fiber.await(run) and expects a { _tag: "Failure" } result when interrupted while the summary provider is running.
• 🟡 (16:13) Reasoning/tool projection coverage includes hosted call ID "call-provider" marked providerExecuted: true and a file result { type: "file", mime: "image/png", uri: "data:image/png;base64,aGVsbG8=", name: "hello.png" }.
• 🟡 (16:13) Test "never commits continuation before unresolved hosted tools are reconciled" exercises a provider turn ending with LLMEvent.finish({ reason: "tool-calls" }), requiring hosted-tool reconciliation before continuation is committed.
• 🟡 (16:13) Step-publication defect coverage directly queries durable event types "session.next.step.ended.2" and "session.next.step.failed.2"; one test installs a database trigger condition WHEN NEW.type = 'session.next.step.failed.2' to simulate failure publication defects.
• 🟡 (16:13) Settled-tool continuation recovery coverage emits a recovered text fragment ending with LLMEvent.textEnd({ id: "text-recovered" }).