Distillation

Date: Sep 16, 2026

• 🔴 (11:23) packages/core/test/session-run-coordinator.test.ts verifies that after interruption cleanup is released, the resumed execution starts with force flags exactly [false, true].
• 🔴 (11:23) packages/core/test/session-run-coordinator.test.ts test "starts one follow-up when a wake races with failure" blocks the first drain, registers wake("session"), then fails with new Error("failed"); the original joined fiber receives that exact failure and one successor starts, making runs === 2.
• 🔴 (11:23) packages/core/test/session-run-coordinator.test.ts test "does not cancel execution when a joined waiter is interrupted" starts two joined callers for the same key, interrupts the second caller, and verifies the underlying drain continues for the first caller with runs === 1.
• 🔴 (11:23) packages/core/test/session-run-coordinator.test.ts test "runs different keys concurrently" runs keys "first" and "second" simultaneously and waits until both drains are active before releasing them.
• 🔴 (11:23) packages/core/test/session-run-coordinator.test.ts test "trampolines synchronous self-waking execution" repeatedly self-wakes key "session" to the exact limit 20_000 and expects runs === 20_000.
• 🔴 (11:23) packages/core/test/session-runner.test.ts uses SessionRecovery, ConfigCompaction, SessionRunnerModel, a recovery model with id: "recovery", and OpenAIChat.route.with({ limits: { context: 4_000, output: 50 } }); its project directory is /project.
• 🔴 (11:23) packages/core/test/session-runner.test.ts defines FragmentKind as "text" | "reasoning" | "tool input" and preserves the order ["text", "reasoning", "tool input"].
• 🔴 (11:23) User specified the invariant "never dispatches a provider after ownership changes" in packages/core/test/session-runner.test.ts: the run is paused during model resolution, execution ownership changes, and provider requests must remain exactly [].
• 🔴 (11:23) User specified the invariant "never publishes streamed output after ownership changes" in packages/core/test/session-runner.test.ts: output from a provider stream that began before execution ownership changed must not be published afterward.
• 🔴 (11:23) User specified the invariant "never starts a local tool after ownership changes" in packages/core/test/session-runner.test.ts; the fenced call is LLMEvent.toolCall({ id: "call-fenced", name: "echo", input: { text: "must not run" } }).
• 🔴 (11:23) User specified the invariant "never fails interrupted tools after ownership changes" in packages/core/test/session-runner.test.ts: interrupted-tool reconciliation must not publish failure after execution ownership changes.
• 🔴 (11:24) User specified the invariant "never commits continuation before unresolved hosted tools are reconciled" in packages/core/test/session-runner.test.ts.
• 🔴 (11:24) Additional packages/core/test/session-runner.test.ts recovery tests identified by exact name and line include: line 598 "redrives only the synthetic recovery continuation"; line 694 "redrives an ordinary prompt after its advisory wake is lost"; line 1362 "rebuilds the baseline directly after completed compaction"; line 1474 "keeps automatic compaction unknown until Compaction.Ended commits"; line 1521 "recovers the continuation committed with Compaction.Ended"; line 1575 "retains only complete serialized messages during compaction"; line 1636 "forces one compaction and retries after provider context overflow"; line 1665 "persists a second context overflow after one recovery"; line 1688 "recovers once from a raw context overflow failure"; line 1716 "publishes the original overflow when recovery summarization fails"; line 1736 "interrupts overflow recovery while the summary provider is running"; line 1761 "preserves effective System updates while compaction rebaseline is blocked".
• 🔴 (11:24) Additional packages/core/test/session-runner.test.ts continuation/tool tests identified by exact name and line include: line 2028 "recovers the continuation committed with a settled tool step"; line 2093 "reloads a model switch before a tool-driven continuation turn"; line 2195 "replays durable provider-executed tool results inline in a second-turn request"; line 2481 "promotes queued input after continuation ends"; line 2528 "preserves durable queued input for a later wake after interruption"; line 2657 "promotes queued input after steering continuation ends"; line 2756 "coalesces multiple active steering prompts into one continuation turn"; line 3325 "interrupts runner continuation when permission approval is declined"; line 3430 "interrupts runner continuation when a question is dismissed"; line 3786 "does not recover context overflow after durable assistant output".
• 🔴 (11:24) packages/core/test/session-projector.test.ts recovery tests identified by exact name and line include: line 67 "atomically reconciles one abandoned execution into one continuation"; line 221 "records a recovery continuation without forging an assistant identity"; line 258 "recovers each abandoned execution once across repeated startup scans"; line 304 "never recovers a live execution lease"; line 342 "never recovers an execution superseded by a newer assistant"; line 389 "redrives a durable recovery wake after admission"; line 440 "quarantines a malformed recovery row without blocking a valid wake"; line 494 "quarantines a full batch of poison recovery rows before redriving valid work"; line 559 "fairly redrives recovery and ordinary work beyond a full bounded batch"; line 637 "quarantines a continuation that belongs to another Session".
• 🔴 (11:24) Further packages/core/test/session-projector.test.ts recovery tests identified by exact name and line include: line 687 "redrives durable ordinary input execution intent after a lost advisory wake"; line 729 "redrives ordinary input without recovering a claim that crashed before promotion"; line 788 "redrives one existing recovery after its replacement claim crashes before promotion"; line 843 "never recovers a pre-promotion claim when input commits at the recovery boundary"; line 910 "rejects a recovered assistant from another Session"; line 957 "quarantines a recovery wake behind a newer assistant"; line 1072 "never promotes a recovery superseded before scanner validation"; line 1136 "quarantines an assistant-less recovery superseded before scanner validation"; line 1194 "never promotes a quarantined recovery continuation as an ordinary steer"; line 1250 "quarantines a malformed claim without blocking a valid claim"; line 1302 "bounds each deterministic recovery scan"; line 1347 "rolls back promotion and Context Epoch mutation after ownership loss"; line 1407 "never advances the Context Epoch after ownership changes"; line 1500 "rolls back claim renewal when Context Epoch advance fails".
• 🔴 (11:24) packages/core/test/session-prompt.test.ts tests identified by exact name and line include: line 147 "delegates execution continuation through SessionExecution"; line 310 "wakes execution when an exact prompt retry recovers a committed message"; line 330 "never wakes execution when retrying an already promoted prompt"; line 374 "durably records execution intent before advisory wake"; line 408 "replays historical prompt admission without execution intent as admit-only"; line 514 "never arms durable wake for a conflicting execution retry".
• 🔴 (11:24) packages/core/test/session-create.test.ts line 211 contains the test "replays one prompt lifecycle into a fresh target database".
• 🔴 (11:24) Newly identified packages/core/test/session-run-coordinator.test.ts test names and lines include: line 31 "joins a wake-started execution without forcing a successor"; line 141 "coalesces wakes received during active execution"; line 218 "interrupts active execution and clears its pending wake"; line 366 "starts one follow-up when a wake races with failure".