Distillation

Date: Sep 17, 2026

• 🔴 (09:23) User asserted the session-recovery invariant: “A newer turn supersedes stale incomplete rows; never resume an older assistant projection.”
• 🟡 (09:23) The database projection code defines SessionAlreadyProjected extends Error, maps session/message data into SessionTable, MessageTable, and SessionMessageTable, and updates token usage through applyUsage(...); cache-read usage is accumulated with tokens_cache_read: sql\${SessionTable.tokens_cache_read} + ${value.tokens.cache.read * sign}``.
• 🟡 (09:23) The database projection adapter exposes message retrieval/update behavior including decodeRow(...) and getAssistant(messageID), handles shell rows where SessionMessageTable.type is "shell", and enforces the newer-turn supersession rule when selecting incomplete assistant projections.
• 🟡 (09:23) Event projectors shown include SessionV1.Event.Updated, message upserts using .onConflictDoUpdate({ target: MessageTable.id, set: { data } }), SessionEvent.ModelSwitched, SessionEvent.Prompted, SessionEvent.Tool.Called, and session deletion via db.delete(SessionTable).where(eq(SessionTable.id, event.data.sessionID)).
• 🟡 (09:23) Usage replacement logic subtracts an existing part’s usage before applying the replacement: if (previous) yield* applyUsage(db, row.session_id, previous, -1).
• 🟡 (09:23) The in-memory session projection also encodes the invariant “A newer turn supersedes stale incomplete rows; never resume an older assistant projection.” Its state contains messages: SessionMessage.Message[], tracks active shells through activeShellIndex(callID), and only mutates shell entries when current?.type === "shell".
• 🟡 (09:23) The in-memory projector identifies tool drafts with item.type === "tool" && (callID === undefined || item.id === callID) and centralizes owned-assistant mutation in updateOwnedAssistant(messageID: SessionMessage.ID, recipe: (draft: DraftAssistant) => void).
• 🟡 (09:23) The in-memory event handlers shown cover session.next.prompted, session.next.step.started, step completion/failure/interruption, session.next.text.started, tool-input lifecycle events, tool completion/error, and shell updates; failed assistant steps set draft.finish = "error", while recovery output uses IDs of the form recovery:${event.data.executionID} and status "error".
• 🟡 (09:23) Tool completion projection preserves timestamps by setting match.time = { created: match.time?.created ?? event.data.timestamp, completed: event.data.timestamp }.
• 🟡 (09:23) createLLMEventPublisher(events: EventV2.Interface, input: Input) tracks tools in a Map, lazily establishes an assistantMessageID, publishes session text/tool/assistant lifecycle events, and exposes interruptAssistant.
• 🟡 (09:23) The LLM event publisher treats invalid tool-event ordering as fatal defects with exact diagnostics including Tool input end before start: ${callID}, Tool input end before start: ${event.id}, and Tool input delta before start: ${event.id}.
• 🟡 (09:23) Before settling a tool result, the publisher ensures input has ended with if (!tool.inputEnded) yield* endToolInput(event), derives output using settledOutput(event.output, event.result), and associates emitted tool events with both callID and the tool’s assistantMessageID.