Distillation

Date: Sep 16, 2026

• 🔴 (13:28) User’s execution-ownership requirement is that the system "never dispatches a provider after ownership changes".
• 🔴 (13:28) User’s execution-ownership requirement is that the system "never publishes streamed output after ownership changes".
• 🔴 (13:28) User’s execution-ownership requirement is that the system "never starts a local tool after ownership changes".
• 🔴 (13:28) User’s execution-ownership requirement is that the system "never fails interrupted tools after ownership changes".
• 🔴 (13:28) User’s provider/tool durability requirement is that the system "never commits continuation before unresolved hosted tools are reconciled".
• 🔴 (13:28) The inspected session-runner flow builds context with SessionHistory.entriesForRunner(db, session.id, system.baselineSeq) and treats currentStep >= agent.info.steps as the last step; on the last step it omits tool materialization, appends Message.assistant(MAX_STEPS_PROMPT), and sets toolChoice: "none".
• 🔴 (13:28) The LLM request sends headers "x-session-affinity": session.id, "X-Session-Id": session.id, and optional "x-parent-session-id": session.parentID; OpenAI promptCacheKey strips the ses_ prefix only when the session ID matches /^ses_[0-9a-f]{64}$/.
• 🔴 (13:28) Before provider dispatch, the runner calls compaction.compactIfNeeded({ sessionID: session.id, entries, model, request }); a positive result dies with continueAfterCompaction(currentStep).
• 🔴 (13:28) LLM event publication is serialized through Semaphore.makeUnsafe(1).withPermit; provider streams always run publisher.flush() via Effect.ensuring.
• 🔴 (13:28) For non-provider-executed "tool-call" events, the runner sets needsContinuation = true, obtains assistantMessageID through publisher.assistantMessageID(event.id), settles the tool inside Effect.uninterruptibleMask, and publishes LLMEvent.toolResult(...) with settlement.outputPaths ?? [].
• 🔴 (13:28) If tools are unavailable after the maximum agent steps, the runner calls publisher.failUnsettledTools("Tools are disabled after the maximum agent steps").
• 🔴 (13:28) Provider/tool failure handling uses these exact messages: "Provider did not return a tool result", "Tool execution interrupted", "Provider turn interrupted", and `Tool execution failed: ${message}`; user-declined tool settlement clears toolFibers, fails unsettled tools, and interrupts.
• 🔴 (13:28) The runner publishes SessionEvent.Step.Ended only when publisher.stepSettlement() exists and publisher.hasProviderError() is false; the event includes assistantMessageID, finish, cost: 0, tokens, optional ending snapshot, and optional changed files.
• 🔴 (13:28) The runner returns needsContinuation: !publisher.hasProviderError() && needsContinuation; stream failures propagate with Effect.failCause(stream.cause), as do interrupted tool-settlement failures.
• 🔴 (13:28) runAfterOverflowCompaction rejects a second overflow recovery with defect "Post-compaction provider attempt cannot recover another overflow" and otherwise retries after Effect.yieldNow.
• 🔴 (13:28) In the inspected session-event schema, durable session events use aggregate "sessionID" and version 1, while step-settlement events use aggregate "sessionID" and version 2.
• 🔴 (13:28) SessionEvent.Step.Ended is event type "session.next.step.ended" with fields assistantMessageID, finish, finite cost, token counts for input, output, reasoning, and cache read/write, plus optional snapshot and files.
• 🔴 (13:28) SessionEvent.Step.Failed is event type "session.next.step.failed" with assistantMessageID and UnknownError.
• 🔴 (13:28) SessionEvent.Text.Delta ("session.next.text.delta") is live-only, while SessionEvent.Text.Ended ("session.next.text.ended") is the replayable full-value boundary.
• 🟡 (13:29) File discovery identified event-related implementation files: packages/core/src/event.ts, packages/core/src/event/sql.ts, packages/core/src/session/event.ts, and packages/core/src/session/runner/publish-llm-event.ts.