Distillation

Date: Sep 17, 2026

• 🔴 (08:36) User stated that an expired Session execution claim never renews itself.
• 🔴 (08:36) User’s lease guidance says timestamps use node wall clocks and SESSION_EXECUTION_LEASE_MS must exceed the deployment’s maximum clock skew plus one renewal interval; execution IDs immediately fence a node after takeover.
• 🟡 (08:36) packages/core/src/session/execution/claim.ts defines SessionExecutionClaim.Claim with readonly sessionID, executionID, and ownerID; Lost reports Session execution claim lost: ${claim.executionID}.
• 🟡 (08:36) packages/core/src/session/execution/claim.ts fences update, renew, settleCancellation, and clear by matching session_id, execution id, owner_id, and requiring expires_at > now; unsuccessful mutations die with SessionExecutionClaim.Lost.
• 🟡 (08:36) SessionExecutionClaim.update(...) renews expiry to now + SESSION_EXECUTION_LEASE_MS while enforcing phase transitions: "cancelled" requires current "cancelling"; "cancelling" excludes current "cancelled"; all other updates exclude "cancelling" and "cancelled".
• 🟡 (08:36) SessionExecutionClaim.transaction(...) runs update(db, claim) before the supplied effect in a database transaction with { behavior: "immediate" }.
• 🟡 (08:36) packages/core/src/session/execution/local.ts creates a process-specific ownerID with crypto.randomUUID(), tracks claims in Map<SessionSchema.ID, SessionExecutionClaim.Claim>, and claims execution by inserting SessionExecutionTable rows with a fresh execution ID, phase "ready", latest assistant sequence or -1, and expiry now + SESSION_EXECUTION_LEASE_MS; conflicts use .onConflictDoNothing().
• 🟡 (08:36) Local Session execution races input.run(claim, force, recovery?.continuationMessageID) against repeated SessionExecutionClaim.renew(...) calls scheduled every leaseDuration / 3.
• 🟡 (08:36) Local execution cleanup removes the process-local active claim; successful drains call SessionExecutionClaim.clear(...). Failed drains retain claims in cancellation states, retain interrupted non-"safe" claims, and otherwise clear known phases other than "unknown" or "continue".
• 🟡 (08:36) Interrupted Session settlement in packages/core/src/session/execution/local.ts renews the claim, clears already-"cancelled" claims, rejects phases other than "cancelling", clears when no assistant message is attached, and marks a completed assistant as "cancelled" before clearing.
• 🟡 (08:36) If interrupted provider-step outcome is unknown, local settlement publishes SessionEvent.Step.Interrupted with the Session ID, assistant message ID, current DateTime, and error { type: "unknown", message: "Provider Step outcome unknown after interruption" }; its commit changes the execution phase to "cancelled" and clears assistant_message_id, after which the claim is cleared.
• 🟡 (08:36) SessionExecutionLocal.interrupt(sessionID) falls back to coordinator.interrupt(sessionID) when no active local claim exists; otherwise it changes the claim phase to "cancelling" and races coordinator interruption/settlement against lease renewal every leaseDuration / 3.
• 🟡 (08:36) SessionExecutionLocal resolves a Session through SessionStore.Service, dies with Session not found: ${claim.sessionID} when absent, and runs SessionRunner.Service.use((runner) => runner.run({ ...claim, force, recoveryInputID })) under the Location layer returned by LocationServiceMap.Service.
• 🟡 (08:36) SessionExecutionLocal.node is built with makeGlobalNode and depends on Database.node, EventV2.node, SessionStore.node, and LocationServiceMap.node.
• 🟡 (08:36) packages/core/src/session/input.ts recovery quarantine performs commit(row) and then updates the matching pending SessionRecoveryTable row with recovery_error: defect.reason, constrained by execution ID, wake_pending = true, and null recovery_error, inside an immediate transaction; it dies with the original defect when quarantine succeeds or Recovery quarantine failed: ${defect.executionID} when no row is quarantined.