plat 219

← Pulse platform issue index

SQLite backlog reconciled — 2026-09-05

Resolved PUL-5D2B9495 in the corresponding workflow databases after checking the existing implementation and passing focused regression tests. Full evidence and the complete remaining inventory: reconciliation audit. This is internal tracking closure, not a claim of a new deployed end-to-end run. Previous SQLite records are retained in audit events; unrelated findings remain open. No business data or historical schedule outcome was rewritten.

PLAT-219 — scheduled turns now honor a linked full run’s durable failure instead of waiting hours for a missing completion callback

Coordination	Value
Assigned agent	Codex
Ticket state	implemented; runtime reverify
Last synchronized	2026-08-29

• Priority: P0 — a failed workflow remained projected as running for about 59 minutes and was finally mislabeled interrupted: server restarted.
• Owner: scheduled conversation-turn lifecycle and full-workflow launch tracking.
• Related: Sales Outreach PUL-5D2B9495. LinkedIn PUL-3565D07C is an older stale-runtime projection and is not claimed as the same root cause.

Reproduction

Sales Outreach schedule 633c499d started the Dubai full run at 10:31:30Z. The run’s authoritative runs/iteration-0/dubai-real-estate/run_metadata.json recorded status=failed and completed_at=10:44:12Z. The wrapping schedule remained running until a server restart at 11:43:12Z, then recorded the restart rather than the already-durable workflow failure.

waitForConversationTurnTree treated any registered running child as proof of live work and reset its ten-minute inactivity clock repeatedly, up to the three-hour live-child ceiling. A missed full-run completion callback therefore overrode the run’s own terminal record.

Fix

• Full-run launch records now include execution_type, iteration, group and exact run_folder immediately, rather than only adding those fields if the completion callback eventually arrives.
• The tracked execution preserves that launch metadata and run folder.
• When the root conversation turn is complete, a linked child still says running, and no progress has occurred for 30 seconds, the waiter checks that exact run’s run_metadata.json every five seconds.
• An explicit durable failed status ends the turn with the workflow failure. Missing/unreadable/running metadata fails open, and completed is not force-settled because a successful run may still legitimately need its completion notification to resume follow-up work.

Verification

Focused tests prove launch metadata survives into the tracked execution and a durable failed descendant is detected while a completed descendant is not misclassified:

go test ./pkg/orchestrator/agents/workflow/step_based_workflow -run 'TestRunFullWorkflow|TestWorkshopExecution' -count=1
go test ./cmd/server -run 'TestConversationTurn|TestDurableFailedWorkflowDescendant|TestWorkshopExecutionWithoutExplicitParent' -count=1

The next real failed scheduled full run is required to confirm the schedule closes with the workflow failure without a restart.