# Diagnostic Finalization Availability - Optimization ID: `diagnostic-finalization-availability-2026-08-05` - Owner: auto-oncall-services - Source references: branch-local replays of `ST-2865`, `TCID-12973`, and `APID-29231` - Implementation status: deployed to staging - Effective status: `optimized_verified` ## Bounded problem The former reserve was only an abort boundary: it stopped the long scenario turn, read the newest handoff, and immediately converted it to a bounded-failure packet. No parent model actually received the reserved time. After scenario work, all three measured replays also reached the independent 75-second final synthesis cap; that exception failed the analysis even though usable compact packets were already available. ## Implementation - Reserve up to 120 seconds of each scenario stage. At the boundary, stop and reap the acquisition turn, validate the current-attempt handoff, and start one fresh read-only, no-network, tool-free parent finalization turn from inline route and handoff data. - If reserved finalization succeeds, keep its parent-owned diagnosis. If the handoff was partial, preserve `execution.completion=partial` independently; this remains trace/checkpoint metadata and does not downgrade a successful parent judgment or the top-level evidence-acquisition status. - If reserved finalization fails, preserve the handoff observations, gaps, and artifact refs in the bounded scenario packet. - Keep the shared diagnostic pipeline at 20 minutes and the worker envelope at 25 minutes. Increase the no-tool synthesis cap from 75 seconds to 3 minutes; the existing stage allocator deducts that larger downstream reserve from the scenario acquisition window instead of extending the end-to-end deadline. - If final synthesis still times out, fails schema parsing, or uses a forbidden tool, return a deterministic `human_required` result from the preserved packets. Do not rerun successful evidence acquisition and do not turn the model-stage failure into a 500 response. ## Local verification - Runner tests cover complete and partial handoff takeover, no-tool finalizer execution, stream cleanup, partial evidence reporting, finalizer failure, forbidden synthesis tool use, and synthesis timeout fallback. - Configuration tests cover the unchanged 20-minute pipeline and 25-minute worker envelope plus the 3-minute synthesis window in both app-delivery templates. - Staging image `4edcfd9abeadd4c2a47ee25c3d8fb9537d630f13` rendered the expected 180-second synthesis, 1200-second pipeline, and 1500-second worker values. The pod remained Ready with zero restarts. - Deployed write replays completed without synthesis timeout or packet loss: `ST-2865` run `06876c15-de4f-46a5-9a7a-6e48d1e766e5`, `TCID-12973` run `60f1bb4f-3f2b-4382-ba4e-3f1e2c41ea0f`, and `APID-29231` run `de874fd8-c11c-416f-8b85-f57d59996f02`. All three returned complete response generation and complete evidence acquisition, then wrote DB runs `120016`-`120018`, evaluations, Feishu Base updates, and their existing Jira internal comments. - `TCID-12973`, which previously ended at the scenario reserve with only a partial packet, produced a complete handoff, five evidence artifacts, a terminal parent decision, and successful final synthesis. `APID-29231` completed its long Wetech evidence turn and synthesis inside the unchanged total pipeline deadline. ## Unresolved gaps - Larger windows remain headroom rather than the primary latency fix. The separate Clinic batch-orchestration record owns the measured Clinic latency reduction. - No deployed replay exercised the deterministic synthesis-timeout fallback; that branch remains covered by unit tests. The source-case failure mode is nevertheless verified because all three former timeout cases now completed synthesis within the bounded allocation. ## Next validation action Monitor later diagnostic runs for `diagnostic scenario synthesis timed out` or loss of preserved scenario packets. Treat any new matching post-verification evidence as a regression and replay the affected case before changing timeout allocation again.