# Diagnostic Scenario Subagent Isolation - Optimization ID: `diagnostic-scenario-subagent-isolation-2026-07-30` - Owner: auto-oncall-services - Source references: `APID-29194`, `APID-29174`, `TCID-12920` - Implementation status: deployed to staging in revision `3efeb556` - Effective status: `optimized_unverified` ## Bounded problem The current diagnostic multiturn implementation keeps routing, the first provider investigation, and causal-gap acquisition in one Codex thread. Even when final synthesis starts fresh after successful evidence, the scenario work itself shares a large skill and provider transcript. Independent storage, write-path, capacity, TiCDC-stage, and network alternatives can therefore be suppressed or merged before they produce bounded findings. ## Candidate implementation - Build the route only from replay-frozen workspace inputs. Current or future Jira comments outside the replay cutoff never enter route or synthesis. - Select one primary and at most two orthogonal secondary scenarios. Signal guards add write-path plus storage boundaries and TiCDC plus independently observed capacity/network boundaries without issue-key hardcoding. Negated, healthy, normal, recovered, or ruled-out evidence never forces a route. - Use one isolated authentication/readiness thread for Wetech and serialize the complete browser-backed provider session through a per-account lock. - Manus does not use the shared Wetech bootstrap or provider lock. Each independent Manus scenario follows its provider contract with a unique output directory and may execute concurrently. - Run every scenario in a separate evidence-and-summary thread. Non-stateful provider or supplied-evidence scenarios may execute concurrently with unique evidence output paths. - Start final diagnosis in a fresh network-disabled, tool-free thread and pass only the frozen ticket packet, structured route, and compact evidence packets. Never resume an evidence thread after success, timeout, or failure. - Permit `human_required` and provider-none cases to analyze supplied evidence read-only while enforcing `human_required` as a monotonic final decision. Existing read-only providers may continue bounded diagnosis for hard-risk cases; provider-none requires frozen supplied evidence. ## Verification evidence - Contract tests cover the three source-case signal shapes: write-path plus storage; flow-control/storage plus latch and observed capacity; and TiCDC plus independently signaled downstream capacity/network. - A generic downstream transaction timeout does not add a network scenario. - A generic write-path case does not retain `query-execution` without a concrete plan, operator, scan-amplification, or digest-performance problem. - Runner tests require distinct route, evidence, and final thread IDs; reject tool use in route/final turns; preserve final continuation thread semantics; keep Wetech stateful execution serial; and allow Manus scenarios to execute independently without the provider lock. - Replay-isolation coverage injects a future support-only secret into the current `SupportCase` and verifies that neither route nor final prompt can see it, while frozen pre-cutoff evidence remains visible. - Local targeted build and tests pass. This is implementation evidence only; it does not prove runtime effectiveness. ## Unresolved gaps - The listed source cases, including APID-29174, are historical problem evidence only and have not received post-change deployed replays. - Deployed TCID-12973 demonstrated distinct route, scenario, and synthesis threads, but also showed that supplied-evidence scenario threads can scan an audit-only diagnostic summary from the shared workspace. Thread separation is therefore effective, while input isolation remains unverified. - Wetech stateful locking is process-local. Current staging uses one worker replica; a future multi-replica rollout needs a distributed provider-session lock or per-run isolated browser profiles. - Manus scenario concurrency is implemented locally, but Manus skill access and IP allowlisting are currently disabled outside this repository. The code change therefore cannot affect live Manus diagnosis until that external provider path is restored and the candidate is deployed. - Provider capability and query quality still depend on the selected provider skill. Scenario isolation prevents context contamination but does not make an unavailable metric or unsupported field available. - No Jira internal comment has been refreshed. ## Next validation action After deployment to staging, validate the structured route, scenario thread overlap, compact evidence packets, and tool-free final synthesis on fresh or newly eligible diagnostic cases. Do not rerun the already completed source cases solely for this concurrency change. For Manus specifically, wait until the external skill and IP path is restored, then use a fresh/new Manus case or a bounded synthetic staging validation to confirm independent scenarios overlap without a provider lock before promoting the problem to `optimized_verified`.