Approved extension: operation-driven continuation

Status: Backend delivery is complete. PR #1186 was merged to dev by chris-palatial on 2026-08-31 as f7b1c86ad39bdb526535e0fbdfc7223397b33f02. Both the Worker test/deploy workflow and Backend test/build/deploy workflow completed successfully for that revision. All six deployed nonpaid API checks passed at 2026-08-31 05:32 UTC: exact Worker revision; unauthenticated operation rejection; numeric artifact rejection; mixed-input rejection by the new contract; unbound-draft rejection; and no turns or Worker binding after rejection. The fresh Dev fixture is 6a9511de9b7e3692cfaacb64; no provider call or original-case mutation occurred. Frontend adoption and real-model production latency measurement remain separate work. The sections following this approved extension describe the earlier turn recovery release.

Backend support evidence

• Four coherent commits: 49c1809d Worker operation/receipt/continuation; 3908abf8 BFF stream/projection; 6e227093 exact-request stream repair without batch fallback; 2eb1b29d generated-reference camera-view analysis removal.
• Worker generated contracts/typecheck and all 1,793 tests pass. Backend production TypeScript check, 6,716 unit tests (43 existing skips), and 191 HTTP integration tests pass.
• Agent Worker PR CI passed on final head 2eb1b29d.
• All 20 exact-worktree HTTPS API checks pass at https://proxy.com:5174 with fresh conversations and a loopback fake model. Both initial and regenerated selection produce zero extra image-preparation calls and do not resend image bytes to the planner.
• Acceptance confirms the receipt and Mongo/Worker restore before a blocked model responds, continuation while sibling images remain in flight, exact replay without another model call, one identity-only user turn, explicit Newton replacing an old simulator question, and rejection of stale/cross-conversation grids and malformed inputs.
• The final local sample returned the committed receipt in 116 ms; the fake-model result followed in 276 ms. This proves ordering only and is not a production latency claim.
• Isolated real Wrangler HTTP tests additionally cover Worker replacement, newer selections/configuration, late sibling completion, and model failure after a committed selection.
• Self-review found and fixed two issues: a failed receipt/projection must retry the same operation instead of silently waiting on a batch fallback; generated single references were still invoking uploaded-photo camera-view analysis before planning. Uploaded-photo and multi-view analysis remain unchanged.
• The shared stack was restarted with Chris's authorization. Actual Backend/Worker process commands serve backend-agent-operation-continuation; Frontend remains frontend-variant-system-single-switch. Normal model configuration has been restored, the local fake model stopped, and the full stack restarted with healthy API/proxy/Queue probes.
• This Backend release preserves the existing overall framework and does not activate the new Frontend path. CAD, post-submit selection, and existing selection clients remain unchanged. No browser acceptance, paid-provider calls, or original-case mutations were performed.

User path and ownership

• After regenerating references, the user selects an image and expects the Dashboard Agent to respond to that selection using the latest conversation and configuration.
• The Worker owns the accepted reference, configuration, next decision, and readiness. Nest owns authentication, the durable run request, stream replay, and Mongo projection. The browser renders the committed receipt and subsequent Agent response.
• The next response is a fresh decision. Previously displayed questions are history, not a queue to restore. If the user has since said “Use Newton”, the simulator is already answered. Required unanswered choices still block submission.

Confirmed contract

• Preserve the existing overall framework, as Chris reiterated on 2026-08-31: reuse the current Agent loop, run/checkpoint recovery, model/tool execution, and stream. Do not introduce a separate Agent, scheduler, or question-restoration mechanism.