PR #33079 real development verification, Linux, 2026-11-18 baseline github/main: 316c3256a3a09cef7ed68914ebe74c7183ac08e1 PR exact head: f0531e7f3d3b1af6b220867a9cee5ab1f529e619 vp i in ../main and PR checkout. One background dev server at a time. main ports: server 16733, web 8693; stopped before PR startup. PR ports: server 13773, web 5733. Data created through project/thread CLI, real running server, Codex gpt-6.1-sol. 54 completed provider turns, 109 conversation messages, 190 timeline items. One turn executed 25 separate shell tool calls. No mocks or seeded fixtures. Copied this isolated app-created state after stopping main, before starting PR. The older PR reports newer-database migration warnings on this copied state. Actual CLI command: node apps/server/src/bin.ts thread show --base-dir --messages <2|200> --json Fetch instrumentation recorded real MCP calls, HTTP response body bytes and elapsed time. Single observations, not a large-thread latency benchmark: {"build":"main","messagesRequested":2,"returned":2,"items":190,"calls":3,"bytes":98266,"mcpMs":63,"wallMs":998,"first":"Reply exactly TAIL_054. Do not use tools or change","last":"TAIL_054"} {"build":"main","messagesRequested":200,"returned":109,"items":190,"calls":4,"bytes":158544,"mcpMs":74,"wallMs":944,"first":"Reply exactly MAIN_TAIL_01. Do not use tools or ch","last":"TAIL_054"} {"build":"pr","messagesRequested":2,"returned":2,"items":190,"calls":1,"bytes":5516,"mcpMs":13,"wallMs":642,"first":"Reply exactly TAIL_054. Do not use tools or change","last":"TAIL_054"} {"build":"pr","messagesRequested":200,"returned":109,"items":190,"calls":2,"bytes":150662,"mcpMs":42,"wallMs":702,"first":"Reply exactly MAIN_TAIL_01. Do not use tools or ch","last":"TAIL_054"} PASS: latest 2 and all 109 messages identical, including order and full returned text. Live MCP via the same initialized CLI credential/session: Newest messages limit=2: positions [187,188], previousPosition=187, hasMore=false. Before 187 limit=2: positions [184,185], previousPosition=184, hasMore=true. Forward limit=2: positions [0,2], previousPosition=null, hasMore=true. Newest activity limit=2: positions [188,189], assistant reply and checkpoint. Before 0: empty items, previousPosition=null. limit=0: rejected with JSON-RPC -32602 by public tool schema. PASS: backward and forward pages match all 111 messages after PR provider turn, no duplicates or gaps. PASS: item text pagination: 9 chunks reconstruct original 61 characters. PASS: combined exclusive after/before bounds. PASS: PR CLI falls back against main server without previousPosition. PASS: main CLI can read PR server. PR CLI thread send --wait: completed, reply=PR_WAIT_OK; one backward messages read for result. Playwright with /usr/bin/google-chrome: PASS: composer send -> real Codex gpt-6.1-sol -> PR_UI_OK. PASS: full browser reload preserves PR_UI_OK and prior history; no page errors. Uploaded 4.8-second recording: playback to end and seeking passed in T3 player. Blacksmith existing focused tests: vp test run apps/server/src/cli/thread.test.ts apps/server/src/mcp/OrchestratorMcpService.activity.test.ts 2 files, 19 tests passed. vp test run apps/server/src/orchestration-v2/ProjectionStore.test.ts -t fork 2 selected tests passed; 51 deselected. Existing CI: 26 successful, 1 failure (Agent context budget). No code changes required for tested behavior. Not exercised: native mobile/desktop, Windows/macOS, relay/tunnel, 20k-item scale, failed-provider-run result recovery, real pre-beforePosition strict server, model-facing evals.