PR 26664 real dev verification, 2026-11-15 UTC
Main: github/main 47f2e91210 (separate /home/user/main worktree)
PR: 90eb8ec0d16a7d8798237a2fe72ae81da9b8c443, perf/linked-result-scan

Commands: vp i on both heads; vp run dev on main, stopped by captured process group before PR vp run dev --home-dir /home/user/t3code/.t3 --no-seed.
Client: repository Playwright + /usr/bin/google-chrome, actual web UI.
Provider: real Codex gpt-6.1-sol through the local proxy, parent and linked child.
A separate source CLI peer backend (no watcher or second Vite) was needed for linked delegation. Its state was isolated in /home/user/verification-peer.
Project /home/user/verification-repo was registered through each app's Add project flow. Linked environments were enabled and paired through Settings.

Original behavior reproduced on main: retrieving the completed remote task scanned all five message pages and transferred middle history, including text discarded by the follower.
The child generated 110 separate shell command executions and commentary; a steered provider message generated 25,992 characters of middle history. No mocks, fixtures, direct data insertion, or demo provider.
Matching comparison: the real pending parent database was snapshotted before completion, then restored for PR retrieval of the exact same completed 229-item peer history. Current main migration 92 makes a partial index incompatible with the older PR; only that isolated DB index was restored to the PR migration 86 definition. Application records and source code were unchanged. Downgrading a live user's database is not claimed supported.
A transparent localhost HTTP forwarder logged actual t3_thread_read requests/responses without changing bodies or adding latency. Bytes below are decoded whole HTTP JSON response bodies, not compressed network bytes.

MAIN success: 5 reads, 222316 decoded HTTP response bytes
2026-11-15T03:21:47.229Z 90ms view=messages after=start chars=20000 textOffset=- bytes=61935
2026-11-15T03:21:47.373Z 61ms view=messages after=81 chars=20000 textOffset=- bytes=62315
2026-11-15T03:21:47.449Z 39ms view=messages after=165 chars=20000 textOffset=- bytes=24304
2026-11-15T03:21:47.510Z 47ms view=messages after=193 chars=20000 textOffset=- bytes=59729
2026-11-15T03:21:47.540Z 15ms view=messages after=214 chars=20000 textOffset=- bytes=14033

PR success: 2 reads, 87016 decoded HTTP response bytes
2026-11-15T03:28:41.043Z 59ms view=messages after=start chars=20000 textOffset=- bytes=61935
2026-11-15T03:28:41.188Z 34ms view=messages after=199 chars=20000 textOffset=- bytes=25081

Both parent results exactly: LINKED_SCAN_SUCCESS_110_END

Neighbor: later run in the same remote thread, long final answer. The tail probe was inconclusive (hasMore=true), forward scanning resumed from the first page's next position, then an item-ID refill read textOffset 20000.
PR long-result fallback/refill: 7 reads, 313243 decoded HTTP response bytes
2026-11-15T03:36:02.356Z 104ms view=messages after=start chars=20000 textOffset=- bytes=61286
2026-11-15T03:36:02.514Z 116ms view=messages after=199 chars=20000 textOffset=- bytes=28075
2026-11-15T03:36:02.643Z 108ms view=messages after=79 chars=20000 textOffset=- bytes=61662
2026-11-15T03:36:02.667Z 12ms view=messages after=161 chars=20000 textOffset=- bytes=27796
2026-11-15T03:36:02.690Z 12ms view=messages after=193 chars=20000 textOffset=- bytes=60459
2026-11-15T03:36:02.711Z 10ms view=messages after=214 chars=20000 textOffset=- bytes=58633
2026-11-15T03:36:02.728Z 7ms view=item-refill after=start chars=20000 textOffset=20000 bytes=15332

Neighbor: intentional SIGKILL of isolated peer provider PID 99200, captured at subprocess spawn. The active run became failed; the follower selected its error instead of the prior run's long answer or current commentary.
PR real failed run: 12 reads, 655092 decoded HTTP response bytes
2026-11-15T03:42:59.625Z 30ms view=activity after=start chars=20000 textOffset=- bytes=60534
2026-11-15T03:42:59.670Z 24ms view=activity after=17 chars=1 textOffset=- bytes=62611
2026-11-15T03:42:59.745Z 49ms view=activity after=62 chars=1 textOffset=- bytes=61501
2026-11-15T03:42:59.827Z 41ms view=activity after=106 chars=1 textOffset=- bytes=61577
2026-11-15T03:42:59.863Z 18ms view=activity after=150 chars=20000 textOffset=- bytes=60609
2026-11-15T03:42:59.899Z 23ms view=activity after=166 chars=20000 textOffset=- bytes=60609
2026-11-15T03:42:59.931Z 20ms view=activity after=182 chars=20000 textOffset=- bytes=38576
2026-11-15T03:42:59.984Z 31ms view=activity after=193 chars=20000 textOffset=- bytes=60620
2026-11-15T03:43:00.052Z 39ms view=activity after=197 chars=20000 textOffset=- bytes=60609
2026-11-15T03:43:00.113Z 16ms view=activity after=213 chars=20000 textOffset=- bytes=53676
2026-11-15T03:43:00.141Z 10ms view=activity after=229 chars=20000 textOffset=- bytes=58998
2026-11-15T03:43:00.171Z 15ms view=activity after=237 chars=20000 textOffset=- bytes=15172

Read-only persisted-record assertions, checked against the actual child final answer:
{
  "success": "LINKED_SCAN_SUCCESS_110_END",
  "long": {
    "status": "completed",
    "length": 24425,
    "equal": true,
    "sha256": "47c2450c3a1cb8926764335f1c35b3715a1fe81e13a7c9e9b4595a001a32003a",
    "begin": "LONG_RESULT_BEGIN",
    "end": "LONG_RESULT_END",
    "numberedLines": 100,
    "everyLineValid": true
  },
  "failure": {
    "status": "failed",
    "result": "The provider event stream closed unexpectedly. Retry the turn; if it keeps failing, check the provider and server logs.",
    "runId": "run:thread:mcp%3Alaunch%3A70e5247f67eb938fbcd608a09daab307a8e9d028c806800de8aaa2bb0c56011d:ordinal:5"
  }
}

Blacksmith: vp test run apps/server/src/peer/RemoteDelegation.test.ts: 52 passed.
Blacksmith: vp run check:changed 4c2f04d35d371c2b2ecb04b283e282325026be8a: passed (format, lint, knip, GraphQL, 3 package typechecks, 54 tests in 2 files).
The runner needed public contributor branch history fetched because snapshots are shallow.
GitHub checks were all passing at the same head at 03:50 UTC.
No source changes, commits, pushes, merge, rebase, force-push, close, or label changes.

Not verified: native desktop/mobile clients; other providers; relay/tunnel/non-loopback networks; no-run recovery, cancellation races, model mismatch notes, concurrent growth, and long middle-page errors through real provider turns (focused tests cover these branches). No CPU/RSS or broad latency claim.