PR #24087 real-environment verification
Main: 6575ce3250 (github/main)
PR: a3256e5ae51b587c9e16f90dc4aa3e971522f789
Linux, /usr/bin/google-chrome, repository Playwright, real Codex gpt-6.1-sol.
One dev server at a time; separate fresh state. Projects registered in the web UI.
Matching 56-turn conversations (ROUND-00 through ROUND-55), created through t3 CLI and real provider replies.
A read-only fetch observer cloned real responses to count calls/items/bytes; no mock transport or provider.

$ timeout 60 node apps/server/src/bin.ts thread show --base-dir <dev-home> <thread> --messages 10 --json
main: visible items=169, output messages=10, last=ROUND-55; read calls=2, returned items=112, response bytes=151660
pr: visible items=171, output messages=10, last=ROUND-55; read calls=2, returned items=67, response bytes=92596

$ timeout 30 node /tmp/compare.mjs
--messages 10: exact role/text match, 10 messages
--messages 1: exact role/text match, 1 messages
--messages 0: exact role/text match, 0 messages
--messages -1: exact role/text match, 0 messages
--messages 10000: exact role/text match, 112 messages

PR tail request: metadata limit=1, then view=messages, limit=100, afterPosition=70.
Main: view=messages, limit=100 from position 0, then next page; all 112 conversation messages read.
Both outputs are oldest first. Independent real histories have 169 and 171 visible items, with identical conversation text. These are transfer measurements, not a latency benchmark.

$ timeout 60 node apps/server/src/bin.ts thread show --base-dir <pr-home> <long-text-thread> --messages 2 --json
Full-text expansion PASS: all 27162 user characters match the submitted prompt; assistant=FULLTEXT-24087-OK.
Read continuation at textOffset=20000 was observed.

Real thread start --wait and 55 sequential thread send --wait calls per history all returned completed, exact ROUND markers.

Blacksmith checks on PR head:
vp test run apps/server/src/cli/thread.test.ts: 6 tests passed.
vp run check:changed 52c2b38ea41ee2271f5aca527009d36d25432689: format, lint, knip, GraphQL, affected typechecks, and adjacent tests passed.
The explicit base is the PR merge base; the worker initially lacked ancestry, then fetched the PR's history.

Tool-heavy PR thread: 110 real shell tool calls; 127 visible items; 14 conversation messages.
--messages 10/1/0/-1 matched the corresponding tail (or empty result).
--messages 15 widened backward: metadata read, afterPosition=26 window, then from the start; all 14 messages returned once, in order.
--messages 10000 also returned all 14 messages.
Final provider reply: HISTORY-24087-READY.