verification: linux, system chrome + repo playwright, 2026-11-01
main: 8b74a69019; pr: 75408277981a28f431834841ab369efde0b6fefa (unchanged)
vp i in both checkouts; one background dev server at a time.
project and provider instances configured through the actual web app.
CODEX_HOME values: /tmp/verify24664/home-a and /tmp/verify24664/home-b.
explicit home and shadow paths blank. real codex CLI 0.164.0, gpt-6.1-sol, low reasoning, local CLIProxyAPI.

main: ask to remember cedar in home a -> ready cedar.
switch to home b: no handoff row.
ask for remembered word -> cedar, but server logs:
[01:41:21.337] WARN: Recovering provider context in a fresh native session
reason: resume_failed; errorTag: ProviderAdapterResumeThreadError
main recovers automatically; missing proactive handoff and wrong-home resume reproduced.

pr fresh thread, matching prompts/model:
home a -> ready cedar.
switch a to b -> Context handoff pending.
ask remembered word -> cedar; Context handoff retained.
same-home followup -> same home cedar; handoff count stays 1.
switch b to a -> pending handoff, reply cedar; handoff count becomes 2.
rg resume_failed or Recovering provider in pr server log -> zero matches.
actual session files in both configured homes contain gpt-6.1-sol turn contexts.

blacksmith: vp i && vp test run apps/server/src/provider/Drivers/CodexDriver.test.ts
exit 0; 1 file passed, 15 tests passed.
GitHub checks queried at unchanged head: no failed/pending checks.

unverified: native desktop/mobile, macos/windows, direct OpenAI auth/inference,
relay/tunnel, managed/shadow accounts, explicit-home precedence in live turns,
trailing-space home in live turns (covered by existing driver regression).
recording demonstrates interaction, not a latency benchmark.