[rocklambros]

resolveSessionConfig() read the session store twice: once to decide whether to negotiate, and again after the handshake to choose what to return. Claude Code runs hooks in parallel and every hook in a session shares one store, so a second resolution could publish its own config between those reads. The first resolution then returned the other exchange's config, including its on_decision_failure. A step whose own handshake negotiated deny could fail by proceed if its decision delivery later failed. Reported privately.

The store is now read once, before anything awaits. A config already stored ends the call as before, with no handshake. Otherwise the config this exchange negotiated wins, including one it negotiated but could not store, which is what the function's existing comment already promised: the value in hand is authoritative for the step that negotiated it.

The store gets a second look only when this exchange negotiated nothing, for example when its own ServerHello was unusable. A config another resolution stored from the same Guardian is a better posture than the ACS default there. Returning only this exchange's value would have closed the race and failed open in that case instead.

Two tests in packages/host-adapter/test/client.test.ts pin both halves against a real handshake and a real memory store:

• An overlapping resolution publishes proceed right after this exchange stores deny. The call must return deny. It failed on the old code with proceed.
• This exchange's ServerHello is unusable while another resolution stores deny. The call must return deny. It passes on the old code and failed when the last line was changed to return only this exchange's value, so it guards the fallback.

bun run typecheck is clean and bun test gives 1112 pass, 1 skip, 0 fail across 1113 tests. The root README said the suite has 1,111 tests, which these two made false, so it now drops the total and keeps the count that matters: without trash on PATH, 26 tests still fail, verified by running the suite with it hidden. uv run pytest gives 278 passed, 1 skipped, and uv run mkdocs build --strict is clean.

Signed-off-by: Rock Lambros rock@rockcyber.com