Pin package.json, package-lock.json and the literal APP_VERSION together,
add 1.25.4 to the controlled-release version allowlist, and record the
output-truncation recovery in the changelog.
When a turn stops at the output-token ceiling (output_truncated), the run no
longer dies. The truncated turn is still dropped whole - no tool is executed
and the invalid assistant message is never appended - but the model is given a
standalone instruction to split its proposals and the run continues. Recoveries
are capped per run (2); past that the original diagnostic still fails the run.
The connection is not cooled down for a truncation, since it responded fine and
only the model overran its budget, so the recovery turn can reach it. Both the
failed attempt and the recovering turn are accounted; neither is unknown.
Fixture regression covers recover-then-continue and the exhausted cap.
A response stopped at finish_reason=length (or Messages max_tokens /
Responses max_output_tokens) is still rejected whole - a complete leading
tool call is never applied because a later call may be cut mid-argument.
Previously the model only saw a generic incompleteness message and would
repeat the same oversized multi-tool turn. Now these paths raise a shared
output_truncated diagnostic telling the model to split proposals into one
file per response. Regression covers streaming/non-streaming Chat,
Responses and Messages, including the AB sample shape.
Formal 1.25.1 upgrade retained PID 36552/Provider/key and all data exact; measured update 156369ms vs 1.25.0 774098ms (~5x faster). The first upgrade attempt was safely rolled back by the HOST_VERSION_MISMATCH guard before the APP_VERSION constant fix. Sample W network diagnostic and single identity preflight passed; the bounded Node task stopped at no_progress after 6 reading-only requests without proposing the allowed files, so it did not close the loop and did not reach R's reservation wall. v2 benefit remains proven at unit level with R's real receipts; no Purge AB/BA run per plan.
Docs only; evidence stays under gitignored .qa.
Faithful restoration of the sealed R live-delivery driver set (initial/network/preflight/node) targeting formal 1.25.1 and the observed-route-max-v2 reservation fix. Connection revision is read from the live snapshot rather than hardcoded; state helpers reuse the committed retained-host-state module. Same fixed expense-report task, jojocode GPT chat connection and 16K/2048/8-request/90000-token caps. No key value is ever read or re-saved.
Drivers only; no product code change.