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.
The 1.25.0 literal in diagnostics.mjs lagged the package bump, so a hot upgrade served 1.25.0 and the installer correctly rolled back on HOST_VERSION_MISMATCH. Advance the constant, add a release-efficiency guard that fails when the constant or compiled bundles drift from package.json, and add the 1.25.1 formal-release driver that refuses a candidate whose compiled server version is stale.
No product behaviour change beyond the version string; the reservation and install-scan fixes are unchanged.
The experimental-overhead case pushes complete input past the context window, so the engine input guard reports context_limit, not the reservation-branch budget wording. Assert the actual message and limitingReason; the guard still blocks dispatch with zero requests.
The root chain is checked before and after each scan; each file uses one lstat+realpath equality, nlink and stream-name checks, and size is rechecked after read. Directory inventory keeps full checks for directories. Same hashes and manifest scope; junctioned subdirectory, hard link, stream path and escape stay rejected. Warm per-file cost on a real 1.16 tree 2.08ms -> 0.17ms with identical hashes.
observed-route-max-v2 excludes at most one receipt whose reported/local input ratio is >=1.5x every other receipt on the route, with >=3 receipts and a later normal one. Excluded receipts stay in the ledger and are listed in excludedSources with their position; repeated highs, the latest receipt and short histories are never excluded. Sealed v1 reservations still verify byte-for-byte. Regression replays the original R evidence: v1 29623 > remaining 17429, v2 15285.