Calling /compact remote with an OpenAI/Responses active model + a
non-remote compactionModel (e.g. an Anthropic summarizer) used to leave
remoteReady=true on the readiness check via shouldUseOpenAiRemoteCompaction(
this.model), but #compactWithFallbackModel walked the candidate chain
starting from the configured compactionModel and ran a local summary on it
— silently doing the opposite of the explicit mode.
Make the readiness check and candidate selection share one source of truth.
When /compact remote is requested and no compaction.remoteEndpoint is set
(an endpoint short-circuits per-model gating in compact()), filter the
candidate chain through shouldUseOpenAiRemoteCompaction so non-remote
fallbacks are skipped. If the filter empties the chain, warn and fall back
to the unfiltered chain so the operation still completes — matching the
spirit of the prior warning. The filter is threaded through
#getCompactionModelCandidates / #resolveCompactionModelCandidates and the
resolved candidates are passed into #compactWithFallbackModel so both
paths see the same list.
Added a regression test that wires an OpenAI active model with an Anthropic
compactionModel, invokes session.compact({ mode: 'remote' }), and asserts
the OpenAI model — not the configured compactionModel — is the first
candidate handed to compact().
Fixes#3104
roboomp
aa463f9008