Matt Cockayne authored Aug 18, 2026 and Matt Cockayne committed Aug 19, 2026

Google's listing has moved since the table was last generated. Six models are
withdrawn and one is new:

  withdrawn  gemini-2.0-flash, gemini-2.0-flash-001, gemini-2.0-flash-lite,
             gemini-2.0-flash-lite-001, gemini-3-pro-preview,
             gemini-robotics-er-1.5-preview
  new        gemini-3.7-flash

The withdrawn entries are the part that bites: Capabilities() answered for
models Google no longer serves, so a caller checking before a call was told
yes and then failed on the request.

gemini-3.7-flash is generally available (3.7-flash-08-2026, no preview
marker) and reports the same limits and support flags as the current default
gemini-3.6-flash. That makes it what D1 selects on the latest tie-break, but
the default constant lives in the core, so the promotion lands in go/chat and
this module follows. The snapshot records the default this module is pinned
to today, which is still gemini-3.6-flash; it refreshes when the core bump
arrives.

Regenerated, not hand-edited, with the listing read on 2026-08-18.

Closes #1

v0.11.0 adds chat.BoundConversation, the seam a provider applies a history
policy through, and the conformance case that proves it runs.

It also moves DefaultModelGemini to gemini-3.7-flash, which is why this sits
on the capability-table refresh rather than on main: the table has to know the
model this module defaults to, or its own capability tests fail.

Config.HistoryPolicy has had no effect on this provider since it shipped. A
caller setting TruncateOldest(40) got a conversation that grew without bound.

genai owns the transcript once a *genai.Chat exists and exposes no way to
rewrite it — History(curated) reads, and nothing writes. So a bounded
conversation means a new session seeded with the turns that survived.
Chats.Create is purely local, validating the history and filling a struct, so
rebuilding costs no round-trip; and the session is only rebuilt when the
policy actually dropped or replaced something.

The bound runs before every request rather than once per call. The ReAct loop
appends a tool request and its results per step, so an entry-time bound never
sees the transcript that overflows. All three paths run it — Chat, Ask and the
streaming step.

Gemini differs from the other two providers in two ways worth stating:

  - The prompt travels as parts on Send and never enters the history, so the
    turn being answered is excluded without any special handling. What does
    need holding out is a trailing model turn carrying function calls, because
    the parts about to be sent are its responses.
  - Nothing is ever Pinned. AddCached creates a provider-side cache and
    references it by name, so cached content never enters the transcript.
    Where a cache cannot be created the content is staged inline exactly as Add
    would — and then it is not cached, so dropping it invalidates nothing.

askSession extracts the session open and bound together, which keeps Ask
inside the complexity limit rather than suppressing the check.

Refs chat#13