Gemini 400 "Requests ending with a model turn are not supported" when a step is stored out of order: Identifier.create is not monotonic across backward clock steps #55514

Description

Google provider (@ai-sdk/google), gemini-3.8-flash, v1 line (seen through Kilo 7.8.8, which ships the v1 packages/opencode unchanged in the relevant files). Mid-way through a tool-calling turn a step fails with:

AI_APICallError: Requests ending with a model turn are not supported.   (400 INVALID_ARGUMENT)

Not retryable, so the turn stops until a new prompt is typed. Same error as #45359 / #47034, but a different trigger from the max-steps prefill, and #55382 alone does not cover it.

Root cause

1. Identifier.create (packages/opencode/src/id/id.ts) takes Date.now() and only handles two IDs in the same millisecond. When the wall clock steps backward between two parts of one step, the later part gets a smaller ID. Clocks do step backward: NTP corrections, VM resume, and WSL2 (where this box steps by seconds every few seconds).
2. Parts are replayed in ID order (toModelMessagesEffect reads PartTable ordered by id).
3. A Gemini step writes step-start > reasoning > tool. With a backward step in between it is stored as tool > step-start > reasoning (or tool > step-finish > step-start > reasoning).
4. convertToModelMessages opens a new block at every step-start, so the message replays as assistant(tool-call) > tool(result) > assistant(reasoning).
5. @ai-sdk/google emits that trailing reasoning-only block as a model turn with thought: true, and Gemini rejects contents that end on a model turn.

Stored parts of the step right before one failure (UTC, in ID order):

msg_13bb9f89400124loVqN16mcfHL  assistant, finish=tool-calls
  prt_13bba0cca001…  18:43:31.146  tool (bash, completed)   <- before its own step-start
  prt_13bba0e0a001…  18:43:31.466  step-start
  prt_13bba0e0c001…  18:43:31.468  reasoning   time.start=…411468  time.end=…411302  <- end before start
  prt_13bba11a6001…  18:43:32.391  step-finish
msg_13bba12cf001hH63Waabr4oSPP  assistant, error "Requests ending with a model turn are not supported."

In one local database: 74 occurrences across 24 sessions; 72 directly follow a step stored like the one above, the other 2 follow an auto-compaction continuation whose summary assistant message was last. Over three days, 0.5–1.5 % of steps for every provider in use (openai, xai, kimi, google) are stored with a tool or reasoning part ahead of the first step-start. Only Gemini rejects it; the others silently get the reasoning replayed after the tool result.

Replaying that shape against gemini-3.8-flash directly: [user, assistant(tool-call), tool(result)] is accepted, adding the trailing assistant(reasoning) gives the 400, dropping it again gives 200.

V2

packages/schema/src/identifier.ts on dev has the same lastTimestamp/counter logic, so IDs can still go backward there. V2 orders history by a seq column, so I have not shown the replay symptom on dev; I only verified v1.

Fix

PR against v1 follows: make Identifier.create strictly ascending across a backward step, and have ProviderTransform.message drop a trailing reasoning-only assistant message for the Gemini SDKs and prime a user turn behind any other assistant tail (outbound only). Kilo tracks this in Kilo-Org/kilocode#15073 and will pull the fix once it lands here.

OpenCode version

v1 at a95d9c0 (2026-10-14); first seen on Kilo 7.8.8.

Operating System

Linux (WSL2)