Fix: retry a step that ends without a finish reason, and never send Gemini a trailing model turn

[byapparov]

Context

On Vertex, gemini-3.8-flash sometimes ends a model step with no finishReason: the stream carries thought chunks only, then {"error":{"code":500,"status":"INTERNAL"}} inside an HTTP 200 stream. The AI SDK reports the step as finish "unknown" and raises no error. @aictrl/cli@0.4.3 treats unknown like tool-calls and continues the loop, rebuilding the history with the reasoning-only assistant message as the last turn. Vertex rejects that request with 400 Requests ending with a model turn are not supported, which is not retryable, so the session errors and the process exits 1. Root cause and captured request: aictrl-dev/aictrl#5645 (root-cause comment).

New evidence from fault injection with real Vertex (2026-09-29): the fault happened on its own in 3 of 14 local runs, at model requests 20–29, i.e. mid-run. At that point no output exists, so no harness-side output contract (a harness-side output contract) can rescue the run. Code Review is exposed too when a mid-run fault kills the CLI. This fix is the primary fix for #5645 and is independent of ADR-022.

Problem / Goal

A provider step that ends with no finish reason and no tool call must be treated as a failed, retryable step, not as "continue". No request to a Gemini provider may end with a model turn.

Code locations (from the #5645 analysis, CLI 0.4.3):

• session/processor.ts:251 stores finish = "unknown"; :429-431 returns "continue".
• session/prompt.ts:336-341 (top-of-loop exit) and :711 (modelFinished) exclude ["tool-calls","unknown"]; :725-734 continues.
• session/message-v2.ts:549-611 and :663 rebuild history keeping the reasoning-only assistant message last.
• No guard in provider/transform.ts:252 or session/llm.ts:241.

Reproduction

1. Put a logging proxy in front of Vertex (Vertex, location global) and run aictrl run --model google-vertex/gemini-3.8-flash --thinking --format json on a PR-Explanation-style prompt.
2. Have the proxy answer one main-loop request with HTTP 200 SSE: three thought-only chunks, then the bare {"error":{"code":500,"status":"INTERNAL"}} body and no finishReason (same bytes as the captured runs/11.proxy.jsonl).
3. Observed: NDJSON tail step_finish(unknown), message_complete(finish=unknown), then the next request has lastRole: "model", Vertex returns 400, session_error, exit 1.
4. Expected: the step is retried with a history that ends in a user or tool-result turn; the run continues.

Environment

• @aictrl/cli@0.4.3 (pinned in aictrl docker/executor/Dockerfile).
• Two production runs of a PR-Explanation-style task failed this way (details in the internal tracker).

Proposed Approach

1. unknown with no tool call = failed step. In processor.ts, set a retryable error (for example "provider ended stream without finishReason") and route it through the existing SessionRetry path, so the retry resends the same history ending in a user turn. Drop reasoning-only assistant messages from the rebuilt history (message-v2.ts). Cap the retries (for example 2 per step), then fail the invocation through the normal failure lifecycle (non-zero exit, consistent with Fix successful exit on provider error finish reasons #108).
2. Gemini guard (defence in depth). In provider/transform.ts message(), for @ai-sdk/google and @ai-sdk/google-vertex, drop trailing assistant messages that contain no tool-call. This alone would spin on empty steps, so it ships with (1), not instead of it.
3. Leave auxiliary requests alone. The CLI sends a concurrent auxiliary request with empty contents (title generation; Vertex answers 404). The retry counter, the guard and any request indexing must not count or alter it.

Acceptance Criteria

• A deterministic provider-stream fixture (thought-only chunks, embedded 500, no finishReason) triggers a retry whose request ends with a user or tool-result turn; the run then completes when the retried step succeeds.
• After the retry cap, the invocation fails with a structured, retryable-classified error; message_complete, session_complete, invocation_complete and the exit code agree (non-zero).
• For Google/Vertex providers, no outgoing request has a trailing assistant message without a tool call (unit test on transform.ts).
• The auxiliary empty-contents request is not counted by the retry logic and is sent unmodified (test).
• NDJSON records the retry (reason, attempt) so aictrl can measure it (reuse the diagnostics shape from Preserve provider finish reasons and bounded diagnostics #109 if merged first).
• Released as a new @aictrl/cli version; aictrl #5645 bumps the pin and re-runs PR Explanation and Code Review on gemini-3.8-flash.

Out of Scope

• Recovery from malformed tool calls (Evaluate bounded recovery from malformed model tool calls #111) and richer finish-reason diagnostics (Preserve provider finish reasons and bounded diagnostics #109). Link, don't merge.
• aictrl's harness-side output contract (ADR-022, milestone Q4 · File-Based Task Outputs).

Roadmap Alignment

• Pillar: EXEC (aictrl executor harness)
• Quarter: Q4 2026 (Growth Engine)
• Theme fit: Gemini 3.8 Flash is the managed default model, so this failure hits new self-serve users first.
• Decision gate impact: indirect: "self-serve signups converting" depends on first runs succeeding.

References

• aictrl-dev/aictrl#5645 (bug, root cause, production comparison)
• Fix successful exit on provider error finish reasons #108 (closed: non-zero exit on provider error finishes), Preserve provider finish reasons and bounded diagnostics #109 (open: preserve finish reasons), Evaluate bounded recovery from malformed model tool calls #111 (open: recovery evaluation)
• Fault-injection notes: organic fault in 3 of 14 local runs; see the auxiliary-request caveat above.

[byapparov]

Upstream status (anomalyco/opencode, formerly sst/opencode), checked 2026-09-29:

• v1 (packages/opencode, dev branch): not fixed. prompt.ts still treats unknown as "continue" at both checks (:1113 and :1295). The continuation still resends the history ending with the assistant turn. The v1 loop was made more consistent about continuing on unknown by #43892, merged 2026-08-21. The MAX_STEPS assistant prefill is also still there (prompt.ts:1281).
• Same error upstream on v1: #45359 (Gemini 3.7 Flash, v1.18.x) is open.
• #47034 (Gemini 3.8 Flash, the exact error) was closed on 2026-09-29 with "fixed on opencode v2". v2 is the rewritten runner in packages/core/src/session/runner/. A user asked there whether v1 will get the fix, and there's no answer yet. From the v2 source, the runner doesn't auto-continue after an unrecognised finish: continuation recovery is listed as future work in runner/llm.ts. That's probably why v2 doesn't hit this. This is an inference, not confirmed.
• Related open PR: #47864 drops empty reasoning parts and empty model turns on Vertex. It doesn't cover our case, which is a non-empty thought with a signature and no text or tool call.

So the fix has to land in our fork. Once it's proven here, it's a good candidate to contribute back to upstream v1 via #45359.