Provider stream truncation (finishReason="other" with zero output) silently accepted, persisting half-finished assistant messages

[edevil]

Bug

When an upstream provider stream is cut mid-generation, the AI SDK emits
a fallback finishReason: "other" with zero output tokens. opencode's
session processor accepts this as a normal end-of-step, persists a
truncated assistant message, and continues — no error, no retry, no
recovery indicator.

Symptom

The session stores a half-finished message and the user gets a partial
response with no warning. Users typically respond by manually
re-prompting, get a few good turns, then hit the bug again — the
"session degradation" pattern reported in #16214.

Trigger

{ type: "text-delta", delta: "..." }
{ type: "text-delta", delta: "..." }   ← upstream stream cuts here
{ type: "finish", finishReason: "other", usage: { outputTokens: 0 } }
                                                 ↑
                       AI SDK's "no stop reason was given" fallback

The session processor sees finish-step with
finishReason === "other" and usage.tokens.output === 0.

Real-world evidence

I found more than a dozen instances of this exact pattern in my own
opencode session database, spanning two providers (anthropic, openai)
and four models (gpt-5.3-codex, claude-opus-4-6, claude-opus-4-7,
claude-haiku-4-5). All exhibit the shape:

{
  "role": "assistant",
  "finish": "other",
  "tokens": { "input": 0, "output": 0, "reasoning": 0, "cache": {...} },
  "cost": 0
}

In one diagnostic example the reasoning text literally ends mid-word
("...whichlang::detect_language() functi"), proving the upstream stream
was severed before the next chunk arrived — this is not a model decision
to stop.

Why this is distinct from related reports

• Either the < or <= operators are causing responses from the ai to get cut off. #23928 — that thread's investigation concluded the user's symptom
  was an upstream proxy synthesizing finishReason: "stop" after
  truncating content (free-tier MiniMax over OpenRouter), not "other".
  The mechanisms differ.
• fix: handle stream interruption for OpenAI-compatible providers #21727 — addresses a similar truncation pattern but only at the
  @ai-sdk/openai-compatible provider's flush() callback, missing
  Anthropic-direct, Bedrock, and Vertex.

Suggested fix

Detect finishReason === "other" && tokens.output === 0 at the session
processor layer (covers all AI-SDK providers), surface as a retryable
APIError, cap retries at a small number to avoid loops on persistently
misbehaving providers.

Environment

• opencode: dev (post-merge of fix(provider): preserve assistant message content when reasoning blocks present #21370)
• Models confirmed affected: claude-opus-4-7, claude-opus-4-6,
  claude-haiku-4-5, gpt-5.3-codex