Turns truncated at the output limit are recorded as normal completions

[iceteaSA]

Description

When a provider stops a turn because it hit the output-token limit, the stream reports finishReason: "length". The session loop treats that as a normal stop.

Two checks in packages/opencode/src/session/prompt.ts exclude only "tool-calls" (and "unknown" in the second) — never "length":

• :1113 — the entry gate that decides whether a session already has work to do
• :1295 — the in-loop decision after a provider turn

Nothing anywhere in src branches on "length". The message is stored with finish: "length" and no error, so downstream sees a turn that ended cleanly.

The worst case is a turn that produced nothing usable. A subagent on an OpenAI-compatible gateway spent its entire output budget inside the reasoning channel — the stored parts were step-start, reasoning (129,961 characters), step-finish, with no text and no tool call, and finish: "length". The task completed with status ok and empty output. From the caller's side that is indistinguishable from a subagent that had nothing to say.

MessageOutputLengthError already exists for this case and has no producer:

• packages/schema/src/v1/session.ts:36
• packages/core/src/v1/session.ts:48
• packages/opencode/src/session/message-error.ts:4

It is already a member of the assistant error union, and packages/opencode/src/acp/service.ts:842 already maps it to stopReason: "max_tokens". The plumbing was built for this and never connected to a source.

Steps to reproduce

1. Use a model whose output limit is reachable in one turn (the 32K OUTPUT_TOKEN_MAX clamp in packages/opencode/src/provider/transform.ts:18 makes this common).
2. Ask for output large enough to hit the limit — a large file write, or a long reasoning chain on a model that bills reasoning inside the output count.
3. Observe the session end. The stored assistant message has finish: "length" and no error; a task dispatch reports ok.

Reproducible in a test by queueing a reply with finishReason: "length" against the test LLM server.

OpenCode version

dev @ 1882c33

Expected

A truncated turn should be distinguishable from a completed one. A turn that produced text or tool calls before truncating is worth continuing; a turn that produced nothing usable should surface an error rather than reporting success with empty output.

Notes

This is item 4 in the root-cause chain of #18108 (P1 in that issue's checklist), filed here on its own because #18108 covers five interacting problems and this one is separable — it needs no change to repairToolCall, the thinking-budget coordination, the doom-loop detector, or OUTPUT_TOKEN_MAX.

Related: #38747 covers the same problem class on the V2 side (streams that end before their protocol completion boundary). #26167 touches the neighbouring code in prompt.ts for empty-stream truncation retries.

I have a fix with tests and will link it here.