server: turn silently ends when model response is truncated at max tokens (finish=length) #53989

Summary

When a model response is truncated by the generation token cap (finish: "length"), OpenCode ends the turn silently: no error, no assistant message, no auto-continue. The session goes idle with outcome: "succeeded" while the agent is mid-task. With always-reasoning models, the entire generation budget can be consumed by reasoning tokens alone (output: 0), so the turn ends before any tool call or text is produced — the agent appears to just "stop" partway through a task.

Environment

• opencode version: 2.0.18 (latest)
• OS: Windows_NT 10.0.26200 (win32 x64)
• Terminal: Unavailable: issue is server-side session behavior, not client-specific
• Shell: C:\WINDOWS\system32\cmd.exe
• Install/channel: desktop app CLI (%APPDATA%\ai.opencode.desktop\cli\2.0.18\opencode-cli.exe)
• Active plugins: none found

Reproduction

1. Use an always-reasoning model whose endpoint applies a modest default max_tokens when none is sent — observed with nebius/zai-org/GLM-5.3-Flash, variant high, no explicit max_tokens override (default cap observed: 8192).
2. Start a multi-step agentic task requiring many tool calls (e.g. iterate on a design and implement it).
3. Watch turns end: the assistant message contains only truncated reasoning_content, finishes with length, and the session silently goes idle. The agent does not continue and nothing indicates why.

Also reproducible with a raised cap: after adding providers.<provider>.models.<model>.body.max_tokens = 65536 (which worked and hot-reloaded without restart), the model consumed the entire 65536-token budget on reasoning (tokens: {input: 8275, output: 0, reasoning: 65536}) and the turn again ended silently with finish: "length".

Expected Behavior

• A truncated turn should be visible to the user — e.g. an assistant/system message stating the response was cut off at the max token limit — rather than a silent idle with outcome: "succeeded".
• Ideally, when a turn ends with finish: "length" and no tool calls, OpenCode would auto-continue the step (or at least prompt to continue), since the agent is by definition incomplete.

Actual Behavior

Session message data for each stop shows an assistant message with content: [{type: "reasoning", ...}] (truncated mid-sentence), finish: "length", rawFinish: "length", tokens {output: 0, reasoning: <cap>}, followed immediately by an idle event with outcome: "succeeded". Server logs contain no ERROR/WARN for these turns, so there is no signal a turn was truncated.

Frequency: consistent — 5 occurrences across 2 sessions on 2026-10-08 (4 at the 8192 cap, 1 at the 65536 cap after the config override).

Additional Context

• Model context: GLM-5.3-Flash is "always-on reasoning" (reasoning cannot be disabled, only effort lowered), so at high effort the generation budget is exhausted by reasoning_content before any tool call is emitted.
• Workarounds that help but do not fix the UX: explicit providers.<provider>.models.<model>.body.max_tokens override (verified: cap moved 8192 → 65536 with hot reload), lower reasoning-effort variant, or a stronger model. The silent stop itself remains whenever the budget is exhausted by reasoning.
• The outcome: "succeeded" idle event after a truncated turn is what makes this hard to diagnose: users see the agent stop with no error, and logs show nothing.