anthropic: adaptive-thinking turn silently ends with max_tokens and zero output tokens when reasoning exhausts the 32k budget #51517

Summary

An assistant turn using adaptive thinking can exhaust its entire response budget on reasoning and then end with max_tokens, delivering zero visible output tokens. The session goes idle with outcome: "succeeded", no error is shown, and the user only sees the agent "stop" — while reasoning tokens are still billed. Recurs across models, agents, and sessions (11 affected turns in the 12 most recent sessions).

Environment

• opencode version: 2.0.18 (channel: latest)
• OS: Windows 10.0.26200 (win32 x64)
• Terminal: VS Code integrated terminal (TERM_PROGRAM=vscode, xterm-256color)
• Shell: Git Bash (C:\Program Files\Git\usr\bin\bash.exe)
• Install/channel: release (latest)
• Active plugins: none configured

Reproduction

1. Configure an agent for an Anthropic model with adaptive thinking (e.g. anthropic/claude-opus-5-5, variant #max).
2. Ask a question that triggers a long reasoning phase (open-ended design/naming decisions, large context).
3. The turn streams for several minutes, then ends with finish: "length", rawFinish: "max_tokens", tokens.reasoning: 32000, tokens.output: 0, no text part in the message content.
4. The session marks idle with outcome: "succeeded" and the user receives no answer and no error.
5. Re-sending "continue" often succeeds, because the retry's reasoning phase happens to fit the budget — masking the failure.

Expected Behavior

• The thinking/reasoning allocation should be capped so that part of the max_tokens budget is always left for the answer, or
• a budget-exhausted, zero-output turn should produce a visible, retryable error/state instead of a silent success.

Actual Behavior

• Silent empty response. No entry in the server log (checked the incident window — only normal INFO events).
• Cost is incurred for the pure reasoning (observed $0.34–$2.69 per wasted turn).
• The observed per-request output ceiling is exactly 32,000 tokens (reasoning + output combined) — every failure shows reasoning: 31997–32000 and output: 0–3.
• Notably, the model catalog lists limit.output: 128000 for claude-opus-5-5, and no limit, outputTokens or related setting exists in the user's global or project config — the 32k cap therefore comes from request handling, not user configuration.

Additional Context

Affected turns observed (date UTC, agent, reasoning tokens, output tokens, cost):

Plus further near-misses with 2–3 output tokens (e.g. 2026-09-25 16:03, reasoning 31998, output 2).

Healthy control turns in the same sessions fit under the budget (e.g. reasoning 24089 + output 2999 → end_turn), confirming the cap is the differentiator.

Workaround: lower the reasoning level for thinking-heavy agents (#max → #high); "continue" retries when a turn comes back empty.