Client busy-loops (~85% CPU, zero syscalls) on long Fireworks reasoning streams — probabilistic, reproducible on 1.17.6 and 1.18.31

Summary

opencode run busy-loops (sustained ~85% CPU, zero syscalls, no output, no error) when a Fireworks reasoning model emits a long reasoning stream. The process never completes and never fails — it spins in userspace until killed.

Affected versions: 1.17.6 and 1.18.31 (latest at time of writing). Linux x64, Bun runtime, llm.runtime=ai-sdk.

Repro

timeout 90 opencode run --agent instruct \
  --model fireworks-ai/accounts/fireworks/models/glm-5p3-flash < prompt.txt

With a complex ~33KB prompt (a demanding extraction-style task), the process hangs until the timeout, printing only the agent banner. The identical prompt against the same model via plain HTTP (curl/httpx streaming) completes in ~20s, every time.

Characterization

• Mechanism: strace -f -p <pid> for 20s during the hang shows zero syscalls; the process sits at ~85% CPU. Single-threaded userspace spin. The log shows stream ... agent=instruct and then nothing — no error, no retry, no parts written.
• Probabilistic per run, scaling with reasoning volume:
  • Complex 33KB prompt: ~10/10 hangs (model emits ~9.3K reasoning deltas via API capture).
  • Same content with a "reason briefly" instruction in the prompt: 0/4 hangs.
  • Smaller slice (~7.4K reasoning deltas): ~50% hangs.
  • Synthetic large prompt with shallow reasoning (~600 deltas): never hangs, at any input size.
• Models: Fireworks reasoning models tested (glm-5p3-flash, glm-5p3, kimi-k3, minimax-m3; deepseek-v4p1-flash intermittently — its reasoning length varies per prompt, matching its intermittent hang rate). Anthropic models not affected (no per-delta reasoning_content events).
• Not the provider: ~100 direct API calls with identical prompts — including system+user messages, temperature 0, and 6-wide concurrency — all complete normally. SSE stream capture shows structurally normal events, finish_reason: stop, for both hanging and clean prompts.
• Not local config: reproduced with the project's plugin directory emptied; provider.<id>.models.<m>.reasoning: false and reasoning-briefing system prompts do not prevent it.

Best guess from the outside: something in the stream/reasoning-delta accumulation degrades (superlinearly?) with event count and, past a threshold, never terminates — the hang probability tracks reasoning delta count cleanly across all experiments.

Repro prompt available on request.

Reproduce with the provided timeout 90 opencode run command against the Fireworks model and a complex prompt, then observe the process with strace to confirm the zero-syscall spin. Trace the TypeScript ai-sdk Fireworks streaming and reasoning-delta handling used by opencode run; compare it with a clean short-reasoning run. Done means long reasoning streams complete normally with output and without sustained CPU spinning.