Summary

The claude-cli runtime in execute.runtime-rara7Weg.js hardcodes a 256KB stdout buffer limit (CLAUDE_LIVE_MAX_STDOUT_BUFFER_CHARS = 256 * 1024). When claude-cli emits a single stream-json line larger than this — common with extended thinking on Opus 4.7 / xhigh / max — the live session is aborted with "Claude CLI stdout buffer exceeded limit." and falls back to the next configured model. From the user's perspective this looks like a timeout.

Repro

• Provider: claude-cli/claude-opus-4-7
• agents.defaults.thinkingDefault: "high" (also reproduces with xhigh / max)
• Trigger: any turn where the assistant emits a large thinking block; we see this most often on multi-tool, multi-step turns

Observed today

8 overflow aborts in a single day on one host, all on claude-opus-4-7. Every abort triggered a fallback to openai-codex/gpt-5.5 (or xai/grok-4), so the user-visible symptom was inconsistent model/voice + perceived timeouts.

{"event":"model_fallback_decision","decision":"candidate_failed",
 "candidateProvider":"claude-cli","candidateModel":"claude-opus-4-7",
 "reason":"format","status":400,
 "errorPreview":"Claude CLI stdout buffer exceeded limit.",
 "nextCandidateProvider":"openai-codex","nextCandidateModel":"gpt-5.5"}

Source

execute.runtime-rara7Weg.js:

const CLAUDE_LIVE_MAX_STDOUT_BUFFER_CHARS = 256 * 1024; // line 306

// handleClaudeStdout, line ~601
if (session.stdoutBuffer.length > CLAUDE_LIVE_MAX_STDOUT_BUFFER_CHARS) {
  closeLiveSession(session, "abort", createOutputLimitError(session, "Claude CLI stdout buffer exceeded limit."));
  return;
}

// parseClaudeLiveJsonLine, line ~532
if (trimmed.length > CLAUDE_LIVE_MAX_STDOUT_BUFFER_CHARS) {
  closeLiveSession(session, "abort", createOutputLimitError(session, "Claude CLI JSONL line exceeded output limit."));
  return null;
}

Both checks share the same constant, so a single big assistant-message event with embedded thinking is enough to trip it.

Ask

Expose this as a config knob, e.g. agents.defaults.runners.claude-cli.maxStdoutBufferChars (or via env var OPENCLAW_CLAUDE_CLI_MAX_STDOUT_BYTES), with a sensible default bump (1–4 MB) for Opus 4.7-era thinking outputs. Would also help to log the actual line/buffer size on overflow so operators can tune.

Workaround

Setting agents.defaults.thinkingDefault: "adaptive" on Opus 4.7 (Anthropic's recommended setting per the extended thinking docs — fixed budgets were removed in 4.7) reduces the frequency, but does not eliminate the underlying issue: any large turn can still exceed 256KB.

Environment

• OpenClaw 2026.4.22 (00bd2cf)
• Node 22.22.2
• Linux
• Runner: claude-cli (cli)