Bundled CLI v2.1.88 broken on Cloud Run — subprocess transport timeout

[sugi-chan]

Problem

The Claude CLI bundled with claude-agent-sdk (v2.1.88) has a broken subprocess transport that causes Control request timeout: initialize when spawning judge board panelists on Cloud Run.

shutil.which("claude") returns None on Cloud Run (no system CLI installed), so simmer-sdk falls back to the bundled binary, which fails intermittently or consistently depending on the environment.

Observed behavior

• Judge board panelists time out during initialization
• The error is flaky — sometimes works, sometimes doesn't, depending on resource contention (2 CLI processes competing for CPU/memory)
• When it does "work" with the broken CLI, judges may return stale/default scores (possibly related to Regression detection should use primary criterion, not just composite #1's flat scoring issue)

Workaround

Install the system Claude CLI in the Docker image:

RUN apt-get update && apt-get install -y nodejs npm && \
    npm install -g @anthropic-ai/claude-code

This gives shutil.which("claude") a path to v2.1.90+ which works correctly.

Suggested fixes

1. Detect broken version — if the bundled CLI is < v2.1.90, warn and suggest installing system CLI
2. Add cli_path parameter to refine() — let callers specify a known-good binary path
3. Update bundled CLI — ship a working version with claude-agent-sdk
4. Fallback to direct API — if CLI init times out, fall back to AsyncAnthropic.messages.create() without tool access (degrades quality but doesn't crash)

Environment

• Cloud Run Jobs (1 CPU, 2GB memory)
• Python 3.13
• claude-agent-sdk bundled CLI v2.1.88
• System CLI v2.1.90 (installed via npm) fixes the issue