llm-council: no retry on 429/5xx — single transient failure kills the entire session

[23r2efewvcs]

Bug

callOpenAICompat and callAnthropic make a single HTTP request per model call. If the upstream returns a transient 429 (rate-limit) or 5xx (server error), the call fails immediately and the model's response is recorded as an error in Phase 1. There's no retry mechanism.

Free-tier endpoints (NVIDIA NIM, OpenRouter :free models) are especially prone to transient 429s when load is high. A single blip kills an entire council session with no recovery.

Current behavior

const res = await postJSON(url, payload, headers);
if (res.status >= 400) return { success: false, content: `[ERROR ${res.status}: ...]`, ... };

One shot. No retry. No backoff.

Proposed fix

• Introduce --max-retries N CLI flag (default 1, preserves current behavior — one attempt, one retry)
• Implement bounded retry with exponential backoff (2s, 4s, 8s...) for 429 and 5xx only
• After retries exhausted, fail fast with clear error message
• Connection-level errors (ETIMEDOUT, ECONNRESET) also retried once before failing

Usage after fix

node council.js run "<query>" --models "..." --chairman "..." --max-retries 3