fix: SDK built-in retries stack with chat_with_retry causing silent 10+ min hangs

[xzq-xu]

Summary

OpenAI Python SDK and Anthropic Python SDK both have built-in retry logic (max_retries=2 by default). nanobot's LLMProvider.chat_with_retry / chat_stream_with_retry adds another retry layer on top (_CHAT_RETRY_DELAYS = (1, 2, 4), up to 3 retries + 1 final attempt). These two layers stack independently, resulting in up to (2+1) × (3+1) = 12 requests for a single transient failure — and zero user feedback during the entire wait.

Environment

• Components: LLMProvider (base.py), OpenAICompatProvider, AnthropicProvider, AgentLoop
• SDK defaults:
  • OpenAI: max_retries=2, timeout=Timeout(connect=5s, read=600s)
  • Anthropic: max_retries=2, timeout=Timeout(connect=5s, read=600s)

Problem

When an LLM backend returns a transient error (504, 503, timeout, etc.):

1. The SDK retries the same HTTP call up to 2 times (with its own exponential backoff).
2. After the SDK exhausts its retries and returns an error response, nanobot's chat_with_retry sees the error content, matches it against _TRANSIENT_ERROR_MARKERS, and retries at the application level (1s, 2s, 4s delays).
3. Each application-level retry triggers a fresh SDK call — which itself may retry up to 2 more times internally.

Worst-case scenario (504 with ~60s upstream timeout):

• 12 total HTTP requests × ~60s each = ~12 minutes of silent hanging
• No progress indication to the user
• Steering/interruption cannot break in (stuck in synchronous retry loops)

Additionally, the SDK's default read=600s timeout means even a single stuck request blocks for up to 10 minutes before the SDK gives up.

Expected Behavior

1. No stacked retries: nanobot should be the sole retry controller. SDK-level retries should be disabled (max_retries=0).
2. Reasonable timeouts: Read timeout should be bounded (e.g., 180s configurable) instead of the SDK's 600s default.
3. User feedback on retry: When retrying, the user should see a progress message (e.g., "AI service temporarily unavailable, retrying (1/3)…").

Affected Providers

Provider	Client	max_retries set?	Custom timeout?
OpenAICompatProvider	AsyncOpenAI(...)	No (SDK default 2)	No (SDK default 600s read)
AnthropicProvider	AsyncAnthropic(...)	No (SDK default 2)	No (SDK default 600s read)
AzureOpenAIProvider	Uses httpx.AsyncClient(timeout=60) directly	N/A (no SDK retry)	Yes (60s)

Suggested Fix

1. Set max_retries=0 on AsyncOpenAI and AsyncAnthropic client construction.
2. Set explicit timeout (configurable, default ~180s read + 10s connect) on both clients.
3. Add an on_retry callback parameter to chat_with_retry / chat_stream_with_retry so callers (e.g., AgentLoop) can surface retry progress to users.

Reproduction

1. Configure nanobot with an OpenAI-compatible provider pointing at a backend that returns 504.
2. Send a message — observe the agent hangs silently for several minutes.
3. Check logs: SDK-level retry lines (_base_client.py Retrying request) interleave with nanobot-level retry warnings.