model_context_window_exceeded is treated as retriable transient error

[vsumner-spacebot]

Bug Description

When a worker hits the model's context window limit (model_context_window_exceeded), Spacebot treats it as a transient API error and retries with exponential backoff (5, 10, 20, 40, 80 seconds). Since the context is the same size on each retry, it fails every time, wasting ~3 minutes of retries and consuming API credits.

Reproduction

1. Spawn a worker that performs commands with large output (e.g., journalctl with hundreds of lines)
2. Worker accumulates enough tool results to exceed the model context window
3. The API returns stop_reason: model_context_window_exceeded with empty content
4. Spacebot retries 5 times with the same bloated context, failing each time

Expected Behavior

• model_context_window_exceeded should be detected as non-retriable — this is a context length problem, not a transient API error
• On context overflow, either: (a) truncate/summarize conversation history and retry once, or (b) fail fast
• No exponential backoff for a condition that won't resolve by waiting

Actual Behavior

• Treated as transient error, retried 5 times with delays of 5, 10, 20, 40, 80 seconds
• All retries fail with the same error
• Total wasted time: ~155 seconds
• Total wasted API calls: 5

Log Evidence

WARN spacebot::llm::model: unexpected empty assistant_content from Anthropic stop_reason="model_context_window_exceeded"
WARN spacebot::agent::worker: transient provider error, backing off and retrying attempt=1..5 delay_secs=5..80
ERROR spacebot::agent::worker: worker transient error retries exhausted retries=5

Impact

• Wastes API credits on guaranteed-to-fail retries
• Adds ~3 minutes of delay before the user gets the failure
• Can cascade: the main agent then tries to retrigger with the same bloated context

Suggested Fix

1. Match on stop_reason == "model_context_window_exceeded" and classify as non-retriable
2. Optionally: implement automatic context truncation (summarize oldest messages) and retry once
3. Alternatively: add a max_input_tokens check before API calls to proactively prevent overflow

Environment

• Spacebot version: 0.3.3
• Model: zai_anthropic/glm-5-turbo
• Observed when workers run journalctl with large output