v1.14

Fixed

Anthropic provider failed every query with Anthropic API error 400 on larger installs (#80)

Root cause (confirmed live against the real Anthropic API): data requests such as get_entity_registry dumped the entire install (800KB+ on ~2000+ entity setups) into a single conversation message, exceeding Claude's 200k-token context window. The API rejected it with 400 prompt is too long, the integration discarded the error body, retried the identical payload 10 times, and the oversized message persisted in conversation history — making every subsequent query fail too.

• Data responses are now capped (~50k chars) with a truncation notice instructing the model to request narrower data; the per-request message window is also capped so multiple data messages can't stack past context or rate-limit budgets
• Provider error bodies are now surfaced to the UI (e.g. Anthropic API error 400: prompt is too long: 205547 tokens > 200000 maximum) instead of a bare status code
• Deterministic 4xx errors are no longer retried (previously 10 futile retries and ~45s of backoff); HTTP 429 now honors the server's retry-after header so per-minute token rate limit windows (e.g. Anthropic tier 1: 30k input tokens/min) can actually reset between attempts
• Failed queries are rolled back from conversation history so they can't poison subsequent queries; the history window can no longer start mid-turn after slicing
• Anthropic request timeout raised from 30s to 300s, matching every other provider
• Agent loop allows up to 8 iterations (was 5) since capped data responses can require additional narrower data requests

Full Changelog: v1.13...v1.14