What's New in v2.2.0
Added
- Stateful conversations: when
POST /message (or /message/stream) supplies a conversation_id, history is persisted via Strands FileSessionManager under <workspace>/.blueclaw/sessions/<id>/. Subsequent requests with the same id replay prior turns. Omitting conversation_id keeps stateless behavior.
conversation_id field on RunTrace and RunRecord (also exposed in /api/traces summary) so traces and history rows can be grouped by conversation.GET /playground — single-page chat UI bundled with blueclaw serve for manually exercising stateful + streaming conversations. Defaults its server URL to the current origin; bearer token entered in the sidebar. Unauthenticated like /health.docs/models.md — detailed model-support reference: per-provider blueclaw.yaml and .env samples (Anthropic, OpenAI, Ollama with tool-calling shortlist, LiteLLM for Gemini/Bedrock), CLI override precedence, cost tracking, and "adding a new provider" recipe.- Prompt-cache token billing:
calculate_cost() accepts optional cache_read_tokens / cache_write_tokens and bills them at 0.1× / 1.25× of the input rate (Anthropic prompt-caching multipliers). All three call sites (session.py, server.py, testing.py) pull cacheReadInputTokens / cacheWriteInputTokens from accumulated_usage, avoiding cost overstatement on cache-heavy sessions.
PRICING_UPDATED constant in blueclaw/models.py — explicit "last reviewed" date for the pricing table; bump it whenever MODEL_PRICING_PER_M is edited.
Changed
build_trace_and_record(...) accepts an optional conversation_id kwarg.build_system_prompt(...) accepts include_history. create_agent automatically passes False whenever a session_manager is attached, so the system prompt no longer narrates a "Recent History" recap that overlaps with the messages the session manager replays. Stops the model from prefacing each stateful reply with a conversation summary.build_system_prompt(...) accepts channel ("terminal" or "api"). create_agent(channel=...) threads it through; blueclaw serve passes "api" so HTTP responses follow chat-client tone rules ("answer only the new question, do not recap, no terminal-only constraints") instead of the CLI's strict plain-text rules. Fixes drift where stateful API replies grew progressively chattier and recapped earlier turns once the model had its own markdown-formatted prior messages replayed back to it.- Pricing table renamed to
MODEL_PRICING_PER_M and switched from per-1k to per-1M token units to match Anthropic's pricing page 1:1 (e.g. Sonnet input is now 3.0 rather than 0.003). Old MODEL_PRICING name kept as an alias for backward compatibility.
Notes
- Concurrent requests for the same
conversation_id are serialized by an in-process per-id lock, acquired before the global concurrency semaphore. Different conversation ids run in parallel (subject to the existing max_concurrent_runs cap).
- Session directories are purged on server start by
purge_old_sessions(trace_retention_days) (no new config knob).
Installation
pip install blueclaw==2.2.0
Links
