Features
- Amazon Bedrock support — Full Bedrock authentication (IAM credentials, bearer token, environment) with automatic model ID normalization to US cross-region inference profiles. Bedrock connections are routed through the Pi SDK for reliable model resolution. (Fixes #451)
- Extended prompt cache — New setting to cache prompts for 1 hour instead of 5 minutes. Reduces cost for long sessions on Claude models via Anthropic API. (Settings → AI → Performance)
- Docker headless server — Multi-platform Docker build for running Craft Agent as a headless server with automatic dependency resolution.
Improvements
- 1M context toggle moved — "Extended Context (1M)" setting relocated from Workspace → Advanced to AI → Performance for better discoverability.
- Headless server docs — Rewritten setup guide with install steps, startup configs, and TLS requirements.
Bug Fixes
- Bedrock model IDs — Fixed
400 ValidationExceptionand inference profile errors by mapping bare Anthropic model IDs to Bedrock-native US inference profile IDs (us.anthropic.claude-*). Also fixes storage migration for misconfiguredproviderType: 'bedrock'connections. (Fixes #451) - Bedrock bundle error — Fixed
Cannot find module './amazon-bedrock.js'by registering the Bedrock provider module for both copies of the Pi SDK in the bundled output. - Branch fork context — When SDK-level branch fork fails, parent conversation is now summarized via mini model as fallback context instead of silently losing history. (Fixes #454)
- MCP tool schemas — Stripped
$schemaproperty from external MCP tool input schemas that caused validation failures. - Pi SDK compaction race — Guarded against auto-compaction race condition in the Pi SDK subprocess. (Fixes #464)
- MCP server URLs — Stopped incorrectly appending
/mcpto MCP server URLs. - Interceptor cleanup — Network interceptor now actively strips cache TTL and context-1M beta headers to prevent conflicts with provider-specific handling.