v0.2.0

Added

• InputHandler interface — human-in-the-loop for agents (built-in ask_user tool + programmatic gates via InputHandlerFromContext)
• Workflow agent primitive — deterministic DAG-based task orchestration
  • Step types: Step, AgentStep, ToolStep, ForEach, DoUntil/DoWhile
  • DAG validation at construction (duplicate, unknown dep, cycle detection)
  • Shared WorkflowContext with concurrent-safe Get/Set
  • Implements Agent interface for recursive composition
• Ingestor — end-to-end extract → chunk → embed → store API with batched embedding
• Extractor interface with built-in PlainTextExtractor, HTMLExtractor, MarkdownExtractor
• Chunker interface with RecursiveChunker (improved sentence boundaries) and MarkdownChunker (heading-aware)
• StrategyParentChild — two-level hierarchical chunking with Chunk.ParentID linking
• Store.GetChunksByIDs for batch chunk retrieval (parent-child resolution)
• PDF extractor stub subpackage (ingest/pdf/)
• ObservedAgent in observer/ — OTEL lifecycle spans, metrics, and logs for any Agent execution
• Spawn() + AgentHandle — background agent execution with state tracking, cancellation, and select-based multiplexing
• ~110 new test cases across framework packages

Changed

• Switch ID generation from xid to UUIDv7 (RFC 9562) — standard format, native Postgres UUID type support
• ContentType is now string (MIME type) instead of int enum
• Remember tool uses Ingestor instead of manual embed+store boilerplate
• Search tool uses Chunker interface instead of ChunkerConfig

Removed

• Pipeline, PipelineResult, NewPipeline — use Ingestor
• ChunkerConfig, DefaultChunkerConfig(), ChunkText() — use NewRecursiveChunker()
• ExtractText() — use Extractor implementations

Fixed

• parseScheduledToolCalls partial unmarshal contamination causing duplicates