Skip to content

CHANGELOG.md

Latest commit

 

History

History
210 lines (158 loc) · 8.67 KB

CHANGELOG.md

File metadata and controls

210 lines (158 loc) · 8.67 KB

Changelog

All notable changes to Bedsheet Agents will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.4.8] - 2026-01-26

Changed

  • GCP Model Default - Updated bedsheet init --target gcp to use gemini-3-flash-preview (was gemini-2.5-flash)
    • Reflects latest Gemini model availability

Added

  • Demo Recording Plan - Added comprehensive demo recording guide (docs/DEMO_RECORDING_PLAN.md)
    • 3-4 minute multi-cloud deployment demo script
    • Pre-recording checklist and dry-run tests
    • Contingency plans for common issues
  • GCP Deployment Deep Dive - Added link to gcp-deployment-deep-dive.html in README and CLAUDE.md

[0.4.7] - 2026-01-22

Changed

  • make ui Improvements - Better first-time experience
    • Checks if cloud-run-proxy component is installed before starting
    • Clear instructions if component missing: gcloud components install cloud-run-proxy
    • Opens browser directly to /dev-ui/ path
    • No more interactive prompts during make ui

[0.4.6] - 2026-01-22

Added

  • make ui Command - One-command access to deployed agent's Dev UI
    • Automatically handles Cloud Run authentication via gcloud run services proxy
    • Opens browser to http://localhost:8080 with authenticated tunnel
    • No manual token handling required

[0.4.5] - 2026-01-22

Added

  • Project Consistency Check - Preflight now detects mismatches between terraform.tfvars and gcloud config
    • Warns when projects don't match
    • Offers to auto-fix by running gcloud config set project
    • Prevents silent deployment to wrong project
  • Project Consistency Documentation - Added warning section in DEPLOYMENT_GUIDE.md

[0.4.4] - 2026-01-22

Added

  • Credential Preflight Check - New make preflight command warns if GOOGLE_APPLICATION_CREDENTIALS is set
    • Detects potential cross-project credential issues before deployment
    • Prevents silent 403 errors at runtime
    • Interactive prompt to continue or abort
  • Credential Warning in Docs - Prominent warning box in DEPLOYMENT_GUIDE.md explaining SDK credential priority
  • Troubleshooting Guide - Added "#1 gotcha" section for credential issues

Changed

  • make deploy and make dev now automatically run credential checks
  • Updated preflight checks list in deployment guide

[0.4.3] - 2026-01-22

Fixed

  • CLI Version Display - Now uses importlib.metadata to dynamically read version from package metadata instead of hardcoded string

[0.4.2] - 2026-01-22

Added

  • Comprehensive GCP Deployment Documentation - Deep dive guide with Mermaid diagrams
    • docs/gcp-deployment-deep-dive.md and .html
    • Architecture diagrams, authentication flows, troubleshooting guides
    • "The Great Debugging" story documenting credential priority issues

Fixed

  • GCP ADK Dev UI - Dockerfile template now uses web mode instead of api_server
    • Dev UI accessible at /dev-ui/ on Cloud Run deployments
    • Local development with make dev includes interactive UI
  • GCP Authentication Documentation - Clarified SDK credential priority:
    1. GOOGLE_APPLICATION_CREDENTIALS environment variable (highest)
    2. Application Default Credentials (ADC)
    3. Compute Engine / Cloud Run service account
  • GCP Terraform - Improved ADC handling and removed google_project_service resources
    • APIs now enabled via gcloud CLI (avoids ADC permission issues)
    • Terraform focuses on resource creation only

Changed

  • Gemini Model - Updated to gemini-2.0-flash (was gemini-1.5-flash)
  • Global Endpoint - GCP target uses global location for Gemini 2.0+ models

0.4.0 - 2026-01-18

Added

  • CLI Tool - New bedsheet command-line interface
    • bedsheet init <name> - Initialize a new agent project
    • bedsheet generate --target <target> - Generate deployment artifacts
    • bedsheet validate - Validate bedsheet.yaml configuration
    • bedsheet deploy --target <target> - Deploy to cloud platform
  • Deployment Targets - Multi-platform deployment support
    • Local (Docker) - FastAPI server with hot reload for development
    • GCP (Terraform) - Google Cloud Run with ADK (Agent Development Kit)
    • AWS CDK - Amazon Bedrock agents via AWS CDK
    • AWS Terraform - Amazon Bedrock agents via Terraform
  • Debug UI - Real-time agent visualization for development
    • Embedded React UI with event streaming
    • Thinking/rationale event extraction from LLM traces
    • Deduplication of repeated thinking events
    • Tool call and collaborator activity tracking
  • Configuration System - YAML-based project configuration
    • bedsheet.yaml schema with Pydantic validation
    • Multi-agent support with supervisor patterns
    • Per-target configuration (region, model, memory)
  • Resource Management - Cloud resource identification
    • Consistent bedsheet- prefix for infrastructure resources
    • Tagging: ManagedBy, BedsheetVersion, Project, Environment, AgentType
  • Multi-Environment Support - dev/staging/prod deployment patterns
    • GitHub Actions CI/CD workflows for each target
    • Environment-specific configuration

Changed

  • Filter delegate action for supervisors with collaborators (AWS)
  • Supervisors use native platform collaboration instead of Lambda handlers

Fixed

  • Duplicate thinking events in Debug UI
  • <answer> tag content incorrectly appearing in thinking panel
  • Delegate action translation for multi-agent AWS deployments

0.3.0 - 2025-12-03

Added

  • Structured Outputs - New OutputSchema class for enforcing JSON schema compliance
    • OutputSchema.from_pydantic(Model) - Create from Pydantic BaseModel
    • OutputSchema.from_dict(schema) - Create from JSON schema dict
  • Anthropic Structured Outputs Beta - Integration with Claude's constrained decoding
    • Uses structured-outputs-2025-11-13 beta header
    • Guarantees 100% schema compliance via constrained token generation
  • LLMResponse.parsed_output - New field for validated structured data
  • MockLLMClient structured output support - For testing agents with output schemas

Fixed

  • Optional Redis Import - RedisMemory is now conditionally imported, allowing the package to work without redis installed (redis remains an optional dependency)
  • Lint Compliance - Fixed ruff lint warning for optional import pattern

0.2.0 - 2024-11-28

Added

  • Supervisor Agent - New Supervisor class extending Agent for multi-agent coordination
  • Supervisor Mode - Orchestrate multiple agents and synthesize their results
  • Router Mode - Direct handoff to a single agent without synthesis
  • Parallel Delegation - Delegate to multiple agents simultaneously with asyncio.gather
  • Built-in Delegate Tool - Automatic delegate action for agent coordination
  • New Event Types:
    • RoutingEvent - Emitted when router picks an agent
    • DelegationEvent - Emitted when supervisor delegates task(s)
    • CollaboratorStartEvent - Emitted when a collaborator begins
    • CollaboratorEvent - Wraps events from collaborators for visibility
    • CollaboratorCompleteEvent - Emitted when a collaborator finishes
  • Orchestration Templates - $collaborators_summary$ variable for system prompts
  • Investment Advisor Example - Full multi-agent demo in examples/
  • Multi-Agent Guide - Comprehensive documentation in docs/

Changed

  • Event union type now includes 10 event types (was 5)
  • Updated project structure with supervisor.py

0.1.0 - 2024-11-25

Added

  • Agent Class - Core agent with ReAct loop and tool calling
  • ActionGroup - Tool definitions with @action decorator
  • Automatic Schema Inference - Type hints converted to JSON schema
  • Streaming Events:
    • ThinkingEvent - LLM thinking/reasoning
    • ToolCallEvent - Tool invocation
    • ToolResultEvent - Tool results
    • CompletionEvent - Final response
    • ErrorEvent - Error handling
  • Parallel Tool Execution - Multiple tools run concurrently
  • Pluggable Memory:
    • InMemory - Dict-based for development
    • RedisMemory - Redis-based for production
  • Orchestration Template - Customizable system prompt with variables
  • AnthropicClient - Claude integration
  • Error Recovery - Tool errors passed to LLM for retry
  • Max Iterations Safety - Prevents infinite loops
  • MockLLMClient - Testing utilities