0.5.0

Improved context window management, important system prompt fixes for guidance, improved reporting.

• dependency updates
• add web_recon module for reconnaissance without exploitation
• make reporting work with only observations for non-exploitation use cases
• reporting uses all findings when MEMORY_ISOLATION=shared
• increase PROMPT_TELEMETRY_THRESHOLD to more reasonable value of 85% to allow for more input context
• fix sliding conversation manager to preserve first messages: initial user prompt was getting lost
• improve handling of failure cases
• patch OllamaModel usage reporting: input and output tokens are swapped
• apply CYBER_AGENT_OUTPUT_DIR everywhere instead of hardcoded “outputs” directory
• set context window message limit based on prompt token limit: 100 lines default, 200 lines for >= 128,000, 300 lines for >= 400,000
• use full paths with LLM content, some models prepend hallucinated filesystem roots
• add operation_paths information to system prompt to control LLM filesystem scope
• add reflection_snapshot information to system prompt (was already referenced by execution prompts)
• run execution prompt optimizer before system prompt rebuilding to load the optimized prompt in the same step
• improve agent continuation message with budget, check point and actions
• update bedrock models to global.anthropic.claude-opus-4-5-20251101-v1:0 / us.anthropic.claude-sonnet-4-5-20250929-v1:0

0.4.2

• prompt budget considers output token limit, lessens exceeding context limit
• clamp output token limit at 12,000 for standard models, 24,000 for thinking models giving more room for conversation history
• improve tracing of sub-agents: swarm, prompt_optimizer, validation_specialist
• additional instructions for running validation_specialist in web module

0.4.1

• add back erroneously removed python_repl and sleep tools
• fix incorrect model parameters (i.e., max output tokens) when swarm model == main model
• validate swarm agent model and fall back to primary model
• fix broken tool calling (ollama, gemini) in report, validation_specialist agents
• relax prompt optimizer validation for line count increase
• minor efficiency updates

0.4.0

• Estimate tokens for system prompt and tools instead of using constants
• Rename 'general' module to 'web'
• swarm tool allows model selection using selected provider or ollama
• Allow modules to specify which built-in tools to use
• Refactor XBOW benchmark script to python

0.3.1

What's Changed

• toolUseId is used as tool name for Ollama and Gemini. Causes false no-progress detection. Fixed to provide a unique ID.
• swarm agents get correct prompt token length, apply reasoning setting and trace attributes
• plan format instruction improvements
• misc fixes from code review

Full Changelog: v0.3.0...v0.3.1

0.3.0

What's Changed

• Fixed browser tool performance on lesser models. (Fixes #37, #38)
  • Add browser instructions for element format.
  • Fix some bad json output.
• Add web search tools.

Full Changelog: 0.2.0...v0.3.0

0.2.0

What's Changed

• Add model rate limiting
• Add forward and reverse channel tools
• Add OAST tools
• Save history of optimized prompts
• Fix evaluation prompt JSON bug

Full Changelog: 0.1.5...0.2.0

0.1.5

What's Changed

• Python 3.11+ now required (was 3.10+)
• Dockerfile size and build optimization
• Install missing tools in docker image, including verification script
• Fix incorrect import for http_request, python_repl, stop so they work again
• Add OLLAMA_TIMEOUT environment variable, specified in seconds
• Add tool_catalog with full tool output.
• Get Ollama model configuration from /api/show, such as context limit
• Improve react CLI output, mostly to see tool start/end
• Fix browser summarization by including all network requests
• Fix browser hangs by moving to separate thread
• Tool artifacts now consider json, image and document
• Introduce custom swarm tool to use our model configs, conversation manager and hooks

Full Changelog: https://github.com/double16/Cyber-AutoAgent/commits/release/0.1.5