Repository navigation
Releases
v0.3.1 — Security & correctness fixes
Patch release fixing a SQL guard bypass and several correctness bugs found while auditing the package. Upgrading is recommended for anyone using the SQLite tools.
pip install --upgrade explainable-agent
**Full Changelog**: https://github.com/emredeveloper/explainable-agent-lab/compare/v0.3.0...v0.3.1Explainable Agent v0.3.0
Explainable Agent v0.3.0
This release hardens Explainable Agent as a Python library while keeping the existing CLI workflow intact.
Highlights
- Promoted the public library API with root-level exports for
ToolRegistry,ToolSpec,define_tool,run_tool, trace dataclasses, and artifact writers. - Added
py.typedso type checkers can recognize the package as typed. - Added isolated per-agent
ToolRegistrysupport for library, testing, and multi-tenant use cases. - Hardened SQLite tooling with read-only query execution, stricter write command allowlists, and safer explicit tool routing.
- Improved multi-agent orchestration with schema-backed JSON planning and clearer diagnostics when no valid delegation plan is produced.
- Reused the shared relaxed JSON parser inside the OpenAI-compatible client.
- Updated README, architecture docs, release metadata, and PyPI-facing package guidance.
- Expanded regression coverage for public API exports, isolated tool registries, SQLite guards, orchestrator planning, explicit tool routing, and eval defaults.
Validation
pytest -q: 16 passedruff check .: passedruff format --check .: passedtwine check: passed- Built and published both wheel and sdist to PyPI.
PyPI
Install with:
pip install explainable-agent==0.3.0
**Full Changelog**: https://github.com/emredeveloper/explainable-agent-lab/compare/v0.2.5...v0.3.0Explainable Agent 0.2.5
This release improves project readiness for open-source use.
Updated package version and PyPI metadata to 0.2.5
Removed committed .env values and kept .env.example
Added GitHub Actions CI for compile, lint, format, tests, and build
Added smoke/unit tests for tools, settings, path safety, and faithfulness scoring
Refreshed examples for Ollama/local-first workflows
Improved numeric and path-based tool support scoring in diagnostics
Normalized Python formatting with Ruff
Full Changelog: v0.2.0...v0.2.5
Explainable Agent v0.2.0
This release focuses on making the local-first agent runtime cleaner, more reliable, and easier to ship.
Highlights
- Added OpenAI-native tool calling support
- Added streaming decision support
- Added per-step token accounting in traces and reports
- Refactored the single-agent runtime for better maintainability
- Improved Ollama compatibility with safer plain-decision fallbacks
- Switched DuckDuckGo integration dependency from
duckduckgo-searchtoddgs
Release Notes
- The core agent loop was reorganized to separate message building, decision requests, heuristic correction, step recording, and tool follow-up prompts.
- Local-provider behavior was tuned to avoid slow or fragile structured/native paths where needed.
- Packaging and project metadata were updated for
v0.2.0. - Legacy test artifacts and older release leftovers were cleaned out to prepare a simpler release state.
Full Changelog: v0.1.4...v0.2.0
v0.1.4 – Developer Verbose Mode & JSONL Eval
Developer-focused verbose mode: When verbose=True, the agent now prints an “Agent tools flow roadmap”, per-step decision source/latency, and a compact developer run summary.
Concise output mode: When verbose=False, runs end with a short flow summary (tool steps + final_answer), step counts, a brief answer preview, and key warnings.
JSONL evaluation improvements: Added examples and docs for running HF-style tool-calling evaluation using data/evals/hf_complexfuncbench_first_turn_100.jsonl, validated with LM Studio and zai-org/glm-4.6v-flash.
Docs & tooling: Updated README.md, docs/ARCHITECTURE.md, and a test_pypi_install.py helper to reflect the new 0.1.4 behavior.
v0.1.3: Multi-Agent Orchestration & SQL Reports
What's New
- Multi-Agent Orchestration (Team of Thoughts):
TeamOrchestratorcoordinates specialized agents with transparent delegation plans, sub-agent traces, and orchestration diagnostics. - SQL in Reports: Orchestration reports now include executed SQL statements and results for sqlite tools.
- ddgs Migration: Web search now uses
ddgspackage (duckduckgo_search deprecated). - ResourceWarning Suppression: Cleaner output in multi-agent demo.
Install
pip install --upgrade explainable-agent
Full Changelog: v0.1.2...v0.1.3
v0.1.2: Efficiency Diagnostics & Chaos Engineering 🛡️
✨ What's New
- Chaos Engineering Mode (
AGENT_CHAOS_MODE=true): Simulates a 20% failure rate (Timeout, Permission Denied, Malformed Data) to stress-test your agent's self-healing capabilities. - Efficiency Diagnostics: Automatically measures
model_output_lengthandtool_output_lengthfor every step. - Context Window Protection: Generates a diagnostic warning in the
report.mdif a tool returns excessively long data (e.g. >4000 chars), preventing the agent from getting "overwhelmed." - Loop Detection: Warns you if the agent is stuck taking too many steps (e.g. >5) without reaching a conclusion.
- Comprehensive Showcase Script: Added
examples/showcase_all_features.pywhich demonstrates SQLite tools, Self-Healing, Chaos Mode, and Evaluation pipelines all in one place.
🐛 Fixes & Improvements
subprocess.runnow safely usessys.executableto run evaluations in the correct Python virtual environment.- Various terminal UI fixes to support standard UTF-8 console encodings smoothly.
📦 Installation
pip install --upgrade explainable-agent
🚀 v0.1.1 - The Explainable Agent Lab is now on PyPI!
We are thrilled to announce the first official PyPI release of the Explainable Agent Lab! 🎉
This release transforms the project from a local repository into a fully installable Python library (pip install explainable-agent) and introduces powerful new features designed to make AI agents transparent, robust, and easy to build.
✨ Key Features in This Release
- Self-Healing Capabilities: Agents can now automatically catch their own errors (like schema violations or tool failures), analyze the root cause (
error_analysis), propose solutions (proposed_fix), and recover by calling alternative tools. - Visual Terminal Tracking: Added
--verbosemode powered byrich, providing beautiful, colorful, step-by-step interactive logs right in your terminal. You can watch the agent's "Rationale", "Confidence", and "Tool Output" in real-time. - Actionable Diagnostic Reports: At the end of every run, the framework generates a detailed
report.mdthat highlights hallucination risks, potential infinite loops, and gives concrete suggestions to improve your system prompts. - Python API & Extensibility: You can now import the agent directly into your own Python codebases. Added the
@define_tooldecorator, making it incredibly simple to attach custom Python functions as agent tools. - Global CLI Access: Run agents directly from your terminal using the global
explainable-agentcommand, supporting both LM Studio and Ollama right out of the box.
📦 Installation
It is now easier than ever to get started:
pip install explainable-agent
Run with Ollama
explainable-agent --base-url http://localhost:11434/v1 --model ministral-3:14b --task "calculate_math: (215*4)-12" --verbose
Run with LM Studio
explainable-agent --model qwen/qwen3-vl-4b --task "duckduckgo_search: What are the latest updates in Python 3.13?" --verbose
For more details on programmatic usage and custom dataset evaluations, check out the updated README and the examples/ directory!