adapt-agent 0.2.0__tar.gz → 0.3.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/CHANGELOG.md +282 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/PKG-INFO +1 -1
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/__init__.py +1 -1
- adapt_agent-0.3.0/adapt_agent/adapters/_governed.py +579 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/crewai/__init__.py +1 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/langgraph/__init__.py +1 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/openai_agents/__init__.py +24 -1
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/pydantic_ai/__init__.py +1 -0
- adapt_agent-0.3.0/adapt_agent/core/governance.py +350 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/evaluation/__init__.py +6 -0
- adapt_agent-0.3.0/adapt_agent/integrations/__init__.py +55 -0
- adapt_agent-0.3.0/adapt_agent/integrations/_common.py +162 -0
- adapt_agent-0.3.0/adapt_agent/integrations/agent_framework.py +137 -0
- adapt_agent-0.3.0/adapt_agent/integrations/claude_agent.py +126 -0
- adapt_agent-0.3.0/adapt_agent/integrations/crewai.py +114 -0
- adapt_agent-0.3.0/adapt_agent/integrations/google_adk.py +161 -0
- adapt_agent-0.3.0/adapt_agent/integrations/langgraph.py +124 -0
- adapt_agent-0.3.0/adapt_agent/integrations/openai_agents.py +239 -0
- adapt_agent-0.3.0/adapt_agent/integrations/pydantic_ai.py +171 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/__init__.py +10 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/evals.py +57 -9
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/evaluation.py +264 -46
- adapt_agent-0.3.0/adapt_agent/optimization/extractors.py +643 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/metrics.py +148 -8
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/optimizers.py +232 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/skills/adapt-agent/SKILL.md +75 -5
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/skills/adapt-agent/references/evals.md +91 -3
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/skills/adapt-agent/references/guardrails.md +115 -2
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/skills/adapt-agent/references/optimization.md +41 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent.egg-info/PKG-INFO +1 -1
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent.egg-info/SOURCES.txt +14 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/adapters.md +24 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/evals.md +86 -0
- adapt_agent-0.3.0/docs/integrations.md +190 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/optimization.md +36 -0
- adapt_agent-0.3.0/tests/test_async_governance.py +786 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_governed_adapters.py +12 -2
- adapt_agent-0.3.0/tests/test_integrations.py +1140 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_metrics.py +15 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_skills.py +125 -0
- adapt_agent-0.3.0/tests/test_structured_evals.py +944 -0
- adapt_agent-0.2.0/adapt_agent/adapters/_governed.py +0 -496
- adapt_agent-0.2.0/adapt_agent/optimization/extractors.py +0 -346
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/CODE_OF_CONDUCT.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/CONTRIBUTING.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/LICENSE +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/MANIFEST.in +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/SECURITY.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/__main__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/base.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/claude_agent/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/google_adk/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adapters/microsoft_agent_framework/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/adversarial/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/cli/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/cli/__main__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/core/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/core/memory.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/core/middleware.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/core/policy.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/core/trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/core/types.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/exceptions.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/observability/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/config.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/dataset.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/claude_agent.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/crewai.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/google_adk.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/langgraph.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/microsoft_agent_framework.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/openai_agents.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/introspection/pydantic_ai.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/judge.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/judges.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/parameters.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/performance.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/proposers.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/providers.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/runners.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/optimization/target.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/patches/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/py.typed +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/security/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/security/firewall.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/security/taint_tracker.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent/skills/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent.egg-info/dependency_links.txt +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent.egg-info/entry_points.txt +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent.egg-info/requires.txt +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/adapt_agent.egg-info/top_level.txt +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/api.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/cli.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/claude_agent.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/crewai.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/google_adk.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/index.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/langgraph.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/microsoft_agent_framework.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/openai_agents.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/frameworks/pydantic_ai.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/index.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/observability.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/policy.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/quickstart.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/releasing.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/security.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/docs/skill.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/01_firewall_and_policy.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/02_adversarial_defense.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/03_langgraph_guarded_agent.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/04_observability.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/05_multi_framework_adapters.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/06_optimize_with_golden_dataset.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/07_train_from_yaml.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/08_agent_evals.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/claude_agent/01_basic_guarded.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/claude_agent/02_policy_observability_trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/claude_agent/03_evaluate_and_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/claude_agent/04_multi_agent_and_training.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/claude_agent/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/claude_agent/claude_agent.train.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/config.example.json +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/crewai/01_basic_guarded.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/crewai/02_policy_observability_trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/crewai/03_evaluate_and_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/crewai/04_multi_agent_and_training.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/crewai/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/crewai/crewai.train.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/google_adk/01_basic_guarded.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/google_adk/02_policy_observability_trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/google_adk/03_evaluate_and_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/google_adk/04_multi_agent_and_training.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/google_adk/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/google_adk/google_adk.train.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/langgraph/01_basic_guarded.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/langgraph/02_policy_observability_trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/langgraph/03_evaluate_and_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/langgraph/04_multi_agent_and_training.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/langgraph/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/langgraph/langgraph.train.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/microsoft_agent_framework/01_basic_guarded.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/microsoft_agent_framework/02_policy_observability_trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/microsoft_agent_framework/03_evaluate_and_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/microsoft_agent_framework/04_magentic_team_and_training.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/microsoft_agent_framework/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/microsoft_agent_framework/magentic.train.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/openai_agents/01_basic_guarded.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/openai_agents/02_policy_observability_trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/openai_agents/03_evaluate_and_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/openai_agents/04_multi_agent_and_training.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/openai_agents/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/openai_agents/openai_agents.train.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/pydantic_ai/01_basic_guarded.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/pydantic_ai/02_policy_observability_trust.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/pydantic_ai/03_evaluate_and_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/pydantic_ai/04_multi_agent_and_training.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/pydantic_ai/README.md +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/pydantic_ai/pydantic_ai.train.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/examples/train.example.yaml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/pyproject.toml +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/setup.cfg +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/__init__.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_action_refs.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_adapters.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_adversarial.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_adversarial_extra.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_cli.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_cli_optimize.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_cli_skills.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_core.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_evaluation.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_firewall_extra.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_memory.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_middleware.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_observability.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_observability_extra.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimization.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_config.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_dataset.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_evals.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_evaluation.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_extractors.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspect_claude_agent.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspect_crewai.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspect_google_adk.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspect_langgraph.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspect_microsoft_agent_framework.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspect_openai_agents.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspect_pydantic_ai.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_introspection.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_judge.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_judges.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_optimizers.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_parameters.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_proposers.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_providers.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_runners.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_optimize_target.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_patches.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_policy.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_security.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_taint_tracker.py +0 -0
- {adapt_agent-0.2.0 → adapt_agent-0.3.0}/tests/test_trust.py +0 -0
|
@@ -7,6 +7,288 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
|
|
|
7
7
|
|
|
8
8
|
## [Unreleased]
|
|
9
9
|
|
|
10
|
+
## [0.3.0] - 2026-08-21
|
|
11
|
+
|
|
12
|
+
### Added
|
|
13
|
+
|
|
14
|
+
- **Async is now a first-class path, not a leaf-node rescue.** Every governed
|
|
15
|
+
agent gained `aexecute`, the awaitable twin of `execute` with identical
|
|
16
|
+
governance. It awaits the framework in the *caller's* event loop, so
|
|
17
|
+
concurrent requests stay concurrent and `contextvars` -- how OpenTelemetry
|
|
18
|
+
propagates the active span -- survive. Previously an async-native framework
|
|
19
|
+
(Pydantic AI, the Claude Agent SDK, Microsoft Agent Framework) could only be
|
|
20
|
+
governed from a synchronous caller, and the advice to use a worker thread cost
|
|
21
|
+
both concurrency and trace parentage. `aresolve_runner` and
|
|
22
|
+
`EvaluationHarness.aevaluate` mirror it upward; the governance stages are
|
|
23
|
+
shared, so the two call styles cannot drift.
|
|
24
|
+
- **Native governance hooks for every supported framework**
|
|
25
|
+
(`adapt_agent.integrations`). An adapter wraps an agent from the outside, which
|
|
26
|
+
governs only a graph's boundary; these plug into the framework's own
|
|
27
|
+
interception point so rules nest *per agent* inside a multi-agent graph,
|
|
28
|
+
compose with middleware the app already stacks, and don't fight the workflow
|
|
29
|
+
runtime. Each factory was written against the installed SDK's own source:
|
|
30
|
+
Microsoft Agent Framework (`Agent(middleware=[...])`), Google ADK
|
|
31
|
+
(before/after model callbacks, with an `on_block="refuse"` mode that
|
|
32
|
+
short-circuits the model instead of aborting the tree), OpenAI Agents SDK
|
|
33
|
+
(input/output guardrails), Claude Agent SDK (`UserPromptSubmit` and
|
|
34
|
+
`PreToolUse` hooks -- the latter reaching tool inputs no outer wrapper can
|
|
35
|
+
see), LangGraph (`pre_model_hook`/`post_model_hook`), CrewAI (kickoff
|
|
36
|
+
callbacks plus a task guardrail), and Pydantic AI (output validator only --
|
|
37
|
+
it has no native pre-run hook, and the docs say so rather than inventing one).
|
|
38
|
+
- `GovernanceGate` (`adapt_agent.core.governance`): the single, framework-free
|
|
39
|
+
implementation of screen -> policy -> screen, now shared by both the adapters
|
|
40
|
+
and the native hooks so a fix reaches every framework at once.
|
|
41
|
+
- **Concurrency for evals.** `EvaluationHarness.evaluate(..., concurrency=N)`
|
|
42
|
+
runs examples in a thread pool for synchronous agents; `aevaluate(...,
|
|
43
|
+
concurrency=N)` uses a bounded async worker pool for async ones.
|
|
44
|
+
`evaluate_agent` takes the same knob. A serial harness made optimization
|
|
45
|
+
impractical -- a 60-eval sweep over a 113-case split is 6,780 LLM round-trips.
|
|
46
|
+
Both paths preserve index ordering, non-fatal per-example errors, and the
|
|
47
|
+
`max_results` memory bound.
|
|
48
|
+
- `extract_output_payload`: unwraps the framework envelope but **keeps the
|
|
49
|
+
structure** -- a mapping passes through, a declared structured output (Pydantic
|
|
50
|
+
model or dataclass) becomes a dict, and a JSON string (including a
|
|
51
|
+
```` ```json ```` fenced one) is parsed back into an object. Previously a
|
|
52
|
+
structured answer was either flattened to `.text` by `extract_output_text` or
|
|
53
|
+
left as a `repr()` by `output_extractor=None`. Three shapes that would each
|
|
54
|
+
have scored 0.0 across every column are handled by name: a LangGraph
|
|
55
|
+
`response_format=` **state** (the answer is under `structured_response`,
|
|
56
|
+
beside `messages`; the state is identified by what LangGraph guarantees --
|
|
57
|
+
`add_messages` coerces every entry to a `BaseMessage` -- so an answer that
|
|
58
|
+
merely has fields of those names is not peeled); a **single-field answer** like
|
|
59
|
+
`{"result": "granted"}` or `{"answer": {"city": "Paris"}}`, which is the
|
|
60
|
+
answer rather than an envelope around one. Shape alone cannot decide that --
|
|
61
|
+
both readings occur for `{"result": [...]}` -- so `execute` now marks its own
|
|
62
|
+
wrapper (`GovernedEnvelope`, a plain `dict` in every respect that matters but
|
|
63
|
+
identifiable), and extraction stops guessing; and a declared output
|
|
64
|
+
whose field happens to be *called* `answer` or
|
|
65
|
+
`result`, which the generic attribute peel would have reduced to its own
|
|
66
|
+
value. `extract_output_text` likewise leaves a declared output unchanged, so a
|
|
67
|
+
per-row `field_match` dispatched by `checks` still sees its fields. A **list of
|
|
68
|
+
records** is the answer rather than a message stream, so a list of dataclasses
|
|
69
|
+
or models comes back whole instead of collapsing to its last element.
|
|
70
|
+
- `field_match(field, ...)` and `field_metrics([...])`: score a structured
|
|
71
|
+
output per field, reported under the field's own name, so a report aggregates
|
|
72
|
+
as a per-column table (`{"lane": 0.94, ..., "pack": 0.0}`) rather than one
|
|
73
|
+
blended number that hides which column moved. `evaluate_agent` switches to
|
|
74
|
+
payload extraction automatically when every metric is structural.
|
|
75
|
+
- `OptimizationResult.to_config(path)` reserves the `_provenance` namespace.
|
|
76
|
+
`load_tuned_config` skips that key on the way in, so a tuned parameter named
|
|
77
|
+
for it could be exported but never reloaded -- and in JSON the real provenance
|
|
78
|
+
block overwrote it first. It is described rather than written, so the loss is
|
|
79
|
+
visible instead of silent.
|
|
80
|
+
- `OptimizationResult.to_config(path)` handles a bare parameter name that is
|
|
81
|
+
also a component prefix (`{"agent": 1, "agent.temperature": 0.2}`): the two
|
|
82
|
+
cannot share the top level, and which one survived depended only on dict
|
|
83
|
+
ordering -- one order raised `TypeError`, the other silently dropped the
|
|
84
|
+
component's knobs. The qualified knobs are exported and the bare name is
|
|
85
|
+
described.
|
|
86
|
+
- `OptimizationResult.to_config(path)` and `load_tuned_config(path)`: export the
|
|
87
|
+
winning configuration as reviewable `{component: {parameter: value}}` YAML and
|
|
88
|
+
load it back. The optimizer applied its result in place, to live objects, so
|
|
89
|
+
it died with the process; now the loop is optimize -> diff -> review -> commit
|
|
90
|
+
and a machine-rewritten prompt cannot reach production unread.
|
|
91
|
+
|
|
92
|
+
### Security
|
|
93
|
+
|
|
94
|
+
- **A tool result carrying a prompt injection reached the model unscreened.**
|
|
95
|
+
Google ADK returns a tool's output under `Part.function_response.response`, a
|
|
96
|
+
mapping, never `Part.text` -- and `extract_texts` neither walked that
|
|
97
|
+
attribute nor reached it within its recursion bound, so the firewall was blind
|
|
98
|
+
on exactly the path that carries untrusted content back from the open web. The
|
|
99
|
+
identical string as a plain text part was blocked. Tool-response attributes
|
|
100
|
+
are now walked explicitly, and the bound is sized from the deepest real
|
|
101
|
+
payload (a governed ADK tool result is eight hops) rather than a guess,
|
|
102
|
+
because a security scan that stops early fails open.
|
|
103
|
+
|
|
104
|
+
### Fixed
|
|
105
|
+
|
|
106
|
+
- **A Claude tool *result* was never screened.** Only `UserPromptSubmit` and
|
|
107
|
+
`PreToolUse` were governed by default, so whatever a tool fetched from the
|
|
108
|
+
open web reached the model unscreened unless it happened to be copied into a
|
|
109
|
+
*subsequent* tool call. `PostToolUse` joins the defaults -- the same gap
|
|
110
|
+
closed on the ADK side.
|
|
111
|
+
- **A tool matcher silently disabled prompt screening.** `matcher=` is a tool
|
|
112
|
+
name, and it was attached to every governed event including
|
|
113
|
+
`UserPromptSubmit`, describing a prompt event that can never match. It now
|
|
114
|
+
applies to tool-scoped events only, which is what the parameter's own
|
|
115
|
+
documentation already said.
|
|
116
|
+
- **A coroutine returning a stream was not drained.** Both `execute` and
|
|
117
|
+
`aexecute` awaited once and handed the still-live async generator on, so
|
|
118
|
+
output screening found no text -- a firewall bypass for streamed content --
|
|
119
|
+
and the caller received a generator where the envelope documents a list. The
|
|
120
|
+
awaited value is resolved recursively, and the sync path now routes through
|
|
121
|
+
the same resolver as the async one rather than keeping a parallel copy that
|
|
122
|
+
could drift.
|
|
123
|
+
- **A handoff target's input governance never ran.** The OpenAI Agents SDK runs
|
|
124
|
+
input guardrails for the *starting* agent of a run only (`run.py` gates them
|
|
125
|
+
on `current_turn == 0`), so a specialist reached by a handoff had its
|
|
126
|
+
firewall, defense and policy rules configured, documented and silently
|
|
127
|
+
skipped for transferred content. Confirmed by driving a real handoff against
|
|
128
|
+
the installed SDK. New `openai_agents.governance_agent_hooks()` binds to
|
|
129
|
+
`AgentHooks.on_llm_start`, which fires per agent and per model call -- so it
|
|
130
|
+
also screens tool results on their way back to the model -- and `inner=`
|
|
131
|
+
keeps the app's own lifecycle hooks running. `on_end` screens the
|
|
132
|
+
specialist's answer, so a handoff target is governed in both directions.
|
|
133
|
+
- **`aexecute` blocked the event loop on a synchronous framework.** The
|
|
134
|
+
resolver falls back to the sync runner so an async app can use one entry
|
|
135
|
+
point uniformly, but calling it did the work before the first await: a
|
|
136
|
+
heartbeat task got zero ticks and three concurrent calls serialised (0.45s
|
|
137
|
+
for 3 x 150ms). The sync fallback runs in a worker thread now --
|
|
138
|
+
`asyncio.to_thread`, so `contextvars` and the active span still reach the
|
|
139
|
+
framework -- and the same three calls take 0.15s.
|
|
140
|
+
- **A shared gate's label did not reach every message.** The Claude refusal
|
|
141
|
+
reason and the OpenAI tripwire's `output_info` interpolated the factory
|
|
142
|
+
parameter rather than the resolved gate id, so a binding using a shared gate
|
|
143
|
+
reported the default while applying that gate's controls. Both use
|
|
144
|
+
`resolved.agent_id` now, matching the MAF span.
|
|
145
|
+
- **An ADK refusal did not say which agent refused.** `on_block="refuse"`
|
|
146
|
+
returns an ordinary `LlmResponse`, so unlike the raising path it carried
|
|
147
|
+
nothing for the surrounding graph to inspect -- two specialists produced
|
|
148
|
+
byte-identical objects, though the factory documented `agent_id` as
|
|
149
|
+
identifying which one refused. The id and the threats now travel in
|
|
150
|
+
`custom_metadata["adapt_agent"]`, leaving `refusal_text` as the caller's copy
|
|
151
|
+
for the end user.
|
|
152
|
+
- **One problem was reported several times.** A payload yields many texts -- a
|
|
153
|
+
request's parts, a message list, a model's fields -- so a single blocked
|
|
154
|
+
request raised `["firewall", "firewall", "firewall"]`. Threat labels are
|
|
155
|
+
de-duplicated, keeping first-seen order; the multiplicity counted texts
|
|
156
|
+
scanned, not distinct problems.
|
|
157
|
+
- **A shared gate was labelled by the wrong agent.** Passing
|
|
158
|
+
`gate=` alongside `agent_id=` -- the advertised multi-agent setup -- returned
|
|
159
|
+
the gate unchanged, so a violation raised an error naming the *shared* gate
|
|
160
|
+
while the hook traced the same invocation under the binding's id. The binding's
|
|
161
|
+
id now wins where it is given, the gate's own label applies where it is not,
|
|
162
|
+
and the span carries whichever the error does.
|
|
163
|
+
- **A blocked output was traced as a successful run.** The observer span closed
|
|
164
|
+
as `completed` before post-middleware and output screening ran, so a caller
|
|
165
|
+
received `SecurityBlockedError` while telemetry recorded success -- hiding the
|
|
166
|
+
output-policy failures monitoring exists to surface. The span now closes last,
|
|
167
|
+
and as an error when either stage raises, on `execute`, `aexecute`, and the
|
|
168
|
+
Microsoft Agent Framework middleware (the only native hook that opens a span).
|
|
169
|
+
- **A bare parameter name did not survive the config round trip.** `to_config`
|
|
170
|
+
filed a name with no `component.` prefix under a synthetic `agent` section,
|
|
171
|
+
renaming it on the way out; `load_tuned_config` could not recover the original,
|
|
172
|
+
so `apply()` silently skipped it and the export/reload round trip did not
|
|
173
|
+
restore the winner. Bare names now stay at the top level, under their own name.
|
|
174
|
+
- **A per-row `field_match` scored a false 0.0 on a model result.** A row's
|
|
175
|
+
`{"check": {"name": "field_match", ...}}` is dispatched by `checks`, which
|
|
176
|
+
cannot be marked structural (the row decides at run time), so extraction leaves
|
|
177
|
+
a Pydantic AI result as a model object. Mapping coercion now accepts models and
|
|
178
|
+
dataclasses, matching what payload extraction already did.
|
|
179
|
+
- **A tuple parameter was exported as a list.** Both encoders read a sequence
|
|
180
|
+
back as a `list`, so reloading changed the winning value's type and a setter
|
|
181
|
+
expecting a tuple would receive a list. Tuples are now described rather than
|
|
182
|
+
exported, keeping the invariant that the config body is exactly what applies
|
|
183
|
+
cleanly.
|
|
184
|
+
- **A drained event stream defeated structural scoring.** An async framework's
|
|
185
|
+
result is materialised into a list, and `extract_output_payload` returned that
|
|
186
|
+
list as the payload -- so a Claude Agent SDK `ResultMessage` carrying
|
|
187
|
+
`{"lane": "NOS"}` left every `field_match` scoring a false 0.0. Recognised
|
|
188
|
+
streams are now scanned from the end for their final payload, while a genuine
|
|
189
|
+
structured list (containing nothing a registered extractor recognises) is
|
|
190
|
+
returned unchanged.
|
|
191
|
+
- **OpenAI guardrail policy never saw the runtime context.** Authorization data
|
|
192
|
+
passed as `Runner.run(..., context=...)` arrives on `RunContextWrapper.context`
|
|
193
|
+
and was not merged into the policy state, so a rule gating on
|
|
194
|
+
`state['trust_score']` found the key absent and failed open.
|
|
195
|
+
- **`defense` was silently inert on the Pydantic AI validator**, for the same
|
|
196
|
+
reason as `policy_enforcer`: `scan_output` runs the firewall only, deliberately,
|
|
197
|
+
since adversarial-*input* detection over an answer flags an agent legitimately
|
|
198
|
+
quoting an instruction. It is now refused alongside policy rather than accepted
|
|
199
|
+
and ignored.
|
|
200
|
+
- **A `policy_enforcer` given to the Pydantic AI validator was silently
|
|
201
|
+
ignored.** That seam sees only the output, so a state-gating rule could never
|
|
202
|
+
fire; it now raises with a pointer to the adapter rather than accepting a
|
|
203
|
+
control it cannot honour.
|
|
204
|
+
- **ADK policy never saw session state.** The callback synthesised policy state
|
|
205
|
+
from the model request alone, so a rule reading `state['trust_score']` found
|
|
206
|
+
the key absent and a fail-open enforcer read that as "no violation". The
|
|
207
|
+
callback context's state is now merged in.
|
|
208
|
+
- **A cancelled native hook left its observer span open**, the same
|
|
209
|
+
`except Exception` gap as the adapter path.
|
|
210
|
+
- **The threaded eval pool stalled behind a slow example.** Refilling waited on
|
|
211
|
+
the *oldest* future, idling the other workers until it finished; with variable
|
|
212
|
+
LLM latency that collapses the achieved concurrency (measured: 0.96s against
|
|
213
|
+
an 0.65s ideal). It now refills from whichever future completes first, with
|
|
214
|
+
index ordering restored by the accumulator.
|
|
215
|
+
- **`aexecute` still blocked on a directly-wrapped OpenAI SDK `Agent`.** Such an
|
|
216
|
+
agent exposes neither `run` nor `run_sync`, so the async preference list could
|
|
217
|
+
not match and the adapter fell back to its synchronous `Runner.run_sync`
|
|
218
|
+
lambda. It now has an async SDK runner using `Runner.run`.
|
|
219
|
+
- **`aexecute` called the framework's *synchronous* entry point.** The runner
|
|
220
|
+
was resolved once from the sync-first `run_method_names`, so on LangGraph,
|
|
221
|
+
CrewAI, Pydantic AI and the OpenAI Agents SDK the async path invoked
|
|
222
|
+
`invoke`/`kickoff`/`run_sync` and blocked the very event loop it exists to
|
|
223
|
+
cooperate with -- while `ainvoke`/`kickoff_async`/`run` went unused. Adapters
|
|
224
|
+
now declare `async_run_method_names` and `aexecute` resolves against it,
|
|
225
|
+
falling back to the sync runner for a framework that has no async twin.
|
|
226
|
+
- **A cancelled `aexecute` left its observer span open forever.**
|
|
227
|
+
`asyncio.CancelledError` derives from `BaseException`, so `except Exception`
|
|
228
|
+
never saw it and neither the error nor the completion path ran.
|
|
229
|
+
- **Renaming a structural metric stripped its `structural` flag.**
|
|
230
|
+
`metrics={"lane_acc": field_match("lane")}` fell back to text extraction, so a
|
|
231
|
+
model-returning agent scored 0.0. The flag now survives both rename paths --
|
|
232
|
+
the harness's and `evaluate_agent`'s, the latter of which had been missed.
|
|
233
|
+
- **Report-only mode silently disabled policy auditing.** With
|
|
234
|
+
`block_on_violation=False` the adapter skipped `policy_violations()` entirely,
|
|
235
|
+
and since `PolicyEnforcer.check_state` is what records violations and fires
|
|
236
|
+
warn/log handlers, the documented rollout mode recorded nothing at all rather
|
|
237
|
+
than recording without refusing. Policy is now evaluated whatever the blocking
|
|
238
|
+
mode; only the refusal is conditional.
|
|
239
|
+
- **`to_config` crashed on tool and skill parameters.** Those hold live
|
|
240
|
+
callables, which no YAML or JSON encoder can represent, so exporting a run
|
|
241
|
+
from the default optimizer's tool stage raised. They are now listed by
|
|
242
|
+
`module:qualname` for the review diff and kept out of the config body, so the
|
|
243
|
+
file stays reloadable and `apply()` can never write a stand-in string over a
|
|
244
|
+
real tool list.
|
|
245
|
+
- **A `policy_enforcer` passed to the Microsoft Agent Framework or Google ADK
|
|
246
|
+
hooks was silently inert.** Both called `review_input` without a `state`, and
|
|
247
|
+
the gate only evaluates policy when given one -- so the control was accepted,
|
|
248
|
+
documented, and did nothing, while `call_next()` ran on. Every hook now passes
|
|
249
|
+
a state, and a test drives all four input-screening hooks against a recording
|
|
250
|
+
enforcer so one binding cannot forget again.
|
|
251
|
+
- **A structured output nested under a wrapper attribute went unscreened.** Text
|
|
252
|
+
extraction recursed only into `dict`/`list`/`tuple`, so a Pydantic AI
|
|
253
|
+
`AgentRunResult.output` holding a `BaseModel` -- the ordinary shape -- had its
|
|
254
|
+
wrapper walked and the answer inside it skipped. Any non-primitive is now
|
|
255
|
+
followed.
|
|
256
|
+
- **A governed adapter's envelope defeated structural metrics.** `execute`
|
|
257
|
+
returns `{"result": <payload>}`, which `extract_output_payload` treated as the
|
|
258
|
+
payload, so every `field_match` scored 0.0 against the real fields. A single
|
|
259
|
+
conventional key is now peeled; a multi-key mapping is still treated as the
|
|
260
|
+
answer, so a structured result that happens to contain `result` is unharmed.
|
|
261
|
+
- **The threaded eval path was not memory-bounded.** `evaluate(concurrency>1)`
|
|
262
|
+
materialised the whole dataset and handed it to `ThreadPoolExecutor.map`,
|
|
263
|
+
which itself submits every example up front. Replaced with bounded submission
|
|
264
|
+
that keeps at most `concurrency` runs in flight and pulls lazily, matching
|
|
265
|
+
what `aevaluate` already did.
|
|
266
|
+
- **Structured outputs were almost entirely unscreened.** Text extraction only
|
|
267
|
+
probed a handful of conventional attribute names (`text`, `content`,
|
|
268
|
+
`output`, ...), so a Pydantic model or dataclass with fields like `lane` or
|
|
269
|
+
`note` -- exactly what a Pydantic AI `output_type` produces -- passed through
|
|
270
|
+
the firewall untouched, and an injection smuggled into one field was never
|
|
271
|
+
seen. Pydantic (v1 and v2) model and dataclass fields are now walked.
|
|
272
|
+
- `execute` raising inside a running event loop abandoned the framework's
|
|
273
|
+
coroutine un-awaited, emitting a `RuntimeWarning` that pointed at the
|
|
274
|
+
framework instead of the real problem. The coroutine is now closed, and the
|
|
275
|
+
error message points at `aexecute`.
|
|
276
|
+
- `get_metric` surfaced a raw `TypeError` for a built-in needing arguments; it
|
|
277
|
+
now explains that `field_match` is reachable through a mapping check spec.
|
|
278
|
+
|
|
279
|
+
### Documentation
|
|
280
|
+
|
|
281
|
+
- `references/guardrails.md` carried no async warning at all, while the evals
|
|
282
|
+
gotchas did -- the one page omitting it described the code path that raises.
|
|
283
|
+
It now documents `aexecute`, why a worker thread is the wrong fix, and the
|
|
284
|
+
native-hook matrix.
|
|
285
|
+
- New `docs/integrations.md`, plus async/concurrency/structured-scoring and
|
|
286
|
+
config-export sections across the skill and the docs site.
|
|
287
|
+
- Two more tests that *execute the documentation verbatim* rather than
|
|
288
|
+
inspecting it: the structured-scoring recipe is run, and every integration
|
|
289
|
+
factory named in the matrix is resolved.
|
|
290
|
+
|
|
291
|
+
|
|
10
292
|
### Added
|
|
11
293
|
|
|
12
294
|
- **Bundled agent skill** (`adapt_agent/skills/adapt-agent/`): the wheel now
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: adapt-agent
|
|
3
|
-
Version: 0.
|
|
3
|
+
Version: 0.3.0
|
|
4
4
|
Summary: ADAPT-Agent: Adversarial Defense & Policy Training for LLM Agents - A comprehensive library for LLM agent optimization and security
|
|
5
5
|
Author-email: Daniel Halwell <danielhalwell@gmail.com>
|
|
6
6
|
Maintainer-email: Daniel Halwell <danielhalwell@gmail.com>
|