hummbl
Structured reasoning framework for AI agents — turns plans, hypotheses, observations, evaluations, decisions, and reflections into durable, inspectable artifacts
Файлы1
Открыть в GitHubSKILL.mdskills/hummbl/SKILL.md
GitHubHUMMBL
Structured reasoning framework for AI agents — turns plans, hypotheses, observations, evaluations, decisions, and reflections into durable, inspectable artifacts. Provides composable reasoning engines, inspectable traces, and domain-specific protocols. It formalizes how agents think, not just what they do.
When to Use
- Building structured reasoning traces (steps, topologies, step types)
- Applying reasoning protocols (Scientific Method, Structured Tool Use)
- Analyzing and scoring reasoning traces
- Planning experiments with trace-driven methodology
- Capturing autoresearch and tool-use patterns for evaluation
Usage
[hummbl] ReasoningTrace(steps=[ReasoningStep(...)])
[hummbl] TraceAnalyzer().analyze(trace)
[hummbl] ScientificMethod().run(hypothesis, observations)
Python API
from hummbl import (
# Reasoning core
ReasoningTrace,
ReasoningStep,
ReasoningTopology,
StepType,
# Protocols
ReasoningProtocol,
ScientificMethod,
StructuredToolUse,
# Analysis & scoring
TraceAnalyzer,
TraceScore,
DimensionScore,
StructuredToolUseScorer,
# Planning
ExperimentPlan,
PlannedExperiment,
TracePlanner,
# Capture
AutoresearchCapture,
ToolUseCapture,
# Typed tuples
TypedTuple,
AttestTuple,
ContractTuple,
DCTTuple,
DCTXTuple,
EvidenceTuple,
SystemTuple,
)
Key Concepts
- Reasoning traces: Ordered sequences of
ReasoningStepobjects withStepTypeclassification andReasoningTopology(linear, branching, etc.). - Protocols: Composable reasoning engines —
ScientificMethod(hypothesis → test → evaluate) andStructuredToolUse(plan → execute → verify). - Trace analysis:
TraceAnalyzerinspects traces;StructuredToolUseScorerscores tool-use quality withDimensionScoreandTraceScore. - Experiment planning:
TracePlannergeneratesExperimentPlanobjects containingPlannedExperimententries. - Capture patterns:
AutoresearchCaptureandToolUseCapturerecord agent behavior for later analysis. - Typed tuples: Governance-aware tuple types (DCT, DCTX, Contract, Evidence, Attest, System) for structured artifact provenance.
Install
cd /work/active/oss/packages/python/hummbl/
pip install -e ".[test]"
Package
- Repo:
hummbl-io/oss - Path:
packages/python/hummbl/ - License: Apache 2.0
- Dependencies: stdlib only