Terms used throughout the documentation.
The single façade that holds all business logic. The CLI, SDK, and REST API are thin adapters over it. See Architecture.
The structured input model for engine.analyze(): content plus optional technology,
source_kind, filename, enrich, and max_root_causes. See SDK guide.
The canonical, interface-agnostic output of an analysis — summary, root causes, diagnostic commands, fixes, references, warnings, best practices, prevention, optional enrichment, and metadata. See Output format.
A thin layer translating between the outside world and the engine. The CLI, SDK, and API are interface adapters; AI providers are outbound adapters.
The Protocol every LLM backend implements (name, model, available(), complete()). The
engine depends on this protocol, never a vendor SDK. See AI providers.
The data passed to a provider's complete(): system, prompt, max_tokens, temperature.
A root cause's 0.0–1.0 likelihood. Bands bucket scores into very_low, low, moderate, high,
very_high. See Output format.
A read-only command a human can run to confirm a hypothesis. read_only is always true; the
toolkit never executes it.
Optional LLM-generated narrative and extra hypotheses layered on top of deterministic findings,
present only when enrich=True and a provider is available.
A view over the knowledge base used by explain and list to browse signatures.
The input line snippets that caused a signature to match — used to justify a finding.
The loaded collection of signatures the engine matches against. Packaged by default; can be custom. See Knowledge base.
The default offline provider. Always "unavailable", which keeps the toolkit deterministic and network-free unless you configure a real provider.
The core guarantee: the toolkit reads text and emits guidance, and never executes commands or mutates state. See Security.
A ranked candidate explanation for the observed symptoms, with title, description, confidence, category, and evidence.
A declarative YAML rule: match patterns plus the root causes, commands, fixes, references, and advice to emit when it fires. See Knowledge base.
The component that evaluates signatures (case-insensitive regex, any_of/all_of, weighting)
against an input and returns scored matches with evidence.
The shape of the input: log, yaml, terraform, compose, kubernetes_manifest,
command_output, error_string, unknown.
A supported platform the engine reasons about (e.g. kubernetes, terraform, openstack). See
Supported technologies.
The output of read-only manifest validation, and a single problem within it (severity, message, optional line/path, hint).
A caution surfaced to the operator (e.g. a fix that risks data loss), with a severity.