Glossary¶
Plain-English definitions of the terms used across these docs. Most link to the page where the concept is explained in depth.
| Term | Meaning |
|---|---|
| Coordinator | The main reasoning agent. It receives your question, decides how to investigate, calls tools, and writes the final answer. Runs on the larger model (gpt-4o by default). See Agent Behaviors. |
| Subagent | A domain specialist (one of pod, metrics, logs, events) that the coordinator dispatches in parallel during a root-cause analysis. Each returns a structured finding. Runs on the cheaper model (gpt-4o-mini). |
| Context fetcher | The workflow step that runs before the agent thinks — it pre-fetches a cluster snapshot (pods + Warning events) so the agent starts with situational awareness. |
| Memory loader | The step that loads pinned, cross-session context (user prefs, past root causes, failure hints) into the prompt. Active only on PostgreSQL. |
| Cluster snapshot | The pre-fetched picture of cluster health (pod list + Warning events) built at the start of every turn. Drives playbook matching and the snapshot sufficiency gate. |
| Snapshot sufficiency gate | A soft bias that lets the agent answer healthy, list-shaped questions straight from the snapshot instead of re-querying. Modes: off / lenient / strict. See Agent Behaviors. |
| Playbook | A YAML investigation recipe for a known failure (CrashLoopBackOff, OOMKilled, …). When the snapshot matches, the playbook is injected into the prompt to guide the agent. 18 ship by default. See Playbook library. |
| Investigation plan | A visible, up-front checklist the agent emits for queries needing 3+ steps, streamed as a plan event so the UI can show it. |
| RCA | Root-Cause Analysis — the deep investigation path where four subagents run in parallel and the coordinator synthesizes their findings into one conclusion. |
| RCAResult | The structured output of an RCA: root cause, confidence, supporting evidence, conflicting evidence, reasoning, and a recommended fix. See What you can ask. |
| AgentFinding | A single subagent's structured contribution to an RCA (signals, hypothesis, confidence, evidence). |
| HITL | Human-in-the-Loop — the approval gate that pauses before any write operation runs, showing you the exact command. See HITL gate. |
| Always-confirm gate | A set of cascading-blast actions (delete namespace, delete pv/crd, set image/resources, drain) that always require explicit approval, even in auto-approve mode. |
auto_approve |
A per-request flag that skips routine approval gates for trusted automation. The always-confirm gate still fires. |
| Reflexion | The cross-session learning loop: KubeIntellect records verified root-cause outcomes and promotes recurring, confirmed patterns back into future prompts. PostgreSQL only. See Reflexion Subsystem. |
| Failure pattern | A recurring, verified problem signature that reflexion has promoted; it is recalled (cluster-scoped) to speed up future diagnoses. |
TARGETED |
A routing outcome: one specific resource is failing, so the agent runs a focused three-read investigation on it before answering. |
RCA_REQUIRED |
A routing outcome: the failure is ambiguous or cross-cutting, so the agent fans out to the four subagents for a full RCA. |
ki_event |
The side-channel message type on the SSE stream that carries progress (status, tool calls, tool results, plans) without disturbing OpenAI-compatible clients. See API Reference. |
| Tool | A capability the agent can invoke: run_kubectl, run_helm (read-only), query_prometheus, query_loki. |
| Role / RBAC tier | The permission level of an API key: readonly, operator, admin, superadmin — determines which kubectl verbs the agent may run. See Security. |
| Demo key | A short-lived, read-only HMAC API key minted on demand for time-boxed access (e.g. the public browser demo). |
| Checkpointer / thread | LangGraph's state store. Each conversation (keyed by X-Session-ID) is a thread whose state is persisted so a paused HITL turn can resume. |
kq / kube-q |
The command-line query client you talk to KubeIntellect with. Separate package (kube-q). See CLI Reference. |
| LangGraph | The agent-orchestration framework KubeIntellect is built on — it defines the workflow graph (memory → context → coordinator → subagents → synthesis). |
Related¶
- Agent Behaviors — how these pieces fit together each turn.
- Architecture — the developer-level reference.
- What you can ask — these concepts in practice.