Skip to main content

​
How Siclaw Investigates

Siclaw’s core is a 4-phase hypothesis-driven investigation engine: 
Phase 1: Context Gathering    → Collects cluster state, events, logs, metrics
Phase 2: Hypothesis Generation → Generates 3-5 ranked hypotheses from evidence
Phase 3: Parallel Validation   → Up to 3 sub-agents independently test hypotheses
Phase 4: Conclusion            → Structured report with root cause and remediation
Every investigation is read-only — Siclaw never modifies your cluster. See Deep Investigation for details.

​
Runtime Modes

Siclaw runs in three modes sharing one agent core:
ModeUse CaseStart Command
CLI (TUI)Personal terminal diagnosticsnpx siclaw
GatewayMulti-user with Web UI, Slack, Larksiclaw-gateway
CronScheduled health patrolssiclaw-cron
For team use, Gateway spawns an isolated AgentBox per user — either in-process (local dev) or as a Kubernetes pod (production).

​
Skills

Skills are reusable diagnostic playbooks your team can create and share: 
skills/
├── core/              Built-in (kubectl checks, log analysis, etc.)
├── team/              Team-managed via Web UI
└── user/<userId>/     Personal skills
Each skill contains a SKILL.md description and optional scripts (.sh, .py). Scripts go through mandatory security review (static analysis + AI review + human approval) before they can execute. See Skills for details.

​
Investigation Memory

Siclaw learns from every investigation:
  • Stores structured records (root cause, causal chain, confidence) in a local database
  • Hybrid search (vector similarity + keyword matching) retrieves relevant past incidents
  • Past findings feed into Phase 2 to improve hypothesis generation
Requires an embedding provider to be configured. See Memory for details.

​
Security Model

Siclaw uses a whitelist-only security model — commands must be explicitly approved to run:
  • Shell: ~80 approved binaries only. sed, awk, nc are intentionally excluded.
  • kubectl: Read-only (13 safe subcommands). All write operations permanently blocked.
  • Skills: Scripts exempt from the binary allowlist, but require the security review gate.
  • Files: Agent can only write to its own data directory.

​
MCP Integration

Siclaw supports the Model Context Protocol (MCP) for extending capabilities with external tools — Prometheus metrics, GitHub issues, custom data sources. See MCP Servers for setup and examples.