lingtai-ai Lifecycle and recording
This is a nested lingtai-portal-guide reference. It covers how the portal interprets agent lifecycle states and how it records/rebuilds topology history.
5-state lifecycle
| State | Meaning |
|---|---|
ACTIVE |
Agent is running and processing |
IDLE |
Agent is running but waiting for input |
STUCK |
Agent encountered an error or is unresponsive |
ASLEEP |
Agent is in sleep mode (.sleep signal or explicit lifecycle transition) |
SUSPENDED |
Agent process is not running (kernel/manifest reports it as such) |
Heartbeat is the portal's liveness ground truth, but it is observed independently from the manifest lifecycle state. BuildNetwork() calls IsAlive(..., AgentAliveThresholdSec()), which resolves the shared TUI/Portal/kernel liveness window from the LINGTAI_AGENT_ALIVE_THRESHOLD_SEC env var (default 10 seconds; missing, blank, malformed, non-finite, zero, or negative values fall back to the default). A stale or missing heartbeat only clears AgentNode.Alive — it does not rewrite .agent.json's stored state, so a genuinely SUSPENDED manifest stays SUSPENDED while an ACTIVE/IDLE/etc. agent with a stale heartbeat keeps its real state and is simply reported not alive.
Recording
The portal starts recording immediately on launch: a background goroutine calls BuildNetwork() every 3 seconds and appends the result as a JSONL line to topology.jsonl. This tape grows indefinitely.
Reconstruction
On startup, if the tape needs reconstruction (missing, empty, or old format), the portal rebuilds it from replay chunks. While that runs, .lingtai/.portal/reconstruct.progress may contain a transient current/total progress string, which the /api/topology/progress endpoint parses and exposes to the browser UI.
A manual POST /api/topology/rebuild instead reconstructs from source data and rewrites the replay chunks. The topology-and-api reference documents both endpoints' exact response shapes.