Instance Lifecycle

Tentarc models execution units as instances, addressed by:

workspaceId:agentType:instanceId

Example: ws_design:task:t_20260305_01

Lifecycle states

Typical lifecycle:

cold -> ready -> running -> waiting_approval -> running -> completed
                            \-> interrupted -> running
                            \-> failed
                            \-> cancelled

Session vs task instance

• session: long-lived conversational execution thread.
• task: bounded unit, often with eviction/TTL semantics when idle.

Use session for continuity, task for isolated job execution.

State transition rules

Key constraints:

• transition must be explicit and persisted
• non-terminal states can resume
• terminal states (completed|failed|cancelled) are immutable except replay/fork workflows

This avoids “zombie execution” after crashes or UI refreshes.

Control intent interaction

Control intents (approval response, interrupt, steer) are applied as first-class events, not ad-hoc flags.
That design keeps lifecycle transitions observable and testable.

Operational advice

• keep instance granularity business-meaningful (not too coarse)
• avoid long uncheckpointed tool chains
• write status as soon as external side effects commit