[BUG] Agent: "invalid transition from Idle to Idle" error logged on every run

[edenreich]

Summary

Every agent run (once per user message/turn) logs an error-level line with a full stacktrace:
failed to transition to Idle state / invalid transition from Idle to Idle. It is harmless - the
agent's state machine is constructed in Idle and Start() redundantly asks it to transition to Idle
again - but because it is logged at error level (and zap's production config attaches a stacktrace to
every error), it reads as a real failure and spams the log on every turn.

Steps to Reproduce

1. infer chat.
2. Send any message (triggers an agent run).
3. Inspect the day's log:
   grep -c "invalid transition from Idle to Idle" ~/.infer/logs/app-$(date +%F).log - one error (plus a
   multi-line stacktrace) per run.

Expected Behavior

Starting the agent while it is already in Idle is a no-op, not an error. Nothing should be logged at
error level (and no stacktrace) for this.

Actual Behavior

{"level":"error","caller":"agent/agent_event_driven.go:200","msg":"failed to transition to Idle state","error":"invalid transition from Idle to Idle","stacktrace":"github.com/inference-gateway/cli/internal/agent.(*EventDrivenAgent).Start\n\t.../agent/agent_event_driven.go:200\ngithub.com/inference-gateway/cli/internal/agent.(*AgentServiceImpl).RunWithStream.func1\n\t.../agent/agent.go:647"}

Root Cause

• NewAgentStateMachine initializes currentState: domain.StateIdle (internal/agent/agent_state_machine.go:59),
  and a fresh state machine is constructed for every agent run (internal/agent/agent_event_driven.go:69,
  called from AgentServiceImpl.RunWithStream per request).
• EventDrivenAgent.Start() then unconditionally calls
  a.stateMachine.Transition(a.agentCtx, domain.StateIdle) (agent_event_driven.go:199).
• The only outgoing transition registered from Idle is Idle -> CheckingQueue
  (agent_state_machine.go:76); there is no Idle -> Idle self-transition, so findTransition returns
  nil and Transition returns invalid transition from Idle to Idle (agent_state_machine.go:202-205).
• Start() logs that error (agent_event_driven.go:200).

The subsequent MessageReceivedEvent drives the real Idle -> CheckingQueue flow, so there is no
functional impact - only a misleading, per-turn error + stacktrace in the log.

Proposed Fix

Either:

• Guard the redundant call in Start() - only transition when not already in Idle
  (if a.stateMachine.GetCurrentState() != domain.StateIdle { ... }), or drop the initializing
  transition entirely since the machine is always constructed in Idle. Minimal, localized.
• Or make Transition treat a same-state target as a successful no-op
  (if sm.currentState == targetState { return nil } before findTransition), which resolves this class
  of redundant self-transition everywhere.

Recommend the localized Start() guard unless a general same-state no-op policy is preferred.

Environment

inference-gateway/cli, main. Surfaced right after #722 removed the unrelated keybinding log noise, so
pre-existing agent-level errors are now visible.

[inference-gateway-releaser]

🎉 This issue has been resolved in version 0.128.0 🎉

The release is available on:

• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀