[BUG] Agent: "total_turns" log always reads 1 in chat (per-request counter, not session total)

[edenreich]

Summary

The total_turns field in the agent completion logs always reads 1 in normal back-and-forth chat.
It is populated from the per-request AgentContext.Turns (which resets on every user message)
instead of the session-cumulative counter, so it never reflects the conversation's real turn count.

Steps to Reproduce

1. infer chat.
2. Send a message and wait for the reply; send a second message and wait for the reply (two simple
   turns, no tool calls).
3. Inspect the day's log:
   grep total_turns ~/.infer/logs/app-$(date +%F).log - every completed run logs "total_turns":1,
   including the second (and every later) turn.

Expected Behavior

A field named total_turns should reflect the cumulative number of model turns in the session (e.g. 2
after two conversational turns) - or be renamed so it is clearly a per-request iteration count. The
session already tracks the cumulative value via sessionTurns.

Actual Behavior

Every completed run logs total_turns:1 for simple chat. The value only exceeds 1 within a single run
that loops through tools; it resets on each new user message, so it never accumulates across the
conversation.

Root Cause

• AgentContext.Turns starts at 0 per agent run (internal/agent/agent_event_driven.go:77) and is
  incremented once per LLM streaming iteration in startStreaming (internal/agent/agent_streaming.go:20).
  A fresh EventDrivenAgent / AgentContext is built for every RunWithStream (i.e. every user
  message), so Turns is per-request and resets each turn - as documented at
  internal/agent/agent.go:41-47.
• The completion logs label this per-request value total_turns:
  internal/agent/states/completing.go:48 and :80, plus internal/agent/agent_event_driven.go:257
  ("agent reached Idle state - turn complete").
• The session-cumulative counter that would match "total" - AgentServiceImpl.sessionTurns
  (internal/agent/agent.go:48), incremented alongside Turns at agent_streaming.go:21 - is already
  maintained (it drives reminder cadence via injectDueReminders, agent_utils.go:854) but is never
  surfaced in these logs.

Proposed Fix

Either (observability-only, no behavioural change):

• Log the session-cumulative count under total_turns (or add a distinct session_turns field sourced
  from sessionTurns) at the completion / reached-Idle log sites, and/or
• Rename the existing per-request field to something unambiguous (e.g. run_turns / iterations) so it
  is not read as a session total.

Turn limiting is unaffected - MaxTurns continues to gate the per-request Turns.

Environment

inference-gateway/cli, main. Surfaced while reviewing agent logs after #722 / #725 cleared unrelated
log noise.

[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 📦🚀