fix(chat): pass conversation history to agent for multi-turn memory

[behealthy28]

Bug: jarvis chat is stateless across turns when an agent is active

Summary

In jarvis chat, multi-turn memory is lost whenever an agent (e.g. the default simple agent) is active. Each turn is answered in isolation — the model never sees prior turns.

Reproduction

$ jarvis chat            # default agent = simple
You> What is 17 times 3?
17 times 3 is 51.
You> Now add 9 to that.
I see you want me to add 9 to something. Could you please specify what you want me to add 9 to?

Expected: "60". The model has no access to the previous turn.

Root cause

src/openjarvis/cli/chat_cmd.py maintains a history: List[Message] and appends each user/assistant turn, but the agent branch calls:

response = agent.run(user_input)        # no context passed

BaseAgent._build_messages(input, context) only includes prior conversation when context.conversation.messages is populated. With no context, every call is single-turn. The no-agent branch (engine.generate(history, ...)) already passes full history, so the bug only manifests when an agent is configured — which is the default.

Fix

Build an AgentContext from the prior history (excluding the just-appended user turn, which _build_messages re-adds from input) and pass it to agent.run:

from openjarvis.agents._stubs import AgentContext
ctx = AgentContext(conversation=Conversation(messages=list(history[:-1])))
response = agent.run(user_input, context=ctx)

Verified: with the patch, the default simple agent chains correctly — 51 → 60 → 120 — and agent tools remain available.

Environment

• OpenJarvis main (post v1.0.3.dev857)
• macOS, Apple M1, engine = mlx (mlx_lm.server), model mlx-community/Qwen2.5-3B-Instruct-4bit
• Behavior is engine-independent (logic bug in chat_cmd.py)