Chat Sessions & the Universal Harness

One executor, every surface

ChatSessionExecutor is the single engine behind every conversational turn in Agentis — the dashboard composer, a Slack thread, a WhatsApp message, and the agent terminal all resolve to the same executor running the same loop. That's a deliberate collapse: a slash command like /plan means exactly the same thing whether you typed it in a browser or a phone, because there's only one place that interprets it.

What a turn actually assembles

A turn isn't "send the whole conversation history to the model and hope." The executor reconstructs a bounded context every time: the agent's persona, the active task, a live tool manifest ranked to a workflow-relevant subset (capped at 12 tools — WORKFLOW_TOOL_CAP — so the model isn't handed a wall of irrelevant capabilities), and fresh observations. Tool calls run in rounds up to AGENTIS_CHAT_MAX_TOOL_CALLS (default 2000, generously high because the real stopping logic is progress-based, not a hard count — see Chat Reliability).

CLI harnesses and streaming runtimes get different timeouts, on purpose

A CLI harness forwarding tools natively (via marker protocol or native MCP) is doing real, sometimes slow, build work — it gets a longer multi-round deadline and a realistic per-round timeout so a build that streams continuous canvas progress doesn't get cut off mid-work. A streaming HTTP runtime keeps a tighter budget that protects interactive latency instead. Same executor, different pacing for a fundamentally different kind of work.

Continue