hermes - 💡(How to fix) Fix [RFC] Token budget system with auto-continuation for bounded-cost sessions

Motivation

In long autonomous sessions (delegate_task, cron jobs), per-turn token consumption can spiral. A single bad loop can burn through the entire context budget in a few turns.

Currently there is no per-turn budget control. The only safeguard is the hard context window limit, which kills the session when hit.

Proposal: Per-turn Token Budget

An optional budget system that caps token consumption per turn:

1. Configurable max_tokens_per_turn — default around 8000, overridable per task
2. Soft cap at 80 percent — warn the model it is approaching the limit
3. Hard cap at 100 percent — truncate the response and force a result
4. Auto-continuation — if the model indicates it needs more, allow a follow-up turn
5. Per-session tracking — cumulative spend across continuation turns

Use cases

• Kanban workers — guarantee bounded cost per task
• delegate_task subagents — prevent runaway spend on a single delegated job
• Cron jobs — keep recurring tasks at predictable token consumption
• Interactive sessions — optional cost-limiter for power users

Design notes

• Token budget should be per-turn, not per-session (the context window handles per-session)
• The soft-cap warning should be injected into the system prompt for that turn
• Truncation at hard cap should be done on the assistant response, not on tool outputs
• Auto-continuation should be opt-in: the model explicitly requests it