tools-only Human Touchpoint Model
ARL-derived escalation decision model. Not every stage transition requires human approval. The harness uses constraint analysis to escalate only when the agent cannot safely proceed.
Escalation Format
When escalating, present:
- What stage produced the escalation
- What triggered it (specific constraint, risk factor, or loop limit)
- What the agent knows (bound constraints, available context)
- What the agent does not know (unbound constraints, missing information)
- Decision options — concrete choices the human can make to unblock
Never present vague "please review" requests.
Pre-Scheduled Gates
Gate 1 — After S1 Discovery, before S2 Planning:
- Triggered when: unbound constraints, domain knowledge gaps
- Skipped when: all constraints bound, sufficient context
Gate 2 — After S4 Task Decomposition, before S5 Execution:
- Triggered when: high complexity, novel architecture
- Skipped when: routine patterns, existing codebase precedent
Dynamic Escalation Points
- NEEDS_WORK loop limit — 3 iterations in S6 without resolution
- NOT_CERTIFIED loop limit — 2 iterations in S7 without resolution
- Agent failure, quality gate cascade failure, contradiction detected
Constraint Types
| Type | Meaning | Action |
|---|---|---|
| Bound | All information available, deterministic | Proceed autonomously |
| Unbound | Information missing, ambiguous, or requires external knowledge | Escalate |
| Mixed | Some bound, some unbound | Isolate unbound; proceed with bound if possible |