| name | cave-gate |
| description | Loop an agent until a checker approves (a bounded fixpoint that always terminates). Use for 'draft then review until approved', refinement loops, write→critique→revise, retry-until-pass. `gate()` / `loop_refine()`. |
Gate — loop an agent until a checker approves
Run the body, then an evaluator; repeat until the evaluator sets context['approved'] truthy, or max_cycles is hit. Two proven laws: P1 termination (halts in ≤ max_cycles) and P2 gate-soundness (SUCCESS ⟺ the evaluator approved). Wrap a critic agent so its verdict becomes the approved flag.
Native API (how you program)
from cave_teams import gate, loop_refine
draft = gate(writer >> critic, approver, max_cycles=3)
draft = loop_refine(writer, critic, max_cycles=3)
The evaluator must set context['approved']. A plain agent writes text, not the flag — wrap it so its verdict (e.g. contains 'APPROVE') becomes context['approved'] = True.
Via cave() (drive it from data)
{"op": "gate", "body": <spec>, "evaluator": <spec>, "max_cycles": 3}
If the spec is malformed, cave() returns {"status": "construction_error", "hint": "read the cave-gate skill"} — the error names this skill.
See also
cave-sequential · cave-conditions · cave-teams
Part of the cave-teams plugin — the language + full DSL is the cave-teams skill; drive any pattern from data with the cave skill.
