decision-gates

Persist the verbatim objective and raw evidence anchors (commands, exit codes, file hashes) at each checkpoint, then run event-driven adversarial audits with an independent subagent that never reads summaries. Recovery always reads anchors before summaries.

Run the gate_consistency.py script on the workers' JSON reports before acceptance. It detects overlapping file conflicts, hash mismatches, coverage gaps, and surfaces SKIP/WARN/PENDING markers so unresolved items cannot be folded into a PASS.

Yes, gates for consistency, cost reconciliation, self-check, scope lock, and deadband detection are pure Python standard-library scripts with zero token cost. Only the adversarial audit gate uses an LLM, at the cheapest model tier.

The gate_deadband.py script fingerprints each failure and trips DEADBAND_TRIPPED after two consecutive identical failure signatures. This forces hypothesis invalidation and rollback to the previous checkpoint instead of blind retrying.

Skip the gates for single-turn small edits or ephemeral Q&A that needs no persistence. The Skill defers scale classification to a task-mode router so lightweight tasks avoid process overhead.

The gate_scope_lock.py script compares changed files against allowed and forbidden glob patterns. Any file matching a forbidden pattern or falling outside the allowed scope returns SCOPE_VIOLATION with exit code 1, blocking delivery.