dsh-policy-strict-gate

The moment a model needs a check most is the moment it is least likely to run one voluntarily: when it is confident, rushing, or already looping. dsh-policy-strict-gate elevates strict_check and the failure journal from tools the agent may call into policy the harness enforces, so a refusal lands before the next write rather than after the fact.

Three independently configured gates do the heavy lifting. The repeated-failure gate detects the same failure signature recurring in a session and injects a targeted remediation notice, deliberately keyed to tools/result so even pipeline-level denials count toward the threshold. The critical-path gate refuses writes to glob-protected paths until a covering Lean specification is accepted, while keeping the repair path open so the fix to the very error it reported is never blocked. The post-write syntax gate runs the language native checker immediately after every accepted write, catching a broken edit at the step that broke it instead of three commands later through an unrelated failure.

Built for DeepSeek Harness users who want security validation without relying on model self-discipline. No build step, no runtime dependencies; the strict-check package is optional at load time, and the repeated-failure gate remains active even without it, so policy can be tightened incrementally.