dsh-completion-gate

The most frustrating failure mode of LLM coding agents is rarely a wrong line of code; it is a premature declaration of success. The model says "I will now run the tests" and the turn silently closes, or it never runs a build, lint, or typecheck at all yet still reports "all done." Completion Gate is a host-side safeguard plugin built to block exactly that.

It does two things. First, the Premature Stop Guard watches for the model naturally halting while its own latest output still contains clear signals of pending work—phrases like "I still need to" or "Next I will run" trigger an automatic follow-up turn that picks up where the agent left off, with a bounded recovery chain so the loop never spins forever. Second, the Production Readiness Gate enforces a hard boundary: once code has changed, the root agent must execute the project's available checks (test, build, lint, typecheck, or custom commands) and bind the results to a SHA-256 workspace fingerprint derived from Git HEAD, staged and unstaged changes, and untracked files. Change a single line after the fact and the old evidence is instantly void. The plugin auto-detects Node.js, Python, Go, Rust, and PHP ecosystems, includes a narrow security-regression scanner for newly added lines, and supports a structured completion attestation so that "passed" carries machine-backed weight rather than just a model's self-report.

It is aimed at teams running autonomous orchestration, multi-subagent pipelines, or model failover on DeepSeek Harness: subagents work freely, but the root agent must clear one non-negotiable readiness gate before closing the turn. It is equally useful for engineering teams that need auditable, reproducible evidence in production repositories—because a fingerprint-bound check log proves "these checks ran against this exact state," not a vague "it should be fine."