baoer_signal_grep 1.6.6 Adds a Jev Semantic Judgment Layer for Coding Agents via MCP

Deep_Bus_7488 · reddit · 2026-09-21

baoersignalgrep, a local-first search and evidence layer for AI coding agents, adds an optional Jev semantic judgment layer (via MCP and native integrations) to fix searches that technically succeed but mislead the agent.

Key design: Jev never replaces local search. The pipeline searches the local workspace first, keeps exact matches and revision info authoritative, runs Jev only on bounded hybrid-search candidates, and returns judgments as structured metadata for prioritizing inspection. When an agent asks where an auth rule is implemented, Jev classifies candidates as implementation, caller, documentation, test-only, mention-only, or irrelevant — a routing signal, not proof of correctness.

Safety boundary: Jev is opt-in; API keys come from env vars, not config files; request/response sizes are capped with explicit timeouts, retries and cancellation propagation; malformed responses stay observable as partial results instead of becoming fake empty successes.

Why MCP: the same search/evidence contract works across Pi, OMP, Claude Code, Codex, Kimi Code and direct MCP clients, with compact human-facing output and evidence-rich model-facing results. The author verified the path with a real provider request, not mocks.

Original post →

More from coding & agent

coding & agent channel →