🧠 Automated AI Knowledge Base for Product Managers — 4-pillar system with self-evolving knowledge graph. Updates every 4h.

A CPU-style scheduler for an agent's token budget — pre-allocate by value-per-token, predict overrun, down-tier or preempt before the window blows.

TypeScript dependency scanner that flags imperative prose aimed at coding agents hidden in package files, docstrings, and MCP manifests

Curated Hermes Agent skills & runbooks — personal knowledge workflows, ready to drop into your own agent. | 精选 Hermes Agent Skill 合集，记录和分享 hermes-agent 使用经验，开箱即用。

Forensic timeline that shows which bash command your AI ran at 3am

Adapt GLM-5.2 tool-calling to the OpenAI tool_calls schema so coding-agent harnesses stop silently mis-parsing GLM function calls

Score the AI-generated code slop your agents left behind — a zero-config static SlopScore (0-100) + heatmap for JS/TS repos

CiteGuard — catch fabricated citations & CVEs in papers and PRs, in CI. Deterministic citation/identifier verifier (DOI/arXiv/CVE/commit/GitHub issue) with a GitHub Action.

Surgical in-session context editing for coding agents — delete failed turns from Claude Code / Cursor sessions instead of /clear-and-restart

Stripe's Idempotency-Key, but for your browser agent — wrap a transactional step so a self-healing retry is deduplicated, not double-submitted.

skillprov v0.3.0 — value-level exec command allowlist: declaring commands:[git] no longer lets a skill quietly curl | sh. Closes the last class-level capability hole.

RefusalScope — probe and classify how an LLM refuses (incl. disguised/soft refusals), and diff refusal scope across model versions

AgentFuse — per-task spend circuit-breaker for LLM agents

Signed manifest + cross-version scope-drift detector for Claude Code skills

AgentFuse — per-task spend circuit-breaker for LLM agents; halts before the next over-budget call

CacheGuard（缓存卫士）— a drop-in proxy that keeps DeepSeek's server-side prefix-cache stable in front of any coding agent, so cache-hit pricing never silently breaks.

Drive Windows desktop apps with an LLM agent via the UI Automation accessibility tree — no pixel/CV automation

Runtime seccomp-style per-call capability sandbox for installable agent skills — declare what each skill may touch, enforce it at the call site, trap+log violations.

cgroups for your LLM bill — hierarchical, preemptive, multi-provider token budget enforcer for coding-agent swarms, one binary

Structured verification gate for AI agent edit loops — AST-level proof that edits actually landed