grok-bot-jev
Python router that has Jev classify Grok Bot requests before costly work, returning actions like reuse_cache, stop_retry or ask_human.
Overview
grok-bot-jev connects Jev to Grok Bot as a cheap decision layer. Before expensive research, browser, retry or subagent work, a Python router around the TypeSafe SDK's system_one call classifies the request and returns an explicit action: reuse_cache, stop_retry, run_deterministic, chat_only, research_capped, allow_subagent or ask_human. This lets Grok Bot reuse a fresh artifact, stop a failing retry, cap research or ask for approval.
It is a small reference implementation that does not change Grok Bot's foundation model. It has a shadow mode for measurement and an active mode for a skill that honours the returned action, plus a kill switch through config or a bypass marker. The router is conservative: it never sends, publishes, pays, deletes or changes permissions without a human approval path, and it logs decisions without logging the API key. A skill template and worked examples are included.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action should Grok Bot take for this request? | reuse_cache · stop_retry · run_deterministic · chat_only · research_capped · allow_subagent · ask_human |
How it works
- Describe the pending request as state
- Router asks Jev for an action
- Shadow mode logs; active mode lets the skill honour the action
- Consequential actions always go through human approval
Strengths
- Avoids needless tool work
- Shadow mode for measurement
- Kill switch and conservative defaults
Caveats
- Reference implementation
- Requires a TypeSafe key
Getting started
git clone, create a venv, pip install -r requirements.txt, copy config.example.yaml, export TYPESAFE_API_KEY, then python scripts/dry_run.py.At a glance
- Type
- Agent skill
- Status
- prototype
- Language
- Python
- Runtime
- router + agent skill
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Agent bridges: connects Jev to Grok Bot as a cheap decision layer, with usage gates, a skill template, and worked examples.
yibie/awesome-jevConnect TypeSafe Jev to Grok Bot as a cheap decision layer - usage gates, skill template, examples
logicrw/awesome-jev-projectsConnect TypeSafe Jev to Grok Bot as a cheap decision layer - usage gates, skill…
wh000wh000/awesome-jev-liveConnect TypeSafe Jev to Grok Bot as a cheap decision layer - usage gates, skill template, examples.
hellogumbo/awesome-jevClassifies a Grok Bot request before expensive research, browser, retry, or subagent work, so it can reuse a fresh artifact or stop a failing retry.
MrJev/awesome-jev
Frequently asked questions
Does it change Grok's model?
No, it only adds a decision layer before costly work.
What actions can it return?
reuse_cache, stop_retry, run_deterministic, chat_only, research_capped, allow_subagent and ask_human.
Can I measure before enforcing?
Yes, shadow mode logs decisions without acting.
Related projects
DeepSeek Harness plugin that registers a native jev_decide tool and declares its own policy for when the agent should ask Jev.
PluginbetaJavaScriptHermes skill that calls Jev as a typed tiebreaker, returning Choice, Score or Noul probabilities so the agent can act or escalate.
Agent skillbetaPythonnpm library that rate-limits, spend-caps, caches and redacts every Jev request made in a loop.
LibrarybetaJavaScriptPortable decision layer for agent harnesses: Jev picks from host-supplied capabilities while the host keeps permissions and execution.
ProjectbetaTypeScriptRust MCP server for Codex that hands classification, relevance, candidate selection and evaluation to Jev with accept/verify policies.
MCP serverbetaRustSelf-hosted Jev MCP server for Claude Code and Codex with confidence gating, cost guardrails, caching and decision logs.
MCP serverbetaPython