Jekyll-Hyde

What Jekyll-Hyde does

Jekyll-Hyde is a Hermes plugin that looks for completion gaps in a session, meaning the difference between the work requested and the artifacts, tool evidence and verification that actually exist. Every Nth non-trivial user turn (default 7, set with the JEKYLL_HYDE_RATIO environment variable or the jekyll_hyde.ratio config key) it forks disposable audit roles. One produces an evidence review, a second proposes the smallest concrete continuation or names an external blocker, and an arbiter labels the evidence genuine, sandbagged or uncertain.

Depending on the operating mode, the result is silence, a focused directive, untrusted plugin guidance injected into the session, or two plans for the user to choose between. The clone exchanges are discarded after each response. The implementation is three files, three Hermes hooks (pre_llm_call, transform_llm_output and post_llm_call) and one slash command, with state kept in $HERMES_HOME/jekyll-hyde/state.json. Firmer wording is reserved for a directly observed promise followed by a goal shift.

Key features

• Audit gate that opens every Nth non-trivial user turn, default 7
• Disposable audit roles for evidence review, smallest next step and arbitration
• Arbiter labels of genuine, sandbagged or uncertain
• Several operating modes, from silent to offering two plans
• Persistent state file under $HERMES_HOME/jekyll-hyde/
• Three Hermes hooks and one slash command

When to use it

• Catching long coding sessions where work is described but never shipped
• Getting a concrete next step when an agent stalls without a real blocker
• Tuning how often a session is audited with the ratio setting

Who it is for: Hermes Agent users running long agentic coding sessions who want a mechanical check on whether work is actually completed.

How it fits with Hermes Agent

It is a plugin built for Hermes Agent and uses its pre_llm_call, transform_llm_output and post_llm_call hooks plus a slash command.

FAQ