Quick Start

Quick Start · v0.11.0

Install into the Python environment that runs Hermes. Replace python below with that interpreter when needed; a system-wide install in another environment will not make the plugin visible to Hermes.

1. Install the release

Back up the TT evidence vault before changing packages. The older rtk-hermes-plus distribution must not coexist with token-terminator: both own the same import package. Disable/remove that legacy distribution first if installed.

python -m pip install --upgrade \
  'git+https://github.com/AronAxe/[email protected]'
python -m rtk_hermes_plus.cli install-context-engine
hermes plugins enable token-terminator --no-allow-tool-override

The managed installer also includes the optional Desktop/backend files. It does not edit config.yaml, select an engine, enable Desktop, or start a server. Do not enable a duplicate RTK rewrite adapter. The optional rtk binary is only needed for terminal-command rewriting, not native compression, recovery or IR.

2. Choose middleware or ContextEngine

Keep another engine: leave your existing Context Engine selection unchanged; TT runs as middleware. For exact old-dialogue workloads in legacy middleware, set TOKEN_TERMINATOR_CONTEXT_COLLAPSE_AFTER_TURNS=0.

Replace LCM/compressor: in hermes plugins, choose Provider Plugins → Context Engine → Token Terminator. Equivalent YAML, while preserving your other enabled plugins:

plugins:
  enabled:
    - token-terminator
context:
  engine: token-terminator

Allow context_engine in restricted toolsets. Only one context engine owns the lifecycle. Restart Hermes/gateway, then verify token_terminator_history with {"action":"status"} in the running session.

3. Enable optional semantic attention and IR

export TOKEN_TERMINATOR_MODE=balanced
export TOKEN_TERMINATOR_JEV=true
export TOKEN_TERMINATOR_CONTEXT_IR=true

Reuse your existing OpenRouter OR direct TypeSafe key and provider setting. Never commit keys. JEV uploads bounded semantic state only after explicit opt-in; IR itself is local. ContextEngine and IR require the actual target tokenizer. Supported tiktoken models are automatic; other models need a matching configured counter/tokenizer, not an unrelated model's tokenizer.

4. Open the dashboard

token-terminator dashboard

Open http://localhost:7474. For dashboard-only managed installation, token-terminator install-dashboard does not select a context engine. Enable the Desktop half under Capabilities → Plugins, restart the gateway as needed, and click TT ↓ … in the bottom status bar. The popover does not need the standalone server. Install the adapter in each intended Hermes profile home.

5. Keep learning off until a policy is reviewed

The learned layer is off by default. Training needs explicit labelled experiments, optional [learning] dependencies and either replay or consented live collection. A private artifact and approved SHA-256 are required before shadow or active. Nothing trains or activates merely because it is installed.

ContextEngine guide · Dashboard · Learning · Configuration · Migration